Enhancing the Detection of Coastal Marine Debris in Very High-Resolution Satellite Imagery via Unsupervised Domain Adaptation

Abstract

In this study, we propose a robust debris estimation model applied to satellite imagery that is suitable for practical applications. In our previous study, we proposed a coastal marine debris estimation model using semantic segmentation applied to very high-resolution satellite images. We identified limitations when applying the model to various lower spatial and spectral resolution satellite images or to areas with fewer satellite images cases. To overcome these limitations, we now employed unsupervised domain adaptation (UDA) techniques to transfer the earlier model to these lower resolution or fewer satellite images. These domain adaptation techniques consider differences in spatial feature distributions and/or satellite sensor characteristics. We confirmed the ability of UDA to classify Planet Skysat and Airbus Pleiades images using MAXAR WorldView images to generate an accurate segmentation map. The UDA, then, allows us to analyze the lower satellite images without the need to independently generate new segmentation labels. We conducted statistical analyses and demonstrated the high correlation between the local debris cleanup data and entropy metrics computed using our UDA approach. Our method enhances the sampling frequency of satellite images by analyzing lower resolution imagery, allowing monthly to weekly, or even daily intervals, and facilitates rapid estimation utilizing fewer images, thereby providing an invaluable tool for coastal debris characterization and assessment.