Lightweight Vision Transformer Compression for On-Device Plant Disease Detection in Resource-Constrained Agricultural Field Conditions

Deploying Vision Transformers onto edge devices in agriculture remains constrained by memory and compute. This study introduces a pipeline integrating Hessian-balanced adaptive block pruning, quantization, and attention distillation, shrinking a chilli crop disease ViT from 327.42 MB down to 6.01 MB while preserving 95.13% accuracy across cross-village test conditions. Notably, an ablation shows a directly trained compact student reaches 94.87% at identical footprint, transparently delineating where multi-stage compression warrants its computational overhead.