Identifying cropped areas in small growers agricultural regions using data mining for food security

The present study aimed at testing the potential of the future mission SENTINEL-2 (European Copernicus program) to map croplands in a region of Madagascar characterized by small size fields and frequent cloud covering. Two approaches were tested and compared : i) a classical remote sensing method (RS) using image object-based analysis, expert rules and supervised classification, and ii) a data mining (DM) approach consisting of the extraction of frequent patterns from the database and the use of these patterns in different algorithms (Naive Bayes, Random Forest, Decision Tree and Support Vector Machine) to build classification rules. Both methods used SPOT images and a ground data set of 324 GPS waypoints collected during the 2012-2013 cropping season. The remote sensing and data mining approaches showed equivalent overall accuracies (82% vs 84% for RS and DM methods respectively). However, the DM approach showed its ability to handle a large volume of data and to do so in a time manner. This approach has also the advantage to extract all the information at its disposal, even temporal behaviors, unlike the object-based RS approach which requires significant participation of the expert. Data mining tools are thus recommended for their considerable potential for the classification without a priori of remotely sensed data, mixing multisource information and consequent time series, especially for the upcoming Sentinel-2 images that are expected to generate a large volume of data to store and process.