Edge-optimized multimodal cross-fusion architecture for efficient crop disease detection

Accurate and timely crop disease detection is critical for reducing agricultural losses and ensuring food security in low-resource settings. Traditional diagnostic methods, such as manual inspections, are often inefficient and error-prone. Existing deep learning models (e.g., ResNet50, Inception V3) struggle with computational inefficiency and poor generalizability in real-world farming contexts. This study proposes a lightweight multimodal fusion model integrating EfficientNetV2 and MobileNetV2, optimized for edge deployment. The architecture leverages compound scaling and feature fusion to recognize subtle disease patterns, and it was fine-tuned on a globally diverse dataset (PlantVillage and field-collected leaf images). The proposed model achieved state-leading metrics (99.0% accuracy, 0.993 precision, 0.990 F1-score, AUC = 0.999997), outperforming benchmarks like ShuffleNet and DenseNet50 (ranked 2nd–6th). Statistical validation via the Kruskal-Wallis test confirmed significant performance differences across models (H=614.90, p=1.4237e−129), with Bayesian analysis showing a 100% superiority probability over DenseNet50. Notably, the model exhibited the lowest confidence variance (0.000012) compared to alternatives (0.000014–0.000032), demonstrating unmatched prediction stability. Deployment on low-end mobile devices posed challenges such as computational constraints and offline usability. However, the TensorFlow Lite-powered mobile app addressed these limitations, offering real-time, offline disease classification with 0.094-second inference latency on devices with ≤2GB RAM. Validated on 249 unseen field images (95.98% accuracy), this solution bridges the gap between high-performance deep learning and real-world agricultural needs, empowering smallholder farmers with an accessible and scalable tool.