A Neural Network with Multiscale Convolution and Feature Attention Based on an Electronic Nose for Rapid Detection of Common Bunt Disease in Wheat Plants

Common bunt disease in wheat is a serious threat to crops and food security. Rapid assessments of its severity are essential for effective management. The electronic nose (e-nose) system is used to capture volatile organic compounds (VOCs), particularly trimethylamine (TMA), which serves as a key marker of common bunt disease in wheat. In this paper, the GFNN (gas feature neural network) model is proposed for detecting VOCs from the e-nose system, providing a lightweight and efficient approach for assessing disease severity. Multiscale convolution is employed to extract both global and local features from gas data, and three attention mechanisms are used to focus on important features. GFNN achieves 98.76% accuracy, 98.79% precision, 98.77% recall, and an F1-score of 98.75%, with only 0.04 million parameters and 0.42 million floating-point operations per second (FLOPS). Compared to traditional and current deep learning models, GFNN demonstrates superior performance, particularly in small-sample-size scenarios. It significantly improves the deep learning performance of the model in extracting key gas features. This study offers a practical, rapid, and cost-effective method for monitoring and managing common bunt disease in wheat, enhancing crop protection and food security.