Development of a spatially explicit active learning method for crop areas mapping from satellite image time series

With the increasing availability of remotely sensed data, there is a pressing need to develop efficient and scalable methods for converting satellite images into useful information. Supervised methods achieve better classification results when there are large quantities of labeled training data available. Despite important efforts to collect and label training data using either conventional methods or crowdsourcing initiatives, it has been suggested that “remote sensing will never have enough training data or the cost of collecting these data might not be realistic” (Ball et al. 2018). Remote sensing, therefore, requires methods that are able to cope with the relatively small amount of training data available. One possibility is to use active learning methods. These methods target at the identification of the so-called informative (unlabeled) samples that can improve the classification model. The selection of informative samples is usually done by considering spectral information (i.e. feature space domain), while ignoring spatial information such as spatial proximity of the available samples. Therefore, innovative solutions are required to incorporate spatial domain in the active learning methods.