UTILIZATION OF TEMPORAL DIMENSION IN SATELLITE IMAGERY: BETTER SEMANTIC SEGMENTATION WITH LOW DATA RESOURCES

Abstract

Time series image processing, a subfield of computer vision, enhances the accuracy of applications by leveraging temporal context. While this advantage is commonly utilized in video-based tasks, satellite imagery can also be treated as time series data when geospatial coordinates and timestamps are considered. Semantic segmentation, a key task in remote sensing, can benefit significantly from this temporal information. However, acquiring high-quality labeled datasets for such tasks remains a major challenge. In this study, we propose a novel temporal-aware domain adaptation framework for semantic segmentation, specifically targeting the detection of oil spills in the Caspian Sea. Our approach integrates time series information to improve cross-domain generalization. We evaluate our method on the synthetic SynthOil dataset, and a custom-labeled real-world dataset provided by Azercosmos and ArcGIS. Furthermore, we enhance the backbone of the Segformer model using a super-resolution dataset curated from Azercosmos and open data from the Esri ArcGIS platform. Experimental results demonstrate the effectiveness of our approach in improving segmentation performance across domains.