Quantifying corn emergence using UAV imagery and machine learning

Corn (Zea mays L.) is one of the important crops in the United States for animal feed, ethanol production, and human consumption. To maximize the final corn yield, one of the critical factors to consider is to improve the corn emergence uniformity temporally (emergence date) and spatially (plant spacing). Conventionally, the assessment of emergence uniformity usually is performed through visual observation by farmers at selected small plots to represent the whole field, but this is limited by time and labor needed. With the advance of unmanned aerial vehicle (UAV)-based imaging technology and advanced image processing techniques powered by machine learning (ML) and deep learning (DL), a more automatic, non-subjective, precise, and accurate field-scale assessment of emergence uniformity becomes possible. Previous studies had demonstrated the success of crop emergence uniformity using UAV imagery, specifically at fields with simple soil background. There is no research having investigated the feasibility of UAV imagery in the corn emergence assessment at fields of conservation agriculture that are covered with cover crops or residues to improve soil health and sustainability. The overall goal of this research was to develop a fast and accurate method for the assessment of corn emergence using UAV imagery, ML and DL techniques. The pertinent information is essential for corn production early and in-season decision making as well as agronomy research. The research comprised three main studies, including Study 1: quantifying corn emergence date using UAV imagery and a ML model; Study 2: estimating corn stand count in different cropping systems (CS) using UAV images and DL; and Study 3: estimating and mapping corn emergence under different planting depths. Two case studies extended Study 3 to field-scale applications by relating emergence uniformity derived from the developed method to planting depths treatments and estimating final yield. For all studies, the primary imagery data were collected using a consumer-grade UAV equipped with a red-green-blue (RGB) camera at a flight height of approximate 10 m above ground level. The imagery data had a ground sampling distance (GSD) of 0.55 - 3.00 mm pixel-1 that was sufficient to detect small size seedlings. In addition, a UAV multispectral camera was used to capture corn plants at early growth stages (V4, V6, and V7) in case studies to extract plant reflectance (vegetation indices, VIs) as plant growth variation indicators. Random forest (RF) ML models were used to classify the corn emergence date based on the days after emergence (DAE) to time of assessment and estimate yield. The DL models, U-Net and ResNet18, were used to segment corn seedlings from UAV images and estimate emergence parameters, including plant density, average DAE (DAEmean), and plant spacing standard deviation (PSstd), respectively. Results from Study 1 indicated that individual corn plant quantification using UAV imagery and a RF ML model achieved moderate classification accuracies of 0.20 - 0.49 that increased to 0.55 - 0.88 when DAE classification was expanded to be within a 3-day window. In Study 2, the precision for image segmentation by the U-Net model was [greater than or equal to] 0.81 for all CS, resulting in high accuracies in estimating plant density (R2 [greater than or equal to] 0.92; RMSE [less than or equal to] 0.48 plants m-1). Then, the ResNet18 model in Study 3 was able to estimate emergence parameters with high accuracies (0.97, 0.95, and 0.73 for plant density, DAEmean, and PSstd, respectively). Case studies showed that crop emergence maps and evaluation in field conditions indicated an expected trend of decreasing plant density and DAEmean with increasing planting depths and opposite results for PSstd. However, mixed trends were found for emergence parameters among planting depths at different replications and across the N-S direction of the fields. For yield estimation, emergence data alone did not show any relation with final yield (R2 = 0.01, RMSE = 720 kg ha-1). The combination of VIs from all the growth stages was only able to estimate yield with R2 of 0.34 and RMSE of 560 kg ha-1. In summary, this research demonstrated the success of UAV imagery and ML/DL techniques in assessing and mapping corn emergence at fields practicing all or some components of conservation agriculture. The findings give more insights for future agronomic and breeding studies in providing field-scale crop emergence evaluations as affected by treatments and management as well as relating emergence assessment to final yield. In addition, these emergence evaluations may be useful for commercial companies when needing justification for developing new technologies relating to precision planting to crop performance. For commercial crop production, more comprehensive emergence maps (in terms of temporal and spatial uniformity) will help to make better replanting or early management decisions. Further enhancement of the methods such as more validation studies in different locations and years as well as development of interactive frameworks will establish a more automatic, robust, precise, accurate, and 'ready-to-use' approach in estimating and mapping crop emergence uniformity.