Evidence Project Final Report - GOV.UKrandd.defra.gov.uk/Document.aspx?Document=14198_SD1705... ·...

Evidence Project SD1705 Final Report 1

General Enquiries on the form should be made to:

Defra, Strategic Evidence and Analysis E-mail: [email protected]

Evidence Project Final Report

Note

In line with the Freedom of Information Act 2000, Defra aims to place the results of its completed research projects in the public domain wherever possible. The Evidence Project Final Report is designed to capture the information on the results and outputs of Defra-funded research in a format that is easily publishable through the Defra website An Evidence Project Final Report must be completed for all projects.

This form is in Word format and the boxes may be expanded, as appropriate.

ACCESS TO INFORMATION

The information collected on this form will be stored electronically and may be sent to any part of Defra, or to individual researchers or organisations outside Defra for the purposes of reviewing the project. Defra may also disclose the information to any outside organisation acting as an agent authorised by Defra to process final research reports on its behalf. Defra intends to publish this form on its website, unless there are strong reasons not to, which fully comply with exemptions under the Environmental Information Regulations or the Freedom of Information Act 2000.

Defra may be required to release information, including personal data and commercial information, on request under the Environmental Information Regulations or the Freedom of Information Act 2000. However, Defra will not permit any unwarranted breach of confidentiality or act in contravention of its obligations under the Data Protection Act 1998. Defra or its appointed agents may use the name, address or other details on your form to contact you in connection with occasional customer research aimed at improving the processes through which Defra works with its contractors.

Project identification

1. Defra project code SD1705

2. Project title

Production of living maps to support the Defra 25 Year Plan Natural Capital related Pioneer Areas and support their application in models

3. Project originator Natural England

4. Total project costs £41,666

5. Project start date 3rd

October 2016

6. Project end date 21st March 2017

2 Living Maps: Satellite-based Habitat Classification

6. It is Defra’s intention to publish this form.

Please confirm your agreement to do so. .................................................................. YES NO

(a) When preparing Evidence Project Final Reports contractors should bear in mind that Defra intends that they be made public. They should be written in a clear and concise manner and represent a full account of the research project which someone not closely associated with the project can follow.

Defra recognises that in a small minority of cases there may be information, such as intellectual property or commercially confidential data, used in or generated by the research project, which should not be disclosed. In these cases, such information should be detailed in a separate annex (not to be published) so that the Evidence Project Final Report can be placed in the public domain. Where it is impossible to complete the Final Report without including references to any sensitive or confidential data, the information should be included and section (b) completed. NB: only in exceptional circumstances will Defra expect contractors to give a "No" answer.

In all cases, reasons for withholding information must be fully in line with exemptions under the Environmental Information Regulations or the Freedom of Information Act 2000.

Evidence Project Final Report SD1705 3

Evidence Project SD1705 Final Report

Living Maps: Satellite-based Habitat Classification

Alexandra Kilcoyne, Richard Alexander, Paul Cox & Jonathan Brownett

Natural England


Project details

This report should be cited as:

KILCOYNE, A.M., ALEXANDER, R., COX, P. & BROWNETT, J. (2017). Living Maps: Satellite-

based Habitat Classification. Evidence Project SD1705

Project manager

Paul Cox

Natural England

3rd Floor

Ceres House

2, Searby Road,

Lincoln

LN2 4DT

[email protected]

Acknowledgements

We would like to thank our colleagues at Natural England and throughout the Defra Group for

their assistance during the project. We are particularly grateful to representatives working on

the 25 Year Environment Plan Pioneer Areas for their feedback on the interim outputs.

In addition to the core project team we would like to thank Matthew Parker (University of

Sheffield) for his significant contributions to the development of Living Maps. Thanks also to

Gwawr Jones (JNCC) for reviewing the report and providing valuable feedback.

mailto:[email protected]


Executive Summary

‘Living Maps’ is an initiative to produce detailed habitat maps derived from satellite imagery for

England. The purpose of this phase was to test and develop a cost effective model for the

creation of Living Maps, focusing on the Defra 25 Year Plan Catchment and Landscape Pioneer

Areas. The project aimed to:

Replicate image processing and object-based classification steps currently done using

commercial earth observation (EO) software with open source alternatives.

Develop and test a cost effective model for the creation of Living Maps that could be

rolled out across England.

Assist Pioneer areas in translating habitat maps into decision support systems on

Natural Capital.

The Living Maps project builds on work undertaken through Making Earth Observation Work for

UK Biodiversity (MEOW; Medcalf et al., 2011, 2013, 2015) to develop new earth observation

methods for mapping habitats and assessing their condition.

The project has made use of freely available Sentinel-1 and Sentinel-2 imagery as well as

ancillary datasets including the WorldClim bioclimatic data, the Environment Agencies (EA)

Integrated Height Model (IHM) and Rural Payment Agencies’ (RPA) Land Parcel Information

System parcels. Pre-processed Sentinel-2 imagery was sourced through DEFRA’s Earth

Observation Centre of Excellence collaborative node, processed by the Joint Nature

Conservation Committee (JNCC). Training data were primarily derived from Farm Environment

Plan (FEP) records. Additional training data were derived from a combination of Ordnance

Surveys VectorMap layers, EA’s Saltmarsh Zonation layer and visual interpretation using

Google Earth imagery. It should be noted that eCognition was used for the segmentation, open

source alternatives are available. All other software used by the project was open source.

Data were collated for the Catchment and Landscape Pioneers and segmented into objects.

Automated parallel processing was used to extract zonal statistics for each satellite image layer

and ancillary datasets for the segmented polygons. Training points were ascribed values from

their surrounding segments to create a set of zonal statistics with known habitat classifications.

A statistical classification process developed in R was used to predict the likelihood of each

habitat class for each segment. A number of different machine learning algorithms were trialled,

with Random Forest achieving the highest accuracies. Segmented polygons were then

assigned habitat classes based on the highest probability predicted across habitat classes. The

second most likely habitat class was also attributed along with its probability.

The resulting classifications achieved accuracies for detailed habitats of 66% in the Catchment

Pioneer and 67% in the Landscape Pioneer, whilst broad habitats achieved accuracies of 74%

in the Catchment Pioneer and 75% in the Landscape Pioneer.

The project has demonstrated the effectiveness of open source software and machine learning algorithms in undertaking object-based habitat classification. The results are readily comparable to that achieved using commercial software but with significant cost savings.


Contents

1 Introduction 12

Habitat Mapping 12

What is the Living Maps project? Why is it needed? 13

Key Earth Observation Concepts 14

Introduction to EO 14

Current Earth Observation Classification Techniques 14

Supervised Classifications 14

Unsupervised Classifications 15

Object-based Image Analysis 15

EO Data Sources Available 15

New EO opportunities 15

2 Study Area 17

Catchment Pioneer 17

Temperature 18

Precipitation 18

Landscape Pioneer 19

Temperature 20

Rainfall 21

3 Datasets 22

Zonal Statistics Layers 22

Sentinel-1 22

Sentinel-2 22

Integrated Height Model 23

Bioclim 23

Moorland Boundary 23

OS VectorMap District 23

Training Data 23


Farm Environment Plan 23

OS VectorMap District 24

Additional Data Sources 24

4 Classes 25

Class Descriptors 25

Urban 25

Mud, Sand or Shingle 25

Coastal 25

Species Rich Grassland 25

Improved Grassland 26

Arable 26

Semi-improved Grassland 26

Heathland and Moorland 27

Wetland 27

Water 28

Deciduous and Mixed Woodland 28

Scrub and Bracken 28

Inland Rock and Scree 28

Deciduous and Mixed Woodland 29

National Ecosystem Assessment 29

5 Method 31

Summary of Classification Steps 31

Classification Steps 31

Random Forest Explanation 32

6 Results and Analysis 33

Results Summary 33

Mapped Outputs 33

Key Findings 41

Limitations 41

Results of Field Validation 42


7 Discussion 44

Knowledge Transfer and Dissemination 44

The possible impact of the evidence on the specified Defra policy outcomes 45

A cost effective method 45

8 Conclusions 46

Summary 46

Further work 46

Annex 47

1 Dataset Information 48

Sentinel-1 48





2 Method 55

Segmentation Steps 55

Classification 56

Load training data and segmented polygons 57

Calculate zonal stats from input layers for the segmented polygons 57

Impute missing values and calculate vegetation indices 57

Vegetation indices 58

Extract zonal statistics for the training data from segmented polygons 58

Sample training data and split into training and test datasets 58

Fit Random Forest model to training data 58

Rerun Random Forest using most significant variables 58

Calculate user and producer accuracy 58

Use model to predict class for each segmented polygon 59

Merge results to shapefile 59

References 60


Figures

Figure 1: Tier diagram for the Crick Framework 13

Figure 2: Map of the location of the Catchment Pioneer area of interest 17

Figure 3: Annual Mean Temperature of Catchment Pioneer region 18

Figure 4: Average Monthly Rainfall of Catchment Pioneer region 19

Figure 5: Map of the location of the Landscape Pioneer area of interest 20

Figure 6: Annual Mean Temperature of Landscape Pioneer region 20

Figure 7: Average Monthly Rainfall of Landscape Pioneer region 21

Figure 8: Processing chain for Sentinel-1 data and example of product 22

Figure 9: Example of samples for productive vegetation in SNAP during LSU 23

Figure 10: Flowchart of classification process 31

Figure 11: Catchment Pioneer Detailed Habitat Map 34

Figure 12: Landscape Pioneer Detailed Habitat Map 35

Figure 13: Probability values attributed by segment – Catchment Pioneer 36

Figure 14: Probability values attributed by segment – Landscape Pioneer 37

Figure 15: Detailed habitat accuracy values – Catchment Pioneer 38

Figure 16: Detailed habitat accuracy values – Landscape Pioneer 38

Figure 17: Broad habitat accuracy values – Catchment Pioneer 39

Figure 18: Broad habitat accuracy values – Landscape Pioneer 39

Figure 19: Broad habitat producer error values – Catchment Pioneer 40

Figure 20: Broad habitat producer error values – Landscape Pioneer 40

Figure 21: An example of errors created from imputing values 41

Figure 22: Examples of training points 43

Figure 23: Full processing chain for Sentinel-1 data using the ESA SNAP Toolbox. 50

Figure 24: An example diagram showing a mixed pixel composed of grass, trees and soil 52

Figure 25: Example Spectrum View 53

Figure 26: Example Spectrum View - zoomed in 53

Figure 27: Spectral Unmixing window in SNAP 54


Tables

Table 1: Training point tier summary 24

Table 2: Conversion between Living Maps and NEA broad habitats 29

Table 3: R packages used in the Living Maps project 56

Table 4: Variable importance 59


Glossary

BAP Biodiversity Action Plan

EO Earth Observation

ESA European Space Agency

FEP Farm Environment Plan

JNCC Joint Nature Conservation Committee

LSU Linear Spectral Unmixing

MEOW Making Earth Observation Work for UK Biodiversity

OBIA Object-based Image Analysis

RPA Rural Payment Agency


1 Introduction

Habitat Mapping

1.1 Habitat maps are a key component to understanding the distribution and extent of habitats across the environment. Knowledge gained serves to establish holistic landscape-scale approaches to the conservation needs of each habitat. Moreover, habitat maps are a recognised stepping stone in facilitating better management practices, biodiversity monitoring, and conservation planning, and in delivering Government policy, such as meeting the UK’s commitments relating to the Convention on Biological Diversity.

