Deep learning model for detection and classification of banana diseases based on leaf images

This title will be presented on Friday, December 15, 2023 at 09.15 - 09.25 GMT+7.

Abstract

This title will be presented on Friday, December 15, 2023 at 09.15 - 09.25 GMT+7. 

Fungal diseases are one of the main reasons for low productivity in banana farming. Early detection of fungal diseases is important. One possible approach is using machine vision. Due to its high accuracy, deep learning is the most widely used algorithm in machine vision for many solutions. Deep learning's ability to model the data into multiple levels of abstraction makes it suitable for many agricultural solutions. However, deep learning requires a high computational resource, challenging many agricultural solutions primarily implemented on devices with low computing resources. Therefore, this study proposes a deep-learning model for detecting and classifying banana diseases based on leaf images. The aim is to train lightweight deep learning and compare their performance for detecting and classifying banana diseases. The study used a dataset of images of a total (16092), which represent three classes: Black Sigatoka (5767), Fusarium Wilt Race 1 (4697), and Healthy (5628). The dataset was trained with four lightweight deep-learning algorithms: MobileNetv2, MobileNetv3-small, ShuffleNetv2, and SqueezeNet. The results showed that SqueezeNet outperforms all other models with an accuracy of 97.1%, precision of 97.1%, recall of 97.1%, and F1-score of 97.1%. MobileNetv2 performed poorly with an accuracy of 73%, precision of 75%, recall of 73%, and F1-score of 71%. For complexity, MobileNetv3-small is the heaviest model with a size of 14 MB and took a short time of 2.465 minutes to train. MobileNetv2 is the lightest model, with a size of 2.51MB, while SqueezeNet took a longer time to train, with an estimated time of 14.76 minutes. Generally, lightweight deep learning performed well in classifying banana diseases based on leaf images. However, improvement must be made to enhance the generalization capability of the algorithms using techniques such as automatic hyperparameter tuning, segmentation, and adding an attention layer.