Applying Deep Hybrid Neural Network for Image Classification

With advancements in sensor technology, classification of hyperspectral image (HSI) has gained popularity in the field of research. The graph convolutional neural network (GCNN) approach is widely used for applications such as learning over graph representation and semi-supervised learning. Research studies demonstrate learning results of graph filers and CRNN having superior results compared to other linear models in extracting hyperspectral image features. However, when applied on real world pattern learning situations CRNN performance have shortcomings. In this study, deep hybrid-multi-graph neural network (DHMN) approach is used to facilitate pattern learning using convolutional network approach. To solve issues related to (i) smoothing-spectral filter approach is used to extract spectral features and (ii) autoregressive moving average (ARMA) filter approach is used to prevent noising. HSI datasets are used to conduct experiments. To smooth and refine deep hybrid network features, GraphSAGE-based network approach is used. Results of experiments demonstrate DHMN performance is better than CRNN model.