Semantic Segmentation on Landslide Containment Devices

This thesis analyzes the application of semantic segmentation models with different approaches and aims to do a model selection and find the best optimal model with a grid search. The research focuses applying semantic segmentation on main component of landslide containment devices that are mesh and wire. Then it gives brief introduction to researches that are done so far related to semantic segmentation and explains Neural networks and how deep learning models are implemented for semantic segmentation problems. Then the data set is analyzed and exploited, as well as the necessary pre-processing steps to prepare the data for model training. Pre-processing phase includes data annotation, creating different data-sets with and without data augmentation techniques employed and data splitting for generating different separate data for different purposes such that training, validation and testing. Various model structures are explored, along with the corresponding metrics and loss functions used to refine the models for comparison. The performance of the final models on test sets is evaluated, and their overall accuracy is determined. In this thesis, binary and multi-class semantic segmentation models are explored and, U-Net and U-Net++ architectures are compared. Additionally, a grid search is conducted to find the best suitable parameters for the employed models, different combination of hyper-parameters are evaluated. The research provides insights into the use of semantic segmentation for landslide containment devices, and highlights the importance of appropriate data pre-processing and model selection for achieving accurate and reliable results. Model performance is evaluated on test sets and reported as well as the prediction images are demonstrated with the original image and ground truth image for better visualize the performance of the model.