Ai-Powered Multimodal Crop Disease Detection Using Deep Learning

Plant diseases drastically cut yield and profit. Although CNNs generalize well to image-labeled leaves in the lab, performance suffers in fields because of inconsistent lighting, occlusions, distracting scene background, and overlapping symptoms due to multiple coinciding stresses. Here we introduce an AI-enabled multimodal monitoring platform that integrates leaf imagery with sensor (temperature, humidty, soil moisture, and pump/motor ON/OFF) observations and image features optionally extracted with UAVs to enhance robustness. The pipeline encompasses four model families: (i) an image-focused ResNet-50 CNN baseline; (ii) image-only transformer (ViT-B/16); (iii) late multimodal fusion of the top image features with sensor-inspired MLP; and (iv) Attentionfusion model learning image-ViT embeddings and sensor context in unison. Tested on standardized PlantVillage/PlantDoc corpora within a balanced set of$\mathbf{6 5}$-class corpora and an expanded 65 -class corpora set, multimodal methods reach$97-98 \%$test accuracy with better macro-F1, especially among rare and visually ambiguous classes, outperforming image-baseline methods. Interpretabilty comes in the form of Grad-CAM maps combined with attention-rollout providing lesion location with context attribution. All artifacts (PyTorch implementation, Colab notebooks, logs, confusion matrices, checkpoints) become available for reproducibility. Our findings demonstrate multimodality fortifying field-scale disease surveillance at scale, enhancing the potential for edge deployability with farmer-accepted transparent diagnostics with context awareness.