Cross-scale detection and cross-crop generalization verification of tomato diseases in complex agricultural environments

Global agricultural diseases cause over 220 billion US dollars in annual economic losses. Early detection of crop leaf diseases is crucial for precision plant protection. While deep learning has advanced single-crop disease detection, models often struggle with cross-crop scenarios due to domain specificity and data scarcity, particularly in resource-limited developing countries. Traditional image processing methods relying on manual feature design lack generalization in dynamic field environments, facing issues like leaf occlusion, diverse lesion shapes, fluctuating light, and weak early symptoms. The advent of CNNs has shifted the paradigm, but challenges like obscured disease spots in complex backgrounds, similar disease textures, and difficulty in identifying low-contrast early lesions persist. This paper addresses model generalization bottlenecks caused by domain differences, proposing a cross-crop transfer learning framework to overcome domain shift limitations by sharing low-level features and optimizing domain adaptation strategies. The primary contributions include a dual-branch adaptive alignment module (TAAM) for robust detection of small and early disease spots, a Faster-GLUDet unit to enhance noise suppression while remaining lightweight, and a multi-scale decoupled detection head (MDH) for balanced detection of various lesion sizes.

The study utilized a "Tomato Leaf Diseases Detect" standardized dataset from Roboflow, comprising six disease categories and healthy leaves across different disease stages with 3,469 high-resolution RGB images. These diseases are prevalent globally, causing significant yield losses. Additional datasets for common beans and potatoes were included to validate transferability. To enhance model robustness, the dataset was augmented to 7,370 images using techniques like horizontal/vertical flipping, grayscale conversion, contrast/brightness adjustment, and synthetic weather features (rain, fog, solar flare, overexposure, snow) based on an atmospheric scattering model. The proposed Tomato Multi-scenario Adaptive Scale Detector (ToMASD) is a lightweight architecture specifically designed for high-precision disease detection in complex field settings. It incorporates three core modules: the Two-branch Adaptive Alignment Module (TAAM) for dynamic multi-scale feature alignment, the Faster-Gated Linear Unit (Faster-GLUDet) for adaptive feature refinement and background noise suppression, and the Multi-scale Decoupling Head (MDH) for independent optimization of classification and regression tasks at different feature scales. The TAAM addresses insufficient feature extraction of small-scale targets and redundant shallow computations by employing a PSAStem for initial feature extraction and dual pathways integrating C3k2 modules and a PSABlock, followed by an Adaptive Alignment Module (AAM) for dynamic weight adjustment and multi-scale feature complementarity. Faster-GLUDet enhances disease feature extraction in complex backgrounds through a gating mechanism, utilizing FasterNetBlock with Partial Convolution for efficient spatial feature extraction and a Convolutional Gated Linear Unit (ConvGLU) to suppress background noise and enhance lesion areas. The Multi-scale Decoupling Head (MDH) receives P3, P4, and P5 feature maps from the FPN, enhancing them through grouped normalized convolutional modules and then decoupling into independent classification and regression branches for precise target categorization and location.

The experiments were conducted on a Linux server with 2*A100 GPUs, Intel Xeon Gold CPU, PyTorch 2.0.1, and CUDA 11.8. Image input size was 640×640 pixels, with an initial learning rate of 0.01, SGD optimizer, weight decay of 0.0005, momentum of 0.937, 200 epochs, and a batch size of 32. Evaluation metrics included Precision, Recall, and mAP. ToMASD achieved a precision of 84.3% and mAP of 81.7%, significantly outperforming thirteen mainstream object detection models. Compared to the YOLOv11n baseline, ToMASD increased P, mAP, and Recall by 6.6%, 7.8%, and 5.9%, respectively, with a compressed computational load of 7.1 GFLOPs. Comparative experiments showed ToMASD maintained superior performance even under challenging weather conditions (foggy and strong light), controlling false detection rates at 6.3% and 9.8% respectively, where other models misidentified fog as disease or missed lesions. The TAAM module's performance was validated by testing different attention modules, with PSA achieving the best balance of accuracy and computational efficiency. ToMASD also demonstrated efficient generalization, achieving 92.1% accuracy for bean leaf diseases and 93.5% for potato leaf diseases through transfer learning. Ablation studies confirmed the effectiveness of TAAM, Faster-GLUDet, and MDH, with the full ToMASD model achieving the optimal balance between parameters, computational efficiency, and performance, improving accuracy by 9.4% over the baseline with only a marginal increase in computational cost. Heatmap visualization confirmed the model's adaptive feature capabilities across different crops, indicating successful lesion localization and adaptive adjustments through weight transfer between tomato, common bean, and potato datasets.

ToMASD effectively addresses complex agricultural environment challenges for tomato disease detection by integrating multi-scale feature fusion and dynamic attention mechanisms. The Two-branch Adaptive Alignment Module (TAAM) mitigates feature attenuation and misalignment, contributing significantly to a proven mAP of 81.7%. The Faster-GLUDet module enhances noise suppression with a local context-aware gating mechanism, reducing false positives in foggy (6.3%) and strong light (9.8%) conditions while maintaining computational efficiency at 7.1 GFLOPs. The Multi-scale Decoupling Head (MDH) ensures balanced detection across different lesion sizes through group normalization and independent task-specific branches, demonstrating stable performance across all six disease categories, even with class imbalance. While ToMASD shows superior performance, further optimization is needed for similar texture diseases like Black Spot (mAP of 76.1%). Future research directions include exploring knowledge distillation for model compression, integrating spatio-temporal features for disease spread prediction, and deploying the model on low-power edge computing devices with LoRaWAN for real-time field monitoring. The study's lightweight architecture and transfer learning paradigm offer a scalable solution for intelligent agricultural disease diagnosis, promoting digital transformation from single-crop to multi-species and multi-environment collaborative management, thus supporting sustainable agricultural development.