Evolving Domain Granulation Based Graph Convolutional Neural Network with Co-Teaching for Remote Sensing Image Classification

Domain adaptation (DA) is an important area of research in remote sensing image classification. Advancement in the remote sensing (RS) technology, produced images at different resolutions and from different domains within a class. The deep feature extractors like Resnet-50 are used to obtain features from the images. These features possesses uncertainty within. It is an important challenge for a classifier to deal with the uncertainty in the feature vectors. The uncertainty in the feature vectors is maximum when the images are obtained from multiple domains within a class. Motivated with this, we propose an evolving domain granulation (EDG) to deal with the uncertainty in the feature vectors from multiple domains with-in a class. Graph convolutional neural network (GCN) is known for its ability to process the feature vectors and implement DA in a non-regular grid structure. We introduce the concept of co-teaching with ensemble of classifiers in GCN to adapt the evolving domain granulated feature for un-supervised RS image classification. The model is named as evolving domain granulation based graph convolutional neural network with co-teaching (EDG-GCN). The superiority of proposed model over state-of-the-art models is tested with multispectral RS scene images and hyperspectral RS patch images.