Hyperspectral image classification using a new deep learning model based on pseudo-3D block and depth separable 2D–3D convolution

Abstract

In the task of classification of hyperspectral images (HSI), standard 2D and 3D convolutions increase the number of trainable parameters, making the model computationally complex, prone to overfitting when trained on smaller datasets, and demanding a large amount of memory. Thus, the cost of computation and memory consumption are bottlenecks of deep hybrid 2D–3D CNN models. This inspired us to propose a light weighted hybrid CNN model with fewer trainable parameters for efficient and accurate classification of HSI by introducing the concept of pseudo-3D block and depth separable 2D–3D convolution. The major components of the proposed framework are standard 3D-CNN, pseudo-3D block, depthwise separable 3D CNN, depth separable 2D CNN, and spatial pyramid pooling (SPP). Pseudo-3D block and depth separable 2D–3D convolution make the model computational and memory efficient by drastically reducing the number of trainable parameters. Further, the SPP layer is placed between the convolution layers and fully connected layers to generate a feature vector of fixed length at different scales. The proposed model is applied to three publicly available datasets. Experimental analysis proves the efficacy of the proposed approach over the state-of-the-art methods.