BananaSqueezeNet: A very fast, lightweight convolutional neural network for the diagnosis of three prominent banana leaf diseases

Abstract

All over the world, bananas are one of the most common fruits. It accounts for nearly 16% of global fruit production. However, every year, a large amount of banana yield losses occur due to different diseases of the banana leaf. It is essential to identify these diseases at an early stage in order to increase banana production. A visual inspection is the most common method of identifying banana leaf diseases. With a visual inspection, errors are common, time is a factor, and expertise is required. This study shows how deep learning and Bayesian optimization can be used to effectively diagnose banana leaf diseases from images without any human intervention. We collected the Banana Leaf Spot Diseases (BananaLSD) dataset from various locations in Bangladesh. The dataset consists of images of three banana leaf diseases: Pestalotiopsis, Sigatoka, and Cordana. Our proposed BananaSqueezeNet model performed exceptionally well in diagnosing banana leaf diseases from the images with an overall accuracy of 96.25%, precision of 96.53%, recall of 96.25%, specificity of 98.75%, F1-score of 96.17%, and MCC of 95.13%. The BananaSqueezeNet model outperforms some state-of-the-art convolutional neural networks that include EfficientNetB0, MobileNetV3, ResNet-101, ResNet-50, InceptionNet-V3, and VGG16. The BananaSqueezeNet model also detected seven other diseases that affect banana leaves, fruits, and stems, including banana fruit scarring beetle, black sigatoka, bacterial soft rot, pseudo stem weevil, yellow sigatoka, banana aphids, and panama disease, with an accuracy of 95.13%. BananaSqueezeNet will enable banana growers to detect banana diseases early, and we hope that it will ultimately lead to an increase in banana production in Bangladesh and around the world.

Keywords

• Deep learning
• Convolutional neural networks (CNN)
• Pestalotiopsis
• Sigatoka
• Cordana
• Banana leaf diseases