Transfer Learning Approaches for Crop Stress Classification With Limited Training Data: A Systematic Review and Meta-Analysis

The early, accurate classification of biotic and abiotic crop stress is vital to food security. Deep learning excels here but demands extensive data, yet annotated field images for specific crop–stress combinations remain rare and costly. Transfer learning has become the default remedy. Following the PRISMA ۲۰۲۰ protocol, we conduct a systematic review of literature from ۲۰۱۶ to ۲۰۲۶ on transfer learning for crop-stress classification under limited data. We extract reported accuracy alongside moderators—transfer paradigm (fine-tuning, feature extraction, few-shot meta-learning, self-supervised, and foundation-model adaptation), backbone family, dataset scale, and whether evaluation used laboratory or real field conditions. Transfer learning confers a large, consistent advantage over training from scratch in low-data regimes, with the steepest gains under the most severe scarcity—a few to a few dozen labelled images per class. However, an under-reported laboratory-to-field gap persists: accuracy on curated single-domain benchmarks often exceeds ۹۵%, yet drops ۲۰–۳۵ percentage points under cross-domain field evaluation. We argue this gap cannot be closed by incremental benchmark gains, and propose a standardized, domain-shift-aware evaluation protocol to make future claims comparable, reproducible, and agronomically meaningful. We note (Section ۳.۶) that this synthesis is indicative rather than a fully weighted meta-analysis—the natural next step.

کلیدواژه ها:

transfer learning ، crop stress classification ، few-shot learning ، domain shift ، precision agriculture ، meta-analysis ، PRISMA ، vision transformers