Multiple classifier combination for crop types phenology based mapping

Currently in Morocco, crop plantation information is mostly collected by three ways: (1) farmer communications, (2) spatially limited land survey and (3) manually photo-interpretation of a newly registered digital image. These procedures provide limited and subjective information with unguaranteed consistency. Land survey could map accurately crop types but it's too time, high cost and labor-intensive which limits its use as a periodic process to monitor crop changes. Remote sensing imagery is shown to be a cost-effective crop mapping approach which can be regularly used to produce an accurate and up-to-date crop map at the different temporal and spatial resolution. In this paper, a phenology based-time series-multiple classifier combination approach is developed instead of a classical one-image-one classifier approach to map crop types. The whole process is performed mainly on four steps. First, all images were radiometrically and atmospherically corrected and the specific ETM+ gap had been resolved. Then, a phonological metrics are extracted from annual Enhanced Vegetation Index (EVI) profiles. In the third step, two classical supervised learning algorithms: Decision Tree (DT), K Near Neighborhood (KNN) and four advanced ones: Support Vector Machines (SVM), Bagging, Random Forest (RF) and Extremely Randomized Trees (Extra Trees) are used in ascending experimental protocol of 3 levels of crossed validation to (1) adjust classifiers' parameters, (2) select the best three classifiers to combine and (3) find the best linear combination from five ones tested. All these three optimization operations are done according to the best error rate computed based on f-measure of omission and commission errors. In the last, the final pixels' prediction is deducted thanks to average decision given by (SVM, RF and Extra Trees) which outperforms the best individual classifier score and all other tested combiners. We show the efficiency of the proposed scheme with experiments carried out with 11 LANDSAT free cloud images depicting Gharb region, one of the largest agriculture plain in Morocco.