A Comprehensive Review on Deep Learning Approaches for Cotton Disease Detection in Low-Resource Environments

Cotton farming faces persistent yield losses from leaf diseases that can be reduced significantly through early automated detection. While convolutional neural network (CNN)-based approaches have demonstrated benchmark accuracies exceeding 95%, translating those results into practical tools for rural farmers remains challenging. This review analyses 30 peer-reviewed studies published between 2019 and 2025, retrieved from IEEE Xplore, Scopus, Web of Science, Google Scholar and Elsevier ScienceDirect. Four methodological directions are evaluated: transfer learning, lightweight CNN architectures, ensemble and hybrid methods and super-resolution (SR) Pre-processing. Transfer learning backbones such as VGG16 achieve up to 98.1% accuracy but carry model sizes exceeding 500 MB that are incompatible with budget Android hardware. Lightweight architectures notably MobileNetV3-Small (9.4 MB, ~95 ms CPU inference) and EfficientNet-Lite0 (15.6 MB, ~120 ms), substantially reduce computational cost but degrade under low-quality field images. SR Pre-processing improves classification accuracy by 5–10% on degraded inputs, yet the combined SR-classifier pipeline has not been validated on real mobile devices. Three critical gaps are identified and prioritised: the absence of geographically representative, field-collected cotton leaf datasets from Indian growing regions; the lack of end-to-end pipelines that jointly optimise lightweight SR with mobile-ready classifiers; and the near-complete absence of on-device evaluation under authentic agricultural conditions. Pairing MobileNetV3 or EfficientNet-Lite with an efficient SR stage and deploying via TensorFlow Lite represents the most technically credible near-term path to a tool that smallholder farmers can genuinely use. Future experimental validation of this pipeline on real-device hardware with field-collected cotton images is explicitly identified as the primary recommended next step, directly addressing the predictive nature of this review.