Research questionHow can Vision Transformers retain plant-disease detection accuracy when compressed for resource-constrained field devices?Plant-disease detection must run on devices with limited storage and computation, while field images can differ across villages and devices. It is unclear how much accuracy compression can preserve and whether combined techniques justify their added cost over a directly trained small model. Latest papersRecent research connected to this question, newest first.Lightweight Vision Transformer Compression for On-Device Plant Disease Detection in Resource-Constrained Agricultural Field ConditionsThe study evaluates three-class Capsicum annuum disease detection using Hessian-guided block pruning, INT8 quantization, and attention-based knowledge distillation. Evidence comes from a 327.42 MB FP32 baseline, compressed models as small as 6.01 MB, four tested configurations, and a same-size directly trained student on a cross-village, cross-device out-of-distribution split; broader crops, datasets, and deployments are not established.research paper · Sep 4, 2026