Reference Mean and Low Streamflow for all Brazilian Catchments Generated Using Machine Learning Models

This dataset provides comprehensive hydrological information for river networks in Brazil, focusing on reference streamflows, specifically long-term mean flows (qm) and low flows exceeded 95% of the time (q95). Covering over 400,000 ungauged river points, the dataset was developed using advanced machine learning models trained on environmental descriptors and validated against data from 1,069 gauging stations spread across the country. The machine learning pipeline evaluated six regression models to achieve high predictive accuracy (R² > 0.8 for qm and > 0.7 for q95). The 62 environmental descriptors - encompassing climate, topography, land cover, lithology, and water storage characteristics - that were used as features for the models are also included.

Key features:

Spatial Coverage: Brazilian territory and the Amazon River basin, based on the BHO 5k dataset of officially adopted river networks.

Outputs: Predicted qm and q95 values for each river stretch, with 90% and 75% confidence intervals to account for prediction uncertainty.

Environmental Descriptors: Aggregated from upstream catchment area.