Enhanced Agricultural Productivity: Dragon Fruit Leaf Disease Detection Using Deep Learning Models

Abstract

We propose a deep learning approach to identify dragon fruit leaf diseases, namely Anthracnose, Stem Canker, and Scale Insect, through image classification. Agricultural disease diagnosis has been a manual, time-consuming task with high possibilities of inaccuracy, particularly for smallholder farmers lacking the assistance of professionals. To overcome these limitations, a dataset of 2,042 images under natural field conditions from Dragon Fruit Garden, Trishal, Mymensingh, and Fiber Plus Agro, Ashulia, Bangladesh, was collected. Through data augmentation, the dataset was enriched to 6,126 training images. We investigated four deep learning models InceptionV3, DenseNet201, MobileNetV2, and a custom CNN based on transfer learning. Among them, DenseNet201 achieved the highest accuracy of 96.48% with improved feature reuse and classification capability. MobileNetV2, though slightly less accurate, has a promising light model for mobile-based applications. The system suggested herein enables early stage disease detection, which can avoid up to 40% of crop loss and excessive pesticide usage, thereby promoting more sustainable agriculture. This research addresses a real-world problem by bridging the gap between AI research and farming needs on the ground, offering an affordable, scalable, and mobile-compatible solution that can empower farmers with timely and accurate disease diagnosis in resource poor regions.