Hyperspectral Image Fusion Algorithm Based on Improved Deep Residual Network

Abstract

High-resolution hyperspectral images are rich in both spatial and spectral information and have numerous applications for target detection, recognition, and tracking. The method proposed in this paper consists of three specific modules: multi-scale input, improved residuals, down-sampling, and two-channel pooling. By using wider and deeper blocks of residuals, the classification accuracy of the model is significantly improved. To increase the depth of the network, 1×1 convolutions were added to the 3×3 convolutions compared to the basic residual block. This makes it easier to control the number of channels so that the information sensed and the features extracted are different during each convolution. In addition to the Elu activation function, the Dropout layer and the batch normalization layer for every 3×3 convolutions in the network, the residual network is enhanced by using two stacked residual blocks after the two-channel pooling layer. Extensive experiments on many well-known benchmark datasets have shown that the proposed method achieves competitive performance.