Using Multispectral Imaging and Artificial Intelligence to Detect Crop Diseases and Pests Early
Аннотация
Crop disease and pest infestation is important for improving agricultural production and for the reduction of chemical use. Nevertheless, traditional RGB-based vision systems can still lack sensitivity for early detection of subtle physiological changes before symptomatic manifestations. This work presents a hybrid multisensor fusion framework that relies on CNN-Attention for multispectral and thermal automated crop-health evaluation. The anticipated model employs 10,800 multisensor observations divided into five categories of crop health: healthy, fungal disease, bacterial disease, viral disease, and pest infestation. Stratified sampling was utilized on the dataset which was split into 70% for training (7560), 15% for validation (1622) and 15% for testing (1618). The framework combines attention-based feature fusion, hybrid feature extraction, and multiclass prediction with multispectral and thermal branches based on two parallel CNNs. In an experimental evaluation against Support Vector Machine, Random Forest, RGB-CNN, multispectral CNN, thermal CNN, and CNN fusion models, the proposed scheme is showing top performance with 96.48% accuracy, 96.21% precision, 96.05% recall, 96.13% F1 score and 0.98 AUC. According to the results of the ablation analysis, we observe the contribution of each of the different models. The models of multispectral-only, multisensor fusion, attention-enhanced, and complete fusion resulted in a performance of 92.18%, 94.12%, 95.61%, and 96.48% respectively. The performance evaluation under different illumination conditions, early disease incidence, low disease severity, and at different growth stages is above 94.27%. The proposed system is expected to deliver an accurate, interpretable and scalable AI for precision agriculture. This would be achieved by fusing spectral, spatial, and physiological crop information for the early management of diseases and pests. Graphic abstract