3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Space Research InstituteDemonstration objectives
� To deliver consistent information at field scale for Ukraine
using Sen2Agri system: � Crop mask;
� Crop type;
� Crop status;
� Early crop area indicators, detection of crop anomalies
� To validate the products of Sen2Agri system
� To introduce the products to the authorities and to be
prepared for preoperational use of the system
Ukraine
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Use case (additional)
Winter wheat crop production estimation
based on crop area estimation and crop
state monitoring with LAI and NDVI
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Site features
• Location: Ukraine
• Intensive agriculture area. Main crop types: winter wheat, winter rapeseed,
spring barley, maize, soybeans, sunflower, sugar beet, and vegetables
• Field size: from 30 to 250 ha
• Crop calendar: Winter: September – July;
Summer: April – October
• Cloud coverage can be very frequent during the growing season
• Topography: mostly flat, slope: 0% to 2%
• Soils: different kinds of chernozems
• Soil drainage is ranging from poor to well-drained. Irrigation infrastructure is
limited
• Climate and weather: humid continental
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
EO and in-situ dataset -1
• S2 acquisitions along the season – up to 1800 scenes (March – October, 2016)
About 15% of S2 images cloudfree
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
EO and in-situ dataset -1
• S1 acquisitions along the season – up to 1000 scenes (March – October, 2016)
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
EO and in-situ dataset -1
• Field campaign and in-situ dataset (LC/LU and crop
types)
Train set – 5536 parcels
Test set – 2153 parcels
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
EO and in-situ dataset -1
• Field campaign and in-situ dataset (LAI validation)
• Indirect method for biopar estimation (DHP imagery, CAN-EYE);
• Total amount – 121 samples
– April – July 2016;
– 9 expeditions;
– JECAM test site (Pshenichne);
– 3 main crops:
• Winter wheat (42 ESU);
• Maize (37 ESU);
• Soy beans (42 ESU)
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Sen2-Agri products assessment L4A Mykolaiv region, 2016
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Sen2-Agri Winter Crop Map
Crop\Are
a (103 ha)
Official
statistics*
Sen2-
AgriSRI
Winter
wheat5367.2 7027.3 6481.9
Winter
rapeseed344 451.3 448.8
*without AR_Crimea
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
GEE Winter Crop Map (derived by SRI)
Crop\Are
a (103 ha)
Official
statistics*
Sen2-
AgriSRI
Winter
wheat5367.2 7027.3 6481.9
Winter
rapeseed344 451.3 448.8
*without AR_Crimea
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Comparison Sen2-Agri & SRI winter crop maps
Sen2-Agri Winter Crop Map Sentinel-2, Natural Color 01/05/2016
Comparison Sen2-Agri & SRIGEE Winter Crop Map (SRI)
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Comparison of winter wheat areas
0
100
200
300
400
500
600
700
800
are
a,
10
3h
a
Regions
Winter Wheat
Statistic Sen2Agri SRI
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Comparison of winter rapeseed areas
0
5
10
15
20
25
30
35
40
45
50
are
a,
10
3h
a
Regions
Winter rapeseed
Statistic Sen2Agri SRI
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Winter wheat production, Sample use case
Winter wheat area estimates based on classification map vs statistics
Winter wheat yield estimates based on time series
of VI and official statistics
Winter wheat production
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Maize production, Sample use case
Maize yield estimates based on Hydrometcenrteforecast and official statistics
Maize area estimates based on classification map vs statistics
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Comparison Sen2-Agri & SRI crop maps
Sen2-Agri crop map 2016 (Kiev Region)
Crop, % PA UA
Winter wheat 96.4 98
Spring cereals 94.2 84.9
Maize 89 95.1
Sunflower 98.9 92.7
Soybeans 90.6 90
OA 93.6
Winter wheat
Spring cereals
Maize
Sunflower
Soybeans
Sen2-Agri crop map 2016 (Kiev Region)
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Comparison Sen2-Agri & SRI crop maps
SRI crop map 2016 (Kiev Region)
Crop, % PA UA
Winter wheat 96.4 98
Spring cereals 94.2 84.9
Maize 89 95.1
Sunflower 98.9 92.7
Soybeans 90.6 90
OA 93.6
Crop, % PA UA
Winter wheat 93.7 95.9
Spring cereals 82.8 74.8
Maize 95.2 93.6
Sunflower 97.3 97.9
Soybeans 92.2 94.3
OA 94
Winter wheat
Spring cereals
Maize
Sunflower
Soybeans
Sen2-Agri crop map 2016 (Kiev Region)
SRI crop map 2016 (Kiev Region)
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Comparison Sen2-Agri & SRI crop maps
Agreed part of two maps
Crop, % PA UA
Winter wheat 96.4 98
Spring cereals 94.2 84.9
Maize 89 95.1
Sunflower 98.9 92.7
Soybeans 90.6 90
OA 93.6
Crop, % PA UA
Winter wheat 93.7 95.9
Spring cereals 82.8 74.8
Maize 95.2 93.6
Sunflower 97.3 97.9
Soybeans 92.2 94.3
OA 94
Crop, % PA UA
Winter wheat97.1 98.7
Spring cereals 94 87.2
Maize 99 96.4
Sunflower 99.8 99.6
Soybeans 94.8 98.8
OA 97.6
Winter wheat
Spring cereals
Maize
Sunflower
Soybeans
Sen2-Agri crop map 2016 (Kiev Region)
SRI crop map 2016 (Kiev Region)
Agreed part of two maps
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Comparison Sen2-Agri & SRI crop maps
Winter wheat Spring cereals Maize Sunflower Soybeans
Statistics 197,2 79,1 268 161,2 173,3
Sen2-Agri 254,63817 22,28391 306,07209 217,55205 190,82718
