Date of Award
Spring 3-2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Information Systems (PhDIS)
First Advisor
Omar El-Gayar
Second Advisor
Ronghua Shan
Third Advisor
Deb Tech
Abstract
In agriculture, the presence of weeds adversely impacts yield. Confronting this challenge manually requires continuous and labor-intensive field monitoring processes. Artificial intelligence, notably deep learning is one the most promising for weed detection, particularly when deployed to edge devices. However, these edge devices have limited computational resources, making it challenging to deploy large Deep Learning (DL) models often resulting in a performance-size trade-off. This research proposes a three-pronged approach that leverages DL lightweight architectures (LWA) to optimize for size and minimize computational resources at the edge, transfer learning to mitigate constraints associated with limited data commonly encountered in this domain, and Bayesian optimization for hyper-parameter tuning.
Evaluation of the proposed approach reveals that, compared to past research, the lightweight architectures utilized in this study surpassed the previously best-performing state-of-the-art ResNet152V2 model. Notably, the LWA EfficientNetV2B0 model achieved an accuracy of 95.79% with a model size of 23 MB, while LWA MobileNetV2 attained an accuracy of 95.09% with a model size of 13.96 MB. Overall, LWA MobileNetV2 attained an accuracy nearly on par with EfficientNetV2B0, while utilizing a substantially smaller model. Such efficiency is crucial for edge devices constrained by their computational capacity.
Implementing advanced deep learning models within weed detection systems can incur substantial costs. The findings from this research contribute towards enhancing both performance and accessibility. The proposed approach demonstrates the efficacy of LWA coupled with transfer learning and hyper-parameter tuning to address the unique needs of the domain while producing better accuracy results and considerable reduction in model sizes compared. The approach can be generalized to other problem domains with similar constraints. In regard to practice, the proposed approach provides a technological solution that is more affordable and accessible to mid and low-tier farmers, thereby enabling them to increase productivity by effectively addressing weed infestation during the critical early stages of crop growth.
Recommended Citation
Shaheen, Ali, "Weed Detection Using Lightweight Deep Learning Models with Transfer Learning and Hyperparameter Optimization" (2024). Masters Theses & Doctoral Dissertations. 446.
https://scholar.dsu.edu/theses/446