Automated classification of land cover for the needs of CAP using Sentinel data. A proactive approach.

Transcription

1 22nd CAP/IACS conference, November 2016, Lisbon - Portugal Automated classification of land cover for the needs of CAP using Sentinel data. A proactive approach. Dimitris KAPNIAS (GAIA) d_kapnias@c-gaia.gr Panagiotis ILIAS (NP) p_ilias@neuropublic.gr

2 founded in 2014 Farmers Organizations Neuropublic Piraeus Bank as joint partnership of 70 Coops Gaia Cloud Platform No 1 GR 150 K Farmers AgriTech Know-How Contract Farming ICT industry 0.75 B Turnover AgriTech Solutions AgriCard Banking services 0.5 M ha Area Smart Farming Young Farmer loans The shareholders legacy to the coalition Agriculture Cooperatives Bank Tech Farmers

3 Digital service provider 3

4 GAIA s involvement in the Aid Application Scheme Nationwide Network of local service advisors (FSP) Society Management Subsidy (CAP) Commerce Smart Farming 92 Farmer Service Points (FSP) certified as AID Application assistants to the beneficiaries GAIA is certified as the coordination & support body ( )

5 Supporting Farmer using Data Advisory Services Weather Forecast Field Sensors Proximity Sensing Remote Sensing Farm Data Data Data fusion Data assimilation Data interpolation Facts How much water and when? When do I have to spray? What is the risk level? What is the precise type and the exact amount of fertilizers needed? Advice Eligibility status? Parcel crop group? Coupled payments & greening requirements? Which choices to make? Which financial instruments to use? How can I support my business? Training Farmer Support

6 IoT and Cloud Infrastructure for Agricultural Monitoring large scale agricultural areas Other Platforms Other Platforms IoT GAIA Sense Data Analysis Decision Making GAIA Smart Farm & GAIA Subsidy (CAP) Advisor API GAIA Cloud Atmospheric Soil Data Data Collection Data Fusion Farmer Satellite RPAS Sensors Weather Other sources Sentinel-2 Sort revisit HR MS Open

7 Subsidies control The role of Sentinels Smart Farming Services CAP eligibility - Compliance On-the-fly automated cross checks of parcels declaration (proactive control): Reducing errors in the location of parcels Discouraging false claims Full scale automatic compliance cross checks (postdeclaration control): crop-wise in terms of modelling and time windows Allows better risk analysis, less RFV, more effective controls Reducing overall error rates Pest management/early hazard warnings Irrigation Fertilisation Parcel Level Sentinel can been used for the multitemporal extraction of vegetation, soil and water indices and the monitoring of the phenological growth.

8 Multi-Temporal Object-based Monitoring Cloud based services that integrate earth observation data with Image Processing, Machine Learning, Spatial Modeling Collect Monitor Actions Store Time Take Actions Space Define Objects Decision Making Assign Properties Parcel = Object Create Models [ ] Vegetation Soil water Indices at parcel level, for (single or multi) crop growing period Crop Models

9 Temperature data for area Leukonas - Prespes Indices in Agricultural Monitoring Indices allows us too visualize the behavior of specific crops at parcel level through time. Rainfall data for area Leukonas - Prespes Extracted Indices patterns agree with phenological growth for the specific crop (beans) S2 Collection dates Multitemporal Visualization of the Prespa area, for the growth period, using indices values (R=NDVI, G=NDWI, B=SAVI)

