Files
Abstract
Remotely sensed satellite data provide essential monitoring of Earth’s ecosystems continuously through time to supplement field level studies. The implementation of uncrewed aerial system (UAS) imagery collection to study agricultural trends in ecological research studies has fueled interest in techniques to bridge spatial scales between remotely sensed images to drive more effective land management. How to combine spatially and temporally incompatible data remains an ongoing effort. This dissertation studies the connections between multispectral imagery at different scales by gauging the effectiveness of two statistical downscaling approaches motivated by climate downscaling: regression kriging (RK) and artificial augmentation (AA). While these statistical downscaling approaches typically make predictions at a discrete number of ground surface points, we propose adaptations of the methods to construct a grid of predictions by treating each UAS pixel as a spatial site. RK makes use of variograms as part of the spatial prediction process, so we employ Spherical and Matérn variograms when constructing our downscaling models. We form two novel sets of statistical downscaling predictions by combining the RK and AA procedures (RKAA and Joint). We implement these downscaling techniques to translate Normalized Difference Vegetation Index (NDVI) values from the planetary to the field scale. We also incorporate field level measurements of above ground biomass of cotton (Gossypium hirsutum L.) crops collected from the United States Department of Agriculture (USDA) study site in Ashburn, GA. The above ground biomass data is collected at six sampling plots across the study site. We construct linear regression models to predict above ground biomass as a function of each downscaled predictor in order to extrapolate the above ground biomass data from the six sampling plots to the entire study site. We assess the downscaled predictions by comparing them to ground truth predictions created from UAS data. The AA predictions show the highest accuracy, followed by the RKAA and Joint predictions. The RK predictions are significantly hindered by the concentrated sampling of the field level data, but have the capacity for improvement with additional modifications. Simulation study results show similar performance among the downscaled predictors.