1.2 While there are numerous habitat maps available for England, many of these have gaps due to the limitations of traditional field survey techniques and available resources. Natural England’s Priority Habitats Inventory is an example of a national habitat map derived from field-based surveys collated from across the country, many of which date back over 30 years. Landcover Map 2007, produced by the Centre of Ecology and Hydrology (CEH), overcomes many of the limitations of traditional terrestrial based habitat maps by mapping land cover for the whole of the UK from satellite imagery. The constraints on the wider use of Landcover Map 2007 include the resolution of the satellite imagery used (20 - 30 m), licensing restrictions and the difficulty of translating between land cover and habitat classes.

1.3 EO has been extensively used to provide a synoptic view of land use, cover and change. With the increasing availability of very high resolution (VHR) satellite and airborne imagery, new methods have developed to allow rapid site evaluation and interpolation of ecological knowledge across wider areas. The combination of EO data with in situ data is progressively being applied to the classification of habitats with the potential to fill the gaps left by previous field surveys, whilst also increasing the temporal and spatial frequency at which data can be collected. The Habitat Inventory of Wales (Lucas et al, 2001) demonstrates strong evidence that this integrated approach can create a valuable product over large areas.

1.4 In addressing the role of EO data in habitat monitoring and surveillance needs in the UK, the Making Earth Observation Work for UK Biodiversity (MEOW) programme has provided an evidence base for both the advantages and limitations of the use of EO. Phase 1 (Medcalf et al., 2011) of MEOW showed that EO techniques together with the development of geoinformatics can make a contribution towards quantifying habitat extent, composition and condition. Object-based image analysis (OBIA) was employed and advocated as a useful approach to EO-based habitat mapping.


1.5 The Crick Framework (Figure 1) was developed in Phase 1 of MEOW as a tool to assess the extent to which EO could identify BAP priority habitats and European Habitats Directive Annex 1 habitats (recognised as the basis of much biodiversity work at country level). Five tiers enable the user to establish how possible it is to map a habitat of interest and suggest what quantity and type of EO data is needed.

1.6 Phase 2 (Medcalf et al. 2013) applied the learning from Phase 1, which had been developed in Wales, to a lowland situation on two areas of Norfolk, a very different biogeographical environment. An OBIA approach was applied and deemed to be fit for purpose for identifying areas supporting high priority habitats. The Crick Framework was further viewed and updated in Phase 2.

1.7 Phase 3 (Medcalf et al. 2015) worked to further develop lessons learnt from Phases 1 and 2 across upland areas, also looking at the possibility of a national EO service for habitat mapping. The Living Maps project may be seen as a continuation of this proposal in its attempt to create a cost effective model of OBIA which could be rolled out across England.

What is the Living Maps project? Why is it needed?

1.8 In part, the concept behind the Living Maps project was to provide a continuation of MEOW, creating a product that could be rolled out nationally and evolve through local feedback. However, fundamentally it was recognised that current approaches to habitat classification were reliant on expensive commercial software, time-consuming manual fitting of rules and computationally costly data processing. These limitations combined seriously constrain the feasibility and economic viability of rolling out the method to the whole country. The Living Maps project has been designed to challenge these restrictions through:

Use of open source software and scripting.

Adopting a machine learning approach rather than a rules-based approach to image

classification.

Figure 1: Tier diagram for the Crick Framework, categorising habitats based on the EO and ancillary data required to map them. VHR = Very High Resolution (<10 m per pixel)


Maximising the efficiency of iterative development through pre-processing of

statistical data.

Combining the use of multi-spectral and synthetic aperture radar imagery.

1.9 These approaches have cumulated in a method which would be significantly more cost effective to roll out nationally.

Key Earth Observation Concepts

Introduction to EO

1.10 Earth observation is the acquisition of information about the Earth’s surface and atmosphere using sensors on spaceborne or airborne platforms. The interaction between the electromagnetic radiation and the Earth’s surface is characterised by certain properties at different frequencies of electromagnetic energy. Sensors with different technology provide complimentary information about the surface. Rapid advances in the spatial, temporal and spectral resolution of these sensors have rendered them a valuable resource for a wide variety of mapping requirements.

1.11 Due to the interaction between vegetation and light reflecting and absorbing at different wavelengths, it is possible (with an appropriate selection of data) to make accurate deductions to infer vegetation classes. In particular, the interaction of the near infrared (NIR) and short wave infrared (SWIR) bands with, for example, leaf structure, enable species distinctions to be made that would otherwise be impossible from red green blue (RGB) imagery alone. This differs from data from low frequency synthetic aperture radar (SAR) which is largely a function of the interaction with, for example the woody component and the ground surface and does not observe the leaf material. The combination of these data together can provide a dynamic range of information to contribute to classifying vegetation into habitat classes. Moreover, the use of varying imagery acquisition dates, i.e., summer and winter imagery, can further be used to aid habitat classification.

1.12 Spectral relationships between bands are also frequently used in earth observation analysis. Chlorophyll pigments in plants strongly absorb red and blue light, giving plants their characteristic green appearance. Conversely, plant material reflects strongly in the near-infrared region. This spectral relationship, which can be characterised using indices such as the Normalised Difference Vegetation Index (NDVI), is unique to vegetation, and varies depending on plant health, biomass, productivity and seasonality, etc. It is therefore an extremely useful index in the discrimination of different species, and is used extensively in remote sensing applications.

Current Earth Observation Classification Techniques

1.13 A wide variety of EO classification techniques have been developed through a variety of disciplines and applications. Typically, these can be grouped into supervised image classification versus unsupervised image classification and pixel-based versus object-based image analysis.

Supervised Classifications

1.14 In a supervised classification, the user selects representative samples of each class they wish to map. These selections form the training data. The image classification software then uses the training data to identify the classes across the entire image, thus the classification is based on the spectral signatures defined in the training data. The algorithms in a supervised classification ‘learn’ the patterns in the data to predict the associated discrete class. Common


supervised machine learning techniques include Maximum Likelihood Estimation, Support Vector Machines, Random Forest and k-Nearest Neighbour.

Unsupervised Classifications

1.15 Unsupervised classifications attempt to find regularities in unclassified data by grouping pixels into clusters, based on their reflectance properties. The user identifies the number of clusters to generate and which information to use to form the clusters. Manual identification of each cluster is then used to allocate a desired class. Often multiple clusters represent a single class and thus clusters must be merged. This approach is frequently enlisted where no sample data exists. Examples of clustering algorithms include K-means, hierarchical agglomerative clustering and ISODATA. A common issue associated with unsupervised classification is determining the number of clusters. Additionally, unsupervised classifications rarely show a 1-to-1 relationship with classes derived from in situ data and therefore, to map habitats according to a classification scheme, aggregation and cross-tabulation must be used.

Object-based Image Analysis

1.16 An object-based approach utilises an image segmentation to produce homogenous image objects by grouping pixels. After the segmentation process is complete, the user identifies sample sites for each class and summary statistics used to classify image objects are extracted from the component pixels. Objects are then classified based on their resemblance to the training classes and associated statistical values. The advantage of an object-based approach is that image objects contain more information than a pixel (e.g., shape, heterogeneity) with objects resembling the shape, size and variation of the real world objects it is aimed to map. The integration of this additional knowledge is paramount when attempting to distinguish ecologically meaningful habitat classes that do not necessarily have distinct non-overlapping spectral features (Bock et al., 2005).

EO Data Sources Available

1.17 EO data sources may be divided into three platforms; spaceborne, airborne and terrestrial. This project is primarily concerned with spaceborne and airborne sensors. Sensors typically deployed on these platforms include film and digital cameras, light-detection and ranging (LiDAR) systems, synthetic aperture radar (SAR) systems, multispectral and hyperspectral scanners.

1.18 Satellite imagery is collected by a host of national and international government, and private agencies. Access is possible through collaboration with NASA and NASA funded institutions and more recently, through Copernicus; previously known as GMES (Global Monitoring for Environment and Security programme). This is a partnership between the European Commission (EC) and European Space Agency (ESA). The EC is guiding the overall initiative, including setting requirements for satellites and managing the services, through Copernicus. ESA is responsible for the delivery of the satellites and receiving the data.

1.19 Other sources of data include derived data, e.g. topological data derived from LiDAR, bioclimatic variables derived from climate data. Derived data can be hugely valuable to a classification, helping to define the ecological niche in which individual habitat classes may occur.

New EO opportunities

The recent launch of the Sentinel-1 and Sentinel-2 satellites offers new opportunities to exploit free at point of use data within EO. Sentinel-1 is a polar-orbiting, all-weather, day-and-night radar imaging mission for land and ocean services. The mission consists of a 2 satellite constellation each in a 12 day orbit, together providing orbit revisit times of 6 days at the


equator. Sentinel-2 is a polar-orbiting, wide-swath, high-resolution, multi-spectral imaging mission for land surface conditions. The mission consists of a 2 satellite constellation providing orbit revisit times 10 days at the equator with one satellite, and 5 days with 2 satellites under cloud-free conditions which results in 2-3 days at mid-latitudes. Furthermore, the release of EA's LiDAR data as open data and with the intention to provide England-wide coverage, the costly licencing constraints on EO-based habitat maps may be negated, making it possible to release outputs as open data.


2 Study Area

2.1 The Living Maps project has targeted the Catchment and Landscape Pioneer Areas of Cumbria and North Devon. The latter area was extended to include mapping of habitats for Dartmoor. The pioneer project areas were selected by the Natural Capital Committee (NCC) to identify good practice and innovative solutions to inform the Defra 25 Year Environment Plan.

Catchment Pioneer

2.2 The study area used consisted of the county of Cumbria, which is inclusive of the Catchment pioneer area (Figure 2). A 1 km buffer was applied to this area.

2.3 The Lake District is situated on a volcanic extrusion. The surface geology is a mixture of granite and sedimentary rock, including mudstone and limestone. Much of the landscape is shaped by its mining and farming heritage.

Figure 2: Map of the location of the Catchment Pioneer area of interest


Temperature

2.4 The coastal location of Cumbria means that it experiences relatively low variation in temperature throughout the year, though this also varies locally by altitude. Mean annual temperatures range from 4 – 9 oC across the region, with an annual range of around 19 – 20 oC (minimum temperatures in range -4 – 0 oC, maximum temperatures in range 15 – 20 oC).

