Semi-supervised Segmentation-driven Classification Pipeline for Grading Cassava Leaf Diseases

Cassava, a vital food security crop widely cultivated in tropical regions, serves as a major carbohydrate source globally. However, its growth and productivity are consistently hindered by the prevalence of viral, bacterial, and fungal diseases. This paper introduces a novel semi-supervised segmentation-driven classification framework designed to identify cassava leaf diseases with high precision. Utilizing a modified Color Index Vegetation Extraction (CIVE) technique, our framework generates pixel-level pseudo labels that facilitate the training of a segmentation unit without the need for densely annotated data. This semi-supervised approach significantly reduces the reliance on extensive human-annotated resources and effectively segments cassava leaves from complex backgrounds, enabling accurate classification of four major cassava diseases: cassava mosaic disease (CMD), cassava bacterial blight (CBB), cassava brown streak disease (CBSD), and cassava green mottle (CGM). The proposed framework is validated on the publicly available Kaggle Cassava 2019 dataset, on which it outperforms the state-of-the-art works by 7.69%, 8.98%, and 7.77% in terms of precision, recall, and F1 score, respectively. Additionally, it yielded a mean accuracy of 89.73% and an F1 score of 0.89 on the Cassava Leaf Disease 2020 dataset.