Biological Systems Engineering


First Advisor

Dr. Yeyin Shi

Date of this Version

Fall 12-2-2022


P. Singh, Semantic Segmentation based deep learning approaches for weed detection, M.S. thesis, University of Nebraska-Lincoln, 2022


A THESIS Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfillment of Requirements For the Degree of Master of Science, Major: Agricultural and Biological Systems Engineering, Under the Supervision of Professor Yeyin Shi. Lincoln, Nebraska: December, 2022

Copyright © 2022 Puranjit Singh


Global increase in herbicide use to control weeds has led to issues such as evolution of herbicide-resistant weeds, off-target herbicide movement, etc. Precision agriculture advocates Site Specific Weed Management (SSWM) application to achieve precise and right amount of herbicide spray and reduce off-target herbicide movement. Recent advancements in Deep Learning (DL) have opened possibilities for adaptive and accurate weed recognitions for field based SSWM applications with traditional and emerging spraying equipment; however, challenges exist in identifying the DL model structure and train the model appropriately for accurate and rapid model applications over varying crop/weed growth stages and environment. In our study, an encoder-decoder based DL architecture was proposed that performs pixel-wise Semantic Segmentation (SS) classifications of crop, soil, and weed patches in the fields. The objective of this study was to develop a robust weed detection algorithm using DL techniques that can accurately and reliably locate weed infestations in low altitude Unmanned Aerial Vehicle (UAV) imagery with acceptable application speed. Two different encoder-decoder based SS models of LinkNet and UNet were developed using transfer learning techniques. We performed various measures such as backpropagation optimization and refining of the dataset used for training to address the class-imbalance problem which is a common issue in developing weed detection models. It was found that LinkNet model with ResNet18 as the encoder section and use of ‘Focal loss’ loss function was able to achieve the highest mean and class-wise Intersection over Union scores for different class categories while performing predictions on unseen dataset. The developed state-of-art model did not require a large amount of data during training and the techniques used to develop the model in our study provides a propitious opportunity that performs better than the existing SS based weed detections models. The proposed model integrates a futuristic approach to develop a model that could be used for weed detection on aerial imagery from UAV and perform real-time SSWM applications

Advisor: Yeyin Shi