Supervised Classification of Satellite Images with Spatially Inaccurate Training Field Data

The use of satellite images for environmental monitoring has shown a great potential to monitor large areas at relatively low costs. Classically, domain experts identify natural habitats in small areas, and construct habitat distribution maps over larger areas using a supervised classification. Since each pixel may correspond to several habitats, we are in a multi-target classification framework. Instances are pixels. Each pixel is associated with several target variables (i.e. habitats) simultaneously. Each variable is a set of category labels, i.e. a set of intervals (bins) representing habitat proportions. In such context, training data often comes from field data acquired by domain experts. However, location of these data may be approximate (e.g. due to accuracy of handheld GPS). This spatial inaccuracy is particularly problematic when these data are used to train a classifier on very high resolution (VHR) satellite images. Indeed, in some cases, spatial accuracy of field data may be much lower than the one of images. In this paper, we propose a preprocessing approach to correct this spatial inaccuracy of field data w.r.t. VHR satellite images. First, our process extracts candidate sequences of pixels related to a given field inventory. Then, it extracts the corresponding sequence of habitats in the field data and compare its similarity with the corresponding candidate sequence of pixels. Finally, it ranks candidate sequences of pixels and select the best one as training data. Two similarity measures are studied in this work. To validate our approach, we compare the performances of 46 multi-target supervised classification algorithms on a dataset dealing with coral reef monitoring. We study accuracy of classifiers with and without our preprocessing approach. We also compare performances of the two proposed similarity measures. Results show that the percentage of pixels whose labels were accurately predicted is much higher with our preprocessed data than the one with raw data.