Joint Use of Remote Sensing and Volunteered Geographic Information for Exposure Estimation

Abstract

The frequency and intensity of natural disasters have been increasing in the last decades due to raising vulnerability and an increasing amount of elements at risk with respect to hazardous conditions. Considering a high spatio-temporal variability of elements at risk, detailed information is costly both in terms of time and economic resources and therefore often incomplete and outdated. To alleviate these restrictions, the availability of very high resolution satellite images promotes accurate and detailed analysis over various scales, covering large geographical areas. To extract relevant information from the data, supervised classi�cation and regression approaches are very popular due to their accuracy, robustness, and �exibility. The idea of such methods is to infer a rule (e.g., a decision function) from limited but properly encoded prior knowledge to assign e.g., a class label to unseen instances of the domain under analysis. Thereby, collection of appropriate prior knowledge (by e.g., detailed in-situ surveys) is normally the most time-consuming and expensive aspect with respect to data processing. Meanwhile, the development of the Web 2.0 enabled Volunteered Geographic Information (VGI) to emerge as a new category of worldwide available geodata, which represents a new potential source of information. In this thesis, we develop a new methodology where VGI contributions to the OpenStreetMap (OSM) project are utilized as an alternative data source for gathering of appropriate prior knowledge. Based on a very high resolution image from theWorldview-2 sensor with 0.5 m spatial resolution, we provide a fusion scheme to automatically use the VGI information for a statistical learning approach. Thereby, we identify several land use land cover classes for the region of Valparaíso, Chile - an area highly prone to diverse natural hazards. Based on several regression techniques as well as disaggregation algorithms incorporating OSM data, additional information concerning exposed buildings and population is estimated. The outcomes prove the applicability of VGI for remote sensing data processing. The exploitation of VGI is in particular important when proper prior knowledge is lacking or incomplete and could be imperative in the scope of an integrative risk assessment. The incorporation of VGI as source of essential prior knowledge disengages the common approach, which relies on site-speci�c dependencies, and promotes worldwide transferability and applicability.