Antsirabe, Madagascar - 2015, land cover map

This work was conducted before the launch of ESA Sentinel-2 mission, which images are particularly adapted to crop monitoring and characterization thanks to their high spatial (10 – 60m) and temporal (5 days) resolutions. We worked with Landsat-8 and Spot5 images to create a time series simulating Sentinel-2 data. Considering the very small size of the fields in the study area, we also used a very high spatial resolution image (VHSR) Pléiades on which a segmentation was applied to delimit objects boundaries. Each object was then classified following a predefined nomenclature. To do so, Random Forest algorithm was used and trained used a learning database, by considering spectral variables from the time series, textures calculated from the VHSR image, and altitude and slope from a digital elevation model (SPOT DEM).
<br>
Through this work, the objective of the CIRAD – TETIS joint research unit is to develop a mapping approach adapted to southern cropping systems constraints (small field size, heterogeneity of cropping practices, landscape fragmentation, cloudy conditions during the cropping season). From this approach, an automatic processing chain for land use mapping was then built: Moringa.
<br>
Details on the approach can be found in the paper referenced below.
<br>
The map is distributed in vector Geopackage format. For each object, the attribute table presents the class attributed by Random Forest in 5 nomenclature levels (from 2 classes for level 1 "Cropland" to 31 classes for level 5 "Crop Subclass"). Validation scores for each class and nomenclature level are given in the article. The attribute table also presents for each level the class membership probability of each object (Max columns). These values are given by the Random Forest algorithm.
<br>