A hydrometeorological data-driven study for assessment and subseasonal prediction of flash drought

Abstract

Flash droughts (FD) are high-impacting climate extremes which create environments with inadequate water availability, excessive temperatures, high pollution, and low economic or ecological productivity. The short- and long-term economic impacts of FD can exceed forty billion dollars and have severe impacts on economic, built, social, and natural systems. This dissertation takes a stepwise approach towards a more comprehensive understanding of FD dynamics, trends, and predictability through data-driven studies which incorporate both a diverse set of observational reanalyses and subseasonal numerical weather prediction ensemble forecasts (SEF). Analysis is conducted utilizing statistical, ML, and DL methods to improve understanding FD variability and predictability and increase climate resilience against dry extremes. The manuscript is divided into five main chapters. In Chapter 1, I provide a general introduction of the research. In Chapter 2, soil moisture, evaporative demand, land surface variables, and two FD indices are comprehensively assessed for trends and variability using hierarchical clustering and signal processing methods. Results indicate high lead-lag associations between EDDI and SMPD in specific geographical regions indicating advanced early-warning of SMPD using EDDI by up to three weeks. Additionally, signal processing and conditional probability approaches identified significant relationships between FD severity and ENSO phases which can improve long-term water management strategies. In Chapter 3, I assessed SEF skill for both evaporative demand and SMPD with results indicating that high skill can be achieved for week 4 predictions, but caution must be employed when comparing dataset from different sources. Case study analysis of three major FD events identified significant limitations of current SEF ensemble forecasts for week 3 prediction of FD, precipitation, temperature, evaporative demand, and soil moisture which currently hinder accurate predictions. In Chapter 4, I explore the extent to which DL convolutional neural networks can improve subseasonal predictability of soil moisture. Through a series of experimental runs, the optimal set of predictors was identified which significantly improves soil moisture forecast skill relative to state-of-the-art models for lead weeks 1-5. Finally, Chapter 5 provides concluding remarks and recommendations for building upon previous research. Together, findings underscore the complexity of FD variability, onset, and predictability within the subseasonal timescale which can be disentangled through data-driven hydroclimate studies. Given the important role of soil moisture on long-lead forecasts, it can also be leveraged in the future to improve subseasonal predictions of other climate variables and phenomenon, such as extreme precipitation and heat waves. The results represent a scientific step forward in improving holistic understanding of flash drought index variability, subseasonal forecast skill, and deep learning methods for enhanced week 1-5 predictability.