A geospatially conditioned feature optimization approach for remote sensing-based maize yield and water productivity estimations across Kansas and Colorado

Abstract

Accurate and efficient assessments of crop production and agricultural water use are critical to the sustainability of national and global food systems, economic stability, and water resource management. Remote sensing technologies, particularly radar and optical remote sensing data from spaceborne satellite systems, have been successfully utilized for gathering data required to make these assessments. However, there exists a knowledge gap regarding the synergistic use of the data derived from these methods to account for the climatic and geospatial variations in agricultural regions. The objectives of this study were to 1) determine the optimal temporal acquisition windows, focal statistic neighborhood dimensions, and vegetation indices, for optical and radar remote sensing data’s use in estimating maize yield across climate zones in Kansas and Colorado, and 2) identify environmental and climatic characteristics of high and low water productivity areas using OpenET data integrated with geospatially optimized yield models and weather station data. Nine site-years resulting from five maize fields and four years spanning across the states of Kansas and Colorado in the U.S. were analyzed. Site-years represented varying climatic conditions and geospatial properties. Seasonal imagery from Sentinel-1 radar and Sentinel-2 optical satellite systems was acquired via Google Earth Engine. ArcGIS Pro was used to generate feature sets for five focal statistic neighborhood dimensions (1x1, 3x3, 5x5, 7x7, and 9x9) and ten vegetation indices and bands (RVI, DpRVI, VH/VV, VH, VV, NDVI, NDRE, SAVI, GNDVI, and EVI). The resulting features for each site were then statistically assessed and selected based on their correlation with the yield data, and subsequently incorporated into a random forest regression model. The final models were evaluated for their ability to accurately estimate yield, both intra- and inter-site-year. Results showed that the feature selection process favored diversity over uniformity in vegetation index and focal statistic neighborhood dimension optimizations. Additionally, intra-site-year random forest regression-based yield predictions performed better than the inter-site-year predictions for all but one site-year, highlighting the positive impact of optimized features for specific geospatial and climatic regions. These models were then implemented at the county scale and compared with published USDA-NASS county-average maize yield predictions. Additionally, the USDA-NASS reported county average yield estimates were used in conjunction with OpenET, a remote-sensing-based evapotranspiration dataset, to estimate county-scale water productivity for a total of 32 site-years across 17 counties in Kansas and Colorado. Statistical analysis of each site-year’s water productivity was conducted to determine its correlation with several climatic variables, including air temperature, relative humidity, precipitation, wind speed, soil temperature, solar radiation, elevation, and SSURGO soil data. The findings from this study indicated that the models developed in Chapter 1 generally did not perform well when scaled to the county level but were able to generate multiple county-scale yield predictions within 1 standard deviation of the USDA-NASS reported county average yield estimates. Low average daily wind speed was identified as the most impactful environmental metric for characterizing areas with high water productivity.