Early Detection of Sugarcane Smut Disease in Hyperspectral Images

Abstract

Sugarcane smut, caused by the fungus Sporisorium scitamineum, is a serious sugarcane disease in Queensland, which can cause 30-100% production loss. Early detection of smut disease is a key step towards disease management. However, early-stage smut symptoms are not visible to the human eye. To address this challenge, we leverage the capability of hyperspectral imaging in data acquisition beyond the human visual spectrum and propose a deep Convolutional Neural Network (CNN) to classify sugarcane images as infected with S. scitamineum or healthy. A key component of the CNN is the Dual Self-Attention Block (DSAB) module that is proposed to identify important image features both spectrally and spatially. Experiments on a collected hyperspectral image dataset show the effectiveness of our proposed method in detecting smut disease before visible symptoms appear.