Data that was collected for Haima and for 4 previous Typhoons (Haiyan, Rammasun, Hagupit and Melor), included windspeed data, rainfall data, population, poverty data and number of damaged houses (where the latter is not collected for Haima, but is to be predicted). All data was calculated at the municipality level of the Philippines. This means for example that per municipality the average rainfall over the whole municipality area was calculated. Finally, all information per municipality was joined to each other using the PCODE-system, which assigns a unique identifier to each administrative area in the Philippines.
The prediction model was build using a Machine Learning method, called Random Forest Regressor. As all regression methods it uses historical data and attempts to learn a correlation between input data and their impact on the output.
Random Forest Regressor was selected as it outperformed another, usually also very successful method - Gradient Boosting Regressor. Its power comes from an interesting strategy of building multiple predictors (decision trees) and averaging their outputs. Each tree is built in a slightly different way, using different subset of historical data, and randomly selecting different variables during the process of building the trees. This strategy allows to build a model that can handle well multidimensional data and estimate well importance of each input variable.
It is a highly configurable method so several experiments were held to select parameters that produce the best results on training data.
Evaluation of performance let us estimate average prediction error to be close to 1800 and median error close to 1300 (half of predictions should be closer to real values than 1300).
Errors were estimated using cross validation: averaging errors made by the model trained on all but one typhoon, which was left out for testing. Procedure was repeated 4 times, each time leaving different one for testing.
The model was applied to predict numbers of houses damaged by Haima, the most recent typhoon. Predicted numbers were used to prioritize municipalities in a scale from 1 to 5 (1 of lowest damage, 5 for highest predicted damage). The categories were assigned using following percentiles: 0%, 35%, 65%, 85%, 95%, which should be interpreted as: 35% of all municipalities with lowest damage were assigned priority 1, 5% of all municipalities with highest predicted numbers were given priority 5. This selection of thresholds was rather arbitrary, with a goal in mind to emphasize places with highest damage. Chosen percentiles translate to the following absolute numbers:
Priority 1 - predicted number of houses damaged below 183
Priority 2 - predicted number of houses damaged between [183, 345)
Priority 3 - predicted number of houses damaged between [345, 1205)
Priority 4 - predicted number of houses damaged between [1205, 2498)
Priority 5 - predicted number of houses damaged over 2498
Prepared model is going to be improved in future by adding more input variables related to geographical information and by extending training data. When more data is available also selection of the algorithm will be reconsidered, if simulation results motivate it.
Caveats / Comments
Accuracy: NDRRMC publishes official data; they can only publish the official data when it is validated by their local counterparts (LGU or regional officials).