Band Selection via Explanations From Convolutional Neural Networks

Abstract

Band Selection is a research hotspot in the field of hyperspectral imaging (HSI) processing. This paper proposes a method that selects bands for HSI classification by the explainability of a convolutional neural network (CNN). We design a CNN architecture and use its 1D gradient-weighted class activation mapping (GradCAM) to obtain a gradient-weighted heatmap (GradHM) by the last layer in the well-trained CNN. Since the pooling layer in the CNN leads to a dimension change of the GradHM, cubic spline interpolation (CSI) is used to up-sample the GradHM. To further improve the accuracy of the up-sampled GradHM for the important band, guided backpropagation was adopted to obtain a more detailed GradHM (Guided-GradHM). Finally, based on GradHM and Guided-GradHM, two strategies, named Average Selection (AS) and Total Selection (TS), are proposed to form new and different combinations of methods (GradHM+AS, GradHM+TS, Guided-GradHM+AS, and Guided-GradHM+TS). In experiments, the proposed methods show better performance compared to other methods. In most cases, if the selected bands were fewer, then the Guided-GradHM+AS was more credible than the other methods. Furthermore, different datasets were used to validate the proposed methods.