<p>Flash droughts are characterized by a rapid decline in soil moisture and represent one of the most critical emerging hydroclimatic hazards in semiarid regions. This study integrates meteorological and satellite-derived soil moisture data using a deep learning U-Net model to detect flash drought events in Northeastern Brazil (NEB) during 2015–2023. The U-Net model achieved strong agreement with SMAP Level-4 observations, with spatial correlations exceeding 0.6, RMSD values below 0.04 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(m^{3}/m^{3}\)</EquationSource> </InlineEquation>, and Nash–Sutcliffe Efficiency (NSE) values <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(>0.5\)</EquationSource> </InlineEquation> across most of the domain. Flash drought events were identified based on a rapid soil moisture drop from the 40th to the 20th percentile within four pentads (20 days). The model accurately reproduced the observed spatial and temporal variability of flash drought frequency and duration, particularly under the influence of ENSO and Atlantic SST anomalies. The highest event frequency occurred in the semiarid interior (Sertão), with 4–6 events per year on average. These results demonstrate the potential of deep learning for high-resolution flash drought monitoring and contribute to improving drought early-warning systems and climate-adaptation strategies in data-scarce regions such as NEB.</p> Graphical Abstract <p>Based on the graphical snapshot, this study presents an integrated approach to detect flash droughts in Northeastern Brazil (NEB) through the combination of meteorological variables and SMAP soil moisture data using a deep learning model. The study area, characterized by high climatic variability and susceptibility to rapid-onset droughts, was analyzed using U-Net architecture a convolutional neural network model trained to estimate daily soil moisture fields from precipitation, temperature, evapotranspiration, wind speed, and relative humidity data. The temporal evolution of soil moisture percentiles was analyzed to identify flash drought events defined by abrupt declines from the 40th to the 20th percentile within four pentads. The U-Net model’s performance was evaluated against SMAP observations, capturing spatial and temporal drought patterns with notable accuracy. Despite a tendency to underestimate drought duration and intensity in heterogeneous regions, the model successfully identified critical drought periods and spatial extents. A non-linear relationship between drought intensity and duration was revealed using LOWESS regression, emphasizing the prevalence and severity of short, intense events. These results contribute to the improvement of early warning systems and adaptive strategies for water and agricultural management in drought-prone areas.</p>

Integration of Soil Moisture and Meteorological Data Using Deep Learning for Flash Drought Detection in Northeastern Brazil

期刊论文
发表日期：

2025年11月18日

查看全文

Isela L. Vásquez P.,
Marcelo Zeri,
C. Arturo Sánchez P.,
Adriano P. Almeida,
David Pareja-Quispe,
Juan G. Rejas Ayuga,
Alan J. P. Calheiros

摘要

Flash droughts are characterized by a rapid decline in soil moisture and represent one of the most critical emerging hydroclimatic hazards in semiarid regions. This study integrates meteorological and satellite-derived soil moisture data using a deep learning U-Net model to detect flash drought events in Northeastern Brazil (NEB) during 2015–2023. The U-Net model achieved strong agreement with SMAP Level-4 observations, with spatial correlations exceeding 0.6, RMSD values below 0.04 \(m^{3}/m^{3}\) , and Nash–Sutcliffe Efficiency (NSE) values \(>0.5\) across most of the domain. Flash drought events were identified based on a rapid soil moisture drop from the 40th to the 20th percentile within four pentads (20 days). The model accurately reproduced the observed spatial and temporal variability of flash drought frequency and duration, particularly under the influence of ENSO and Atlantic SST anomalies. The highest event frequency occurred in the semiarid interior (Sertão), with 4–6 events per year on average. These results demonstrate the potential of deep learning for high-resolution flash drought monitoring and contribute to improving drought early-warning systems and climate-adaptation strategies in data-scarce regions such as NEB.

Graphical Abstract

Based on the graphical snapshot, this study presents an integrated approach to detect flash droughts in Northeastern Brazil (NEB) through the combination of meteorological variables and SMAP soil moisture data using a deep learning model. The study area, characterized by high climatic variability and susceptibility to rapid-onset droughts, was analyzed using U-Net architecture a convolutional neural network model trained to estimate daily soil moisture fields from precipitation, temperature, evapotranspiration, wind speed, and relative humidity data. The temporal evolution of soil moisture percentiles was analyzed to identify flash drought events defined by abrupt declines from the 40th to the 20th percentile within four pentads. The U-Net model’s performance was evaluated against SMAP observations, capturing spatial and temporal drought patterns with notable accuracy. Despite a tendency to underestimate drought duration and intensity in heterogeneous regions, the model successfully identified critical drought periods and spatial extents. A non-linear relationship between drought intensity and duration was revealed using LOWESS regression, emphasizing the prevalence and severity of short, intense events. These results contribute to the improvement of early warning systems and adaptive strategies for water and agricultural management in drought-prone areas.