Corn emergence uniformity estimation and mapping using UAV imagery and deep learning

Authors: VONG, CHIN NEE - University Of Missouri; CONWAY, LANCE - University Of Missouri; FENG, AIJING - University Of Missouri; ZHOU, JIANFENG - University Of Missouri; Kitchen, Newell; Sudduth, Kenneth - Ken

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/22/2022
Publication Date: 5/6/2022
Citation: Vong, C., Conway, L.S., Feng, A., Zhou, J., Kitchen, N.R., Sudduth, K.A. 2022. Corn emergence uniformity estimation and mapping using UAV imagery and deep learning. Computers and Electronics in Agriculture. 198. Article 107008. https://doi.org/10.1016/j.compag.2022.107008.
DOI: https://doi.org/10.1016/j.compag.2022.107008

Interpretive Summary: Corn seeds sometimes do not emerge uniformly due to cold and wet soils, planter malfunction, inconsistent seeding depths, and tillage practices. When corn plants do not emerge at about the same time (i.e., within one to two days of each other) and spacings between plants are not consistent, grain yield can be reduced. For farmers to detect corn stand problems over large areas when trying to decide whether to replant, or for making other crop management decisions, automated field-level assessment tools are needed. This research was conducted to explore the ability of unmanned aerial vehicle (UAV) images to estimate corn stand characteristics shortly after plant emergence. One outcome of this research was the adaptation of a modeling method to allow processing of the UAV images for correct estimations of corn stand properties. The modeling outcomes were very good, especially for plant density (i.e., plants per area) and emergence date estimation. The model predicted more than 95% of the variation in these two properties. The model explained variation in spacings with a 73% accuracy. We also found that seeds planted at deeper depths generally resulted in fewer plants, were later emerging, and had poorer uniformity. The final outcomes of this research were field-scale precision maps of stand properties, including plant density, spacing, and emergence date. As this process of plant population assessment becomes more automated, farmers will be able to successfully utilize UAV scouting to help evaluate corn stand over 1000s of acres. This they will be able to use for more effective and economical crop management decisions.

Technical Abstract: Assessment of corn (Zea Mays L.) stand uniformity is important to evaluate crop yield potential. Previous studies have shown the potential of unmanned aerial vehicle (UAV) imagery and deep learning (DL) models in estimating early stand count and plant spacing uniformity, but few have extended further to field-scale mapping. Additionally, estimation of plant emergence date using UAV imagery in field-scale studies has not been achieved. This study aimed to estimate and map corn emergence uniformity using UAV imagery and DL modeling. Corn emergence uniformity was quantified with plant density, plant spacing standard deviation (PSstd), and mean days after first emergence (DAEmean). A UAV imaging system equipped with a red, green, and blue (RGB) camera was used to acquire images at 10 m above ground level at 32 days after planting (20 days after first emergence) in a field planted at four planting depths (3.8, 5.1, 6.4 and 7.6 cm). A pre-trained convolutional neural network, ResNet18, was used to estimate the three emergence parameters. The results showed an estimation accuracy of 0.97, 0.73, and 0.95 for plant density, PSstd, and DAEmean for the testing dataset. When the estimation was done on the whole field, the average plant density and DAEmean decreased and the average PSstd increased with increasing depths, indicating deeper depths caused less and late emergence as well as less spatial uniformity. However, when observing the emergence parameters along transects with landscape variability, the 5.1 cm depth generally had the lowest coefficient of variation, CV (i.e., variability) for all the parameters. The 7.6 cm depth had the highest CV for plant density and PSstd for nearly all the replications but lower CV for DAEmean than 3.8 and 6.4 cm depths. The findings indicate that UAV imagery and the methods presented can capture field-scale variability in different emergence parameters. These results will be applicable for future field-scale agronomic studies relating to crop emergence uniformity as affected by factors such as planting environment, crop management practices, and varieties/hybrids as well as for making planting decisions in commercial production.