Jakarta is highly vulnerable to flooding, mainly due to water flowing from upstream of the Ciliwung River, which aggravates flooding conditions in downstream areas. The Katulampa Barrage serves as a key observation point for early warning, with an average flood travel time to Jakarta of 13–14 hours. This study uses the Long Short-Term Memory (LSTM) framework with satellite-based GSMaP rainfall as input to extend the lead time of the Jakarta flood early warning system. Two modeling scenarios were analyzed. In the first scenario (LSTM-1), GSMaP rainfall data were used to predict observed water levels at Katulampa. In the second scenario (LSTM-2), HEC-HMS was first used to simulate discharge values that the LSTM machine learned from, with GSMaP rainfall also used as an input variable. The LSTM-1 did not perform well because the water level observations at Katulampa are of low quality. On the other hand, the LSTM-2 scenario shows the ability to forecast up to 6 hours ahead. These results demonstrate that physical hydrological simulation results that feed into a machine learning model can improve flood forecasting capabilities and extend the lead time for Jakarta's flood management system's preparedness. This study emphasizes that as long as the LSTM emulation is trained with consistent data, it delivers reliable results.
Kardhana et al. (Sun,) studied this question.