Efficient Hyperspectral Imagery Classification Method with Lightweight Structure and Image Transformation-Based Data Augmentation

With the increasing popularity of deep learning models in the hyperspectral image (HSI) classification field, more and more complex methods have been proposed. However, an increase in accuracy was accompanied by an increase in model size and computational complexity. As a result, application of these models will be limited in solutions with strict hardware specifications, as well as solutions with a limited training set. To reduce the required number of parameters and total FLOPs, we design a lightweight model for HSI classification, based on a ghost module. Specifically, we use 3D ghost modules to build an efficient 3D-CNN network termed TinyNet. To further increase the performance of our TinyNet model, a combination of cross-entropy and contrastive center loss is utilized for training. Additionally, we design a novel augmentation technique based on image transformations. The proposed lightweight model, together with our augmentation technique, can lead to satisfactory HSI classification results. Experimental results on two HSI datasets demonstrate the effectiveness of the proposed HSI classification method when compared to the competing techniques.