Natural Hazards Monitoring Using Multiple Remote Sensing Techniques and Machine Learning Algorithms

Natural hazards often fit into two categories according to the speed and extent of a hazard: rapid onset hazards and slow onset hazards. Both could pose considerable risks to society development, community security, and ecosystems, especially in urban areas with dense populations. Satellite-based remote sensing techniques provide essential information for preventing and mitigating natural hazards. This dissertation has demonstrated both long-term and short-term monitoring methods using multiple remote sensing techniques together with advanced machine learning algorithms. The 2021 Texas winter storm is taken as an example to exemplify the methods of monitoring rapid onset hazards (disasters), and the subsidence in Tianjin, China is used to exemplify the methods of monitoring slow onset natural hazards. In February 2021, an unprecedented winter storm dumped the snow record in Texas. It claimed hundreds of lives and evolved into a major disaster nationalwide. This study uses the differential coherence from Sentinel-1 Synthetic Aperture Radar imagery to characterize the surface disturbance due to this winter storm. Furthermore, machine learning algorithms are applied to quantify Texas statewide snow depth using surface disturbance map, SAR amplitude, precipitation, temperature, surface topography, land cover, and population. The result shows that approximately 90% of Texas accumulated over 30-mm snow depth. A consequent model accuracy of 99% demonstrates that our approach can provide an independent snow depth estimation. Tianjin is one of the large urban regions in China that has suffered from severe land subsidence induced by excessive groundwater withdrawal for approximately half a century. However, since the launch of the South-to-North Water Diversion (SNWD) project, the groundwater withdrawal has significantly been released, thus greatly easing the land subsidence worries here since about 2019. This study quantitively delineated the impacts of the SNWD on land subsidence in Tianjin using Sentinel-1A/B Interferometric Synthetic Aperture Radar (InSAR) (2014-2021), GPS (2010-2021), and groundwater data. The results show that as of 2021, the subsidence area (> 5 mm/year) has reduced by 15% of the subsiding area before SNWD; the areas of rapid subsidence (> 3 cm/year) and extremely rapid subsidence (> 5 cm/year) have reduced to about 70% and 60% of the areas before SNWD, respectively. Besides, the Principal Component Analysis (PCA), an unsupervised machine learning method, is employed to highlight primary factors controlling the recent land subsidence from the large dimension InSAR displacement results. Integrating frontier technologies, i.e., remote sensing and machine-learning algorithms, into existing natural hazards monitoring systems will significantly improve our ability to understand, communicate, and forecast natural hazards and assess risks, thus improving hazard mitigation and disaster management. Two case studies in this study demonstrate the effectiveness of the SAR and machine-learning integrated methods in both slow onset and rapid onset natural hazards monitoring.