WheatDisease-HRY: A Real-Field Wheat Disease Dataset with Baseline Deep Learning Benchmarks for Automated Disease Detection

Abstract

In India, to increase the wheat yield production and support sustainable agricultural practices, timely and correct identification of wheat crop diseases such as yellow rust, brown rust, etc., is very much essential. Further, the majority of deep learning algorithms show high efficiency in plant disease detection and recognition, but these studies rely on handcrafted datasets captured under controlled laboratory conditions and accordingly will not perform well in the real world domain. Furthermore, there is a lack of real field wheat crop disease datasets collected in India, particularly from Haryana state, being the largest producer of wheat crops. Therefore, to address the issue of lack of region specific crop diseases dataset, which are region-specific, the wheat crop disease data set is curated, consisting of 2672 images of healthy and diseased leaves of wheat crops collected from the real field of CCS Haryana Agricultural University, Hisar. The dataset includes typically three classes of leaves of wheat crop, i.e., Yellow Rust (910), Brown Rust (922) and Healthy leaves (840) collected using both a DSLR camera and different smartphone cameras under natural lighting and field conditions. All images were validated and labeled by taking the expertise of wheat pathologists to ensure reliability on the dataset. The study also provides a comprehensive, systematic workflow for transforming raw data into a high-quality benchmark dataset for training using image preprocessing and augmentation techniques. Besides this, a comparative benchmarking analysis is performed under identical experimental conditions using three widely adopted deep learning architectures, which includes ResNet50, MobileNetV2 and EfficientNet-B0. The experimental results show that ResNet50 and EfficientNet-B0 achieve similar performance i.e. approximately 91% classification accuracy on real-field data. However, MobileNetV2 offers a lightweight alternative suitable for mobile and edge deployment. Furthermore, Grad-CAM based explainability analysis was performed to validate model predictions and highlight disease specific regions in wheat leaves. Therefore, this study contributes a practical region specific WheatDisease-HRY dataset and baseline benchmarking framework for developing robust AI based wheat disease diagnosis tools for real world agricultural applications in India.