A Spiking Neural Network for Hyperspectral Image Classification
Zhengda Han, Yu Li · 2025
Hyperspectral image classification methods based on deep learning models suffer from the problems of complex model structure and long training time. According to the characteristics of event driven and low energy consumption of spiking neural network, a network architecture is proposed based on the Leaky Integrate-and-Fire model, which encodes the input image using a direct coding method and applies a Gaussian function for backpropagation. The network framework incorporates spiking standard convolution, depth convolution, and point convolution to significantly reduce computational complexity. Experimental results indicate that the proposed method achieves superior classification performance on the Pavia University and WHU-Hi-LongKou datasets compared to existing mainstream algorithms, attaining overall accuracies of 99.60% and 99.18%, respectively, while also reducing training time. This study verifies the efficiency and practicality of spiking neural network in hyperspectral image classification and provides new ideas for lightweight model design.