Detecting Desertification in Southern Morocco Using a Multi-Sensor, Random Forest Approach

Existing landcover products depicting degraded land do not accurately show the extent of desertification of traditional cultivation systems. These products rely on recent NDVI (Normalised Difference Vegetation Index) time series and statistical data. Historical desertification is missed, although these abandoned fields are distinctive in satellite imagery as reflectant and smooth surfaces. We present our Google Earth Engine workflow for detecting desertification using satellite data (full details are in our recent paper). We used the random forest algorithm to classify five landcover categories including desertified fields, applied to a data stack comprising a 13-band Sentinel-2 composite and derived tasselled cap components, and a Sentinel-1 VV-polarisation composite. We test our approach for case studies of the Skoura and Draa oases in southern Morocco with a resulting accuracy of 74-76% for the desertification class.