Recent advancements in wheat disease detection: A review of image processing, machine learning, and deep learning techniques

Abstract Wheat provides a significant amount of the food calories consumed by humans worldwide, thus diseases of wheat are a major food security concern. Due to climate change, wheat diseases have become more prominent and can reduce crop yields dramatically. Diseases of wheat include rusts, Fusarium head blight, and powdery mildew. Methods for detecting diseases of wheat have ranged from manual inspection to support vector machine and random forest algorithms. Recently, deep learning approaches have been employed, such as convolutional neural networks (CNNs) and transformers to automatically extract features that can diagnose wheat diseases with great accuracy, upwards of 95% accuracy in lab conditions. Pros of deep learning approaches include detection that is early stage, non-destructive, and scalable; ability to be used for precision agriculture practices, as well as limiting pesticide usage when detection deems it unnecessary. Cons include high computational cost, lack of real-field generalizability (changes in light conditions, symptom occlusion, presence of similar-looking backgrounds/symptoms), and lack of model explainability (black box), and requirement of many expert-annotated images. Future directions include training models that have explainable AI (XAI) components, lightweight models that can run on edge computing devices to take this technology to the farm, and more collaboration with researchers, growers, institutions, and open data repositories.