Michael Uzowuihe

DAMAGE ASSESSMENT USING MACHINE LEARNING Disaster Problem It is quite obvious of the occurrence of disasters such as hurricanes, earthquakes, floods etc. that occur in the environment and tend to cause damages to people or buildings. This has been a challenge, in the sense that obtaining accurate but timely data to enable planning for an effective response to such a situation is lagging. Solution to Damages On the response to this crisis, very high-resolution satellite imagery (Remote Sensing) became an important tool that provides data (visuals) at an unprecedented scale while making the extraction of its information slow and labour-intensive for operation. As a result of this, machine learning (ML) was applied to automate models for detecting damages caused by disasters and proffer solution to future ones through consistently training and testing them on different disaster events. Machine Language Application Thus, the Building Damage Detection in Satellite Imagery Using Convolutional Neural Networks came to be, using the ML approach to process the satellite data from Remote sensing, as well as generate assessments of any damage. Subsequently, this has probably reduced the time, energy and effort required of crisis workers to generate disaster reports. Meanwhile, a timely decision on delivering aids that would cover the damages caused in any area are strictly implemented without any delay. Processes & Flows The assessment generated report is a split of two steps i.e. building detection and damage classification. The ML approach uses the detection model to analyze the building images and the classification model to give an output score of 0.0 and 0.1 representing whether a building is damaged or not respectively. In cases where illumination and colour differences tend to be a problem in getting an accurate result, a histogram equalization can be used to normalize between images before testing. Additionally, getting training data with regards to satellites images of buildings has been a bottleneck or challenge to the training of models using the ML approach for damage assessment. Therefore, all images that are currently used were sought from commercially available sources like UNOSAT and REACH. The results obtained so far in using the Models were tested against damage assessment of experts in the field and it showcases a 70% accuracy, which is the threshold for making highlevel decisions in a disaster. As thus a positive ground truth for evaluation is obtained for this study.