Mapping tillage operations over a peri-urban region using combined SPOT4 and ASAR/ENVISAT images

摘要

This study aimed at assessing the potential of combining synchronous SPOT4 and ENVISAT/ASAR images(HH and HV polarizations) for mapping tillage operations (TOs) of bare agricultural fields over a periurban area characterized by conventional tillage system in the western suburbs of Paris (France). The reference spatial units for spatial modeling are 57 within-field areas named “reference zones” (RZs) homogeneous for their soil properties, constructed in the vicinity of 57 roughness measurement locations, spread across 20 agricultural fields for which TOs were known. The total RZ dataset was half dedicated to successive random selections of training/validating RZs, the remaining half (29 RZs) being kept for validating the final map results. Five supervised per-pixels classifiers were used in order to map 2 TOs classes (seedbed&harrowed and late winter plough) in addition to 4 landuse classes (forest, urban, crops and grass, water bodies): support vector machine with polynomial kernel (pSVM), SVM with radial basis kernel (rSVM), artificial neural network (ANN), Maximum Likelihood (ML), and regression tree(RT). All 5 classifiers were implemented in a bootstrapping approach in order to assess the uncertainty of map results. The best results were obtained with pSVM for the SPOT4/ASAR pair with producer’s and user’s mean validation accuracies (PmVA/UmVA) of 91.7%/89.8% and 73.2%/73.3% for seedbed&harrowed and late winter plough conditions, respectively. Whatever classifier, the SPOT4/ASAR pair appeared to perform better than each of the single images, particularly for late winter plough: PmVA/UmVA of 61.6%/53.0% for the single SPOT4 image; 0%/6% for the single ASAR image. About 73% of the validation agricultural fields (79% of the RZs) were correctly predicted in terms of TOs in the best pSVM-derived final map. Final map results could be improved through masking non-agricultural areas with land use identification system layer prior to classifying images. Such knowledge of agricultural operations is likely to facilitate the mapping of agricultural systems which otherwise proceed from time-consuming surveys to farmers.