A Novel Unsupervised Approach for Training Convolutional DSSNs on Remote Sensing

This letter proposes an unsupervised approach for training deep semantic segmentation networks (DSSNs) with convolutional (CNN) layers: Weighted Dual SSN (WDSSN). Feature maps from the encoder-decoder block are grouped into subclusters before passing through the argmax layer. Pseudo labels for guiding the learning process are achieved with a superpixel refinement step. The network weight updates are computed through a batch-weighted cross-entropy (CE) loss proposed to tackle the problem of uneven class distributions usually present in remote sensing tasks. Experimental results show that our approach can learn high-level features from satellite imagery, and performs better than previously published unsupervised convolutional DSSN methods, obtaining accuracies of 63.3% in the Potsdam dataset, 76.3% in the Potsdam-3 dataset, and 39.0% in the Coco dataset. github.com/jpklock2/WDSSN.