Supervised classification of aerial imagery and multi-source data fusion for flood assessment

Floods are among the most devastating natural hazards and the ability to produce an accurate and timely flood assessment before, during, and after an event is critical for their mitigation and response. Remote sensing technologies have become the de-facto approach for observing the Earth and its environment. However, satellite remote sensing data are not always available. For these reasons, it is crucial to develop new techniques in order to produce flood assessments during and after an event. Recent advancements in data fusion techniques of remote sensing with near real time heterogeneous datasets have allowed emergency responders to more efficiently extract increasingly precise and relevant knowledge from the available information. This research presents a fusion technique using satellite remote sensing imagery coupled with non-authoritative data such as Civil Air Patrol (CAP) and tweets. A new computational methodology is proposed based on machine learning algorithms to automatically identify water pixels in CAP imagery. Specifically, wavelet transformations are paired with multiple classifiers, run in parallel, to build models discriminating water and non-water regions. The learned classification models are first tested against a set of control cases, and then used to automatically classify each image separately. A measure of uncertainty is computed for each pixel in an image proportional to the number of models classifying the pixel as water. Geo-tagged tweets are continuously harvested and stored on a MongoDB and queried in real time. They are fused with CAP classified data, and with satellite remote sensing derived flood extent results to produce comprehensive flood assessment maps. The final maps are then compared with FEMA generated flood extents to assess their accuracy. The proposed methodology is applied on two test cases, relative to the 2013 floods in Boulder CO, and the 2015 floods in Texas.