Figure 3: Annual Mean Temperature of Catchment Pioneer region (source: WorldClim)

Precipitation

2.5 The coastal and upland geography of Cumbria also make it one of the wettest regions of England, with an average annual rainfall between 800 – 2000 mm (depending on geographical location). Upland areas have the highest range in rainfall – from 90 – 100 mm in driest month to 200 mm in wettest month (Figure 4).


Figure 4: Average Monthly Rainfall of Catchment Pioneer region (source: WorldClim)

Habitats

2.6 Cumbria high fells contains the most biologically diverse range of upland habitats in England, with internationally important fell habitats, arctic alpine plants, lakes, rivers, woodlands and a few species-rich meadows/pastures. Native broadleaf woodland and conifer plantations are extensive on the valley sides and bottoms, while the fells support scattered trees and scrub, plus a few small, high level and gill woodlands. Woodlands and peatlands, including blanket bog, are important carbon stores, requiring appropriate management to prevent carbon loss to the atmosphere and through water run-off. Coastal areas contain a diverse range of habitats including coastal and flood plain grazing marsh, salt marsh, coastal sand dunes, intertidal mudflats, coastal vegetated shingle, saline lagoons and maritime cliff and slope habitats (Natural England, 2014).

Landscape Pioneer

2.7 The Landscape Pioneer area is based on that of the North Devon Biosphere Reserve (including marine areas). The study area included a 1 km buffer as well as extending the boundary to include Dartmoor National Park (Figure 5).

2.8 The study area consists predominantly of the Culm, an open, rolling plateau divided by an intricate pattern of valleys. The dominant influence on the landscape is the heavy clay overlaying the Upper Carboniferous Culm Measures, giving rise to heavy, wet soils, making cultivation difficult. Reflecting this, and the oceanic climate, the predominant land use is grass production for livestock. To the south, Dartmoor consists of an extensive area of upland granite overlaid with thick deposits of peat.


Figure 5: Map of the location of the Landscape Pioneer area of interest

Temperature

2.9 Devon in general has a mild climate due to its south-westerly location. The temperature near the coast ranges from between 1 – 3 oC in the coldest month, to 19 – 21 oC in the warmest month. On Dartmoor, the annual temperature range is between -1 – 1 oC and 17 – 19 oC.

Figure 6: Annual Mean Temperature of Landscape Pioneer region (source: WorldClim)


Rainfall

2.10 Low lying areas of Devon have relatively low rainfall compared to Dartmoor, which can typically have double the rainfall of coastal regions. The total annual rainfall of Dartmoor around 1100 mm per year.

Figure 7: Average Monthly Rainfall of Landscape Pioneer region (source: WorldClim)

Habitats

2.11 The Culm area is characterised by the species rich pastures, typically on poorly drained acid soils, supporting a suite of purple moor grass and rush pasture communities. On steeper slopes western oak woodland is still dominant. Sea cliffs and slope run along the length of the coast, with the Taw-Torridge estuary comprising areas of mud and sand flats and saltmarsh. Dartmoor supports internationally important blanket bogs surrounded by large expanses of upland heathland and grass moorland (Natural England, 2014).


3 Datasets

Zonal Statistics Layers

3.1 A summary of each of the datasets used for input to the habitat classification is listed below. All data were georeferenced to British National Grid, EPSG 27700 and were resampled to 10 m unless otherwise specified.

Sentinel-1

3.2 Sentinel-1 is a polar-orbiting, all-weather, day-and-night radar imaging mission for land and ocean services. Sentinel-1 data was downloaded from the Copernicus portal (https://scihub.copernicus.eu/dhus/#/home), then processed at the original resolution of 5 m x 20 m. Scenes were acquired on 06/07/2016 (summer) and 08/01/2016 (winter) for the Catchment and Landscape Pioneers. Figure 8 outlines the main steps in the processing chain applied through the ESA’s SNAP Toolbox for Sentinel-1 data.

Figure 8: Processing chain for Sentinel-1 data and example of product

Sentinel-2

3.3 Sentinel-2 is a polar-orbiting, multispectral high-resolution imaging mission for land monitoring to provide, for example, imagery of vegetation, soil and water cover, inland waterways and coastal areas. Sentinel-2 data was provided as Analysis Ready Data (ARD) by JNCC (see Annex 1). Cloud masking was completed through a manual digitising process. Scenes were acquired on 02/06/2016 (summer) and 14/03/2016 (winter) for the Catchment Pioneer and 19/07/2016 (summer) and 16/11/2016 (winter) for the Landscape Pioneer. An additional scene acquired for the 29/11/2016 (winter) was used to for Landscape Pioneer due to a high percentage of cloud on the winter image. Clouds were manually masked by digitising and removing the areas.

3.4 ESA’s SNAP Toolbox (http://step.esa.int/main/toolboxes/snap/) was used to create Linear Spectral Unmixed (LSU) layers for productive vegetation, non-productive vegetation and shade for summer and winter scenes in both Pioneer Areas. By identifying pure pixels of desired criteria, each pixel is attributed with affinity to the training pixels.

Co-Registering

Terrain Flattening and

Terrain Correction

Multi-Temporal Filter

Data Calibration

https://scihub.copernicus.eu/dhus/%23/home

http://step.esa.int/main/toolboxes/snap/


Figure 9: Example of samples for productive vegetation in SNAP during LSU

Integrated Height Model

3.5 The Environment Agencies (EA) Integrated Height Model (IHM) was used as a direct input with derived layers of slope and aspect also included.

Bioclim

3.6 Maximum temperature, minimum temperature and annual rainfall were downloaded as freely available raster layers at ~1 km spatial resolution (30 arc-seconds) from WorldClim (http://worldclim.org/bioclim).

Moorland Boundary

3.7 The RPA’s Moorland Boundary is a dataset dividing the enclosed and unenclosed parcels of land within England. The dataset was rasterised and a proximity raster layer calculated in QGIS.

OS VectorMap District

3.8 OS VectorMap District layers (https://www.ordnancesurvey.co.uk/business-and-government/products/vectormap-district.html) were downloaded from the OS OpenData Supply website. Layers for building, foreshore, road, surface water (area only), tidal water and woodland were rasterised and proximity rasters calculated for each layer in QGIS. A further dataset of the original OS layers combined into one raster layer was also created.

Training Data

Farm Environment Plan

3.9 The Farm Environment Plan (FEP) dataset contains details of features which are considered to have environmental value, recorded to inform Higher Level Stewardship (HLS) agreements. The FEP data were selected as being a readily available national dataset with a set of classes complimentary to the desired output classes. Farm Environment Plan data were collected from 2005 up to 2014. Data were extracted from the Environment Stewardship database, Genesis and pre-processed to remove features not of interest for habitat mapping (e.g., ‘Stone-faced banks’) or considered not detectable by earth observation (e.g., ‘G11: Habitat for

http://worldclim.org/bioclim

https://www.ordnancesurvey.co.uk/business-and-government/products/vectormap-district.html

https://www.ordnancesurvey.co.uk/business-and-government/products/vectormap-district.html


invertebrates’). A tiered system was allocated based on percentage of Land Parcel Identification System (LPIS) parcel coverage by feature; summarised in Table 1. Up to 50 points were randomly selected within each tier to provide a representative sample of each class. During image classification only data from the first two tiers was used, with priority given to data from Tier 1.

Table 1: Training point tier summary

Tier Percentage Coverage

1 > 95

2 90 - 94.9

3 80 - 89.9

4 70 - 79.9

5 60 - 69.9

6 < 60

3.10 Finally, the FEP classes were translated to the nearest Biodiversity Action Plan (BAP) Broad or BAP Priority class with detailed and broad categories defined for each training point. The data were validated against Natural England’s Priority Habitats Inventory (PHI), which also contains data from other field surveys and Natural England’s designated sites database.

OS VectorMap District

3.11 VectorMap District data layers were used to create training points for surface water, building and sea. 50 random points were selected within each of these categories.

Additional Data Sources

3.12 Due to insufficient samples in the FEP data, bracken and coniferous woodland was mapped manually using high resolution aerial imagery and Google Street View. Saltmarsh samples were created from the Environment Agency’s saltmarsh zonation dataset, available as open data on data.gov.uk.

3.13 Training data point samples were manually checked through a visual assessment against the Sentinel-2 data and aerial photography to ensure that selections were representative of the associated habitat. Where more than one point intersected a segment, the habitat covering the largest area was kept and the secondary habitat reallocated to another segment.

http://data.gov.uk/


4 Classes

Class Descriptors

4.1 Where appropriate, classes from the FEP data have been assigned to the BAP Broad and BAP Priority classes.

Urban

4.2 Urban

4.3 This class includes any built structure, including houses, offices, factories, farm buildings, reservoir dams and other built structures, tarmac, concrete or gravel.

Mud, Sand or Shingle

4.4 Mud, Sand or Shingle

4.5 Mudflats are part of the unvegetated intertidal zone that occurs on the seaward or estuary edge. They are regularly inundated and consist of fine sediments that remain saturated.

Coastal

4.6 Coastal Saltmarsh

4.7 Defined as an area of fine muds and silts (occasional small patches of sand or shingle), regularly inundated by the tide and vegetated by salt-tolerant plants.

4.8 Coastal Sand Dunes (including Machair/ Coastal Vegetated Sand Dunes)

4.9 Aeolian deposits of sand, vegetated by specialist plants. A sand dune system may include many zones, through embryonic and fore dunes through to fixed dune grassland which may include dune slacks.

4.10 Maritime Cliff and Slopes

4.11 Characterised by clifftop vegetation that is influenced by salt spray, this broadly refers to heathland or unimproved grassland on maritime cliffs and slopes. Frequently found as a mosaic, with bracken, scrub and wet flushes.

Species Rich Grassland

4.12 Lowland Dry Acid Grassland

4.13 Semi-natural grassland on nutrient-poor, free-draining soils in the lowland and enclosed upland fridge. Primarily managed by grazing, mosses and/or lichens are sometimes frequent. May be a mosaic with lowland heath.

4.14 Lowland Calcareous Grassland

4.15 Species-rich, semi-natural grassland on chalk and limestone primarily managed by grazing, located in the lowlands to upland fringe, generally below 300 m in altitude.


4.16 Upland Calcareous Grassland

4.17 Generally species-rich, semi-natural grassland on calcareous soils over Carboniferous limestone in upland areas. Usually dominated by fine-leaved grasses and primarily managed by grazing. Upland Calcareous Grassland often occurs in parts of large-scale enclosures with other less species-rich grassland types.

4.18 Lowland Meadows

4.19 Species-rich, semi-natural grassland on neutral, free-draining soils in the lowland and upland fringes, including species-rich flood plain grassland. Primarily managed by cutting and/or grazing.

4.20 Upland Hay Meadows

4.21 Enclosed land on free-draining or moist neutral soils, found in Cumbria, in the eastern Lake District, Northumberland, Durham, the Pennines and Dales of Yorkshire. Meadows are cut for hay, with aftermath grazing.

