We worked on the assumption that agricultural systems shaped the landscape through human cropping practices, and that the resulting landscape can be described with Q set of coarse resolution satellite-derived metrics (spectral, textural, temporal, and spatial metrics). A Random Forest classification model was developed at the village scale in South Mali, based on 100 samples, with data on the main type of agricultural system in each village (three-class typologyLand 30 MODIS-derived and socio-environmental metrics calculated on agricultural areas. The model was found to perform well (overall accuracy of 60 percent) and was stable. Class A (food crops) and B (intensive agriculture) displayed good producer's accuracy (70 percent and 6'7 percent, respectively), while class C (mixed agriculture) was less accurate (50 percent). The most important metrics were shown to be the annual mean of NDVI, follo'wed by the phenology transition dates and texture metrics. However, 'when considering each set of metrics separately, texture emerged as the most discriminating factor (with 53 percent of global accuracy). This result, i,e., that even COQl'se resolution imagery contains textural information that can be used for crop mapping, is new, Such maps could be used in food security systems as an indicator of system vulnerability, or as spatial inputs for crop yield models.