Crop performance, aerial, and satellite data from multistate maize yield trials

Accurate genotype-specific early yield estimates at fields and plots offer potential benefits to farmers in optimizing their agronomic practices, breeders in screening hundreds and thousands of varieties, and policymakers in decisions contributing to the overall improvement of agriculture and food production systems. Effective, generalizable approaches to track plant growth and predict yield at the individual plot level require large matched datasets of remote sensing and ground truth data collected across multiple environments. Low-altitude drone flights are increasingly being used to collect data from field evaluations of new crop varieties, while satellite imagery is being explored to track yield and management practices at the regional and field scales. Despite their lower spatial resolution, satellite platforms exhibit multiple logistical and technical advantages in scalability and accessibility, and could facilitate plot-level predictions, especially with steadily improving spatial resolution. However, genotype-specific, plot-level, high-resolution satellite images from multiple environments integrated with the ground truth measurements are not yet publicly available. Here we generated, described, and evaluated a set of more than 20,000 plot-level images of over 80 hybrid maize (Zea mays) varieties grown in six locations across the US corn belt under various management practices collected from (near simultaneous) satellite and drone flights integrated with ground truth measurements of crop yield. Of the six baseline models examined, models employing data collected from satellite images often matched or exceeded the performance of models employing data collected from drones for both within-environment and cross-environment yield prediction. Large, multimodal, multi-environment, genetically diverse training datasets such as those generated in this study, along with more complex models could help unlock the power of satellite imagery as an important new addition to the tool of farmers, plant geneticists, crop breeders, and policymakers.

Methods
UAV Image Acquisition and Processing
UAV visible spectral (RGB) imagery was collected at three time points per location. The goal was to acquire images of maize during the vegetative, reproductive, and post-flowering growth stages from the fields at each location, capturing images at three different time points (Supplemental Data Set S1). In Scottsbluff, NE, images were acquired with DJI Matrice 600 Pro with DJI Zenmuse X3 and a 12 Mega Pixel (MP) RGB (red, green, blue) camera as an image acquisition sensor. Images were acquired at an altitude of 100 ft (30.48 m) with a front overlap of 90% and a side overlap of 65%. In North Platte, images were acquired using DJI Inspire 2 with a Sentra Double 4K AG+ RGB camera as an image sensor at an altitude of 50 ft (15.24 m) with front and side overlap of 70%. In Lincoln, images were acquired with a DJI Phantom 4 RTK with a DJI Zenmuse P1 camera, with a 45 MP RGB camera as an image acquisition sensor. Images were acquired at an altitude of 115 ft (35 m) with front and side overlap of 80%. In Missouri Valley, Ames, and Crawfordsville, IA, DJI Phantom 4 Pro V2.0 with DJI 20 MP RGB cameras was used as an image acquisition sensor, and images were acquired at an altitude of 100 ft (30.48 m) with front and side overlap of 80%. The UAV images were processed and stitched using Pix4D Mapper 4.8.4 (Pix4D 2024) and AgiSoft Metashape 1.8.4 (Agisoft Metashape 2024), photogrammetric software to create RGB orthomosaic images using default parameters during image processing. 

Satellite Image Acquisition
Pléiades Neo was used to capture images at all locations at six different time points (approximately two weeks apart), with the first three time points close to the dates of the three UAV image acquisitions at each location. Table 1 shows the specifications of this satellite constellation. The average widths of the six bands in satellite multispectral images are as follows: Red (620 – 690 nm), Green (530 – 590 nm), Blue (450 – 520 nm), Near-infrared (NIR, 770 – 880 nm), Red Edge (700 – 750 nm), and Deep Blue (400 – 450 nm). Along with multispectral images, a single-band panchromatic raster file with a wide width band of approximately 450-800 nm was generated. Each image captured a total area of 100 km × 100 km per location, covering the entire experimental field at each location simultaneously. Final 16-bit GeoTIFF satellite images with 30-cm resolution were generated and provided to us after panchromatic sharpening or pan-sharpening using panchromatic band image files, manual ortho-rectification, and atmospheric correction by Pleiades Neo. 

Assigning Plot-level Labels to Images
For precise segmentation of small plots from satellite images, a single UAV image captured at a single time point at each location was used as a reference for plot segmentation due to the low resolution of the satellite image.  With available UAV images, before the plot-level segmentation, satellite images were registered to the UAV image captured at the nearest date to the date of satellite image capture using ArcGIS Pro 10.8.2 through georeferencing. Registration was performed using the ground control points installed at either side of each field when available. able field features in satellite and UAV images were used for accurate registration as much as possible. After registration, the first-time point registered UAV and satellite image pair were used for plot segmentation for each location. All computation was performed using the ArcPy Jupyter Notebook environment implemented in ArcGIS Pro V3.2.0. Plot grids generated as described above were used to crop field images into plots. For each plot grid, a minimum bounding rectangle with four corners of rectangles surrounding/enclosing the plot grids was generated. When pixels inside the rectangle lacked values due to the geographical orientation of the fields, those regions were masked with zero pixel values for further data acquisition. Plot grids generated from the images at the first time point were used to segment the corresponding satellite images within the same location from all other time points.