4.22 Purple Moor Grass and Rush Pastures

4.23 Species-rich, semi-natural grassland with abundant jointed rushes (blunt flower rush, sharp flowered rush or jointed rush) and/or purple moor grass found on poorly drained neutral and acidic soils of the lowlands and upland fridge. Purple moor grass and rush pastures can occur on the upland fringes and above the moorland line. However, these should not be confused with semi-improved pastures (where soft rush is often abundant) and species-poor, rush-dominated flushes, or specie-poor purple moor grass wet acid grassland, which lacks most of the wildflower indicator species.

Improved Grassland

4.24 Coastal and Floodplain Grazing Marsh

4.25 Periodically flooded meadows or pastures with ditches that maintain the water levels, containing fresh or brackish water. Ditches are often especially rich in plants; the main grassland is often not very species-rich. All areas are grazed but some are cut for silage or hay.

4.26 Improved Grassland

4.27 Improved grassland is defined as those which are dominated by rye grass species and active management such as fertiliser applications. They could be either grass leys or permanent grassland.

Arable

4.28 Arable and Horticultural

4.29 This includes land used for the production of root crops, combinable crops, field-scale vegetables or maize.

Semi-improved Grassland

4.30 Semi-improved Grassland


4.31 A transitional habitat between improved grassland and species-rich grassland, semi-improved grassland occurs on a wide range of soils where a wide range of grass species may be present. The habitat may have previously had management or inputs applied to it, such as drainage channels created, fertiliser, slurry or pesticides applied, but these no longer happen.

Heathland and Moorland

4.32 Acid Grassland

4.33 Includes enclosed species-poor acid grassland on land designated as being within Severely Disadvantaged Areas, typically dominated by fine-leaved and bent fescue grasses, heath rush, mat-grass and purple moor grass, and unenclosed wet and dry acid grassland in moorland grazing units. Generally found on shallow peat (< 50 cm deep) or mineral soils. Dwarf shrubs generally absent or less than frequent.

4.34 Blanket Bog

4.35 Comprises of upland wetland vegetation, usually on unenclosed moorland and generally on at least 50 cm deep of low incline or flat blanket peat. Characterised by bog-mosses and cotton-grasses (especially hare’s-tail cotton-grass, and a mix of purple moor grass, deer grass and dwarf shrubs, usually with other indicator species present).

4.36 Lowland Heathland

4.37 Lowland heath is usually found below an altitude of 250 m and includes dune heath, wet heath, dry heath and valley mire communities. Generally it occurs on areas with acid soils and shallow peat. Typically it comprises of a complex mosaic of heathers, gorse, fine grasses, wildflowers and lichens, with at least 25% coverage of heathers and other dwarf shrubs.

4.38 Upland Heathland

4.39 Comprises of heath vegetation in moorland grazing with at least 25% cover of dwarf shrubs. Generally found on nutrient-poor, well-drained acid soils, including shallow peat (< 50 cm deep).

4.40 Upland Flushes, Fens and Swamps

4.41 Characterised by acid and base-rich mires in basin or valley topography, and springs, seepages and flushes, generally with water movement, on mineral or peaty soils in moorland grazing units. Typically dominated by cotton-grasses, sedges, spike-rushes and rushes with occasional wetland herbs and/or a carpet of mosses, especially ‘brown mosses’ and bog-mosses. Usually at least seasonally waterlogged they include other generally small features as part of a mire system, e.g., runnels, soakaways, species-rich rush beds, sedge lawns, but excluding purple moor grass and species-poor rush swards. Also excludes narrow (< 5 m wide) fringes of swamp adjacent to open standing water, reedbeds and blanket bog.

Wetland

4.42 Lowland Fens

4.43 Found on floodplains, on the fringes of open water, in valleys and around springs and flushes. The water table is close to or above the surface for most of the year and the soil under fens is waterlogged. They are differentiated from blanket bogs and lowland raised bogs in that they are fed by surface water and ground water (minerotrophic) in addition to direct rainfall.

4.44 Lowland Raised Bog


4.45 Comprises peatland vegetation, dominated by bog-mosses, cotton-grasses and heathers. A rare habitat, limited to areas where there are already acid peat deposits, mostly in the north-west. They develop mainly in cool, humid lowland areas e.g., head of estuaries, river floodplains and depressions where drainage is impeded. The bog is usually higher than the surrounding land and thus it only receives water as rainfall (ombrotropic).

Water

4.46 Sea

4.47 Sea, including brackish water at the mouth of estuaries.

4.48 Surface water

4.49 All other surface water present in the image.

Deciduous and Mixed Woodland

4.50 Traditional Orchards

4.51 This class is defined as five or more trees, although as this has been used for an EO map, a minimum area is more likely to contain a greater number of trees. Traditional orchards are characterised by either standard of half-standard fruit trees, usually planted at low densities (< 150 trees / ha), grown on vigorous rootstocks on permanent grassland. Mature trees should have 90% of their foliage above 1.5 m, with trunks at least 1 m in circumference at the base or form their first major fork at least 1.5 m above ground level.

4.52 Broadleaved, Mixed and Yew Woodland

4.53 Woodland with native and non-native species. Mixed woodland includes woodlands established by natural regeneration and planting. Native semi-natural woodland has also been grouped in this category; thus including native trees that do not obviously originate from planting. These areas are characterised by complete canopy cover.

Scrub and Bracken

4.54 Scrub

4.55 Any areas of scrub vegetation, whether that is birch, hawthorn, thick tall brambles or other scrubby species. The area of scrub has to be dense and continuous; scattered scrub species within a habitat are not included.

4.56 Bracken

4.57 Areas of vegetation dominated by bracken.

Inland Rock and Scree

4.58 Inland Rock and Scree

4.59 This class covers a wide range of rock types, varying from calcareous to acidic. Particular characteristic of higher altitudes, many rock habitats are inaccessible to grazing animals and unmanaged. Where vegetation is present, it is generally dominated by herbs, ferns, grasses, mosses or lichens.


Deciduous and Mixed Woodland

4.60 Broadleaved, Mixed and Yew Woodland

4.61 All broadleaved woodland, regardless of whether it is a plantation, semi-natural and natural woodland is mapped in this class.

4.62 Coniferous Woodland

4.63 All coniferous woodland, regardless of whether it is a plantation, semi-natural and natural woodland is mapped in this class.

4.64 Wet Woodland

4.65 Alder, birch and willows are usually the predominant tree species, but sometimes ash, oak, pine and beech occur on the drier riparian areas. Succession from open herbaceous wetlands results in a wide range of structures and compositions, determined by the composition of the original vegetation, the climate and the nutrient status.

National Ecosystem Assessment

4.66 The National Ecosystem Assessment outlines eight broad habitat types:

Mountains, Moorlands and Heaths

Semi-natural Grasslands

Enclosed Farmland

Woodlands

Freshwaters, Open waters, Wetlands and Floodplains

Urban

Coastal Margins

Marine

4.67 Table 2 outlines the possible conversion from the Living Maps broad habitat types to the National Ecosystem Assessment (NEA) broad habitat types. Where more than one NEA class is considered to match the equivalent detailed habitats, all of relevance has been listed.

Table 2: Conversion between Living Maps and NEA broad habitats

Living Maps Broad Habitat NEA Broad Habitat

Arable Enclosed Farmland

Coastal Coastal Margins


Coniferous Woodland Woodlands

Deciduous and Mixed Woodland Woodlands


Heathland and Moorland Freshwaters, Open waters, Wetlands and Floodplains

Mountains, Moorlands and Heaths


Improved Grassland Enclosed Farmland

Inland Rock and Scree Mountains, Moorlands and Heaths

Scrub and Bracken Mountains, Moorlands and Heaths

Semi-improved Grassland Semi-natural Grasslands

Species Rich Grassland Enclosed Farmland



Urban Urban

Water Freshwaters, Open waters, Wetlands and Floodplains

Marine

Wet Woodland Woodlands

Wetland Freshwaters, Open waters, Wetlands and Floodplains


5 Method

5.1 For a more detailed description and justification of the method used in the Living Maps project, please refer to the Technical Report in Annex 1.

Summary of Classification Steps

5.2 A summary of steps is outlined in the flowchart in Figure 10.

Figure 10: Flowchart of the classification process

Classification Steps

5.3 The classification process can be described in the following steps:

5.3.1 The image segmentation was applied in the commercial package Trimble eCognition version 9.2.1 using the Sentinel-2 summer and winter imagery, alongside the LPIS field parcel data. Area and perimeter values were calculated for each segment in QGIS and then divided to give an indication of object shape.

5.3.2 The training data (as described in Section 3) and segmented polygons were loaded in to the statistical software R (https://www.r-project.org/).

5.3.3 Zonal statistics for median, mean, mode and standard deviation were calculated from the input layers for the segmented polygons.

Extract zonal statistics for the training data

from segmented

polygons

Calculate zonal stats from input layers for

the segmented

polygons

Load training data and

segmented

polygons

Impute missing values and calculate

vegetation indices

Calculate user and producer

accuracy

Use model to predict class for each segmented

polygon

Sample training data to select a similar number of points per class and split into training and

test datasets

Merge results to

shapefile

Weight the

training data

Rerun Random Forest using most

significant variables

Fit Random Forest model to

training data

https://www.r-project.org/


5.3.4 Missing values (due to either cloud or snow) were imputed using nearest neighbour averaging.

5.3.5 Area and perimeter values were calculated for each segment in QGIS and the ratio used to give an indication of object shape.

5.3.6 Vegetation indices were calculated using the Sentinel -1 and Sentinel-2 data following data exploration into which indices have the most significance to each class. A list of indices used can be found in the final script.

5.3.7 Zonal statistics were extracted for the training data from the segmented polygons that contained them.

5.3.8 80% of the sample training data were used as the training dataset with the remaining 20% used to test the classification accuracy. Weighting was given to the training and test data points based on the recorded tier, with lower tiers given a higher weighting. Oversampling was implemented when the training data did not have enough samples.

5.3.9 A Random Forest model was fitted to the training data and then rerun using the top 42 most significant variables from the previous step as this was found to improve the overall classification accuracy.

5.3.10 User and producer accuracy figures were calculated using the independent test data.

5.3.11 The model was used to predict the classes for each segmented polygon. The results were merged and written to a shapefile.

Random Forest Explanation

5.3.12 Random Forest was selected as the most successful statistical classification method. A multinomial classifier, Random Forest is a fast and parallelisable algorithm (compared to other statistical classification methods such as Support Vector Machines). Random Forest has a low computation requirement and can therefore be rerun without large time investment. The method is extremely flexible, able to deal with continuous and categorical variables, and able to discern interactions between variables.

5.3.13 A Random Forest is built out of a large number of decision trees, which have been trained on a subset of the training data and a subset of the training variables. Trees are built by making a series of sequential partitions of the training data, each one based on the value of a single variable, until the data has been separated into groups with identical classes, or a maximum number of partitions are reached. New data points can then be fed through each tree to produce a prediction. The predictions from each tree are aggregated, and the class with the most “votes” is considered the best overall prediction. Each of the individual trees is a weak predictor of class, but together the forest has stronger predictive power.


