Context-Aware Multimodal Transfer Learning for Multi-Crop Plant Disease Detection
Precise identification of a disease in a crop is a critical step in protecting yields, optimizing inputs, and managing this sustainably. Decisions based on image-based deep learning models have achieved high diagnostic accuracy on controlled datasets, but in field variability, similar symptoms, nutrient stress and unfavorable environmental conditions, the decisions based on images could be uncertain. Study compares the performance of classical machine learning baselines, custom-trained convolutional neural networks (CNNs), and pretrained or transfer learning models on the challenging PlantVillage benchmark dataset with 54,305 RGB images of leaves across 38 crop-health classes and 14 plant species. The overall best classical methods was SVC using RBF kernel with high accuracy of 83.58% and F1 score of 82.96%. As depth and regularization increased, it was observed that the accuracy gradually increased and reached 95.71% for the functional CNN with dropout. The best results at the image level were obtained by the Transfer learning method: MobileNetV2 reached an accuracy of 99.30% and an F1-score of 99.30%. Based on these findings, the paper proposes an interpretable multimodal transfer learning decision-support architecture, comprising a MobileNetV2 image encoder, a residual MLP for soil attributes and dual LSTM encoders for a 48-hour and a 168-hour weather window. The evidence image, soil, full-weather and shortweather evidence are respectively assigned explicit weights of 0.80, 0.10, 0.05, and 0.05 in the proposed late-fusion mechanism. The framework is modular, explainable, and as desirable as possible in situations of decision support in the field where a diagnostic aspect is to be enhanced by the agronomic context.