Doctor of Philosophy (PhD)
Geography and Anthropology
Soil moisture is an important factor for accurate prediction of agricultural productivity and rainfall runoff with hydrological models. Remote sensing satellites such as Soil Moisture Active Passive (SMAP) offer synoptic views of soil moisture distribution at a regional-to-global scale. To use the soil moisture product from these satellites, however, requires a downscaling of the data from an usually large instantaneous field of view (i.e. 36 km) to the watershed analysis scales ranging from 30 m to 1 km. In addition, validation of the soil moisture products using the ground station observations without an upscaling treatment would lead to cross-level fallacy. In the literature of geographical analysis, scale is one of the top research concens because of the needs for multi-source geospatial data fusion. This dissertation research introduced a multi-level soil moisture data assimilation and processing methodology framework based on spatial information theories. The research contains three sections: downscaling using machine learning and geographically weighted regression, upscaling ground network observation to calibrate satellite data, and spatial and temporal multi-scale data assimilation using spatio-temporal interpolation.
(1) Soil moisture downscaling
In the first section, a downscaling method is designed using 1-km geospatial data to obtain subpixel soil moisture from the 9-km soil moisture product of the SMAP satellite. The geospatial data includes normalized difference vegetation index (NDVI), land surface temperature (LST), gross primary productivity (GPP), topographical moisture index (TMI), with all resampled to 1-km resolution. The machine learning algorithm – random forest was used to create a prediction model of the soil moisture at a 1-km resolution. The 1-km soil moisture product was compared with the ground samples from the West Texas Mesonet (WTM) station data. The residual was then interpolated to compensate the unpredicted variability of the model. The entire process was based on the concept of regression kriging- where the regression was done by the random forest model. Results show that the downscaling approach was able to achieve better accuracy than the current statistical downscaling methods.
(2) Station network data upscaling
The Texas Soil Observation Network (TxSON) network was designed to test the feasibility of upscaling the in-situ data to match the scale of the SMAP data. I advanced the upscaling method by using the Voronoi polygons and block kriging with a Gaussian kernel aggregation. The upscaling algorithm was calibrated using different spatial aggregation parameters, such as the fishnet cell size and Gaussian kernel standard deviation. The use of the kriging can significantly reduce the spatial autocorrelation among the TxSON stations because of its declustering ability. The result proved the new upscaling method was better than the traditional ones.
(3) Multi-scale data fusion in a spatio-temporal framework
None of the current works for soil moisture statistical downscaling honors time and space equally. It is important, however, that the soil moisture products are consistent in both domains. In this section, the space-time kriging model for soil moisture downscaling and upscaling computation framework designed in the last two sections is implemented to create a spatio-temporal integrated solution to soil moisture multi-scale mapping.
The present work has its novelty in using spatial statistics to reconcile the scale difference from satellite data and ground observations, and therefore proposes new theories and solutions for dealing with the modifiable areal unit problem (MAUP) incurred in soil moisture mapping from satellite and ground stations.
Xu, Yaping, "Mapping Soil Moisture from Remotely Sensed and In-situ Data with Statistical Methods" (2019). LSU Doctoral Dissertations. 4961.