Leaf disease recognition using deep learning methods

The digitalization of crop production has placed leaf-image-based disease recognition among the top research priorities. This paper presents a compact and reproducible system designed for rapid deployment in cloud environments and subsequent adaptation. The proposed approach combines multitask learning (simultaneous prediction of plant species and disease), physiologically motivated channel processing, and error-tolerant data preparation procedures. Experiments were conducted on the New Plant Diseases Dataset (Augmented). To accelerate training, six of the most represented classes were selected, with up to 120 images per class. Images were resized to 192×192 and augmented with geometric and color transformations as well as soft synthetic lesion patches. The ExG greenness index was embedded into the green channel of the input image. The architecture was based on EfficientNet-B0; the proposed HiP²-Net model included two classification heads for disease and species. Training was carried out in two short stages, with partial unfreezing of the base network’s tail in the second stage. Evaluation employed standard metrics, confusion matrices, test-time augmentation, and integrated gradients maps for explainability. On the constructed subset, the multitask HiP²-Net consistently outperformed the frozen baseline model in accuracy and aggregate metrics. Synthetic lesions reduced background sensitivity and improved detection of mild infections, while incorporating ExG enhanced leaf tissue separation under variable lighting. Integrated gradient maps highlighted leaf veins and necrotic spots, strengthening trust in predictions and facilitating expert interpretation. The proposed scheme combines the practicality of cloud deployment with simple, physiology-inspired techniques. Adopting the “species + disease” setup together with ExG preprocessing and soft synthetic lesions improves robustness to lighting, background, and geometric variations, and makes it easier to transfer models to new image collections.