P. et al. (2025) Integration of Soil Moisture and Meteorological Data Using Deep Learning for Flash Drought Detection in Northeastern Brazil

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-11-18
• Authors: Isela L. Vásquez P., Marcelo Zeri, Arturo Sánchez, Adriano Almeida, David Pareja-Quispe, Juan Gregorio Rejas Ayuga, Alan J. P. Calheiros
• DOI: 10.1007/s41748-025-00929-z

Research Groups

• Centro Nacional de Monitoramento e Alertas de Desastres Naturais (CEMADEN), São José dos Campos, São Paulo, Brazil
• University of Padova (UNIPD), Padova, Italy
• National Institute for Space Research (INPE), São José dos Campos, São Paulo, Brazil
• Universidad Nacional Mayor de San Marcos (UNMSM), Lima, Peru
• Department of Space Programs, National Institute for Aerospace Technology (INTA), Madrid, Spain
• Technical University of Madrid (UPM), Madrid, Spain

Short Summary

This study developed and validated a deep learning U-Net model to integrate meteorological and satellite-derived soil moisture data for detecting flash drought events in Northeastern Brazil (NEB) from 2015–2023. The model accurately reproduced flash drought frequency and duration, demonstrating its potential for high-resolution monitoring and improving early-warning systems in data-scarce regions.

Objective

• To develop and validate a deep learning U-Net model for integrating meteorological and satellite-derived soil moisture data to detect flash drought events in Northeastern Brazil (NEB).
• To characterize the frequency, duration, and intensity of flash droughts in NEB and understand their modulation by large-scale climate variability.

Study Configuration

• Spatial Scale: Northeastern Brazil (NEB), with data aggregated to a 9 kilometer spatial resolution (matching SMAP). Input meteorological data originally at 0.1 degree x 0.1 degree resolution, standardized to 256 x 256 pixels for the U-Net model.
• Temporal Scale: Data period from June 2015 to December 2023. Training period from June to December, 2015–2019. Validation period from 2020–2023. Daily data, aggregated into pentads (5-day averages) for flash drought detection.

Methodology and Data

• Models used: U-Net (Deep Learning Convolutional Neural Network), Locally Weighted Scatterplot Smoothing (LOWESS) regression.
• Data sources:
  • Meteorological Data:
    • Daily precipitation: Daily Quadrilateral Weather Data of Brazil (BR-DWGD) (0.1 degree x 0.1 degree resolution).
    • Air temperature at 2 meters, wind speed at 10 meters, relative humidity at 2 meters: ERA5-Land reanalysis (0.1 degree x 0.1 degree resolution).
  • Soil Moisture Data:
    • Level-4 Soil Moisture Active and Passive (SMAP) product (9 kilometer spatial resolution, 3-hour temporal resolution). Used as the target variable (ground truth).

Main Results

• The U-Net model demonstrated strong agreement with SMAP Level-4 observations, achieving spatial correlations exceeding 0.6, Root Mean Square Deviation (RMSD) values below 0.04 cubic meters per cubic meter, and Nash–Sutcliffe Efficiency (NSE) values greater than 0.5 across most of the domain, particularly in the northern NEB.
• Flash drought events were identified by a rapid soil moisture decline 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, which were strongly influenced by ENSO and Atlantic Sea Surface Temperature anomalies.
• The semiarid interior (Sertão) exhibited the highest flash drought frequency, averaging 4–6 events per year.
• U-Net detected flash drought onset approximately 10 days earlier and showed broader spatial coverage for moderate events compared to SMAP, but systematically underestimated the duration and intensity of short-term events by 20–30% due to its inherent spatial smoothing.
• A non-linear inverse relationship was identified between drought duration and the soil moisture decline slope, indicating that shorter events tend to be more intense.
• Flash drought dynamics were significantly modulated by large-scale climate variability (ENSO, ITCZ, SACZ) and local edaphic properties, with sandy soils showing rapid desiccation.

Contributions

• Developed a robust deep learning U-Net framework for high-resolution (9 kilometer), daily, continuous soil moisture estimation and flash drought detection in data-scarce regions like NEB, offering a significant improvement over traditional regression approaches.
• Demonstrated enhanced predictive skill by anticipating flash drought onset by approximately 10 days, thereby substantially improving early warning capabilities for these rapid-onset events.
• Quantified the modulation of flash drought frequency and intensity by large-scale climatic variability (ENSO, ITCZ, SACZ) and confirmed a non-linear inverse relationship between drought duration and intensity.
• Provided operationally relevant, high-resolution drought diagnostics that directly support agricultural planning, climate-resilient water management, and the strengthening of early warning systems, with a flexible architecture transferable to other semi-arid and tropical regions.

Funding

• CNPq grant 444205/2024-1 (Alan J. P. Calheiros)
• European Union – Next Generation EU through the PRIN (Progetti di ricerca di Rilevante Interesse Nazionale) programme (grant 2022ZC2522) for the “raINfall exTremEs and their impacts: from the local to the National ScalE” (INTENSE) project (C. Arturo Sánchez P.)

Citation

@article{P2025Integration,
  author = {P., Isela L. Vásquez and Zeri, Marcelo and Sánchez, Arturo and Almeida, Adriano and Pareja-Quispe, David and Ayuga, Juan Gregorio Rejas and Calheiros, Alan J. P.},
  title = {Integration of Soil Moisture and Meteorological Data Using Deep Learning for Flash Drought Detection in Northeastern Brazil},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-00929-z},
  url = {https://doi.org/10.1007/s41748-025-00929-z}
}