Edge Intelligence for Field-Deployable Plant Disease Detection: A Critical Review of Lightweight Models, Adaptation, and Uncertainty-Aware Learning, with Implications for Pomegranate
Shrihari D. Khatawkar, Manikrao DhoreIntroduction:
While laboratory-trained models are highly accurate, several factors can affect the accuracy of models in the field, such as changes in lighting, background noise, plant growth stage, and imaging conditions. In this review, emphasis will be placed on robust and efficient edge-intelligence approaches for detecting pomegranate diseases.
Methods:
Based on PRISMA, eight aspects were covered: lightweight CNN/YOLO models, Kolmogorov-Arnold networks, Mamba-based models, domain and test-time adaptation, uncertainty quantification, foundation models, model compression, and multimodal sensor fusion. Of these studies, 108 were published within the last six years (2020-2026), and three were fundamental studies.
Results:
Emphasis is placed on the trade-offs between accuracy, computational efficiency, adaptation stability, and uncertainty calibration. While lightweight models fail to capture the small lesions, heavy models are computationally expensive; state-space models and test-time adaptation either lose local information or suffer from compounding errors under distribution shift. Running and managing foundation models on resource-limited edge devices is also challenging.
Discussion:
While most studies have investigated one challenge, there are a couple of studies that have combined visual features with bioimpedance and/or temporal (weather) data to detect presymptomatic features. This shows the need for integrated architectures that involve efficient representation learning, model compression, model adaptation, multimodal sensing, and riskaware prediction.
Conclusion:
PomStage-Net, a conceptual framework (not yet experimentally validated), is proposed for lightweight, temporally ordinal, and uncertainty-calibrated detection of pomegranate diseases. Future research should focus on causally inspired multimodal sensor fusion, adaptive learning, and uncertainty characterization to achieve reliable, real-time, fielddeployable agricultural diagnosis.