Comparative Analysis Of CNN Models For Lettuce Leaf Disease Detection In Hydroponic Farming

Abstract

Hydroponic farming is rapidly growing in the Philippines as a sustainable solution for small-scale farmers especially in regions like Nueva Ecija. Although hydroponically grown lettuce has several advantages, it is highly susceptible to diseases that affects quality and productivity. Thus, prompt and accurate detection of disease is essential in reducing losses and ensuring sustainable production. This study presents a comparative analysis of three commonly use Convolutional Neural Network (CNN) models, namely EfficientNet, InceptionV3, and ResNet50 to identify the best CNN model which detect lettuce leaf diseases in hydroponic systems. A diverse dataset collected from online sources and local hydroponic farms in Nueva Ecija, subsequently augmented, generating a total of 1,056 images. Transfer learning was used to train each CNN model and evaluated across three test scenarios including the distinction of lettuce from non-lettuce, classification of diseased versus non-lettuce, and identification of specific disease utilizing classification accuracy. The results show that EfficientNet achieved the highest overall mean accuracy of 92.31%, outperforming the other models in every test scenarios. On the other hand, InceptionV3 and ResNet50 revealed difficulties in classifying non-disease and healthy leaves. These findings indicate that EfficientNet could be an instrumental component for developing automated diagnostic applications to monitor diseases on hydroponic farms. The study recommends the dataset should be expanded, and that agricultural stakeholders should collaborate to develop CNN-based detection applications that are more practical and useful in real-life settings. This would help hydroponic farming in the Philippines become more sustainable and profitable.