SC-CAN: Spectral Convolution and Channel Attention Network for Wheat Stress Classification

Biotic and abiotic plant stress (e.g., frost, fungi, diseases) can significantly impact crop production. It is thus essential to detect such at an early stage before visual symptoms damage become apparent. To this end, paper proposes a novel deep learning method, called Spectral Convolution Channel Attention Network (SC-CAN), which exploits the difference in spectral responses of healthy stressed crops. The proposed SC-CAN method comprises two main modules: (i) convolution module, consists dilated causal convolutional layers stacked residual manner capture features; (ii) channel attention global pooling layer fully connected that compute inter-relationship between feature map channels scaling them based on their importance level (attention score). Unlike standard convolution, focuses local features, learn both features. These also have long receptive fields, making suitable for capturing dependency patterns hyperspectral data. However, because not all maps produced by are important, we propose module weights according level. We used classify salt (i.e., stress) four datasets (Chinese Spring (CS), Aegilops columnaris (co(CS)), Ae. speltoides auchery (sp(CS)), Kharchia datasets) Fusarium head blight disease biotic dataset. Reported experimental results show outperforms existing state-of-the-art techniques with overall accuracy 83.08%, 88.90%, 82.44%, 82.10%, 82.78% CS, co(CS), sp(CS), Kharchia, datasets, respectively.