Semantic Segmentation for HR Aerial Imagery using UNET model and comparing Loss Functions

UNET Model --- 

Semantic Segmentation has increasingly been used and claimed to thrive in providing better classification results when compared to other classification techniques – pixel-based and objectbased classification. This is because semantic segmentation effectively extracts the objects of the same class and provides label to each pixel of that class. This characteristic of semantic segmentation helps it effectively differentiate between different objects. This project aims to undergo semantic segmentation for a very high-resolution aerial imagery using UNET architecture. The reason for considering UNET architecture over others is associated with its performance to easily yield good results for high resolution images. The considered dataset has six classes in total – water, buildings, land, vegetation, road and unlabeled. In addition, the project also aims to compare between two prominently used loss functions – Binary Cross Entropy and Cross Entropy Extension, which are applied to optimize the model architectures. The cross-entropy extension loss function is a custom loss function that includes both, dice loss and focal loss, which promises better results since it accounts the imbalanced class weights for the given dataset. The results for this project could not completely claim the better performance of cross entropy extension loss function as it still requires more work to gain better results in order to provide a strong comparison between both the considered loss functions.

Keywords – Semantic Segmentation, Loss Functions, UNET architecture, Remote Sensing