Adaptive Granular Neural Networks for Remote Sensing Image Classification

Monitoring and measuring the conditions of the earth surface play important role in the domain of global change research. In this direction, various methodologies exist for land cover classification of remote sensing images. Neural network (NN) is one such system that has the ability to classify land cover of remote sensing images, but most often its accuracy deteriorates with the level of uncertainty. Granular NN (GNN), that incorporates information granulation operation to NN, is one of the solutions to address this issue. However, complexity of the remote sensing dataset demands for a GNN with changeable granular structure (shape and size of granules) based on the requirement. In this study, our objective is to develop a GNN with adaptive granular structure for the classification of remote sensing images. An architecture of this adaptive GNN (AGNN) evolves according to the information present in the incoming labeled pixels. As a result, the AGNN improves the classification performance compared to other similar models. Performance of the model has been tested with hyperspectral and multispectral remote sensing images. Superiority of the proposed model to other similar methods has been verified with performance measurement metrics, such as overall accuracy, users accuracy, producers accuracy, dispersion score, and kappa coefficient.