An Ensemble Learning Approach of Multi-Model for Classifying House Damage

Abstract: In the past decades, natural disasters have caused an immeasurable impact on the development of human society. The original way to measure the damage is to detect and quantify the number of damaged buildings, usually by driving around the affected area and noting down manually, which takes a lot of resources. Therefore, damage assessment is vital to the relief helpers and first responders so that resources and help can be planned and allocated appropriately. To tackle this issue, an ensemble learning model for classifying house damage is proposed in this paper. Firstly, to facilitate the processing of the model, a set of satellite images before and after the destruction of houses is preprocessed using data augmentation. Secondly, the basic models are first trained and used to predict test sets. Their running parameters were recorded to demonstrate their performances. Thirdly, using Bootstrapping to generate multiple sub-datasets and Bagging is adopted to ensemble the basic classifiers. Finally, the ensemble learning model is used to predict the test set, and the final result is determined by voting strategy to achieve the optimization effect. The experimental results on the public dataset show that the proposed model achieves an ideal effect on classification. Furthermore, by comparing the results of the basic classifiers with that of the model using the Bagging strategy, we find that Bagging is significantly helpful to improve the model effect.

Find Paper here
Download here