Forecasting the Damages Caused to Buildings Due to Earthquake using Various Machine Learning Techniques

Abstract

Earthquake is the deadliest natural disaster known to humankind. An earthquake is defined as an unanticipated violent tremble of the ground or earth's surface due to the sudden release of a large amount of energy. Earthquakes are normally created when the rock below abruptly breaks along a fault. This impromptu release of energy creates seismic waves that cause the ground to shake. Earthquake destroys both lives and properties. In fact, earthquakes kill more people than all the other natural disasters put concurrently. Many lives are lost due to the poor quality of building structures.

Our paper is based on an earthquake that occurred in Nepal in 2015. We are using data that was collected by the National Planning Commission Secretariat Kathmandu using the Central Bureau of Statistics and Living Labs of Kathmandu’s surveys. The goal of our paper is to build an accurate Machine Learning which would predict the damage level a building would suffer if an earthquake would happen. We have tried to implement different Machine Learning models. We are considering an F-1 score to get the accuracy of models. We have also taken training and testing time for consideration. Our endgame is to use that model to predict what kind of buildings are safer.

The process begins with first collecting the data and applying Exploratory Data Analysis to it. Exploratory Data Analysis assists us to interpret the data well. By using Exploratory Data Analysis we were able to find the relationships between the parameters. We also found out about the outliers in Age, Height percentage and area percentage. To remove these outliers we have used the Winsorization technique. This technique helps to spread the data evenly. After winsorization, we trained the following models, Decision Tree, Random Forest, Gradient Boosting Classifier, XGBoost, Logistic Regression, Lightgbm, KNN, Adaboost.

From the aforementioned models, the top 3 models were XGBoost, Gradient Boosting Classifier, Lightgbm with 74.73%, 74.54% and 74.52% accuracy respectively. Total training time taken by these models were 1389 seconds, 5980 seconds and 1606 seconds respectively.