Automated production of forestry thematic maps – a concept of
remotely sensed data fusion in the Czech NFI2
Dpt. of Photogrammetry & Remote sensing
Forest Management Institute
Czech Republic
2014 ForestSAT conference, 4-7 November 2014, Riva del Garda (TN), Italy
Photogrammetry and remote sensing
• processing of aerial images from Czech Office for Surveying, Mapping
and Cadastre (COSMC)
• photogrammetric survey in NFI2
• support for the NFI2 field survey
• normalised digital surface model (nDSM) derived from stereo-imagery
• color-infrared (CIR) orthophotomaps
Wall-to-wall maps derived from EO data
• periodical mapping of forest borders (Landsat series)
• multitemporal detection of clearcuts(Landsat series)
• thematic maps based on classification of ortophoto and nDSM
Remote sensing department at FMI
Photogrammetric survey in NFI2
• Landuse/landcover classification of sample plot according to the definitions of FAO (ENFIN) and LULUCF
• Assessment on forest plots:– stand height information
– tree species composition /mix type
– stand development phase
– crown cover /canopy closure
• Identification of trees and bushes
• Estimation of the total number, lengths and areas of objects on transect (linear woody vegetation, solitaire trees, small groups of trees and bushes outside forest, etc.)
Stereo-interpretation executed on a 0.5 x 0.5 km sample grid with a limited set of variables acquired
Stereo-interpretation: 306 000 plots
(statut end of September 2014)
NFI2 grid point assessment
• interpretation quadrate with size of 51 x 51 metres consisting of 16 points
• altitude and landcover type are assessed at each point - coniferous tree, broadleaf tree, bush, land surface, other types
PHONIL – the stereo interpretation
In NFI2, the goal was to develop a stereo-interpretation tool without need of
workstation-based data processing
Both application and the database schema are developed for the effective NFI planning and
management of the field survey workload
PHONIL = custom application developed on platform of SW Photopol• full processing support on side of GIS database server
• storing and data management in PostgreSQL excludes duplicities, ensures integrity,
easy backups, etc…
• management of controls (validation) of visual interpretation
• management of survey priorities
• DSM points generated by image correlation from stereo images in 2 x 2 meter grid
• CIR 8-bit data in compression ECW format
• image correlation in PhoTopoL SW
• terrain normalisation using elevation values from LiDAR based DTM
nDSM and orthophoto from UltraCam images
Aerial data fusion - forest thematic maps
• automated processing using object-based image analysis (eCognition)
• supervised Random Forest classification of CIR orthoimages and nDSM using photogrammetric
survey (grid points /plot LULC categories) as a training data
• country-wide thematic maps with spatial resolution 2m/pixel
• GIS post-processing and manual editing…
Output maps:
₋ high vegetation boundaries – forest / olwtc/ other according to FAO
₋ distribution of coniferous /deciduous vegetation
₋ basic forest stand growth stages
Multi-resolution segmentation / classification
• three segmentation levels
• scale and shape parameters adjusted according to
target landuse/landcover categories
• two cycles of RF classification using NFI2 plot
classification – forest, olwtc, other,….
• two cycles of RF classification using NFI2 grid point
assessment and NFI2 plot classification –
conifers/broadleaved forest
• rule-based classification of nDSM - basic forest stand
growth stages
• rule-based contextual and geometric cleaning
• morphological operations to enhance vector
representation of output
high vegetation boundaries
forest / olwtc/ other land
overall accuracy 96,3% (simulated RF in R stats)
distribution of coniferous /deciduous stands
deciduous
coniferous
other
young up to 8m
waterbodies
impervious
Classificationconiferous deciduous impervious other
Producers accuracy
(simulated RF in R stats)Groundtruth
coniferous 400 33 0 17 88,9%
deciduous 58 180 0 20 69,8%
impervious 6 0 10 16 31,3%
other 24 19 0 478 91,7%
basic forest stand growth stages
other
mature above 20m
young up to 8m
seedlings up to 2,5m
young 8 – 20m
clearcut
Conclusion – future steps
• aerial imagery processed in various manner provide a range of
information on forests
• automated classification of spectral (CIR ortophoto) and height data
(nDSM) based on thematic content from visual interpretation sufficient
for wall-to-wall mapping
• GIS processing needed to ensure LULC classes in accordance with NFI
(ENFIN) definitions
• developed processing workflow transferable to other image (satellite,
LiDAR) data
• higher thematic accuracy expected from ESA Sentinel-2
multi/hyperspectral sensor
2014 ForestSAT conference, 4-7 November 2014, Riva del Garda (TN), Italy