| |
| Web URL(s): | https://wssa.net/wp-content/uploads/ilovepdf_merged-26.pdf#page=209
Last checked: 04/19/2024
Requires: PDF Reader
Notes: Item is within a single large file |
|---|
| Publication Type:
| Report |
|---|
| Content Type: | Abstract or summary only |
|---|
| Author(s): | Joseph, Mikerly M.;
Zhao, Chang;
Schumann, Arnold W.;
Boyd, Nathan;
Petelewicz, Pawel |
|---|
| Author Affiliation: | Joseph, Zhao, and Petelewicz: University of Florida, Gainesville, FL; Schumann: University of Florida, Lake Alfred, FL; Boyd: University of Florida, Balm, FL |
|---|
| Title: | Database development workflow and comparison of different annotation methods for weed detection in turfgrass systems |
|---|
| Section: | Oral SWSS Contests
Other records with the "Oral SWSS Contests" Section
|
|---|
| Meeting Info.: | San Antonio, Texas: January 22 - 25, 2024 |
|---|
| Source: | Weed Science Society of America - Southern Weed Science Society Joint Meeting. 2024, p. 205-296. |
|---|
| Publishing Information: | [United States]: Weed Science Society of America |
|---|
| # of Pages: | 2 |
|---|
| Abstract/Contents: | "Targeted spraying using ground-based equipment enables drastic reduction in herbicide inputs and application costs. Such technologies rely on maximized efficacy of computer vision-based deep learning algorithms to detect weeds in real-time. Their performance can be elevated by expanding and diversifying training dataset. However, the choice of annotation method and model development procedure may impact the minimum amount of data required for successful training. This study evaluated 1) the performance of object detection versus semantic segmentation in spotted spurge [Chamaesyce maculata (L.) Small] recognition in bermudagrass (Cynodon spp.) turf canopy using You Only Look Once (YOLOv8) model trained with highly restricted dataset and 2) single-step training process using manually labeled images only and two-step training in which model was first pre-trained with manually labeled portion of training dataset, and then deployed to automatically predict target weed in remaining images (step 1). Following this both manually and automatically labeled images were pooled together and used to retrain the model (step 2). All models were trained using a limited dataset of 800 training images, 100 validation images and 138 test images, to determine whether the selected approach yields improvements. Semantic segmentation outperformed object detection as denoted by higher precision (0.61 vs. 0.40), recall (0.62 vs. 0.33), and F1 score (0.62 vs. 0.36) and resulted in successful (all aforementioned metrics >0.50) spotted spurge detection despite the low amount of training images used. The equivalent size training dataset was insufficient to successfully train the object detection model. Therefore, further limiting dataset for pretraining step resulted in poor automatic labeling and lack of improvements in final model's performance. This indicates that prerequisite for use of such two-step approach is pre-trained model's ability to already adequately detect target plant (i.e., achieve precision, recall and F1 score of >0.50)." |
|---|
| Language: | English |
|---|
| References: | 0 |
|---|
| See Also: | See also related item "Simulation-based nozzle density optimization for maximized efficacy of a machine-vision weed control system for applications in turfgrass settings" ASA, CSSA, SSSA International Annual Meeting, 2023, p. 152514, R=333544. R=333544 |
|---|
| Note: | This item is an abstract only! |
|---|
| ASA/CSSA/SSSA Citation (Crop Science-Like - may be incomplete):
Joseph, M. M., C. Zhao, A. W. Schumann, N. Boyd, and P. Petelewicz. 2024. Database development workflow and comparison of different annotation methods for weed detection in turfgrass systems. Abstr. Annu. Meet. Weed Sci. Soc. Am. p. 205-296. |
| Fastlink to access this record outside TGIF: https://tic.lib.msu.edu/tgif/flink?recno=336048 |
| If there are problems with this record, send us feedback about record 336048. |
| Choices for finding the above item: |
|---|
| Web URL(s):
https://wssa.net/wp-content/uploads/ilovepdf_merged-26.pdf#page=209
Last checked: 04/19/2024
Requires: PDF Reader
Notes: Item is within a single large file |
| 
MSU catalog number: b2180085 |
| Request through your local library's inter-library loan service (bring or send a copy of this TGIF record) |
