Scale Effects on the Remote Estimation of Evapotranspiration
MetadataShow full metadata
Consideration of spatio-temporal scale in geographic research is often mentioned, but scale effects on geographic modeling are poorly understood. The goal of this research is to explore and test the critical role of geographic relationships as they are represented across multiple spatial scales of analysis and to develop a general methodology for studying scaling relationships. Specifically, this dissertation uses remotely sensed data to investigate scale dependencies of mass and energy fluxes in the hydrologic cycle by developing a methodology to analyze the scaling relationships of evapotranspiration (ET). This research seeks to answer three categories of questions: (1) What effects do different aggregation techniques have on ET estimation as resolution decreases? How does landscape heterogeneity impact the aggregation process? (2) As major inputs to ET models, do the Normalized Difference Vegetation Index (NDVI) and surface temperature have consistent impacts on estimated ET across scales? Which input has a stronger correlation with ET estimation and does that correlation remain across scales of analysis? (3) How does the spatial autocorrelation of ET vary with scale? Does a relationship exist between the variance of spatial autocorrelation and spatial resolution?
Three analyses were performed to answer these research questions, including a comparison of aggregation techniques, an assessment of the primary indicators of ET, and the spatial autocorrelation of ET estimates. First, two aggregation techniques are used to aggregate ET from 30 meters to coarser resolutions. Findings indicate that the two aggregation techniques used produce statistically different ET estimates at resolutions finer than 960m. A pixel-by-pixel comparison of paired ET estimates at each scale reveals that landscape heterogeneity has an important impact on ET estimation accuracy. Most large pixel-by-pixel differences occur in non-vegetated areas and their boundaries with other land cover classes. Second, a correlation analysis was conducted on ET and its primary indicators across scales. Findings indicate that surface temperature is more correlated to ET than NDVI at resolutions finer than 240m. At resolution coarser than 240m, the correlations of ET with NDVI, surface temperature become inconsistent. When scale changes, the landscape heterogeneity alters the influence of ET drivers. Third, a geostatistical analysis was performed to study the spatial autocorrelation of ET across scales. Range and partial sill were used to study the spatial autocorrelation of ET. Findings indicate that spatial resolution determines the spatial autocorrelation we observe. When the resolution is finer than a threshold, ranges are identifiable. Finer resolution can better present the variance of landscape than coarser resolution. However, the finest resolution may not be the best resolution to reflect the spatial variability. The resolution with highest partial sill should be used to best represent the landscape’s spatial variability.
This research provides a way to study the scaling relationship impacted by the heterogeneity of landscape from three aspects, including the aggregating techniques, correlation of phenomenon and its major indicators, and spatial autocorrelation. As a whole, findings in this study improve our understanding of ET modeling and estimation using remotely sensed data. More importantly, this research elucidates general scale considerations that must be assessed prior to geographic modeling because they influence how we represent cause-effect relationships and the modeling of spatial patterns and processes.