Catchment Filtering is not a Neutral Step: Missing-Data Thresholds Alter Basin Composition and Regional LSTM Performance in CAMELS-BR

Código

V-END_II-EBHE0172

Autores

Thiago Masaharu Osawa, Dário Hachisu Hossoda, Elton Vicente Escobar Silva, José Rodolfo Scarati Martins

Tema

19 - Inteligência artificial e big data na gestão de desastres

Resumo

Artificial intelligence and big data are increasingly used to support flood forecasting, early warning systems, and disaster risk management. However, the reliability of regional deep learning models depends strongly on the quality and representativeness of the hydrological data used for training. This study evaluates how streamflow missing-data thresholds and training-period length affect regional LSTM streamflow modeling in CAMELS-BR. Four experiments were conducted using the NeuralHydrology framework and a regional CudaLSTM model trained with CAMELS-BR meteorological forcings and static catchment attributes. Experiments E1?E3 used the same period, from 1990 to 2024, but applied missing-data thresholds of 5%, 10%, and 20%, retaining 618, 720, and 801 basins, respectively. Experiment E4 used a 10% threshold but extended the training period back to 1980, retaining 668 basins. All models were evaluated over the same test period, from 2019 to 2024. Results showed that catchment filtering was not spatially or hydroclimatically neutral: relaxing the missing-data threshold added basins that were more concentrated in the Southeast, smaller, wetter, and more runoff-productive. The best model performance was obtained with the strictest threshold, E1, which achieved the highest NSE (0.57), highest KGE (0.63), lowest RMSE (0.74), lowest MAE (0.39), and nearly unbiased simulations. Relaxing the threshold increased the number of basins but did not improve performance, and adding 1980?1990 records also did not provide gains. These results show that catchment filtering is not a neutral preprocessing step in CAMELS-BR regional modeling.