Abstract:
Early and accurate identification of paddy leaf diseases is crucial for timely intervention and
effective crop management. Although deep learning based approaches have demonstrated strong
performance in paddy leaf disease classification, they often struggle to capture the diverse multiscale
symptom patterns of diseases, resulting in the misclassification of challenging samples. Moreover, the
high computational cost of deep learning based models remains a significant limitation. To address
these challenges, this study proposes a novel Haar wavelet-guided deep Feature Pyramid Network
(FPN) model. In the proposed framework, deep features are extracted using a backbone network, and
a FPN is employed to capture multiscale disease symptoms. In parallel, Haar wavelet features are
extracted to provide complementary texture cues that are not explicitly modelled by deep features.
Guided by these wavelet features, a gating-based dynamic feature selection module is then utilised to
identify image-specific FPN features, which are subsequently used for classification. In the second
phase of the study, a lightweight variant of the proposed model is developed using a knowledge
distillation technique, achieving high classification accuracy while substantially reducing the number
of parameters and FLOPs. In addition, a new benchmark dataset, named NP-LankaPaddy, is
constructed to address the limited availability of datasets representing Sri Lankan paddy leaf disease
conditions. Experimental results demonstrate that both the proposed approach and its lightweight
variant outperform state-of-the-art methods across four benchmark datasets, achieving notably high
accuracies of 98.04% and 97.84%, respectively, on the well known Paddy Doctor dataset. The source
code and dataset supporting this study are openly available at https://doi.org/10.5281/zenodo.18334503.
