Efficient Multi-Crop Disease Detection Systems: A PRISMA-Guided Systematic Review of Deep Learning, Vision Transformers, Explainable AI, and Edge-Deployable Models

Abstract

Automated crop-disease diagnosis have moved beyond the single-crop convolutional classifier to multi-crop systems which include lesion localization, hierarchical classification, severity estimation, explainability and edge deployment. However, much of the existing literature is fragmented: many models have been shown to be close to saturated accuracy when trained on curated leaf images; however few models have demonstrated stable performance under field environments (i.e. background), various devices, unseen crops/diseases or varying levels of computational resources. The most prevalent families of deep learning models used were CNN's and Transfer-Learning based models although Hybrid CNN-Transformer architectures, YOLO based lesion localisation algorithms, Ensemble/XAI systems including knowledge distillation and quantization, and semantic zero-shot approaches all experienced significant growth. Benchmark accuracy rates greater than 95% were commonly reported by many studies; yet, studies conducted in field oriented settings often found disease classification, severity grading, and cross domain transfer to be significantly more difficult than controlled environment background recognition. A total of 22 studies validated their findings with either field or external validation; 17 studies provided evidence related to latency and/or model size; 14 studies examined severity and/or localized lesions at the level of pixels; and 13 studies included some form of explanation. Based upon the synthesis of results an effective multi-crop architecture would consist of a crop aware hierarchical pipeline: image quality gating; a lightweight leaf/lesion localization algorithm; a shared CNN-MobileViT backbone architecture; crop and disease specific heads; calibrated uncertainty estimates; severity estimation; visual explanations; and a quantized edge model connected to a cloud based drift detection and active learning loop. The review concludes that advances will likely depend less on achieving another small margin increase in benchmark accuracy and more on demonstrating cross region validation, open set recognition, calibration, energy consumption measurements and clinically analogous reporting of failures.