A CUSTOMIZED SWIN TRANSFORMER-BASED FRAMEWORK FOR CASSAVA LEAF DISEASE CLASSIFICATION

Abstract

Cassava leaf diseases present significant agricultural challenges due to visual similarity between pathological conditions and variability in field conditions, complicating timely intervention. The accuracy of disease identification in early stages has been critical in preventing crop losses; however, due to symptom overlap and environmental variations, manual monitoring has become increasingly difficult. In this paper, a deep learning approach for cassava disease diagnosis named "Modified Swin Transformer Framework" has been proposed, attempting to enhance classification capability by employing a transformer-based vision approach. In the proposed method, the hierarchical structure of Swin Transformer has been customized based on input dimensionality, adaptive patch embedding, and output targeting for cassava disease classification. In this approach, the input image has been split into adaptive non-overlapping patches and processed using shifted windows and attention within these patches. This process has helped the method link all windows efficiently by avoiding locality issues of non-overlapping regions in attention, while being computationally efficient. The framework has further developed based on Swin Transformer architecture and has included adaptive patch and position embeddings to take advantage of the transformer's global-linking capability by employing multi-head attention in these embeddings. Furthermore, the framework has developed and incorporated multi-scale feature aggregation into this method, which utilizes hierarchical feature fusion with these inclusive designs to address multi-scale symptom representation during processing. The inclusion of multi-scale aggregation has therefore facilitated this method to link global patterns as well as local patterns; hence, its integrity has helped improve disease classification capability by minimizing intra-class variability of cassava diseases and increasing inter-class differences among Cassava Bacterial Blight, Cassava Brown Streak Disease, Cassava Green Mottle, Cassava Mosaic Disease, and healthy leaves. In testing the proposed framework, an accuracy of 96.80% and an F1-score of 96.40% have been achieved on the Kaggle public dataset, which has outperformed standard CNN models and baseline Swin Transformer; the framework has thus proved its effectiveness as a computer-assisted tool for cassava disease observation and classification.