Soares et al. (2026) River intermittency: mapping and upscaling of water occurrence using unmanned aerial vehicle, Random Forest and remote sensing landscape attributes

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-02-13
• Authors: Nazaré Suziane Soares, Carlos Alexandre Gomes Costa, Till Francke, C Mohr, Wolfgang Schwanghart, Pedro Henrique Augusto Medeiros
• DOI: 10.5194/hess-30-849-2026

Research Groups

• Federal University of Ceará, Campus do Pici, Fortaleza, Brazil
• University of Potsdam, Institute of Environmental Science and Geography, Potsdam, Germany
• Federal Institute for Geoscience and Natural Resources, Hannover, Germany
• Federal Institute of Education, Science and Technology, Campus Fortaleza, Fortaleza, Brazil

Short Summary

This study maps and models the spatio-temporal dynamics of an intermittent river using UAV surveys to classify water occurrence and Random Forest models with landscape attributes, dam presence, and satellite indices. Model (a), incorporating Sentinel MNDWI, proved most successful in simulating intermittency both temporally and spatially with approximately 80% accuracy.

Objective

• To map and model the spatio-temporal dynamics of an intermittent river in the Brazilian semi-arid region.
• To identify the most important environmental and anthropogenic variables affecting river intermittency patterns.

Study Configuration

• Spatial Scale: The study focused on the Umbuzeiro River (approximately 105 km long) within the Benguê catchment (approximately 1000 km²) in the Brazilian Semiarid. River reaches were classified at 1.0 m resolution, and modeling units were circular areas with a 100 m diameter, spaced every 50 m.
• Temporal Scale: Twelve UAV surveys were conducted between March and November 2022 (monthly during the rainy season and once in November). Hydrological data were collected daily for 2022. Satellite imagery varied in temporal resolution (Sentinel-2: 5 days; Planetscope: daily).

Methodology and Data

• Models used: Random Forest classification models were implemented using the R package "randomForest" with recursive feature elimination (R package "caret"). Three model variants were tested based on dynamic predictors:
  • (a) Model using Sentinel-2 spectral indices (e.g., MNDWI) and Dynamic World LULC data.
  • (b) Model using Planetscope spectral indices (e.g., NDVI).
  • (c) Model using hydrological data (antecedent precipitation index).
• Data sources:
  • UAV imagery: Phantom 4 pro (RGB) and eBee SQ (multispectral) systems for high-resolution orthomosaics and digital terrain models (3.5-4.5 cm ground resolution). Used for visual classification of 1.0 m river reaches into "Wet", "Transition", "Dry", or "Not Determined".
  • Satellite imagery:
    • Sentinel-2 MSI: 10 m (blue, green, red, NIR) and 20 m (SWIR, red-edge) resolution for spectral indices (NDWI, MNDWI, SWI, NDMI).
    • Planetscope: 3.0-4.1 m resolution for spectral indices (NDWI, MNDWI, SWI, NDVI, NDMI).
    • Dynamic World (Sentinel-based): 10 m LULC product.
    • MapBiomas (Landsat-based): 30 m annual LULC maps.
    • Global Water Surface (Landsat-based): Historical water occurrence data.
  • Topographic data: Shuttle Radar Topography Mission (SRTM) DEM (30 m resolution) for mean altitude, maximum drainage area, mean hill slope, and number of stream cells.
  • Hydrological data: Daily precipitation from Aiuaba Experimental Basin, antecedent precipitation index (30 d), and daily water level (m) and storage volume (hm³) of the Benguê Reservoir.
  • Anthropogenic data: Manually mapped damming structures (45 identified) along the Umbuzeiro River to calculate distance from/to dams.

Main Results

• UAV surveys revealed distinct spatio-temporal patterns: "Wet" conditions were more prevalent downstream and during the rainy season, while "Dry" conditions dominated the dry season. "Transition" and "Not Determined" classes were associated with vegetation and narrower reaches.
• Recursive feature elimination showed that the five most important predictors were sufficient for accurate modeling, achieving approximately 80% overall accuracy.
• The most important static predictors consistently selected across all models were mean elevation, drainage area, distance from the last dam, and distance to the next dam. "Distance from the last dam" was identified as the single most important predictor.
• Dynamic predictors varied by model: Sentinel MNDWI (model a), Planetscope NDVI (model b), and Antecedent Precipitation Index (30 d) (model c).
• All models achieved an overall accuracy of approximately 80% for both training and testing datasets, outperforming a benchmark classifier (57%). "Wet" and "Dry" classes were predicted with higher balanced accuracy than "Transition" and "Not Determined" classes.
• Model (a) and (b) successfully captured the temporal dynamics of river drying and rewetting during extrapolation to the entire river. Model (c) showed limited ability to predict temporal dynamics.
• Spatially, models (a) and (c) better identified areas prone to persistent wet conditions, including the Benguê reservoir, outperforming model (b).
• Model (a), utilizing Sentinel MNDWI, was identified as the most successful in simulating intermittency both temporally and spatially due to its ability to aggregate sufficient spatial information.

Contributions

• Developed a scalable and adaptable modeling framework integrating high-resolution UAV data, remote sensing landscape attributes, and machine learning to map and model spatio-temporal water occurrence in narrow, non-perennial rivers.
• Provided a fine-scale understanding of how anthropogenic flow regulation (dam presence and distance) significantly influences surface water distribution, even for small impoundments.
• Demonstrated the effectiveness of Sentinel MNDWI for mapping water occurrence in narrow non-perennial rivers, even in semi-arid, vegetated environments.
• Addressed a critical knowledge gap on spatial patterns of intermittency by integrating topographic, climatic, and human modification drivers.
• Generated high-resolution ground-truth data (1.0 m reaches) from UAV imagery for intermittent river classification, which is crucial for unmonitored areas.

Funding

• Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), Process 88881.371462/2019-01 (PROBRAL program).
• European Union's Horizon 2020 research and innovation programme, grant agreement no. 869226 (DRYvER project).
• Fundação Cearense de Apoio ao Desenvolvimento Científico e Tecnológico (FUNCAP), project 21300411010023.
• Brazilian National Council for Scientific and Technological Development (CNPq) research productivity fellowship.

Citation

@article{Soares2026River,
  author = {Soares, Nazaré Suziane and Costa, Carlos Alexandre Gomes and Francke, Till and Mohr, C and Schwanghart, Wolfgang and Medeiros, Pedro Henrique Augusto},
  title = {River intermittency: mapping and upscaling of water occurrence using unmanned aerial vehicle, Random Forest and remote sensing landscape attributes},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-849-2026},
  url = {https://doi.org/10.5194/hess-30-849-2026}
}