Deep learning-based optical technique for early identification and detection of seed-borne disease

Abstract

The agricultural sector is the backbone of the country's economy. There are a variety of factors that affect crop production, with seed-borne diseases being one of the most important. Seeds and grains are the agricultural sector's backbone, making them critical assets to manage and preserve both pre- and post-harvest. There are different optical techniques for the analysis and processing of diseased seed samples employing various image acquisitions such as laser biospeckle, infrared imaging, and so on, as shown in the literature research. However, traditional image processing systems have some processing and experimentation limitations. As a result, a deep learning-based optical technique processing pipeline for detecting and identifying diseased and healthy seed samples using laser speckle patterns in this study. Transfer learning and ensemble learning-based algorithms and models are used in the proposed study. AlexNet, VGG16, ResNet18, GoogleNet, and MobileNetV2 are the best five state of-the-art transfer learning-based CNN models, with ResNet18 having the greatest accuracy of 95.33%. Along with TL models, a CNN model with an accuracy of 95.5 % was trained on the target dataset as a baseline model for comparing it to other models. Finally, the (re-)trained models are used in an ensemble learning-based strategy. The algorithms Majority Vote and Early Fusion are used. The accuracy of the Majority Vote model, which included all five (re-)trained models, was calculated to be 96.1 %. The accuracy was calculated as 98.1% for the early fusion model, which incorporated the top three models with the best individual test accuracies. The results demonstrate the use of ensemble learning models exceeds the performance of individual state-of the-art models for the targeted dataset and desired task.