[A-TT35] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT35-11] Machine learning-based streamflow estimation using GSMaP_NRT in Marikina River Basin

*Nelson Stephen Lising Ventura¹, Tsuyoshi Kinouchi¹ (1.Institute of Science Tokyo)

One of the most significant challenges in hydrologic simulation is the availability of streamflow data for both gauged and ungauged catchments. In the Marikina River Basin, publicly available streamflow data can be easily accessed; however, due to various factors leading to instrument failure, some daily streamflow observations may be missing. The study aims to address the gaps in data availability and to provide discharge values by implementing the Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Long-Short Term Memory (LSTM) algorithms for estimating daily mean discharge using satellite precipitation estimates from the hourly GSMaP_NRT in selected streamflow stations within the river basin. The results showed that utilizing all stations as outputs of the models produced better performance in comparison to estimating each gauge values individually. In addition, including the previous streamflow observations as additional input values resulted in lower estimation errors. However, when prior gauge measurements are considered, higher errors can be observed in consequent estimations over large gaps of missing data as depicted in the time series analysis. Moreover, the streamflow estimates generally tend to underestimate higher discharge observations, specifically during flood events. Although the XGBoost algorithm performed better than the other tree-based algorithm RF, the neural network-based LSTM provided the best performance overall. Nevertheless, the streamflow estimation models still require improvements as the resulting errors stream from the lack of consideration for other parameters influencing hydrologic process such as land use, land cover, soil characteristics, and evaporation.