SRI 189,15012 56,55195 318,64464 234,94455 226,81557
0
50
100
150
200
250
300
350
Are
a (
10
3h
a)
Crop Areas 2016 (Kiev Region)
Statistics Sen2-Agri SRI
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
• We compared classification results for different
algorithms and input data:
– Atmospheric correction (Sen2Cor, MACCS, TOA)
– Band combinations
– Best sensors and bands combination
– Classifier
Experiments on algorithms
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Best atmospheric correction method
Satellite data:� 11 Sentinel-2 scenesGround data:� 563 ground samples
(train and test sets)
TOA (OA=80.7%) Sen2cor (OA=80.6%) MACCS (OA=82.7%)
Training setTest set
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Best bands combination for optical data
Combination Overall Accuracy, %
S2 10m (g,r,NIR) 70.9
S2 10m (b,g,r,NIR) 74.5
L8 (without blue band) 75.8
L8 74.6
S2 10m(warp 20) 71.1
S2 10m(warp 20) + band 5 75.9
S2 10m(warp 20) + band 6 72.6
S2 10m(warp 20) + band 7 71.5
S2 10m(warp 20) + band 8A 71.4
S2 10m(warp 20) + band 11 74.6
S2 10m(warp 20) + band 12 73.8
S2 20m 76.3
S2 10m(warp 20) + S2 20m 78.2
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Best sensors and bands combination
Satellite data:� 4 Sentinel-2 scenes� 21 Sentinel-1 scenes� 2 Landsat-8 scenesGround data:� 728 ground samples
(train and test sets)
S1 (OA=77%) S1+S2 (OA=79.4%) S1+S2+L8 (OA=79.9%)
Combination Overall Accuracy, %
S2 + L8 78.6
S2 + L8 without blue bands 76.9
S1 10m 77
S1 10m + S2 10m 79.4
S1 + S2 79.5
S1 + S2 + L8 79.9
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Best classifier
# Class ENN RF CNN
UA,% PA,% UA,% PA,% UA,% PA,%
1 Artificial 74.4 95 53.9 85.2 75.3 51.5
2 Winter wheat 80.2 81.6 73.2 69.6 78.9 86.5
3 Winter rapeseed 95.2 18.9 96.8 73.2 94.7 62.8
4 Spring crops 45.1 32 5.6 4.6 22.7 8.7
5 Maize 93.4 94.6 90 94.7 92.1 94.6
6 Sugar beet 74.7 100 72 99.9 71.6 100
7 Sunflower 88.5 94.9 90.4 88.6 87.7 93.1
8 Soybeans 95.5 79.3 96.8 73.1 94.4 76.7
10 Forest 99.1 99.7 99.2 98.2 98.3 99.7
11 Grassland 87.9 78.4 86.8 80.4 88.5 83.5
13 Water - - 95 100 80.3 98.1
14 Wetland 74.9 66.4 52.5 73.6 70 65.6
15 Winter barley 23.1 31.5 10.7 28.6 28.1 45.5
18 Buckwheat 18 68.5 75.4 99.9 21.8 39.7
OA,% / Kappa 79.9 / 0.77 79.5 / 0.77 82.3 / 0.8
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Discussion
1. MACCS is the best method for atmospheric correction with gain + 2% for OA.
2. Using all available bands for optical data showed the best result compare to other bands combination (OA = 78.2%).
3. Combination of all satellites outperforms all other combination (OA = 79.9), but the obtained classification map has 30 m resolution.
4. CNN is the best classifier (OA = 82.3%) compare to common approaches (Random Forest and ensemble of neural networks).
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
• Did you have the opportunity to operate the Sen2Agri system ?
– Yes, we are operating Sen2-Agri system
• What is your experience?
– Positive, but system still non end-user friendly – UI improvements for data
search and jobs manipulations are necessary (below in details)
– Tricky to manage the system in case of some errors
– 20 Tb is too tiny storage for Ukraine
• What are your recommendations for the future for the system ?
– To add products filtration on dates and extent (starting from L2A) –
especially on ‘Products’ tab and on ‘Custom job’ tab
– To add understandable progress bar for L4A and L4B execution (with
time estimates)
– To improve manipulations with Executing Jobs (workflow management
tools with UI)
Feedback on system and products
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
• What are your recommendations for the future for the system (cont)?
– To add functionality for launch processing for sub-site (for example for Kiev region only when the whole Ukraine as site is specified)
– To optimize system performance - it takes too much time for the territory of Ukraine
– To improve system user guide – especially for dealing with technical troubles
– To add automated checks of product integrity in case of power failure
• What are your recommendations for the future for the Sen2Agri products?
– To include SAR data (Sentinel-1) – most of the images for 2016 were clouded
– To add product merging over specified administrative units (on vector boundaries)
– To consider percentage of cloudiness that Sen2-Agri system downloads -higher results acquisition
Feedback on system and products
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Feedback on system and products
Sites
S2A productsL3B (LAI products)
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Cropland mask and crop type maps (35UPQ tile sample) – April – July, 2016
Sample Products on SRI premices
L4A
L4B
Non-maskedL4B
masked
L4B
Consorcium
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
Sample products from the system installed in SRI
• Vegetation indices (NDVI, LAI)
NDVI (17/07/2016) LAI (17/07/2016)
3rd Sen2Agri User Workshop - Rome, 28-29 June 2017
• Do you consider the demonstration phase relevant for testing the
operational capabilities of the Sen2-Agri system ?
– YES
• Do you consider the demonstration phase relevant with respect
to your objectives?
– YES
• Which improvements do you expect in the future (priority
ranking)?
– To add functionality mentioned in feedback section
• What are the top priority you would recommend to contribute to
the system uptake by your team?
– Integration of SAR data
Recommendations