Accurate delineations of Earth’s surface features are crucial for environmental science, but for dynamic landscapes like thawing permafrost, these datasets rapidly become outdated and inaccurate. Manual refinement of these outdated labels is prohibitively slow, while existing foundation segmentation models often fail to accurately segment complex natural features in satellite images. To address this, we aim to use inaccurate labels as geospatial priors to prompt a vision foundation model for zero-shot semantic segmentation of Earth observation objects with current satellite imagery. We developed an interactive segmentation framework that synergistically combines the Segment Anything Model (SAM) with expert oversight. Our system features two components: (1) a scalable automated pipeline that uses historical delineations as a strong geospatial prior for prompting SAM, and (2) an interactive tool that enables experts to refine outputs and create new, high-fidelity labels rapidly. Using the challenging case study of retrogressive thaw slumps, we demonstrate that this human–AI partnership achieves robust segmentation performance, with a best Intersection over Union of 0.80 ± 0.10 for routine updates. This approach matches the quality of manual expert digitisation but at twice the efficiency. This model- and task-agnostic framework offers a practical, transferable solution for generating the dense, time-series data required to monitor landscape changes, transforming a laborious mapping task into an AI-assisted, expert-overseen workflow. • A novel method that updates geospatial delineations with current satellite imagery • A workflow that effectively integrates machine-scale processing with expert oversight • A model-agnostic workflow for diverse environmental monitoring tasks • Enabling high-fidelity time-series segmentation to quantify landscape change rates
Yang et al. (Fri,) studied this question.