LeafMorph-X: A Morphology-Aware End-to-End Deep Learning Framework for Accurate Tomato Leaf Disease Detection
Abstract
Accurate and scalable tomato leaf disease detection remains a critical challenge in precision agriculture due to complex lesion morphology, non-uniform illumination, background clutter, and subtle inter-class visual similarity. The present paper introduces a new end-to-end deep learning architecture, which combines sophisticated image-restoration, foundation-model-based segmentation, self-supervised representation-learning, and state-space classification to achieve robust disease recognition. The first step is NAFNet++, which removes sensor-induced and environmental noise, then retains fine-grained leaf textures and vascular features, and the second step is RetinexFormer, which normalizes illumination and enhances contrast to ensure that perceptual consistency is maintained regardless of the heterogeneous lighting conditions. An efficient SAM-Adapter that incorporates parameter-dependent task-specific prompt tokens, divides diseased lesion areas and leaf veins, and provides supplementary pathological and morphological features, at the same time. Plant-MAE, a self-supervised masked autoencoder adapted to plant imagery, also enhances semantic embeddings of unlabeled data, significantly enhancing discriminability between difficult disease categories. MambaVision, a state-space vision architecture with linear computational complexity, is used to perform final classification, and makes use of cross-attention pooling to capture the structural interaction between lesion appearance and vein morphology. Assessed on six scales (Early Blight, Late Blight, Septoria Leaf Spot, Tomato Yellow Leaf Curl Virus, Bacterial Spot, and Healthy), the proposed framework classifies with an accuracy of 98.4% and an F1-score of 98.1%, indicating a high probability of real-world classification accuracy upon agricultural application.
