ABSTRACT The Philippines is frequently affected by tropical cyclones (TCs). Among the TC‐associated hazards, rainfall can lead to cascading impacts such as floods and landslides. A robust and computationally inexpensive TC rainfall forecasting method is critical in disaster preparation and risk reduction efforts. We used machine learning (ML) to develop a TC rainfall forecast model from parameters such as TC track and locale‐specific characteristics. Specifically, a self‐organizing map (SOM) was utilized to cluster the TC tracks, which were then fed into a random forest (RF) regression model that used TC position, intensity, translational speed, and other parameters to predict accumulated TC rainfall. The resulting artificial intelligence (AI)‐based TC rainfall model was initially assessed against ground rainfall observations for calibration. Then, the model was evaluated for its prediction skill. Model interpretability of the RF model revealed insights into how the input parameters influence the model response. The RF model determined that distance to TC has the most influence on the variability of the accumulated TC rainfall, followed by TC duration, latitude of land grid, and the type of TC track as clustered by the SOM. The model produced similar rainfall distributions to calibrated satellite rainfall observations. It was able to produce rain predictions well and is particularly skillful in predicting intense rainfall events in comparison with the other statistical or dynamical weather models (i.e., WRF model). The predictive ability of the RF model, together with its low computational power requirement, makes it a potential tool to augment TC rainfall forecasting in the Philippines.
Mesias et al. (Tue,) studied this question.