Costa Filho, Edson, authorChávez, José L., advisorVenayagamoorthy, Karan, committee memberNiemann, Jeffrey, committee memberKummerow, Christian, committee member2024-09-092024-09-092024https://hdl.handle.net/10217/239213This study aimed to advance irrigation water management by developing and evaluating a procedure to improve the multispectral data from sub-optimal remote sensing sensors when using the optimal spectral resolution for a given remote sensing (RS) of crop actual evapotranspiration (ETa) algorithm. Data have been collected at three research sites in Colorado under different irrigation systems, soil textures, and vegetation types. The research site in Greeley (CO) has a five-year dataset (2017-2018 and 2020-2022). The fields in Fort Collins and Rocky Ford (CO) have data from 2020 and 2021. Three categories of ETa algorithms were evaluated in the study: The reflectance-based crop coefficient (RBCC) with three different models based on the normalized difference vegetation index (NDVI), soil-adjusted vegetation index (SAVI), and fractional vegetation canopy cover (fc), the one-source simplified surface energy balance (OSEB) based on a surface aerodynamic temperature approach, and the two-source surface energy balance algorithm (TSEB) using two different resistance approaches (parallel and series). All three ETa modeling categories use either just surface reflectance in the visible and invisible light spectrum (e.g., RED, BLUE, GREEN, Near-infrared) or a combination of multispectral and thermal data as inputs to predict crop ETa, alongside local micrometeorological data from nearby agricultural weather stations. A total of six RS of ETa algorithms were evaluated in this study. A total of five RS sensors were evaluated: three spaceborne sensors (e.g., Landsat-8, Sentinel-2, and Planet CubeSat), one proximal device (multispectral radiometer), and an uncrewed aerial vehicle (UAS). The spatial resolution of the RS sensors varied from 30 m to 0.03 m. The accuracy assessment of the crop ETa predictions considered a statistical performance analysis using, among several statistical metrics, the mean bias error (MBE) and root mean square error (RMSE), and compared estimated ETa values from all seven RS ETa algorithms with observed ETa values obtained from the Eddy Covariance Energy Balance System (Greeley and Fort Collins sites) and a weighing lysimeter (Rocky Ford). The study was divided into three stages: a) the evaluation of different remote sensing (RS) pixel spatial resolutions (scales) as inputs on the estimation of different types of data needed for estimating ETa in hourly and daily time frames; b) the development of a calibration protocol and standards for the use of different imagery spatial resolutions (scales) in RS of ETa algorithms. The calibration approach involved a novel two-source pixel decomposition approach for partitioning surface reflectance into soil and vegetation using a non-linear, physically based spectral model, machine-learning regression, and a novel spatial light extinction model (kp); c) the accuracy evaluation of resulting ETa rates from calibrated/standardized data (for each selected RS of ETa algorithms). Results of stage one of the study indicated that depending on the RS of ETa and RS sensor data (spatial and spectral resolutions), the accuracy (MBE ± RMSE) of estimated ETa predictions varied. For the NDVI and fc RBCC ETa algorithms, Sentinel-2 provided the best RS data for predicting daily maize ETa. Errors were 0.21 (5%) ± 0.78 (18%) mm/d and 0.59 (14%) ± 1.07 (25%) mm/d, respectively. For the OSEB algorithm, Planet CubeSat gave the best RS data since it provided the smallest error for hourly maize ETa, -0.02 (-3%) ± 0.07 (13%) mm/h. For the SAVI RBCC model, the MSR data provided the best results since the maize ETa error was -0.13 (-3%) ± 0.67 (16%) mm/d. For the TSEB in series and parallel, the errors when estimating hourly maize ETa were -0.02 (-3%) ± 0.07 (11%) mm/h and -0.02 (-4%) ± 0.09 (14%) mm/h, respectively when using MSR data. For stage two of the study, the best machine learning regression model for a given RS sensor data and RS of the ETa algorithm depended on the surface reflectance composite (plant or bare soil values). The best machine-learning models for adjusting RS data were the regression tree and the Gaussian Process Regression. Regarding the pixel decomposition approach based on the novel spatial light extinction coefficient model, the novel approach provided reliable predictions of kp using the different RS sensor data. The error in predicting kp was -0.01 (-2%) ± 0.05 (10%) when combining all RS sensor data for the two-year data set at LIRF (years 2018 and 2022). For stage three of the study, results showed improvements in the accuracy of crop ETa estimation after adjusting the RS data using the proposed calibration protocol. At the Greeley site, regarding the RBCC RS of ETa algorithm, adjusted data from Planet CubeSat had better performance when estimating daily crop ETa since the error was reduced from 21% to 16% for the fc-input model. For the SAVI-input model, the RS data that performed better was the UAS. Errors were reduced from -0.42 (-11%) ± 0.76 (20%) mm/d to -0.21 (-5%) ± 0.41 (11%) mm/d. For the NDVI-input model, the adjusted UAS data performed better when estimating daily maize ETa. The improved accuracy was 0.32 (8%) ± 0.40 (10%) mm/d. At the Rocky Ford site, for the fc-input model, adjusted RS optical data from the MSR performed better. Daily maize ETa error was reduced from 17% to 15%. For the SAVI-input model, the RS data that performed better was the Landsat-8, with errors being reduced from -1.84 (-28%) ± 2.61 (39%) mm/d to -1.14 (-17%) ± 1.79 (27%) mm/d. The NDVI-based RBCC model had better performance when using adjusted MSR data daily maize ETa. Regarding the OSEB RS of crop ETa approach, at the Greeley site, the OSEB-adjusted data from UAS performed better. Hourly maize ETa error was reduced from 0.11 (19%) mm/h to 0.07 (13%) mm/h for the OSEB algorithm. For the TSEB parallel algorithm, the RS data that had better performance was the Landsat-8/9 since the error was reduced from 0.19 (34%) mm/h to 0.11 (20%) mm/h. For the TSEB series algorithm, the adjusted UAS data performed better. Daily maize ETa errors decreased from 0.10 (18%) mm/h to 0.05 (9%) mm/h. In summary, this study provided an RS calibration approach to support irrigation water management through the development and evaluation of a method for enhancing optical multispectral data sourced from various RS sensors. This study also highlighted the efficacy of machine learning models, like regression tree and Gaussian Process Regression, in adjusting RS data based on surface reflectance composites. Furthermore, a novel pixel decomposition approach utilizing a spatial light extinction model effectively predicted the light extinction coefficient. Overall, this research showcases the potential of RS data adjustments in improving the accuracy of ETa estimates, which is crucial for optimizing irrigation practices in agricultural settings.born digitaldoctoral dissertationsengCopyright and other restrictions may apply. User is responsible for compliance with all applicable laws. For information about copyright law, please see https://libguides.colostate.edu/copyright.evapotranspirationirrigation water managementremote sensinghydrologyenvironmental biophysicsmachine-learningOptimizing remote sensing data for actual crop evapotranspiration mapping at different resolutionsText