6 Results and Analysis

Results Summary

6.1 To assess the accuracy of the classification, a confusion matrix was produced. This displays the accuracy of class predictions using the test data. The overall accuracy (proportion of correctly classified polygons) from the Catchment Pioneer error matrix shows 66% accuracy for detailed habitats and 74% accuracy for broad habitats while the Landscape Pioneer shows 67% accuracy for detailed habitats and 75% accuracy for broad habitats.

6.2 Individual class scores range more considerably, with some classes achieving 100% accuracy in both the detailed and broad classifications. For broad classes these include Coniferous Woodland, Mud, Sand or Shingle and Urban in the Catchment Pioneer and Arable, Coniferous Woodland, Mud, Sand or Shingle and Water in the Landscape Pioneer. In the Catchment Pioneer, 20 out of 30 detailed classes achieved at least 60% accuracy while in the Landscape Pioneer 16 out of 27 detailed classes achieved at least 60% accuracy. In the Catchment Pioneer, 6 classes (Acid Grassland, Lowland Meadows, Semi-improved Grassland, Traditional Orchards, Upland Flushes, Fens and Swamps and Upland Heathland) recorded accuracies below 50% while in the Landscape Pioneer, 5 classes (Lowland Meadows, Scrub, Upland Flushes, Fens and Swamps, Upland Hay Meadows and Upland Heathland) recorded below 50% accuracy. The probability of the predicted class from the Random Forest model is attributed on per segment basis as part of the output product.

6.3 It should be noted that these accuracy scores are not a perfect indicator of the success of the model, as they are based on prediction on a subset of the training data, which we have found to be imperfect in some cases, due to a number of factors (mislabelling of FEP data, poor overlap of FEP land parcels with segmented polygons, etc.).

Mapped Outputs

6.4 The habitat maps for the Catchment and Landscape Pioneer Areas are shown in Figure 11 and Figure 12. As would be expected, the lowland areas are dominated by Improved Grassland, with this class making up the largest vegetation class in both the Catchment and Landscape Pioneers at 1,054 km2 and 713 km2 respectively. Upland areas are characterised by areas of Blanket Bog and Acid Grassland in the Catchment Pioneer whilst in the Landscape Pioneer Blanket Bog and Upland Heathland dominate. Bracken and Upland Heathland are also prevalent in uplands and upland fringes of the Catchment Pioneer classification. In the Landscape Pioneer, upland fridges are characterised by Purple Moor Grass and Rush Pastures. The areas of highest elevation in the Catchment Pioneer are classified as Inland Rock and Scree. Urban areas and waterbodies in both the Landscape and Catchment Pioneer Areas were readily identifiable using the approach, having distinctive signals.

6.5 Figure 13 and Figure 14 show the classification probability, attributed per segment. These figures suggest that there is greater confidence in the upland areas and around the coast for both the Catchment and Landscape Pioneer Areas. However this is more likely to reflect the wider diversity of lowland habitats, many of which may appear to be spectrally similar. Within the output, a secondary class has also been attributed. This represents the second most likely class predicted by the model. Where the probabilities between the first and the second class predicted are similar, this may be representative of mosaics of different habitat classes occurring within the segment or where classes cannot be easily separated.


Fig

ure

11: C

atc

hm

en

t Pio

neer D

eta

iled

Hab

itat M

ap


Fig

ure

12: L

an

dsc

ap

e P

ion

eer D

eta

iled

Hab

itat M

ap


Fig

ure

13: P

rob

ab

ility v

alu

es a

ttribu

ted

by

seg

men

t – C

atc

hm

en

t Pio

nee

r


Fig

ure

14: P

rob

ab

ility v

alu

es a

ttribu

ted

by

seg

men

t – L

an

dsc

ap

e P

ion

ee

r


Graphed outputs

6.6 The following graphs (Figure 15, Figure 16, Figure 17 and Figure 18) show the breakdown of user accuracies (proportion of segments predicted to be each class that are correct) for the Catchment and Landscape Pioneer Areas.

Figure 15: Detailed habitat accuracy values – Catchment Pioneer

Figure 16: Detailed habitat accuracy values – Landscape Pioneer


Figure 17: Broad habitat accuracy values – Catchment Pioneer

Figure 18: Broad habitat accuracy values – Landscape Pioneer


6.7 Figure 19 and Figure 20 illustrate producer accuracies for broad habitats (proportion of segments that are correctly predicted) across the classes. The similarity in spectral signatures may explain some of the misclassifications shown in the producer error graphs. For example, in both the Pioneers, Scrub and Bracken has been misclassified as Deciduous and Mixed Woodland, a habitat which may have a similar spectral signature and found in similar context to particularly Scrub. Similarly, Wet Woodland shows confusion with Deciduous and Mixed Woodland, also attributable to very similar spectral signatures. Misclassification of Inland Rock and Scree to Mud, Sand or Shingle is also understandable, as is confusion between grassland types. Other producer errors have a more complex explanation and may be to do with the locations and resolution of training samples or the ability of EO data to classify such habitats using the range and resolution of imagery and ancillary datasets used in Living Maps.

Figure 19: Broad habitat producer error values – Catchment Pioneer

Figure 20: Broad habitat producer error values – Landscape Pioneer


Key Findings

6.8 The use of a high resolution topographic model (IHM) was particularly useful in identifying habitats such as Inland Rock and Scree. Moreover the derived layers of height and slope were important for identifying Blanket Bog. It was also noted that the addition of proximity rasters and climate data significantly improved the classification. Traditional Orchards for example, were highly correlated with distance from buildings. The addition of Sentinel-1 data facilitated structural characteristics in the habitats to be identified. For example, Wet Woodland identification was supported by the median of a winter Sentinel-1 VH scene.

6.9 Additional statistics, such as standard deviation, were also found to be beneficial. For example, Sentinel-2 summer blue band standard deviation was identified as one of the most significant variables across multiple habitats. This may be because the blue band is useful for differentiating between soil and vegetation and between deciduous from coniferous woodland, due to the interaction with water. Thus the calculation of standard deviation allowed habitats with a greater range of spectral variation to be identified.

6.10 Finally, the inclusion of specific vegetation indices was found be beneficial to the classification of specific habitats. Whilst Random Forest does have the capability to identify basic interactions between variables, the addition of identified relationships was found to increase classification accuracy and reduce noise.

Limitations

6.11 High cloud cover was a limitation within the Sentinel-2 winter scene with a cloud bank covering a large proportion of the eastern edge of the scene. Moreover, on the summer scene, some of these areas were covered by snow; consequently the modelling used values derived from the other ancillary datasets and the Sentinel-1 imagery to identify the habitat. The primary winter Sentinel-2 scene for the Landscape Pioneer also had a high percentage of cloud cover, concentrated in the north west of the scene. This was however mitigated by a second scene captured 23 days later, only covering the north west of the area. These scenes were used in combination to represent the winter variations in habitats, an approach which is considered to have been successful.

6.12 Whilst imputing values of areas covered by cloud was very effective for most of the scene, there are some examples where it did not predict the correct habitat. One of the more obvious examples of this is a water body in the uplands of the Landscape Pioneer. In this example, there were clouds at the location in both the summer and winter Sentinel-2 scenes; the location was also outside the extents of the secondary winter scene. This example is shown in Figure 21. However, it should be noted that these circumstances are the exception and would be improved by imagery with a lower percentage cloud cover.

Figure 21: An example of errors created from imputing values where clouds from multiple image dates overlap. From left to right: Landscape Pioneer classification, Sentinel-2 summer, Sentinel-2 winter and

Sentinel-2 summer and winter masked scenes showing overlap in no data areas and segmentation.


6.13 It is acknowledged that sample locations are likely to be biased by the coverage of Environmental Stewardship agreements from which they were derived and the data from the FEP maps being summarised at LPIS parcel level. Moreover, small or rare features in the landscape are likely to have fewer training samples and therefore this reduces the potential to identify such features. Upland Flushes, Fens and Swamps for example in the Catchment Pioneer lacked training points. Additionally, when these features were visited on a field validation it was recognised that due to their size there may be difficulties in mapping the class. In the Landscape Pioneer, mosaic or fragmented habitats such as Upland Heathland caused similar problems.

Results of Field Validation

6.14 Field work was undertaken in Cumbria to validate the initial map and create new in situ data for the Catchment Pioneer model. The areas visited were Stone Arthur, north of Grasmere, and Red Screes, north of Ambleside. Prior to the visit, the existing model predicted these areas to be made up principally of bracken, acid grassland, and blanket bog. Some examples can be seen in Figure 22.

6.15 In situ data were collected by photographing areas of interest using a smartphone. Modern smartphones contain very good GPS navigation systems with comparable accuracy to handheld GPS. The GPS data can later be extracted into a shapefile in QGIS and the image data manually labelled to create training points.

6.16 Field data of the area suggested that a large amount of the existing bracken classifications were correct. We found that bracken tended to be situated on slopes, below around 600 m in altitude. Arable and Horticultural classifications were also generally good, however there were no examples of Blanket Bog witnessed, suggesting this was being over classified. A number of examples of small flushes and fens were collected, however these tended to be small areas (25 - 100 m2). As the maximum resolution of the Sentinel-2 scenes are 100 m2, and most of the segmented polygons are much larger than this, the conclusion was reached that it may be difficult to map these habitats. Similar issues are thought present when mapping areas of Scrub.

6.17 Finally, the large amount of bare rock and scree witnessed, particularly at Red Screes, led to the conclusion that an inland rock and scree class should be added to the data; a number of training points were collected for the new class. This has proved more difficult to implement than expected, as the often highly heterogeneous mixture of bare rock and acid grassland in places has caused some grassland areas to classify as rock, and vice versa.


Figure 22: Examples of Bracken, Acid Grassland, Inland Rock and Scree and Flushes training

points collected at Stone Arthur and Red Screes


7 Discussion

Knowledge Transfer and Dissemination

7.1 A central component of the Living Maps project relates to knowledge transfer and dissemination. The platform chosen to host the output classification maps was ArcGIS Online. One advantage of this is that the user does not need desktop GIS or associated skills to view the output. Furthermore, the output can easily be shared to a wide number of people without the transfer of files. Additionally, records can be kept of the number of views the maps receive, which may be an indicator of their usefulness.

7.2 As well as the dissemination of outputs, an objective of the project was to share the methods used to produce the classification so they could be adopted more widely. All of the process has been built on using open source software apart from the segmentation in eCognition; there is therefore great potential for the approach to be applied more widely. As described in the method, the R scripts used in the Living Maps project are available to download under an open source license. Elements of the code are already being applied to other projects, for example looking at changes in productivity on Sites of Special Scientific Interest and how this relates to condition of grassland sites.

