A Semantic Classification Approach on Luffa Aegyptiaca Leaf Diseases Detection Utilizing Multiple Models

Abstract

This paper proposes a semantic classification method for the identification of Luffa aegyptiaca leaf disease utilizing deep models. Since there were not any similar public data available, a highresolution dataset of 1,800 images was captured from actual agriculture fields of Jalkuri, Narayanganj, and Khagan, Ashuliya, Bangladesh. There were six distinct classes in the dataset: Alternaria Leaf Spot, Angular Leaf Spot, Downy Mildew, Fresh (Healthy), Holed, and Mosaic Virus. Random resized cropping, flip horizontal and vertical, rotation, and addition of noise were some of the data augmentation and preprocessing methods used to preprocess the data before training. Some of the CNN-based models ResNet50, VGG19, InceptionV3, ResNet152V2, and a light-weight custom CNN were compared on the performance basis. Out of the above, 97.27% accuracy was achieved by ResNet50, and it was found to be extremely efficient in making a discrimination among a set of patterns of disease leaf. ResNet50 has been employed with web platform in the form of Flask, thereby field-level deployed it among farmers and agricultural experts. Comparison comparison, precision-recall scores, confusion matrices found to establish that all of the models have been classified. The outcome of this research carries a major implication for plant early disease detection and precision agriculture in confined rural settings.