An Edge Efficient and Lightweight Hybrid CNN Vision Transformer Network With CBAM for IoT Based Plant Disease Detection

Abstract

Detecting crop disease early enough to act on it remains a bottleneck for smallholder and precision farming alike, particularly when diagnosis depends on manual inspection or lab analysis that cannot keep pace with an outbreak. We present AgriViTNet, a compact CNN-Vision Transformer hybrid built for exactly this constraint: It extracts local texture cues through convolution, relates them across the whole leaf surface through a single-layer Transformer encoder, and sharpens the resulting features toward disease-relevant regions through a Convolutional Block Attention Module (CBAM) before classification. Rather than layering incremental tweaks onto existing designs, this architecture is built around a single question: How much accuracy can be retained once every component is stripped to what an edge device can actually run, and the CBAM stage in particular is placed after the Transformer rather than as a bolt-on addition, which the ablation results later show matters. Training and evaluation used a merged dataset comprising 45 classes, chosen specifically to stress-test the model beyond the single-crop benchmarks common in prior work. On this dataset, AgriViTNet reaches 99.49% overall accuracy, ahead of the MobileH-Transformer, TwoConV-BiLSTM, and VGG-Inception-DenseNet baselines compared here, while staying at 3.1 million parameters and 1.54 GFLOPs, roughly two orders of magnitude smaller than the heaviest hybrid tested. This robust performance, combined with its minimal resource requirements, enables deployment in real-world agricultural settings, supporting timely disease management and sustainable farming practices. The architecture’s scalability and efficiency make it a practical solution for precision agriculture, bridging the gap between technological innovation and field deployment.