FUTURE-PROOF COFFEE PLANT DISEASE DETECTION BASED ON COUNTER-FACTUAL RECOMMENDATION WITH A HYBRID VISION TRANSFORMER AND CONVOLUTIONAL NEURAL NETWORK MODEL

Abstract

Coffee plantations are vulnerable to several diseases that harm roots, leaves, and cherries, jeopardizing crop productivity and farmer livelihoods. Small-scale farmers lack access to precise and accessible technologies for diagnosing and controlling these diseases. Traditional machine learning methodologies are restricted to single-disease classification and lack the intricacies of multi-disease contexts. In this work, the proposed model has a unique hybrid model that integrates vision transformer (ViT) and convolutional neural network (CNN) architectures for the identification and early detection of several coffee plant diseases. The ViT module identifies global associations in plant images, while the CNN extracts intricate local characteristics, facilitating thorough disease diagnosis. Furthermore, the counterfactual recommendation system models the impacts of several treatments and preventative strategies on the original images, offering practical insights. Our model attains an accuracy of 0.9881 % on a dataset of 1056 images, surpassing current methodologies. The suggested solution is included in the Affogato app, enabling farmers to make educated, customized choices about disease control. This method not only improves disease detection but also promotes sustainable coffee-growing techniques, enhancing crop production and farmer livelihoods.