7.3 Throughout the project, engagement was maintained with the Landscape and Catchment Pioneers, through teleconferences, workshops and presentations. Due to the timing of the two initiatives (with the Pioneers projects still being established), there was limited opportunity to translate the outputs of Living Maps work into products to help the Pioneers with Natural Capital based decision making. A parallel contract, run by Entec undertook analysis to support Natural Capital assessments in Pioneer Project areas. As Living Maps was not available, Entec used Land Cover Map 2007 as the basis of mapping Natural Capital in each area. Following on from both projects there is an opportunity to explore what difference having Living Maps available would have made to this work. With the involvement of Defra’s Environment Analysis Unit in both the Living Maps and the Pioneer projects, these requirements may be followed up.

7.4 One specific requirement identified by the Landscape Pioneer was the ability to translate the outputs from Living Maps into the classes used by the National Ecosystem Assessment (NEA). The correlation between the NEA and Living Maps classes is stated in this report (Table 2). Discussions with Pioneer Leads have also highlighted interest in extending the Living Maps approach to cover the Greater Manchester Urban Pioneer project and possible other areas of the country.

7.5 Whilst the Living Maps project has produced outputs maps for the two areas, it is envisaged that these may be further refined in response feedback received and further in situ data, to help address some of the issues already highlighted in the report. Because the classification process is fully automated, incorporating new training data into the model is a straightforward process.

7.6 Defra are also engaging with the Centre for Ecology and Hydrology (CEH) through involvement in a steering group on land use change in the UK using satellite data, led by the Department for Business Energy and Industrial Strategy (BEIS). Through this engagement, we aim to ensure there is synergy between Living Maps and other land cover and land use change products.


The possible impact of the evidence on the specified Defra policy outcomes

7.7 The Living Maps project brings together two initiatives within Defra. The first is the Earth Observation Centre of Excellence, which aims to bring together expertise from across the Defra, industry and academia to match policy and operational needs to earth observation data. The second is the Defra 25 Year Environment Plan, which includes the use of new technologies and encourages innovation.

7.8 In relation to Natural Capital based decision making, habitat maps can form an integral component. The importance of incorporating the natural environment into national accounting frameworks has been highlighted by The Economics of Ecosystem Biodiversity (TEEB) global reports (TEEB, 2010). Knowing the location and extent of Natural Capital assets is a primary step towards implementing decision support systems. The Living Maps project has shown that a good level of accuracy can be achieved through satellite-based habitat classification systems. By extrapolating ecological knowledge, areas with previously little or no habitat information can be attributed. In these areas, satellite-based habitat maps are the most cost effective solution we have to date.

7.9 Moreover, by defining habitats across large areas and decreasing the time taken to expand to new areas, semi-automated classification methods, like those demonstrated in the Living Maps project, can facilitate the adoption of more holistic Natural Capital based products which can be applied on a landscape-scale approach. Approaches like this will significantly contribute to the creation of resilient landscapes, enabling policy makers and practitioners to provide the type of evidence that can be used to provide a picture of the resource from site to national scale.

A cost effective method

7.10 The Living Maps project represents a significant saving in comparison to previous methods for habitat maps creation. A cost of £630,000 (exclusive of ground validation costs) is estimated (£4.75 / km2) for scaling this project to the extent of England’s terrestrial area. The future availability of Analysis Ready Sentinel-1 data will further reduce this figure. In contrast, MEOW 3 has a projected average cost of £5.4 million (£36.60 / km2). This substantial difference (88%) demonstrates that Living Maps represents an economically robust choice while delivering a comparable end product.


8 Conclusions

Summary

8.1 The aims of the Living Maps project have been achieved through the application of image processing and object-based classification steps. Using predominately open source solutions, a cost effective model has been developed with the potential to be rolled out across England. Advances in automation and use of machine learning techniques has reduced the skilled personnel time needed to configure, refine and re-run the classification. The project has also made use of predominantly open data sources or those without licensing restrictions meaning that the outputs may be more easily disseminated and applied to other uses. The use of open data sources also makes it easier for the Living Maps processes to be replicated by other organisations. Further work has indicated that with more representative training data, higher accuracies would be readily achievable.

8.2 Furthermore, the Living Maps project has facilitated a relationship between the Earth Observation Centre of Excellence and the Catchment and Landscape Pioneer Areas. These relationships stand to benefit both groups, working towards a synergetic approach to Natural Capital and the Defra 25 Year Environment Plan. The benefits also cover the significant increase in the Evidence Earth Observation Service expertise and capability as a direct result of involvement in the Living Maps project.

Further work

8.3 The outputs produced as a result of the project are to be seen as the first step in a dynamic process. Future work includes:

Inclusion of additional data sources, for example the National Forest Inventory and

soils data.

Addition of new sources of training data.

Incorporation of feedback concerning the current classification.

Further investigation of open source methods of segmentation to enable the entire

process to make use of open source software.

Replacing the licence-restricted IHM with an open data alternative.


Annex

This Annex outlines the technical aspects of the Living Maps in greater detail. Section 1 provides additional information on the datasets used and Section 2 provides a more detailed review of the method, including a breakdown of the script.


1 Dataset Information

1.1 This section provides more detail on the Sentinel-1, Sentinel-2 and Linear Spectral Unmixing datasets used in the project.

Sentinel-1

1.2 Sentinel-1 is a polar-orbiting, all-weather, day-and-night radar imaging mission for land and ocean services. The mission consists of a two satellite constellation each in a 12 day orbit, together providing orbit revisit times of six days at the equator. The radar instrument transmits and receives in C-band (5.405 GHz) at a resolution of 5 x 20 m in Interferometric wide-swath mode. This is the mode that is used over land masses in Europe, including England. The data is processed and stored as Level 1 Single Look Complex (SLC) and Level 1 Ground Range Detected (GRD) products by the ground segment of ESA. The SLC product contains the intensity of returns and phase information, whereas the GRD product does not contain the phase information, due to the enhanced processing that it receives. This GRD data is multi-looked and projected to ground range using the earth ellipsoid model. These datasets are available for download through the Copernicus portal (https://scihub.copernicus.eu/dhus/#/home).

1.3 For this project the GRD scenes covering the sites were downloaded from the portal. The following scenes were downloaded:

Landscape Pioneer:

S1A_IW_GRDH_1SDV_20160108T063109_20160108T063134_009399_00D9C6_8DAF

S1A_IW_GRDH_1SDV_20160120T063109_20160120T063134_009574_00DECB_0F51

S1A_IW_GRDH_1SDV_20160201T063106_20160201T063131_009749_00E3EF_85FA

S1A_IW_GRDH_1SDV_20160706T063120_20160706T063145_012024_012923_4AAE

S1A_IW_GRDH_1SDV_20160718T063121_20160718T063146_012199_012EE3_953C

S1A_IW_GRDH_1SDV_20160730T063121_20160730T063146_012374_013496_5EEC

Catchment Pioneer:

S1A_IW_GRDH_1SDV_20160108T062954_20160108T063019_009399_00D9C6_850D

S1A_IW_GRDH_1SDV_20160108T063019_20160108T063044_009399_00D9C6_B8BA

S1A_IW_GRDH_1SDV_20160120T062954_20160120T063019_009574_00DECB_BCE9

S1A_IW_GRDH_1SDV_20160120T063019_20160120T063044_009574_00DECB_7549

S1A_IW_GRDH_1SDV_20160201T062951_20160201T063016_009749_00E3EF_5DE3

S1A_IW_GRDH_1SDV_20160201T063016_20160201T063041_009749_00E3EF_4E5F

S1A_IW_GRDH_1SDV_20160706T063005_20160706T063030_012024_012923_8BD6

S1A_IW_GRDH_1SDV_20160706T063030_20160706T063055_012024_012923_60B2

S1A_IW_GRDH_1SDV_20160718T063006_20160718T063031_012199_012EE3_34DE

S1A_IW_GRDH_1SDV_20160718T063031_20160718T063056_012199_012EE3_153D

S1A_IW_GRDH_1SDV_20160730T063006_20160730T063031_012374_013496_3B77

S1A_IW_GRDH_1SDV_20160730T063031_20160730T063056_012374_013496_794C

1.4 These scenes all come from Sentinel-1A, as this was the only active satellite in the constellation during the capture times required. There was not a single orbital swath covering the Cumbria study area, and so two scenes from each orbit had to be downloaded to cover

https://scihub.copernicus.eu/dhus/#/home)


this site. This meant that an additional processing step was required for these scenes, explained below.

1.5 Each set of scenes was processed through the processing chain shown in Figure 23, using the ESA’s SNAP Toolbox for Sentinel-1 data. Each set is a site, and season, so for North Devon, 3 dates of scenes have been downloaded for winter and 3 for summer. 3 different dates of scenes were used for the Multi-Temporal Filter. Just a single scene from each site and season was used from the processing in the classification.

Cross Correlation

Terrain Correction


Calibration: to

Beta0

Remove GRD

Border Noise

Thermal Noise

Removal

Apply Orbit File

Stack

Warp

Terrain Flattening

Calibration

Co-Registration


Terrain Flattening and Terrain Correction


Figure 23: Full processing chain for Sentinel-1 data using the ESA SNAP Toolbox. The right processing chain shows an overview of the stages in processing the Sentinel-1 data, and is as shown in the summary. The left chain is the full processing chain showing the actual steps run through SNAP.

1.6 For Cumbria, an extra step was required in the processing chain, this was Slice Assembly. This process was also carried out in SNAP, and run before the processing chain shown in Figure 23. It combined the two neighbouring scene rows together, creating a larger seamless scene which could then be processed as normal through the processing chain.

1.7 The first 3 calibration steps: Remove GRD Border Noise, Thermal Noise Removal and Apply Orbit File, all run using the default settings. This includes Applying the Orbit File, using the Sentinel Precise Orbit information and downloading this automatically. The final calibration step, Calibration, creates Beta0 virtual bands, rather than the standard Sigma0. This is done, because the Terrain Flattening steps require Beta0.

1.8 The Co-Registration stage contains 3 steps which stack and correct the data for uniform alignment. This stage is run twice, once for each return type (VV & VH), as well as repeated for each season and each site. The first step, Stack, combines the 3 scenes from each season into a single dataset, using the default settings. Cross Correlation is then undertaken and creates 200 GCP’s automatically across the combined scene, with these tested to ensure they are on land. Two iterations are undertaken, with a GCP tolerance of 0.5. This then leads into the Warp step, which carries out the warp using Bilinear Interpolation, a RMS Threshold pixel accuracy of 1 and polynomial order of 1.

1.9 The Multi-Temporal Filter uses all 3 scenes within the now stacked and corrected dataset. It carries out a Refined Lee filter across the scenes allowing for speckling to be identified and corrected. Using a multi-temporal filter maintains the resolution and sharpness of the data, which would otherwise be reduced through use of a single speckle filter.

