Theses and Dissertations

Advisor

Gudla, Charan

Committee Member

Wang, Haifeng

Committee Member

Chen, Jingdao

Date of Degree

5-16-2025

Original embargo terms

Visible MSU Only 2 Years

Document Type

Graduate Thesis - Campus Access Only

Major

Computer Science (Research Computer Science)

Degree Name

Master of Science (M.S.)

College

James Worth Bagley College of Engineering

Department

Department of Computer Science and Engineering

Abstract

Soybean (Glycine max) plays a vital role in global food systems, yet its productivity is often reduced by diseases that can lead to significant yield losses, ranging from 8% to 25% annually, and costing up to $3.9 billion in the United States alone. To address this, early and precise detection of soybean diseases is essential. In this study, we utilize Neural Architecture Search (NAS) to develop an optimized deep learning model named SoyNASNet, designed for the Soybean Disease Image Dataset (ASDID), containing 9,422 images across eight disease classes. The dataset poses challenges, including class imbalance and visual similarity between classes. SoyNASNet achieves an accuracy of 98.3%, outperforming state-of-the-art models such as DenseNet121, ResNet50, Swin Transformer, and other. Key architectural parameters were optimized through NAS, and model interpretability was enhanced using filter visualization and Grad-CAM. The results highlight NAS’s potential in precision agriculture through scalable and explainable deep learning solutions.

Download

Share

COinS