Abstract: The mycogone perniciosa of Agaricus bisporus caused huge economic losses to mushroom farmers. In view of the unobvious early symptoms and the lack of effective disease diagnosis methods, the microscopic hyperspectral imaging technology was applied to the early detection of the mycogone perniciosa of Agaricus bisporus. Firstly, the BS-Net-FC algorithm based on the attention mechanism and the sparse auto-encoder reconstruction network was used to select the band of microscopic hyperspectral image of chlamydospore by means of the strong nonlinear learning ability of neural network. The MTCEM algorithm based on the morphological specificity of chlamydospore targets was proposed to detect the chlamydospore targets. The results showed that the band selection algorithm ensured the information of the band subset and effectively reduced the redundant bands. The AUC of MTCEM algorithm reached the best value of 0.878 5 on the band subset with 17 bands. This method could effectively reduce the data redundancy and had good detection effect on the chlamydospore targets, thus to provide new idea and technical support for the early detection of the mycogone perniciosa of Agaricus bisporus.