Deep Learning With Attention Mechanism For Sweet Pumpkin Disease Identification

Abstract

Sweet pumpkin (Cucurbita moschata) is a commonly grown plant that gains its importance through the means of nutrition and economy. Nevertheless, other leaf rotations including Downy Mildew, Mosaic, Leaf Curl are the threats to its production. Detection of these diseases is usually slow, tedious and inaccurate in most cases using traditional manual methods especially where there is lack of professional agricultural assistance in the rural setup. The paper presents a new system of classifying sweet pumpkin diseases using deep learning with iteration integration using preprocessing image and attention mechanism to enhance clarity and precision of the outcomes. Preprocessing was applied to a dataset of 5 classes of imagery of leaves, resizing, contrast enhancement using CLAHE enhancement and thorough data augmentation to enhance diverse data representation and decrease overfitting. They created two models: Baseline cnn implemented in TensorFlow/Keras with a 93.4 percent validation accuracy and Attention based transfer learning model with (MobileNetV2/ResNet18) implemented in PyTorch with 97.9 percent validation accuracy. The performance was measured in terms of accuracy, precision, recall, F1-score, and confusion matrices, where the attention-based model outperformed others regarding its generalization and feature concentration. The suggested project can be implemented on mobile or IoT-type solutions to offer farmers fast and secure detection of diseases, minimize the consumption of pesticides, and contribute to the sustainable agricultural practices.