Double Acquisition Neural Network for Matching Marine Debris Patches Across PlanetScope and Sentinel-2 Imagery

By Gabrielė Tijūnaitytė

Abstract
The gradual accumulation of plastic waste in marine environments threatens biodiversity and human welfare. Monitoring this waste can inform policy evaluation and support waste recovery efforts. While existing remote sensing studies focus on detecting marine debris (MD), a proxy for plastic waste, in isolated satellite scenes, tracking MD agglomerations (patches) over time in multi-temporal imagery remains largely unexplored.

This study presents the Double Acquisition Neural Network (NN) designed to match MD patches across PlanetScope (PS) and Sentinel-2 (S2) double acquisitions captured within a one-hour interval. The MD patch matching task was framed as a cross-platform image retrieval problem. A dataset of 3,445 annotated PS-S2 MD patch pairs was compiled to train and evaluate the model for this problem. The study systematically examined different designs of supervised contrastive frameworks and possible retrieval performance gains from restricting the candidate search scope, aided by varying levels of prior knowledge on local drift.

The optimal configuration, featuring two modality-specific ResNet-18 encoders trained with symmetric cross-entropy loss, achieved a top-1 retrieval accuracy of 37.0% in a constrained-to-study-site scope regime when initialised with Earth-observation-domain pre-trained weights. Incorporating prior knowledge of local drift further improved accuracy to 62.2% and reduced the mean position of the true match to 1.9.

These findings demonstrate that the proposed Double Acquisition NN can reliably re-detect MD patches within two retrieval attempts when aided by drift knowledge. This work lays the foundation for automating MD patch tracking and enabling time-resolved monitoring of marine plastic pollution.