Quantifying Mangrove Extent and Uncertainty: An Embedding-Driven Approach for Southeast Asia and Papua New Guinea

Abstract. Mangroves are an important ecosystem across Southeast Asia and Papua New Guinea, and accurate estimates of their occurrence and extent are critical. However, area estimates vary widely across studies, often because they rely on direct pixel counts without stratified sampling or design-based inference, making them sensitive to image and model biases that vary across space and time. This study evaluates whether satellite foundation-model embeddings can provide a more temporally consistent representation for mangrove monitoring while examining spatial transferability. We analyze the spatiotemporal structure of 64-dimensional AlphaEarth Foundations embeddings and apply a gradient-boosted tree classifier to generate annual mangrove probability maps for 2017–2024. These maps are combined with stratified probability sampling and design-based accuracy assessment to derive accuracies and unbiased area estimates with quantified uncertainty. The embeddings show clear mangrove–non-mangrove separability and relative temporal consistency in the embedding space, supporting generally stable regional classification performance (overall accuracy ≈0.80–0.82; AUC ≈0.92–0.93). While performance declines in heterogeneous coastal systems indicate remaining transferability limits, the results demonstrate that foundation-model embeddings provide a temporally consistent and operational basis for large-scale mangrove monitoring with robust, design-based area estimation. More broadly, embeddings offer a compact representation that supports feature-space analysis and more consistent interpretation of changes over time.