TOWARDS PREDICTING RICE LOSS DUE TO TYPHOONS IN THE PHILIPPINES

We improve an existing impact forecasting model for the Philippines by transforming the target variable (percentage of damaged houses) to a fine grid, using only features which are globally available. We show that our two-stage model conserves the performance of the original and even has the potential to introduce savings in anticipatory action resources. Such model generalizability is important in increasing the applicability of such tools around the world.

There is still no full understanding of the relation between drought impacts and drought indices in the Horn of Africa where water scarcity and arid regions are also present. This study assesses their relation in Kenya. A random forest model reveals that each region, aggregated by aridity, has its own set of predictors for every impact category. Water scarcity was not found to be related to aridity. Understanding these relations contributes to the development of drought early warning systems.

Local communities in northern Malawi have well-developed knowledge of the conditions leading to flash floods, spatially and temporally. Scientific analysis of catchment geomorphology and global reanalysis datasets corroborates this local knowledge, underlining the potential of these large-scale scientific datasets. Combining local knowledge with contemporary scientific datasets provides a common understanding of flash flood events, contributing to a more people-centred warning to flash floods.

This research introduces a novel approach to estimate flood damage in Malawi by applying a machine learning model to UAV imagery. We think that the development of such a model is an essential step to enable the swift allocation of resources for recovery by humanitarian decision-makers. By comparing this method (EUR 10 140) to a conventional land-use-based approach (EUR 15 782) for a specific flood event, recommendations are made for future assessments.