Automatic fusion and classification using random forests and features extracted with deep learning

Fusion of different sensor modalities has proven very effective in numerous remote sensing applications. However, in order to benefit from fusion, advanced feature extraction mechanisms that rely on domain expertise are typically required. In this paper we present an automated feature extraction scheme based on deep learning. The feature extraction is unsupervised and hierarchical. Furthermore, computational efficiency (often a challenge for deep learning methods) is a primary goal in order to make certain that the method can be applied in large remote sensing datasets. Promising classification results show the applicability of the approach for both reducing the gap between naive feature extraction and methods relying on domain expertise, as well as further improving the performance of the latter in two challenging datasets.