JoWUA

Abstract

Cotton (Gossypium hirsutum) is a backbone crop of the world textile industry, accounting for more than USD 600 billion of trade worldwide every year. Bacterial Blight (Xanthomonas citri pv. malvacearum), Target Spot (Corynespora cassiicola) and Cotton Leaf Curl Virus (CLCuV) are foliar diseases which impose yield losses estimated at 25% to 70% under endemic conditions, posing threat to food security and farmer livelihoods in South Asia, Sub-Saharan Africa and the Americas. Current deep learning approaches have significant drawbacks: Convolutional networks that are primarily one-stream are not able to collect both macro-level (global) information regarding lesions and the micro-level (local) information regarding tissue texture and texture grading is not possible at the same time, which is needed for agronomic decision support. In this study, a novel dual-stream transfer learning framework, called AgriLens-Net, that uses both parallel EfficientNetV2-S and MobileNetV3-Large backbones for global context extraction and local lesion texture profiling, respectively. A cross-attention based multi-modal fusion layer dynamically combines and weights the spatial feature maps of both streams, and further a dual-head classification architecture simultaneously outputs labels of the disease categories and continuous disease severity scores. The framework reaches a diagnostic accuracy of 98.74%, an F1-score of 98.61% on macro-average and a benchmark MSE of 0.031 on the Kaggle Cotton Plant Disease. When installed on NVIDIA Jetson Orin Nano edge devices, AgriLens-Net maintains real-time inference at 38ms per frame, confirming the feasibility of deployment in a field. To ensure co-optimization of both objectives without task interference, multi-task learning with composite loss function 𝐿𝑡𝑜𝑡𝑎𝑙=𝛼,𝐿𝑑𝑖𝑎𝑔+𝛽,𝐿𝑔𝑟𝑎𝑑 is used, where α and β are weighting coefficients for the diagnostic classification loss and the severity grading loss, respectively. These results validate AgriLens-Net as a production-ready solution for precision cotton disease management.