A deep learning model for detection and classification of coffee-leaf diseases using the transfer-learning technique

Abstract

The early treatment and detection of plant diseases are important processes, as many diseases affecting crops are highly contagious. Recent advancements in deep learning have helped to provide innovative tools that have not only assisted early detection, but also significantly improved the performance and accuracy of Coffee Leaf Disease (CLD) classification and treatment. However, training a deep learning model from scratch can be both resource and time-consuming. To overcome this challenge, the transfer learning technique can take full advantage of pre-trained models for more general tasks on extensive datasets o ameliorate the performance of a new, related task using few-shot training. This paper proposes a deep learning model, coupled with transfer learning, for CLD detection that aims to provide high accuracy forecasting of diseases that could affect coffee leaves. Our method involves 195 different pre-trained deep learning models, including real-time models like MobileNet and dense ones like EfficientNet and ResNet for the detection of four different diseases. The findings suggest that the EfficientNetB0 model, with transfer learning, has the most relevent accuracy (99.99%), and thus offers an effective solution for coffee leaf diseases classification of. This result could be used to develop applications that help coffee growers to improve the productivity and quality of their crops through early and accurate detection of coffee plant leaf diseases. Such an Artificial Intelligence based application would provide growers with timely control measures, preventing the spread of disease, and minimizing crop damage.

Keywords

• transfer-learning
• coffee
• diseases
• few-shot
• real-time