DC-SAM: DILATED CONVOLUTION AND SPECTRAL ATTENTION MODULE FOR WHEAT SALT STRESS CLASSIFICATION AND INTERPRETATION

Abstract

Salt stress can impact wheat production significantly and is difficult to be managed when the condition is critical. Hence, detecting such stress whet it is at an early stage is important. This paper proposed a deep learning method called Dilated Convolution and Spectral Attention Module (DC-SAM), which exploits the difference in spectral responses of healthy and stressed wheat. The proposed DC-SAM method consists of two key modules: (i) a dilated convolution module to capture spectral features with large receptive field; (ii) a spectral attention module to adaptively fuse the spectral features based on their interrelationship. As the dilated convolution module has long receptive fields, it can capture short- and long dependency patterns that exist in hyperspectral data. Our experimental results with four datasets show that DC-SAM outperforms existing state-of-the-art methods. Also, the output of the proposed attention module reveals the most discriminative spectral bands for a given wheat stress classification task.