A Lightweight and Efficient Deep Learning Framework for Rice Crop Disease Detection with Domain-Invariant Feature Extraction

Abstract

Rice fields suffer yearly yield losses from plant diseases. Deep models often struggle to generalize across various disease categories and rely on resource-intensive preprocessing, hindering practical deployment. This study drastically reduces high parameter count by proposing a lightweight and robust deep learning framework “CfPNet” which delivers excellent generalization across various rice disease classes including under-explored sheath rot under natural environmental conditions. A key component of our framework is a domain-invariant feature extractor which integrates dual-path hybrid normalization strategy in the feature extraction layers to suppress domain-specific variations while maintaining class-relevant features. Without preprocessing, the proposed model achieves a success rate of 98.5%, which is higher than the 97.76% accuracy achieved with hybrid preprocessing approach. These results mark evidence of the robustness of proposed framework to generalize effectively under natural environmental conditions. The results are validated using 5-fold cross validation, t-test and benchmarked against state-of-the-art pulling a highly efficient lightweight architecture.