Specific Project Objectives & Deliverables

Cropland and crop type mapping

In the frame of the Sen2Agri ESA project, the JECAM site was selected as one of the demonstration site to demonstrate the Sen2Agri system in Near-Real-Time. The full growing season time series of Sentinel-2 and Landsat 8 images covering the JECAM site was automatically downloaded and processed to generate (1) a suite of Dynamic Cropland Mask and (2) Crop Type map.

The Dynamic Cropland Mask consists in a binary map separating annual cropland areas and other areas, thus corresponding to a mask over annually cultivated areas. The annual cropland is defined as a piece of land with a minimum area of 0.25 ha, actually sowed/planted and harvestable at least once within the year following the sowing date. This binary map is produced along the agricultural season on a monthly basis, to serve for instance as a mask for monitoring crop growing conditions, as basis for sampling stratification and for agricultural extension. Its accuracy is expected to increase as long as additional images are integrated into the development process. 


The Crop Type product is a map of the main crop types or groups in a given region. The top main crop types are considered per region. The main crop types are defined as those covering a minimum area of 5% of the annual cropland in the region representing a cumulated area higher than 75% of the latter. For Belgium, the main crop types are Winter Wheat, Maize, Barley, Potatoes, Sugar beet and rapeseed. The crop types are identified over the previously generated crop mask. 


a)                                                         b)

c)                                                         d)

Figure 5: Composites and crop type classification outputs. (a) L3A September 2016 false-color cloud free composite (NIR,RED,GREEN), 
b) crop type classification output on (a) area,  c) L3A  September 2016 false-color cloud free composite (NIR,RED,GREEN), d) croptype classification output on (c) area


LAI estimation with optical images

The two years field campaign of DHP acquisition allows to validate the LAI retrieved with S-2A images for the winter wheat. The results show the added value of S2-A new spectral bands to estimate the LAI comparing to SPOT5 Take5 which tend to a systematic underestimation of the LAI value retrieved from 3.5. The new S2-A bands have two main impacts: (i) overcome the saturation effect around a LAI of 4 and (ii) decrease the spreading of the point around the 1:1 axis; these two effects allow the decreasing of the RMSE from 1 to 0.7.

Figure 6: LAI estimate over a farmer fields, estimation available in near real time during the all growing season.


Tillage detection

The image processing has been realized in three main steps. First, the preprocessing of S1A SLC images has been realized using SNAP with S1 Toolbox. Second, the hourly rain data of 29 meteorological stations has been gridded at 10km over the whole study area. Third, the pixels of the images which received a rain exceeding 1mm within 3 hours before the acquisition have been masked. The preprocessing chain consisted in: applying the precise S1A orbits, calibrating, removing thermal noise, debursting, multilooking (5x1) and terrain correcting with SRTM 1sec. Since the analysis is performed on an object base, the images have not been filtered. In total, 17 images were used for the orbit 110 (descending) and 16 images were used for the orbit 161 (ascending). The time series ranges from 04/07/2015 to 17/02/2016.

The classifications achieved accuracies of 87% for field use, 86% and above 94% for tillage detection in winter wheat and cover crop respectively. However, those accuracies are relying on the choice of the good discrimination periods. This requires crop calendar and crop growth conditions knowledge to account for timing particularities in the farmers’ behaviour. Furthermore, the global overall accuracy of tillage detection is the product of the field use OA and of the tillage detection OA. This product gives a global OA of 78-80% (depending on the field use). To further improve this discrimination, one could test fields’ stratification between agro-meteorological areas in order to decrease the variability in soils conditions, farming calendar and the way the tillage is realized.


Figure 7: Tillage occurrence detection in Wallonia with S1A - SAR time series. R: Sigma0 VV, G: Sigma0 VH, B : ratio

©2015 Joint Experiment for Crop Assessment and Monitoring