Deep Learning approach for Multiple Source Classification in Remote Sensing Imagery

In this paper, we present a deep learning approach for learning for multiple remote sensing sources. The method starts by eliminating the distribution shift between the different sources and the target dataset using an adversarial learning approach based on min-max entropy optimization. After convergence, the results are aggregated using an average fusion layer. As pre-trained CNN we use in the work the recent state-of-the-art EfficientNet models. In the experiments, we assess the method on four remote sensing datasets acquired over different locations of the earth's surface and are labeled by different experts. The obtained results confirm the promising capability of the proposed method.