1.10 The final stage is Terrain Flattening and Terrain Correction. Now that the advantages of using 3 scenes for the Multi-Temporal Filter have been achieved, only a single date of scene is processed out through terrain flattening and terrain correction. As both sites are covered by the same satellite orbit, this date is the same for both and was the first date scene for each season: 08/01/2016 and 06/07/2016. The terrain flattening took the Beta0 product, and carried out the flattening algorithm, using a bilinear interpolation and a 10 m version of the Integrated Height Model DTM. The DTM was in the WGS 84 projection to match the satellite data and minimise on the fly transformations which would have potentially caused errors in the results. This produced a Gamma0 product, which was then processed through the terrain correction step. The step applied an Earth Gravitational Model and carried out Radiometric normalisation, and again used the Integrated Height Model DTM at 10 m resolution in WGS 84 projection for the terrain corrections. This completed the processing of Sentinel-1 data in the ESA SNAP Toolbox.

1.11 The Sentinel-1 data was then finalised for use in the classification by reprojection into British National Grid Projection and clipping to the study area.

Sentinel-2

1.12 Sentinel-2 is a polar-orbiting, wide-swath, high-resolution, multi-spectral imaging mission for land surface conditions. The mission consists of a two satellite constellation providing orbit revisit times 10 days at the equator with one satellite, and 5 days with 2 satellites under cloud-free conditions which results in 2-3 days at mid-latitudes. The optical instrument payload samples 13 spectral bands: four bands at 10 m, six bands at 20 m and three bands at 60 m spatial resolution. The orbital swath width is 290 km.


1.13 The data is processed and stored at Level-1C as top-of-atmosphere reflectance in cartographic geometry and includes the application of basic geo-referencing and orthorectification as well as assignment of radiance and reflectance values. The Level-2A product requires processing through the Sentinel-2 Toolbox for bottom-of-atmosphere in cartographic geometry. For Level-1C and Level-2A, the granules, also called tiles, are 100 x 100 km2 ortho-images in UTM/WGS84 projection. These data are available for download through the Copernicus portal at the same location as the Sentinel-1 data.

1.14 ‘Analysis Ready Data’ (Level-2A) was provided by JNCC as granules, downloaded in GeoTIFF format in British National Grid (27700). There are many advantages in using these data rather than the Level-1C product that could be downloaded directly from ESA. Firstly, Sentinel-2 records different bands at different resolutions, including 20 m bands and 10 m bands; this is reflected in the data downloaded from ESA. However, data processing is more efficient if the data pixels are of a consistent size, so this task uses resampling to convert the 20 m bands to 10 m (Bands 5, 6, 7, 8A, 11 and 12 are resampled to 10 m using a nearest neighbour transformation in GDAL). Bands are also stacked, which makes the data easier to use.

1.15 The presence of the atmosphere influences the radiation from the ground to the sensor and this process corrects this. JNCC runs a basic automated process using a Dark Object Subtraction (DOS) method using the Remote Sensing and GIS software Library (RSGISLib: http://www.rsgislib.org/index.html). Finally, JNCC has improved the usability of data by adding names to reflectance bands and building image overviews and pyramids into the files. There is currently no cloud masking on these data.

1.16 The following scenes were acquired as Analysis Ready Data:

Landscape Pioneer:

S2A_tiles_20160719_37_5

S2A_tile_20161106_37_5

S2A_tile_20161129_80_5

Catchment Pioneer:

S2A_tiles_20160602_80_3

S2A_tiles_20160602_80_4

S2A_tiles_20160314_80_3

S2A_tiles_20160314_80_3

1.17 There was not a single orbital swath covering the Catchment Pioneer study area, and so two scenes from each orbit had to be downloaded to cover this site. The images were merged using GDAL. Due to a high percentage of cloud on the Landscape Pioneer winter scene (S2A_tile_20161106_37_5), a secondary scene covering a different proportion of the area of interest from 23 day later was also used (S2A_tile_20161129_80_5).

1.18 Clouds and mist prevent satellites from detecting reflectance data from the surface of the earth, and so for analytical purposes data from these areas should be masked out from the raster surface. Polygons were manually identified around a cloud obscured and data values were set to 0.

1.19 The Sentinel-2 data was then finalised for use in the classification and clipped to the study area.

http://www.rsgislib.org/index.html


Linear Spectral Unmixing

1.20 Linear Spectral Unmixed (LSU) images were calculated from summer and winter Sentinel-2 scenes for shade, productive vegetation and non-productive vegetation using the SNAP Toolbox.

1.21 The signal detected by a sensor into a single pixel is frequently a combination of numerous disparate signals. This is known as a mixed pixel – an example diagram is given in Figure 24. Mixed pixels either occur when the resolution is low enough that disparate materials can jointly occupy a single pixel, or when distinct materials are combined into a homogeneous mixture. The resulting spectral measurement in both examples will be the composite of the individual spectra.

Figure 24: An example diagram showing a mixed pixel composed of grass, trees and soil

1.22 Mixture modelling is built on the assumption that within a given scene, the surface is

dominated by a small number of distinct materials with relatively constant spectral properties. These substances are called endmembers, with the fractions in which they appear within a mixed pixel referred to as fractional abundances. There exists a linear relationship between the fractional abundance of the substances comprising the area being imaged and the spectra in the reflected radiation. The aim of spectral unmixing is to find the endmembers that can be used to ‘unmix’ other mixed pixels.

1.23 The SNAP toolbox provides a method for calculating LSU images. The user must first select the endmembers for shade, productive vegetation and non-productive vegetation using the ‘Pin Placing Tool’. These selections were made by continuously checking the ‘Spectrum View’ window to ensure the signature was consistent with what would be expected for each of the pure pixel endmembers.

1.24 For each endmember category, four examples were chosen. The selected endmember was the spectral signature that represented an average of these examples. Example samples for

Grass

Tree

Soil

40 m

30

m

Mixed Pixel


PV are shown in Figure 25 with a zoomed in graph shown in Figure 26. Spectral samples were saved to input to the LSU tool.

Figure 25: Example Spectrum View

Figure 26: Example Spectrum View - zoomed in


1.25 The final stage was to input the endmembers into the LSU tool and select the parameters. Default parameters were chosen, using a Constrained LSU with a minimum spectral bandwidth of 10.0.

Figure 27: Spectral Unmixing window in SNAP


2 Method

2.1 This section describes in further detail the segmentation and classification steps used.

Segmentation Steps

2.2 The segmentation process can be described in the following steps:

2.2.1 The segmentation was applied in Trimble eCognition version 9.2.1. The processing of the segmentation used the Sentinel-2 summer and winter imagery, alongside the LPIS field parcel data.

2.2.2 Segmentation first took place into a Chessboard and was then synchronised to LPIS field parcels. This used the algorithm ‘Synchronise Image Object Hierarchy’ and created segments exactly based on LPIS.

2.2.3 These were then classified into the following classes: Fields greater than 10ha, Fields less than 10ha and Non Fields. This used the ‘Classification’ algorithm, and rules inside the classes based on Overlap with the LPIS thematic layer and object Area. Non Fields are those which contain no overlap with the LPIS, and the field classes that do overlap with the LPIS and are then separated on their area.

2.2.4 Next ‘Multi-Resolution Segmentation’s took place, using the Sentinel-2 summer and winter imagery.

2.2.5 Firstly this was undertaken on fields greater than 10 ha and non-fields. The multi-resolution segmentation used the following band weightings for both summer and winter images:

Blue – 1

Green – 1

NIR – 5

Red – 5

Red Edge5 – 1

Red Edge6 – 1

Red Edge7 – 1

Red Edge8A – 5

SWIR1 – 2

SWIR2 – 2

A Scale parameter of 120, Shape criterion of 0.1 and Compactness criterion of 0.8 were also used. This broke large fields into sub parcels, and enabled splitting of fields which partially contained woodland, or were cropped by 2 different crops. It also created segments in non-fields, breaking these areas down to better represent the variation in habitats occurring.

2.2.6 Secondly Fields less than 10 ha were segmented. This was set up to try to only create additional segments when there were large variations within the field; such as if features like


woodland or scrub were present. To achieve separation of woodlands and scrub the segmentation used layer weightings more sensitive to these features, with highest weightings in the SWIR bands: Blue – 0

Green – 0

NIR – 2

Red – 1

Red Edge5 – 1

Red Edge6 – 1

Red Edge7 – 1

Red Edge8A – 1

SWIR1 – 3

SWIR2 – 3

The Scale parameter was again 120, however with Shape and Compactness criterion of 0.3 and 0.2. This segmentation process achieved separation of areas of scrub and woodlands in these smaller fields however did also in some fields create an extra segment on the field boundaries. This is thought to be due to large field margins sometimes being quite different to the crops and distinctive in the imagery.

2.2.7 These results were then exported from eCognition into a polygon shapefile. Area and perimeter were then calculated for each segment in QGIS and then divided to give an indication of object shape.

Classification

2.3 The image classification for Living Maps was undertaken using the R programming language for statistical computing within the R Studio integrated development environment.

2.4 The R packages used for producing Living Maps were:

Table 3: R packages used in the Living Maps project

Package Description

rgdal Bindings for the Geospatial Data Abstraction Library

raster Geographic data analysis and modelling

rgeos Interface to Geometry Engine - Open Source (GEOS)

randomForest Classification and Regression by randomForest

impute Imputation for microarray data

reshape2 Reshaping Data with the {reshape} Package


ggplot2 Elegant Graphics for Data Analysis

foreach Provides Foreach Looping Construct for R

doSNOW Foreach Parallel Adaptor for the 'snow' Package

plyr The Split-Apply-Combine Strategy for Data Analysis

sqldf Perform SQL Selects on R Data Frames

data.table Extension of data.frame

2.5 The final code for Living Maps is hosted on GitHub at:

https://github.com/NE-EEOS/LivingMaps

Load training data and segmented polygons

2.6 Inputs to the model were the segmented polygons, training data and raster layers. The training data consisted of two shapefiles containing point data: one with habitat classes, derived primarily from Farm Environment Plan data, and one with non-habitat classes, derived primarily from Ordnance Survey VectorMap district polygons.

Calculate zonal stats from input layers for the segmented polygons

2.7 For each of the segmented polygons, statistical data were calculated by extracting values from the raster layers. For the imagery layers (multi-spectral and synthetic aperture radar data) mean, median and standard deviation were calculated for each band. Median was used in addition to mean to help reduce the impact of outliers within each segmented polygon. Standard deviation was used to provide a measure of the heterogeneity within each polygon. For the topography, proximity and climate only mean values were calculated. Modal values were calculated for the categorical land use layer.

2.8 Zonal statistics were extracted for a total of the 93 combined layers and statistics for over 300,000 polygons. To reduce the processing time a rasterised version of the segmented polygon layer was used as calculating zonal statistics between two raster layers, even with the requirement for resampling to align the rasters, was found to be faster than calculating statistics using a polygon layer. The segmented raster layer was split into 25 tiles to enable parallel processing. Zonal statistics were then calculated for each tile, processing the raster input layers in parallel. Because segmented polygons may have been split across tiles, combined values were calculated for such polygons. For mode, median and standard deviation statistics these values were estimated.

2.9 The outputs from the zonal statistic consisted of a table with a row for each segmented polygon and a column for each of combination of raster layer and statistic. This table was then saved to be used for the model processing. The zonal statistics only needed to be recalculated when the input layers changed.

