Improved lightweight convolutional neural network models for the detection and evaluation of Fusarium head blight in wheat.

Fusarium head blight (FHB), a frequent disease in wheat cultivation, can lead to substantial yield losses and the production of mycotoxins in grains. Therefore, the development of wheat varieties resistant to FHB is an important strategy to reduce related losses. In this respect, manual surveys of FHB are time-consuming and labor-intensive. To overcome this issue, this paper proposes a method for detecting and evaluating wheat FHB using color imaging and deep learning. Initially, a lightweight convolutional neural network model based on the You Only Look Once (YOLO) v8s artificial intelligence (AI) model was designed to detect wheat spikes from color images. Testing revealed that the model's mean average precision in spike detection reached 0.964. Moreover, another lightweight model was developed for detecting wheat spikelet and FHB. To enhance the detection capability of the model for small objects, space-to-depth convolution (SPD-Conv) and BiFormer attention modules were integrated. The results indicated that the model can accurately detect spikelet and FHB, with a mean average precision of 0.936. Finally, based on the wheat spikelet detection results, the rate of diseased wheat spikes (RD_S) and the disease index for wheat (DI_W) were calculated to evaluate the severity of wheat FHB. For RD_S and DI_W, the coefficients of determination between phytologists' evaluations and the estimates derived from the proposed method were 0.71 and 0.93, respectively. These results demonstrate that the proposed method facilitates the accurate and efficient detection of wheat FHB and contributes to the quantitative evaluation of FHB in the field.