Mapping the evolution of small reservoirs in Brazil from 1984 to 2025 using deep learning

Technical Summary: Mapping the Evolution of Small Reservoirs in Brazil (1984–2025)

Problem Statement
Water research in Brazil has historically overlooked the widespread damming of small streams for agricultural purposes, such as cattle watering, farm-scale hydropower, irrigation, and aquaculture. While large dams are well-studied, small reservoirs (<1 km²) are ubiquitous yet difficult to map. Their small size leads to mixed pixels in satellite imagery, and distinguishing artificial reservoirs from natural lakes, wetlands, and oxbow lakes remains a significant challenge. Consequently, most regional and global datasets exclude these features, creating a gap in understanding the cumulative impacts of small stream impoundments on water temperature, connectivity, biodiversity, greenhouse gas emissions, and evaporative water losses. Existing mapping efforts have been limited by the difficulty of remote sensing discrimination and a lack of consistent, long-term datasets.

Methodology
To address these challenges, the authors developed a deep learning approach to generate annual, country-wide maps of small reservoirs in Brazil from 1984 to 2025.

• Data Sources: The study utilized a 42-year archive of Landsat imagery (Landsat 5 ETM, Landsat 7 ETM+, Landsat 8 OLI, and Landsat 9 OLI-2). Cloud-free median-value surface reflectance composites were generated using Google Earth Engine, incorporating scenes from adjacent years to minimize cloud contamination. The input bands included Blue, Green, Red, NIR, SWIR-1, and SWIR-2.
• Training Data: Ground truth was established using a 2017 cloud-free Sentinel-2 composite (10m resolution). Researchers manually annotated 1,510 tiles (5km x 5km), identifying 4,474 reservoirs. To improve model robustness, 196 additional tiles covering river floodplains and wetlands without reservoirs were added to the training set to help the model distinguish non-reservoir water bodies.
• Model Architecture: A Multi-Scale Attention Network (MA-Net) with a ResNet-34 backbone was trained from scratch to perform binary semantic segmentation. The model was designed to delineate full reservoir extents without relying on strict rule-based assumptions.
• Cross-Sensor Calibration: To ensure consistency across the four Landsat generations, independent binary prediction thresholds were tuned for each sensor. Landsat 8 served as the reference for Landsat 7, which in turn served as the reference for Landsat 5. Landsat 9 utilized the same threshold as Landsat 8 due to sensor similarity.
• Inference and Post-Processing: The model was applied to the entire Brazilian territory using a sliding window approach. Post-processing involved masking cloud-corrupted pixels and filtering out detections overlapping with known large water bodies (>1 km²) to remove false positives from inlets or bays of major lakes.

Key Results

• Reservoir Growth: The study detected a nearly fourfold increase in the number of small reservoirs, rising from 263,913 in 1984 to 996,245 in 2025. The total surface area expanded from 3,510 km² to 8,550 km².
• Spatial Distribution: Initially, reservoirs were concentrated in the Pampa, Cerrado, Mata Atlântica, and Caatinga biomes. Over the study period, the Amazon biome experienced the most dramatic growth, with reservoir area increasing by 996% (from 127 km² to 1,390 km²). The Pantanal remained relatively free of reservoirs.
• Temporal Dynamics: The rate of reservoir creation remained relatively constant over four decades, with occasional spikes and plateaus. The Caatinga biome showed high variability, with reservoir counts fluctuating significantly in response to drought events.
• Model Performance: The MA-Net achieved high accuracy on the withheld 2017 test set. When comparing overlapping years between sensors, Landsat 8 and 9 demonstrated superior sensitivity to small reservoirs compared to Landsat 7 and 5, detecting significantly more reservoirs (e.g., 43.8% more than Landsat 7) while maintaining similar total area estimates. This suggests Landsat 8/9 detected many very small reservoirs that previous sensors missed or fragmented.
• Sensor Consistency: Despite the Scanline Corrector failure on Landsat 7 in 2003, agreement between Landsat 5 and 7 remained high. However, the study notes that sensor differences introduce discontinuities in the time series, particularly regarding reservoir counts, necessitating careful selection of sensors based on the specific application (e.g., count vs. area analysis).

Significance and Claims
The authors claim this work provides the first country-wide, annual dataset representing the evolution of small reservoirs over four decades. The primary significance lies in:

1. Filling a Data Gap: Providing a consistent, multi-decadal record of small stream barriers, which are critical for understanding the hydrological impacts of agricultural expansion in Brazil.
2. Methodological Advancement: Demonstrating that deep learning semantic segmentation can effectively map small, stream-fed reservoirs across multiple Landsat generations, overcoming the limitations of two-stage classification approaches and manual annotation.
3. Scientific Utility: The publicly available maps enable new research into the cumulative impacts of small dams on greenhouse gas emissions, stream fragmentation, water quality, and biodiversity. The dataset allows for the refinement of emission estimates and the analysis of drivers behind reservoir construction in relation to agricultural policy, infrastructure development, and climate variability.

The paper concludes that while the dataset is robust, it likely underestimates total reservoir extent due to seasonal variations (maps reflect dry season minimums) and potential confusion with rain-fed impoundments (barraginhas). Nevertheless, the study establishes a foundational tool for monitoring small-scale surface water changes in Brazil and offers a framework for applying similar methods to global surface water dynamics.