Adapting with Confidence : Certainty Weighted Guidance for Domain Adaptation in Remote Sensing Semantic Segmentation

Sammanfattning

Deep learning-based semantic segmentation methods have proven effective in many remote sensing tasks. However, their performance degrades when data distributions shift, for example, when applying a model to a domain different from the one on which it was trained. This hinders the applicability of said models. In this thesis, we investigate how knowledge about data can be incorporated into domain adaptation methods to improve semantic segmentation performance on the target domain. We propose two frameworks based on adversarial and contrastive learning, incorporating existing data and prior knowledge through a pixel-wise certainty measure. Experimental results show that our methods outperform the baseline by 6.8 percentage points and conventional domain adaptation methods by 2.2 percentage points. Our findings suggest that the contrastive approach is better suited for the task as the tailored loss favors target domain alignment. We also found that an adaptive certainty measure, based on model predictions, was best suited for guiding adaptation. Furthermore, extending the unsupervised module of the methods with more data degrades performance, suggesting that direct supervision is essential for accuracy and stable training. Our methods are effective in scenarios with geographically aligned source and target domains, enabling better segmentation without requiring target annotations. This work contributes to reducing the reliance on extensive manual annotation, thereby lowering the cost and increasing the scalability and adaptability of deep learning models.