Flash Drought Monitoring and Management: Applications of Remote Sensing and Machine Learning Techniques

Flash droughts, marked by their rapid onset and severe impacts, pose significant challenges to hydrological and ecological systems. Despite growing recognition of their effects, the relationship between flash drought and streamflow, land surface recovery post-flash drought, and prediction of streamflow flash droughts remain poorly understood or unexplored. This dissertation addresses three critical knowledge gaps: the hydrologic response to flash droughts, the resilience of different land use and land cover (LULC) types to drought-induced stress, and the potential for predictive modeling of streamflow flash droughts (SFDs) using deep learning. In the first study, over 1,000 flash drought events were identified across 258 catchments in the Mississippi River Basin from 1980 to 2022 using the Standardized Antecedent Precipitation Evapotranspiration Index (SAPEI). Results showed that catchments with limited upstream regulation experienced sharp streamflow declines during flash drought onset, with Granger causality tests indicating statistically significant lagged responses (p < 0.05) in over 70% of catchments. The strength and timing of the response varied by region, with the central and southeastern MRB exhibiting the most consistent declines. The second study examined ecosystem recovery using MODIS-derived Gross Primary Productivity (GPP) across the ten most severe flash droughts in the MRB. Recovery duration varied substantially by LULC type, with forests taking an average of 60–90 days to return to pre-drought GPP levels, compared to 30–45 days for croplands. Spatial analysis revealed that droughts in 2006, 2011, and 2022 led to prolonged recovery periods in the southwestern MRB, highlighting the influence of land management, soil characteristics, and repeated drought exposure. In the third study, streamflow flash droughts were predicted using three deep learning architectures, baseline LSTM, LSTM with static catchment attributes, and the Temporal Fusion Transformer (TFT), across 671 CAMELS catchments. The TFT model achieved the highest predictive performance, with a mean KGE of 0.88, outperforming the LSTM integrated with static attributes (KGE = 0.81) and baseline LSTM (KGE = 0.78). Incorporating both dynamic meteorological variables and static physiographic features improved prediction skill, particularly in regions with complex terrain and land use heterogeneity. Together, these studies demonstrate that flash droughts propagate through interconnected hydrological and ecological systems and that their impacts and recovery trajectories are shaped by both natural and anthropogenic factors. By integrating remote sensing data, standardized drought indices, and deep learning models, this research provides a novel, multi-dimensional framework for assessing impacts of and predicting flash drought events across diverse landscapes. These findings offer actionable insights for improving drought preparedness, enhancing ecosystem resilience assessments, and advancing early warning systems.