Supplementary material to "Ten years of hydrometeorological observations at 10-minute resolution and its application in machine learning hydrological models"

Le résumé fourni par la source

This README describes the dataset �tled 'TTI-HydroMet: A Decade of High-Resolu�on Rainfall and Streamflow for the Tamanduateí River Watershed, Brazil,' which can be accessed at htps://doi.org/10.5281/zenodo.17654660.The dataset is divided into two primary folders: Code and Data.Further details are given below. Folder: CodeThis folder includes four different codes for data processing.• Create_Radar_Input: this Python script is responsible for construc�ng the final modeling datasets by integra�ng weather radar observa�ons with river stage measurements. It reads a preprocessed radar dataset containing radar data (gridded variables) at 10minute resolu�on and a �me series of observed stage data ("nível"). The code aligns both datasets temporally using the �mestamp column ("datahora"), verifies that radar data exist for all stage observa�on �mes, and converts stage values from millimeters to kilometers. For each specified forecast lead �me (e.g., 10, 60, 120, and 240 minutes), the script generates a target output column that represents the future stage shi�ed forward by the corresponding number of �me steps. The resul�ng datasets contain the �mestamp, all radar grid variables, the current stage, and one forecast target column (e.g., "out00h10m", "out01h00m"). Each lead �me is saved as an independent CSV file following a standardized naming conven�on ("iFast_Radar_Obs_2015_2025_ .csv"), which serves as direct input to the machine-learning modeling framework.• Model_v6_SP: this Python script coordinates the full machine learning experimenta�on process for training, valida�ng, and evalua�ng river stage forecas�ng models. It loads previously generated CSV datasets, dynamically selects input variables based on the experiment configura�on, and prepares the data for modeling by defining input-output structures and temporal splits. The script executes the ML4FF framework (Soares et al., 2025) using nested cross-valida�on (inner and outer loops) and a holdout dataset to ensure robust performance assessment. Mul�ple machine-learning or deep-learning algorithms can be tested, with automa�c hyperparameter op�miza�on performed via Bayesian search. For each experiment, the script saves trained models, predic�on results, performance metrics (NSE, RMSE, and KGE), execu�on diagnos�cs, and full configura�on metadata. Outputs are organized into �mestamped result directories and exported as CSV and Excel summary files, along with configura�on logs.• RadarDataMissingCode: this Python script analyzes missing data in the 10-year �me series from 01/04/2015 to 29/03/2025.It generates a distribu�on of missing data based on dura�on or period -specifically, the number of missing records mul�plied by the length of missing data -considering only precipita�on exceeding 0.2 mm.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.