Explainable Lightweight CNN for On-Device Rice-Leaf Disease Detection in Indonesian Smallholder Farms

Abstract

Background: Rice is the primary staple food crop in Indonesia, yet leaf diseases—including bacterial blight, blast, and brown spot—cause annual yield losses of 10–30%, disproportionately affecting smallholder farmers who lack timely access to plant-pathology expertise; unreliable rural connectivity further limits cloud-based diagnostic tools. High-capacity convolutional neural networks deliver strong accuracy but are too large and slow for low-cost devices, and their opaque predictions undermine farmer trust. Objective: This study designs and evaluates RiceLeaf-Edge, an explainable and lightweight convolutional neural network for on-device rice-leaf disease detection that operates fully offline. Methods: Following Design Science Research methodology, a compact depthwise-separable student network was trained with knowledge distillation from a high-capacity teacher and compressed via INT8 post-training quantization; a Grad-CAM visual explanation module was integrated and evaluated on a rice-leaf dataset comprising five classes (healthy and four disease categories, n = 3,355 images). Results: RiceLeaf-Edge achieved 97.3% accuracy and 97.0% macro-F1—within 0.8 percentage points of the heavy baseline (98.1%) at only 8.9 MB and 34 ms on-device latency versus 92.4 MB and 164 ms for the heavy baseline. Explanations were faithful (insertion score 0.87; deletion score 0.18) with 92.6% symptom agreement. Conclusion: The framework demonstrates that trustworthy, deployable agricultural diagnosis is achievable at the edge on commodity hardware, offering a transferable recipe for edge AI in low-connectivity settings.