Spatial blind source separation for soil moisture data

Abstract

Spatial Blind Source Separation (SBSS) is a recent extension of Independent Component Analysis (ICA) for spatial data. Standard ICA ignores the spatial dependency structure of spatial data, while SBSS uses this information. The goal of this thesis is to evaluate the SBSS method in a new field of application. In cooperation with the Institute of Geodesy and Geoinformation of the Vienna University of Technology, Australian soil moisture data are investigated, consisting of gridded satellite observations from 1998 to 2018. Soil moisture is an essential factor in understanding climate processes and therefore weather extremes and climate change. Understanding space-time patterns of soil moisture facilitates insights in the fields of hydrology, agriculture, and socioeconomics. Principal Component Analysis(PCA), ICA and SBSS are all applied to the data and their results are contrasted with each other and the existing literature on Australian soil moisture data in the context of Blind Source Separation (BSS). Correlations between loadings of the results of PCA,ICA and SBSS and the most relevant climate modes for Australia are investigated via Spearman correlations for concurrent and time-lagged observations. One finding of this work is that the results of SBSS are consistent with existing studies, while ICA, when looking at anomalies, fails to provide new insights or even reproduce known results. Spatiotemporal dependencies of the observations are explicitly taken into account in the novel SBSS approach, while they are ignored in the context of standard ICA. The thesis aims to identify the advantages of SBSS over conventional PCA and ICA in the context of the presented analysis. Notable, higher correlations to the Climate Oscillation Indices(COIs) are obtained for SBSS, and new patterns of SBSS components complement existing knowledge. SBSS is a useful candidate for BSS of climate processes.