Date of Award
Fall 11-2021
Document Type
Dissertation
Degree Name
Doctor of Science in Information Systems
Department
Business and Information Systems
First Advisor
Omar El-Gayar
Second Advisor
Cherie Noteboom
Third Advisor
Austin O'Brien
Abstract
Traditional means of on-farm weed control mostly rely on manual labor. This process is time-consuming, costly, and contributes to major yield losses. Further, the conventional application of chemical weed control can be economically and environmentally inefficient. Site-specific weed management (SSWM) counteracts this by reducing the amount of chemical application with localized spraying of weed species. To solve this using computer vision, precision agriculture researchers have used remote sensing weed maps, but this has been largely ineffective for early season weed control due to problems such as solar reflectance and cloud cover in satellite imagery. With the current advances in artificial intelligence, past research on weed detection in SSWM has used a large deep convolutional neural network (DCNN) for weed detection. These models are, however, computationally expensive and prone to overfitting on smaller datasets. Consequently, although DCNNs have shown continuous accuracy improvements in research settings, they remain relatively unused for practical purposes in precision agriculture due to their large number of parameters and the difficulty to implement on resource-constrained devices. Accordingly, this research investigated the use of model compression to reduce complexity and increase the efficiency of DCNNs in low-resource conditions. The proposed approach involves stacking two pre-trained DCNN models – Xception and DenseNet – to reduce the effect of performance degradation during the model compression process. A performance evaluation of the resulting XD-Ensemble indicated that the model outperformed both state-of-the-art DCNNs and a lightweight EfficientNet-B1 model in a resource-constrained environment in terms of prediction accuracy, model size, and inference speed. The current study contributes to enhancing viability while minimizing the environmental footprint of agricultural technologies as well as maximizing their production efficiency.
Recommended Citation
Ofori, Martinson, "Transfer-Learned Pruned Deep Convolutional Neural Networks for Efficient Plant Classification in Resource-Constrained Environments" (2021). Masters Theses & Doctoral Dissertations. 371.
https://scholar.dsu.edu/theses/371
Included in
Artificial Intelligence and Robotics Commons, Databases and Information Systems Commons, Data Science Commons, Other Computer Sciences Commons, Theory and Algorithms Commons