Multi-Task Deep Learning Framework for Segmentation and Severity Estimation of Leaf Diseases in Multi-Crop Environments

Abstract

Crop diseases pose a major threat to global food security, creating a pressing need for effective and accurate diagnostic mechanisms that can be applied across diverse agricultural settings. This paper proposes a Multi-Task Deep Learning Framework (MTDLF) for the simultaneous segmentation of diseased regions and estimation of disease severity in crop leaves. The framework employs a shared ResNet-50 encoder with two task-specific decoders: a U-Net-based segmentation branch and a regression-based severity prediction head, trained using a composite loss formulation. In addition to the dual-task architecture, two consistency-driven mechanisms are introduced. A Severity-Constrained Segmentation Refinement (SCSR) module aligns predicted lesion-area proportions with estimated severity values, while a Lesion-Area Distribution Matching (LADM) loss enforces distributional consistency between segmentation outputs and severity-based lesion expectations. The model is trained and evaluated on publicly available, severity-annotated datasets of rice, maize, tomato, grape, and cotton leaves. Experimental results demonstrate that the proposed framework achieves a mean Intersection over Union (IoU) of 85.7%, a Dice coefficient of 88.3%, a Mean Absolute Error (MAE) of 7.5, and an   of 0.92, outperforming conventional single-task methods and recent multi-task baselines. Furthermore, the model attains real-time inference performance of approximately 25 ms per image, making it suitable for edge-level deployment. The proposed MTDLF provides a unified and efficient approach to multi-crop disease monitoring, offering a practical pathway toward reliable, data-driven precision agriculture.