Due to its geographical location, the Philippines is highly exposed to Tropical Cyclones (TC). Every year at least one TC will make landfall and cause significant humanitarian impact and economic loss. To reduce the humanitarian impact of TC, the Philippine Red Cross with the German Red Cross and 510, an initiative of The Netherlands Red Cross, designed and implemented a Forecast Based Financing (FbF) system. The early actions in the FbF system are pre-identified and will be triggered when an impact-based forecasting model indicates a pre-defined danger level will be exceeded. This research develops and evaluates multiple ML algorithms for classification and regression with a lead time of 120 to 72hrs before TC landfall. The algorithms are trained on around 40 historical typhoon events and xx predictors on the hazard, vulnerability, coping capacity, and exposure are used. The classification model predicts if 10% of buildings in a municipality are completely damaged or not. The regression model gives the percentage of buildings that are completely damaged in a municipality. The RandomForest algorithm outperformed other algorithms for both classification and regression for both training and validation datasets. The ML models performed better than a baseline model (a wind-damage curve per building type) for the historical typhoon events. The Philippine Red Cross has been using the ML model since 2019, whereby actual forecast information from ECWMF replaces the historical hazard information at landfall. However, the ML impact-based forecasting model cannot be better than the hazard information that goes into it. Those typhoons that rapidly intensify cannot be captured at the cutoff of 72 hrs lead time (the minimum time required to start up early actions). But for the other typhoons, ML is very beneficial as a trigger tool for activating early actions and can support the reduction of the impact of typhoons on vulnerable communities.

Reliable predictions of the impact of natural hazards turning into a disaster is important for better targeting humanitarian response as well as for triggering early action. Open data and machine learning can be used to predict loss and damage to the houses and livelihoods of affected people. This research focuses on agricultural loss, more specifically rice loss in the Philippines due to typhoons. Regression and binary classification algorithms are trained using feature selection methods to find the most important explanatory features. Both geographical data from every province, and typhoon specific features of 11 historical typhoons are used as input. The percentage of lost rice area is considered as the output, with an average value of 7.1%. As for the regression task, the support vector regressor performed best with a Mean Absolute Error of 6.83 percentage points. For the classification model, thresholds of 20%, 30% and 40% are tested in order to find the best performing model. These thresholds represent different levels of lost rice fields for triggering anticipatory action towards farmers. The binary classifiers are trained to increase its ability to rightly predict the positive samples. In all three cases, the support vector classifier performed the best with a recall score of 88%, 75% and 81.82%, respectively. However, the precision score for each of these models was low: 17.05%, 14.46% and 10.84%, respectively. For both the support vector regressor and classifier, of all 14 available input features, only wind speed was selected as explanatory feature. Yet, for the other algorithms that were trained in this study, other sets of features were selected depending also on the hyperparameter settings. This variation in selected feature sets as well as the imprecise predictions were consequences of the small dataset that was used for this study. It is therefore important that data for more typhoons as well as data on other explanatory variables are gathered in order to make more robust and accurate predictions. Also, if loss data becomes available on municipality-level, rather than province-level, the models will become more accurate and valuable for operationalization.

In the Asia–Pacific, where rice is an essential crop for food security and economic activity, tropical cyclones and consecutive floods can cause substantial damage to rice fields. Humanitarian organizations have developed impact-based forecasting models to be able to trigger early actions before floods arrive. In this study we show how Sentinel-1 SAR data and Otsu thresholding can be used to estimate flooding and damage caused to rice fields, using the case study of tropical storm Talas (2017). The current most accurate global Digital Elevation Model FABDEM was used to derive flood depths. Subsequently, rice yield loss curves and rice field maps were used to estimate economic damage. Our analysis results in a total of 475 km² of inundated rice fields in seven Northern Vietnam provinces. Flood depths were mostly shallow, with 2 km² having a flood depth of more than 0.5 m. Using these flood extent and depth values with rice damage curves results in lower damage values than the ones based on ground reporting, indicating a likely underestimation of flood depth. However, this study demonstrates that Sentinel-1-derived flood maps with the high-resolution DEM can deliver rapid damage estimates, also for those areas where there is no ground-based reporting of rice damage, showing its potential to be used in impact-based forecasting model training.

Tropical cyclones (TCs) significantly impact the agricultural sector, particularly rice production. Anticipatory action aims to reduce the impact by triggering early actions based on an impact-based forecast. We present a machine-learning framework to predict rice losses at the municipal level in the Philippines, using data from 51 TCs between 2006 and 2020. The target variable is the percentage of rice area loss, and the features cover hazard, exposure, and vulnerability. We developed regression and binary classification models using Random Forest (RF) and XGBoost (XGB) and integrated them into a two-step model for operational use. In this two-step model, the classification identifies the municipalities to be selected for anticipatory action, while the subsequent regression supports the decision-making on the magnitude of the early actions for the selected municipalities, as it estimates the extent of rice loss. The RF models outperformed the benchmarks and the XGB models in classification and regression tasks. The hazard features—particularly maximum sustained windspeed, rainfall, and minimum track distance—were the most influential predictors, whereas the exposure features and the vulnerability feature hampered model performance. Two methods for extracting rainfall features were compared: a landfall-based approach and an approach that delineates time frames according to the TC’s spatial extent. The landfall-based method had comparable performance and was simpler to understand and implement. While further enhancements in data collection and modelling are encouraged, our two-step modelling approach can be used to trigger anticipatory actions for rice losses from TCs.