Impute missing values and calculate vegetation indices

2.10 Whilst Sentinel-2 images were selected to be as cloud free as possible there were still areas of cloud on the images. To address this, the areas of cloud in each image were set to no data values. In order to classify areas covered in cloud, missing values from the images were imputed from the other Sentinel-1 and Sentinel-2 images and proximity layers. For the Landscape Pioneer an additional winter Sentinel-2 image was used to help to impute areas of


cloud on both the existing summer and winter images. Missing values within the zonal statistics table were estimated using k-nearest neighbour imputation. In short, for each row of data with missing values, the 10 nearest neighbours by Euclidean Distance are found. The missing elements of the row are then imputed by taking the mean of that element across the neighbours. Imputation was done using the “impute” package from Bioconductor, which was designed for gene expression analysis but is generalisable to any large dataset.

Vegetation indices

2.11 In addition to normalised differential vegetation indices (NDVI) and normalised differential wetness indices (NDWI) for the two Sentinel images, a number of indices were added to the zonal statistics table. These were identified by generating combinations of ratio, difference and normalised difference indices for all bands within each of the Sentinel-1 and Sentinel-2 images. A model was then created for each of the classes in the training data using gradient boosted machines to identify the importance of each indices for each class.

Extract zonal statistics for the training data from segmented polygons

2.12 The training points for the habitat and non-habitat classes were combined into a single spatial layer and then linked to the zonal statistics values for the segmented polygons they intersected.

Sample training data to select a similar number of points per class and split into training and test datasets

2.13 Due to the nature of the training data there were more training points for some widespread habitats, such as upland heathland, than the more restricted habitats, such as fens. To prevent over-classification the same sample size was used for all training classes. Random sub-sampling was applied to the training points, weighted by spatial accuracy. A randomisation seed was applied to provide a consistent training data set allowing comparison between model outputs. The training data were then split into a training and test datasets (80% and 20% respectively per class) with the latter used to calculate the accuracy assessment.

Fit Random Forest model to training data

2.14 A Random Forest model was fitted between the detailed classes from the training data and the zonal statistics from the input layers, including all the derived indices.

Rerun Random Forest using most significant variables

2.15 To improve the model fit the Random Forest model was re-run, this time using the most significant variables identified by the first run of the model. Using approximately forty of the most significant variables was found the result in the best model fit for both the Catchment and Landscape pioneer areas. Table 4 shows an example result from gradient boosted machines, indicating the likely influence of each variable within the Random Forest classification.

Calculate user and producer accuracy

2.16 The Random Forest model was used to predict detailed classes for the test dataset. User and producer accuracies were then calculated for each class as well as total accuracy.


Use model to predict class for each segmented polygon

2.17 The Random Forest model was then used to predict the detailed class for all of the segmented polygons using the data from the zonal statistics table. The first and second most likely class were attributed along with their corresponding probabilities.

Merge results to shapefile

2.18 The outputs were then joined to the segmented polygons and saved to a shapefile.

Table 4: Variable importance

Variable Value

height 4.71

slope 3.99

dist_foreshore 3.87

dist_tidalwater 3.46

dist_woodland 3.43

max_temp 3.41

min_temp 3.21

dist_building 3.09

annual_rainfall 2.61

S2_summer_blue_sd 1.95

di_S2_summer_green_median_S2_summer_rededge5_median 1.92

S2_summer_rededge5_median 1.86

di_S2_summer_blue_S2_summer_rededge5 1.85

vectormap 1.84

dist_surfacewater 1.78

S2_summer_rededge5 1.70

di_S2_summer_red_median_S2_summer_rededge5_median 1.64

S2_summer_swir1 1.59

S2_summer_green 1.57

ndi_S2_summer_rededge6_S2_summer_rededge7 1.56

di_S2_summer_red_median_S2_summer_swir2_median 1.55

sar_summer_vv 1.49

ri_S2_summer_rededge6_S2_summer_rededge7 1.48

ndi_S2_summer_rededge6_median_S2_summer_rededge7_median 1.45

S2_summer_green_median 1.45

S2_summer_nir_median 1.45

di_S2_summer_red_S2_summer_rededge5 1.43

di_S2_summer_red_S2_summer_swir2 1.41

di_S2_summer_blue_S2_summer_green 1.40

ri_S2_summer_blue_S2_summer_red 1.40

di_S2_summer_blue_S2_summer_swir1 1.39

di_S2_summer_red_S2_summer_swir1 1.38

S2_winter_blue_sd 1.38

sar_winter_vv_median 1.38

di_S2_summer_green_S2_summer_rededge5 1.37

di_S2_summer_red_median_S2_summer_swir1_median 1.36

ri_S2_summer_green_S2_summer_swir1 1.35

S2_summer_swir1_median 1.35

sar_summer_vv_median 1.34

di_S2_summer_rededge7_S2_summer_rededge8a 1.33

ndi_S2_summer_green_S2_summer_swir1 1.33

ndi_S2_summer_blue_S2_summer_swir1 1.31

di_S2_summer_green_median_S2_summer_nir_median 1.30

di_sar_summer_vh_median_sar_summer_vv_median 1.30

ndi_S2_summer_green_median_S2_summer_swir1_median 1.30

ri_S2_summer_rededge6_median_S2_summer_rededge7_median 1.30

ndi_S2_summer_blue_S2_summer_red 1.29

di_S2_summer_green_S2_summer_rededge8a 1.29

ndi_S2_summer_green_S2_summer_rededge5 1.29


References

ALLEN, R. (2003) New Forest Valley Mires: Hydrogeological Site Accounts in Wheeler, B.D., SHAW,

S. & TANNER, K. (2009) A wetland Framework for impact assessment at statutory sites in England

and Wales. Environment Agency Science Report: SC030232/SR1.

BIVAND, R., KEITT, T. AND ROWLINGSON, B. (2014) rgdal: Bindings for the Geospatial Data Abstraction Library. R package version 0.9-1. http://CRAN.R-project.org/package=rgdal

BIVAND, R AND RUNDEL, C. (2014) rgeos: Interface to Geometry Engine - Open Source (GEOS). R package version 0.3-8. http://CRAN.R-project.org/package=rgeos

BOCK, M., XOFIS, P., MITCHLEY, J., ROSSNER, G. AND WISSEN, M. (2005) Object-oriented

methods for habitat mapping at multiple scales – Case studies from Northern Germany and Wye

Downs, UK. Journal for Nature Conservation, 13: 2-3, 75-89.

CLARKE, M. (1988) Past and present mire communities of the New Forest and their conservation.

PhD thesis, University of Southampton.

COX, J., JANES, M. AND ÅBERG, U. (2015) New Forest Wetland Restoration Review. Higher Level

Stewardship Agreement, the Verderers of the New Forest AG00300016. Available at:

http://www.hlsnewforest.org.uk/hls/info/80/completed_projects

DOWLE, M., SHORT, T., LIANOGLOU, S. AND SRINIVASAN, A. WITH CONTRIBUTIONS FROM SAPORTA, R AND ANTONYAN, E. (2014) data.table: Extension of data.frame. R package version 1.9.4. http://CRAN.R-project.org/package=data.table GROTHENDIECK, G. (2014) sqldf: Perform SQL Selects on R Data Frames. R package version 0.4-10. http://CRAN.R-project.org/package=sqldf HASTIE, T., TIBSHIRANI, T., NARASIMHAN, B. AND CHU, G. (2016) impute: impute: Imputation for microarray data. R package version HIJMANS R. J. (2015) raster: Geographic data analysis and modeling. R package version 2.3-24. http://CRAN.R-project.org/package=raster LIAW, A. AND WIENER, M. (2002) Classification and Regression by randomForest. R News 2(3), 18-22. LUCAS, L., MEDCALF, K., BROWN, A. AND BLACKMORE, P. (2001) Updating the Phase 1 habitat map of Wales, UK, using satellite sensor data. ISPRS Journal of Photogrammetry and Remote Sensing, 66(1): 81-102. MEDCALF, K., PARKER J., TURTON, N., AND FINCH C. (2011) Making Earth Observation Work

for UK Biodiversity Conservation – Phase 1. Report to the JNCC and Defra.

MEDCALF K., PARKER J., TURTON, N., AND BELL, G. (2013) Making Earth Observation Work for

UK Biodiversity Conservation – Phase 2. JNCC Report 495 Phase 2, JNCC Peterborough 2014

http://cran.r-project.org/package=rgdal

http://cran.r-project.org/package=rgeos

http://www.hlsnewforest.org.uk/hls/info/80/completed_projects

http://cran.r-project.org/package=data.table

http://cran.r-project.org/package=sqldf

http://cran.r-project.org/package=raster


MEDCALF, K., PARKER, J., BREYER, J., AND TURTON, N. (2015) MEOW Phase 3: Cost effective

methods to measure extent and condition of habitats. A report produced by Environment Systems

Ltd., for Defra and the JNCC.

NATURAL ENGLAND (2014) National Character Area profiles: data for local decision making.

Available at:

https://www.gov.uk/government/publications/national-character-area-profiles-data-for-local-decision-

making

R CORE TEAM (2014) R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/. REVOLUTION ANALYTICS AND WESTON, S. (2015) doSNOW: Foreach Parallel Adaptor for the 'snow' Package. R package version 1.0.14. http://CRAN.R-project.org/package=doSNOW RSTUDIO TEAM (2015) RStudio: Integrated Development for R. RStudio, Inc., Boston, MA URL http://www.rstudio.com/. TEEB. (2010) The Economics of Ecosystems and Biodiversity: Mainstreaming the Economics of Nature: A Synthesis of the Approach, Conclusions and Recommendations of TEEB WICKHAM, H. (2007) Reshaping Data with the reshape Package. Journal of Statistical Software, 21(12), 1-20. URL http://www.jstatsoft.org/v21/i12/. WICKHAM, H. (2009) ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York, Revolution Analytics and Steve Weston (2015). foreach: Provides Foreach Looping Construct for R. R package version 1.4.3. http://CRAN.R-project.org/package=foreach WICKHAM. H. (2011) The Split-Apply-Combine Strategy for Data Analysis. Journal of Statistical Software, 40(1), 1-29. URL http://www.jstatsoft.org/v40/i01/.

https://www.gov.uk/government/publications/national-character-area-profiles-data-for-local-decision-making

https://www.gov.uk/government/publications/national-character-area-profiles-data-for-local-decision-making

http://www.r-project.org/

http://cran.r-project.org/package=doSNOW

http://www.rstudio.com/

http://www.jstatsoft.org/v21/i12/

http://cran.r-project.org/package=foreach

http://www.jstatsoft.org/v40/i01/

Date post:	14-Aug-2020
Category:	Documents
Upload:	others
View:	4 times
Download:	0 times

Evidence Project Final Report - GOV.UKrandd.defra.gov.uk/Document.aspx?Document=14198_SD1705... ·...

Documents