10 Land Cover Classification using +Indices PHASE PHASE 1: 1: Indices Indices (i) (i) extraction extraction using using atmo atmo spherically spherically corrected corrected bands bands (b) (b) PHASE 1: Indices (i) extraction using atmo spherically corrected bands (b) Crop Crop Code Code & Crop Crop Description Description b3 b3 b2 b2 b1 b1 b3 b2 b1 bn bn bn i3 i3 i2 i2 i1 i1 i3 i2 i1 in in in Crop Code & Crop xxxxxxxxxxxx Description xxxxxxxxxxxx Image Image (b1, (b1, b2, b3,, b2, b3, bn) bn) Image Image (i1, (i1, i2, i3,, i2, i3, in) in) Object Object (reference (reference parcel parcel or or holding holding LPIS) LPIS) Land Land Cover Cover Class Class System System PHASE PHASE Image 2: 2: (b1, Descriptor Descriptor b2, b3,, bn) Extraction Extraction Image (i1, for for i2, i3, each each, in) object object Object per per Image Image (reference parcel or holding LPIS) Land Cover Class System Multi-Dimensional Multi-Dimensional PHASE 2: Descriptor Extraction for each Dimension Dimension object per varies varies Image Multi-Dimensional SIFT SIFT Feature Feature Extraction Extraction Dimension varies Method(s) Method(s) t22 t22 SURF SURF SIFT [ 32, 32, 0,.,., 123, 123, 0 ] Feature a mapping Extraction between different Method(s) object properties t22 SURF ORB ORB [ Descriptor Descriptor 32, 0,(scaling) (scaling)., 123, 0 ] Segment Segment All All pixels pixels CHIST CHIST ORB Descriptor (scaling) (Polygon: (Polygon: x1y1, x1y1, x2y2, x2y2, x3y3, x3y3, e.g x4y4) x4y4) Parcel Location, All All Bands Bands or/ or/ and and Indices Indices EO Indices, PGS, Soil -> Crop. Segment All pixels CHIST PHASE PHASE (Polygon: 3: 3: x1y1, Model Model x2y2, x3y3, Creation Creation x4y4) (data (data driven) driven) All Bands or/ and Indices Train Train set set That process has 4 main phases: PHASE 3: Model Creation (data driven) Input Input (time (time series series data data for for polygons) polygons) Output( Output( Labe)l Labe)l SVM SVM Train set Training Training Evaluation Evaluation [ Input (time 22, 22, series 5, 5,.,., data for 12/04/2016] 12/04/2016] polygons) Output( [1.1.1] [1.1.1] Land Land Cover Cover Labe)l [ 32, 32,.,., 123, 123, 07/05/2016] 07/05/2016] [1.1.1] [1.1.1] SVM RF RF Classification Classification Model Model Training Evaluation.. 5, 10 12/04/2016] Land Cover [ 4, 4, 1, 1, 1, 1, 32, 22, 22,.,., 123, 123, 07/05/2016] 22/10/2016] 22/10/2016] [1.1.1] [1.1.1] NN NN RF Classification Model. PHASE PHASE [ 4, 1, 1, 22, 4: 4: Crop Crop., estimation estimation 123, 22/10/2016] for for other other [1.1.1] parcels parcels using using the the data-driven data-driven NN input input -output -output model(s) model(s) PHASE 4: Crop Input Input (time (time estimation series series data data for for for other a polygon) polygon) parcels using the data-driven Model Model input -output model(s) [ Input (time 22, 22, 0 series, 5, 5,.,., data for a 12/8/ /8/2016 polygon) ].. Model [ 4, 4, 1, 1, 1, 1, 22, 22, 0, 5,.,., 123, 123, 10 12/8/ /8/2016 ]. [ 4, 1, 1, 22, 0,., 123, 12/8/2016 ] xxxxxxxxxxxx At the classification process, we want to create models that perform

11 Land Cover Classification using Indices TRAIN SET Crop Cover (ha) Parcels No. Wheat Stone fruits Legumes Maize Trees Fallow Pasture 1, Total 3,873 1,816 Total Area (ha): 282,600 Agricultural Areas (ha): 53,580

12 Land Cover Classification using Indices TEST SET Crop Cover (ha) Parcels Crop No. Acc. (%) Wheat 2,164 Wheat % Stone fruits 2,400 Stone 5045 fruits 81% Legumes 1,344 Legumes % Maize 2,917 Maize % Forest Trees 549Forest 894 Trees 61% Fallow 3,258 Fallow % Pasture 1,167 Pasture 51 78% Total 13,799 Total 25,264 74%

13 GAIA Smart Farm & GAIA Subsidy (CAP) Advisor GAIA Cloud API Two back-end web services that rely on EO-S2 data: Polygon, time period -> Times series of indices, Estimated Crop Pattern Polygon, time period -> Crop classification Farmer That allows: On-the-fly automated cross checks of parcels declaration (proactive control) Full scale automatic compliance cross checks (post-declaration control - locate the outliers) Crop monitoring Smart farming (Advices based on Facts) Plus Smart Farming data: Digital data (Valid Train set). Other type of data describing soil and atmospheric conditions (better mapping).

14 Next steps: Period : Subsidy and Smart Farming services for 16 large scale areas (pilot sites) and for 15 significant crops. Collaboration for specific sites and crops with the partners of DataBio H2020 project (3 years from 2017 to 2020). Enrich the modelling process with more Descriptors (SL1 + L8). Deep Learning (for specific time windows using VHR data). Big Data Analytics (Outliers) Thoughts - Discussions : IACS will take advantage of other Cloud Platforms (Food Security).

15 22nd CAP/IACS conference, November 2016, Lisbon - Portugal Thank you for your attention Any questions? Dimitris KAPNIAS (GAIA) d_kapnias@c-gaia.gr Panagiotis ILIAS (NP) p_ilias@neuropublic.gr 15