Accounting for Satellite Precipitation Uncertainty: The Development of a Probabilistic Landslide Hazard Nowcasting System

Abstract

Satellite multisensor precipitation products (SMPPs) provide valuable precipitation information for hydrological modeling, water resources management, natural disaster assessment, and other applications on a global scale and can be especially beneficial to data-limited regions. Uncertainty in SMPPs presents a challenge to fully utilizing these near real time sources of global precipitation. As the key forcing variable in hydrological processes, precipitation and any associated variability will impact model output. While various models have been developed to characterize SMPP uncertainty, methods of modeling SMPP uncertainty for applied use in models have been limited. This study explores the challenges of modeling SMPP uncertainty for applied use and develops a probabilistic landslide hazard nowcasting system that demonstrates one method of incorporating SMPP uncertainty in a decision tree framework. The improved performance of a probabilistic landslide hazard model which accounts for SMPP uncertainty confirms the value of converting SMPP data from an unknown source of uncertainty to a known source of uncertainty. Even without knowing the true antecedent rainfall, representing the known uncertainty of SMPP input within the model improves model performance. Chapter 2 will review the accuracy of SMPPs, discuss common metrics of SMPP error, and explore SMPP uncertainty models that have been developed over the past two decades. The limitations of current error models and needs for applied use in environmental models is discussed and methods of accounting for these limitations are proposed using a CSGD-based error model. Chapter 3 details the development of a probabilistic landslide hazard model that incorporates SMPP uncertainty using a CSGD-based error model. The advantages of probabilistically representing SMPP data as input to the model are demonstrated by increased true positive rate and more realistic variability in model output. A summary of findings and discussion of future work is included in Chapter 4.