EnviroSpect-guided ConvMixer–ViT framework for environment-robust potato leaf disease detection

Abstract

Potato is one of the world’s most important staple crops, yet its productivity is persistently compromised by foliar diseases particularly early blight (Alternaria solani) and late blight (Phytophthora infestans) which together account for substantial annual yield losses. Conventional visual diagnosis by agronomists is laborintensive, subjective, and ill-suited to large-scale deployment, while prevailing deep learning solutions tend to degrade under real-world variability in illumination, background, and leaf orientation, and typically require large, annotated datasets that are rarely available in field conditions. To address these limitations, we propose a hybrid deep learning framework that couples ConvMixer, for localized lesionscale feature extraction, with a Vision Transformer (ViT), for modeling long-range spatial dependencies, and aggregates their representations through an elementwise ensemble. The pipeline is preceded by EnviroSpect, a novel preprocessing scheme that independently remaps hue, enhances saturation, and normalizes intensity in a custom hue–saturation–intensity (HSI)-based color space to preserve disease-discriminative cues under heterogeneous imaging conditions. Classification is performed by a prototypical-network-based head, which leverages class prototypes computed from training features to produce similarity-based decisions that complement standard softmax classification. The framework was trained and evaluated on three datasets: PlantVillage, Potato Plants, and a custom field-acquired AgriPlant set, together comprising 5,304 images across three disease classes. The proposed model attains classification accuracies of 99% on PlantVillage and 98% on Potato Plants, with a macro F1-score of 99.02%, precision of 99.20%, and recall of 98.85%, consistently outperforming individual ConvMixer, ViT, ResNet18, MobileNetV2, and InceptionV3 baselines. Ablation and statistical significance analyses (p< 0.05) confirm that each component — the hybrid backbone, EnviroSpect preprocessing, and the prototypical network head contribute meaningfully to overall performance, with EnviroSpect providing the largest single improvement. Importantly, the model achieves these results with only 8.9 M parameters and 19 ms per-image graphics processing unit (GPU) inference, striking a favorable accuracy– efficiency balance for deployment on resource-constrained edge devices. The proposed framework achieves a strong accuracy–efficiency balance for potato leaf disease detection and offers a transferable foundation that could be extended toward scalable, environment-robust plant-disease monitoring in other crops and agricultural scenarios.