Crop Mapping through an Adaptive Multiview Active Learning Strategy

In this paper we present preliminary results for a new methodology based on an active learning (AL) approach for crop mapping from hyperspectral remote sensing images. The proposed solution is based on an adaptive multiview (MV) AL strategy which improves state-of-The-Art methodologies in three main ways: 1) view sufficiency is increased by a spectral-spatial view generation approach which incorporates spatial features coming from segmentation maps; 2) diversity across views is guaranteed by generating a dynamic view at each iteration by selecting important features from the predefined views; 3) further improvements are obtained using an ensemble of classifiers instead of a single one. The method is validated experimentally on the Indian Pine dataset with the aim of discriminating among different crop types. We give evidence of improvements in terms of classification accuracies with respect to state-of-The-Art AL strategies.