COMPARATIVE ANALYSIS OF MULTIPLE MACHINE LEARNING ALGORITHMS FOR POST-EARTHQUAKE BUILDING DAMAGE ASSESSMENT IN HATAY CITY FOLLOWING THE 2023 EARTHQUAKE

Date

2024-4-26

Author

Kaya, Ömer

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

150
views

259
downloads

Cite This

This thesis presents earthquake damage assessment methodologies, focusing on integrating satellite imagery and machine learning (ML) algorithms. The study aims to enhance the accuracy and efficiency of identifying damaged buildings using only post-earthquake satellite images by combining remote sensing and ML techniques. ML algorithms, like Multivariate Adaptive Regression Splines (MARS), Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forest (RF), and ensemble learning methods, are evaluated for their effectiveness in automatic feature extraction from satellite imagery. Principal Component Analysis (PCA) is employed with the derived texture image bands, augmenting the discriminative power of the models. A ten-fold cross-validation process is used in the study. The study also tackles challenges in classifying building footprints from non-nadir imagery, where intricate building shapes and densities impede shadow detection, requiring a customized approach integrating vector data refinement. Examining building footprint areas across folds uncovers considerable variations in size distribution, particularly evident in Fold 5. Evaluating the highest accuracy fold with the Damage Proximity Map emphasizes the criticality of aligning ML classification outcomes with ground truth data. The results of the study highlight several key findings. SVM with a linear kernel emerges as the top-performing algorithm, mainly exhibiting superior accuracy, achieving 57% to 64.82% accuracy. Additionally, MARS demonstrates stable performance across folds, maintaining accuracies around 60% to 62%. In contrast, SVM with 2nd-degree polynomial kernel and ensemble methods exhibit different inaccuracies across folds.

Subject Keywords

Multispectral Imagery, Post-Earthquake Building Damage Assessment, Machine Learning

URI

https://hdl.handle.net/11511/109446

Collections

Graduate School of Natural and Applied Sciences, Thesis