|FENG, AIJING - University Of Missouri|
|ZHOU, JIANFENG - University Of Missouri|
|Sudduth, Kenneth - Ken|
Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/10/2020
Publication Date: 8/17/2020
Citation: Feng, A., Zhou, J., Vories, E.D., Sudduth, K.A. 2020. Evaluation of cotton emergence using UAV-based imagery and deep learning. Computers and Electronics in Agriculture. 177. Article 105711. https://doi.org/10.1016/j.compag.2020.105711.
Interpretive Summary: Crop emergence is an important factor for making field management decisions, such as replanting, that need to be made at very early stages. Unmanned aerial vehicle (UAV) systems have been an efficient tool for crop scouting and in this study, a novel image processing method was used to map cotton stand count and canopy size in near real time at two weeks after planting. Results showed that the developed method could estimate both stand count and canopy size accurately and the process was more efficient than traditional methods. This study has demonstrated the potential of using UAV-based imaging to estimate cotton stand density quickly and early enough in the season for remedial action. This approach may be useful to researchers and to farmers who are interested in obtaining early-season plant population estimations at relatively low cost and high resolution.
Technical Abstract: Crop emergence can be evaluated using plant population, stand count and uniformity of seedlings. Crop emergence is an important agronomic factor for making field management decisions, such as replanting, that are time-sensitive and need to be made at very early stages. Unmanned aerial vehicle (UAV)-based imaging systems have been an efficient tool for crop scouting, including stand count, however, the efficiency and accuracy need to be improved. The goal of this study was to develop and validate a novel image processing method using deep learning technology for processing individual image frames from a UAV-based RGB imaging system that was used to map cotton stand count and canopy size in near real time. Image frames were pre-processed to correct distortions, calculate ground sample distance and geo-reference cotton rows in the images. A pre-trained deep learning model, resnet18, was used to estimate stand count and canopy size of cotton seedlings in each image frame. Results showed that the developed method could estimate stand count accurately, and estimated stand count could explain 95% of the variation in manually measured stand count (R2 = 0.95) of cotton in the test dataset. Similar results were achieved for canopy size with R2 = 0.95 in the test dataset. The processing time for each image frame of 20 M pixels was 2.22 s, which was more efficient than traditional mosaic-based image processing methods. An open-source automated image-processing framework was developed for cotton emergence evaluation and is available to the community for efficient data processing and analytics.