Robust Edge-AI for Explainable Maize Leaf Disease Diagnosis using CNN-Transformer Fusion

Foliar fungal pathogens—Northern Leaf Blight, Common Rust, and Gray Leaf Spot—routinely reduce global maize harvests by over 30%. While vision algorithms exhibit high diagnostic accuracy in laboratory benchmarks, practical field adoption is impeded by rural connectivity gaps, uninterpretable black-box decisions, and compute-heavy architectures on low-cost mobile hardware. We propose an edge-native, explainable diagnostic framework uniting an EfficientNet-B0 convolutional stream with a Data-efficient Image Transformer (DeiT-Tiny) via a fused 1472-dimensional bottleneck. To enable real-time interpretability without backpropagation, a Linearized Class Activation Mapping (Linear-CAM) surrogate is embedded directly into the forward ONNX inference graph, generating 224 × 224 heatmaps in < 1.8 ms. Dynamic post-training INT8 quantization compresses storage by 42.48% (down to 21.84 MB), lowers inference energy by 45.59% (47.08 mJ/frame), and accelerates CPU execution to 48.04 ms on commodity smartphone silicon while sustaining 98.75% diagnostic accuracy. In-situ field trials across commercial maize plots in Vehari, Pakistan, validated 96.0% top-1 accuracy. The pipeline is packaged into an offline mobile application (Asan Kissan) supporting multi-region canopy touch cropping, calibrated Sigmoid triage gating, and audio guidance. Model checkpoints, code, and Android binaries are open-sourced at GitHub - muslimraza9989-cmyk/Asan-Kissan-EdgeAI