Impact-based forecasting for anticipatory action to typhoons in the Philippines
To reduce the humanitarian impact of TC, both the Philippines Red Cross and United Nations OCHA Philippines have designed an agency-specific protocol, respectively in 2019 and 2021, which can be used to trigger early actions and release funding based on an impact-based forecasting model.
Due to its geographical location, the Philippines is highly exposed to Tropical Cyclones (TC). On average the Philippines receives about 20 TCs a year and they can cause significant humanitarian impact and economic loss. To reduce the humanitarian impact of TC, both the Philippines Red Cross and United Nations OCHA Philippines have designed an agency-specific protocol, respectively in 2019 and 2021, which can be used to trigger early actions and release funding based on an impact-based forecasting model. Both agencies use the same impact-based forecasting model, as developed by 510, an initiative of the Netherlands Red Cross.
The early actions (such as distributing house-strengthening kits) are pre-identified and triggered when the impact-based forecasting model indicates a pre-defined danger level is exceeded (with a lead time of 120 to 72 hours before landfall). The machine learning model consists of a classification and regression component and is trained on over 60 historical events. Around 40 predictors of hazard, vulnerability, coping capacity and exposure are used. The target is the damage to houses. The classification model predicts whether 10 per cent of buildings in a municipality will be completely damaged or not. Subsequently, the regression model gives the percentage of buildings that are completely damaged in a municipality. The machine learning model performed better than a baseline model (a wind-damage curve per building type) for historical typhoon events. The operational machine learning model uses actual forecast information from the European Centre for MediumRange Weather Forecasts to replace the historical hazard information at landfall. However, the machine learning impact-based forecasting model cannot be better than the hazard information that goes into it. T
hose typhoons that intensify rapidly cannot be captured at the cut-off of 72 hours lead time (the minimum time required to start up early actions). But for other typhoons, machine learning is very beneficial as a trigger tool for activating early actions and can support the reduction of the impact of typhoons on vulnerable communities. Another limitation of the initial machine learning model is that the spatial resolution is at the municipal level because the disaster losses and damages data are only available at that level. To tackle this issue, the ISI Foundation, UN OCHA Centre for Humanitarian Data and 510 The Netherlands Red Cross transformed the target variable (percentage of completely damaged houses) and not yet grid-based predictors to a 0.1° grid resolution by de-aggregation using Google Open Buildings data. In this way, the impact-based forecasting model could achieve a higher performance. Furthermore, the grid-based model increases the resolution of the predictions, which may allow for a more targeted implementation of anticipatory action.
Subsequent research focused on replicating and testing the approach in other TC-prone countries. However, in some countries, the disaster losses and damages data had such a poor spatial resolution that deaggregation was not possible. If the resolution of damage data is too coarse (for example, the provincial level), de-aggregation will assign damage data to parts of the province that are too far away from the track of the TC to have had damage in reality. This clearly shows the importance of having a minimum level of spatial resolution for damage data. Furthermore, in terms of transferring the model to a different country, a minimum number of historical events in the new country is required to retrain the model, ideally over more than five years. Ultimately, a transferable model combined with impact data with sufficient quality will facilitate the scaling up of anticipatory action for TCs.