The agricultural sector in South Asia, especially countries like Bangladesh, China, and India, relies heavily on cotton, a crop vital to both the economy and the global textile industry. However, cotton production faces significant challenges from leaf diseases, which can drastically reduce crop yield, impacting farmers’ livelihoods and the region’s economy. Traditional methods for detecting these diseases involve manual inspection, which is labor-intensive, time-consuming, and prone to error. Although machine learning and deep learning approaches have been developed for automated disease detection, most existing models require large datasets and high computational resources, limiting their applicability in resource-constrained areas. This paper presents a novel, lightweight ensemble model optimized for detecting cot- ton leaf diseases on mobile devices using a small dataset. Leveraging transfer learn- ing, our approach enables farmers to use as few as 40 known images to fine-tune the ensemble model directly on their devices, achieving high classification accuracy even with minimal data. We evaluate four pre-trained models—ResNet50, MobileNetV2, EfficientNetB0, and InceptionV3—and compare their effectiveness. After trans- fer learning, ResNet50 achieved the highest accuracy of 97.87%, and InceptionV3 achieved 93.61%, demonstrating their suitability for small-scale, mobile-friendly ap- plications. The ensemble model achieved a classification accuracy of 95.7% through averaging and 97.87% using majority voting. This approach empowers farmers in de- veloping regions by providing a practical, adaptable tool for early disease detection, potentially reducing crop loss and enhancing yield with limited resources.
