Explainable Machine Learning Insights into Wetland Dynamics and Carbon Storage in the Irtysh River Basin

作者：Kaiyue Luo, Alim Samat, Tim Van de Voorde, Weiguo Jiang, Jilili Abuduwaili · 发表于：Earth Systems and Environment · 年份：2025 · DOI：10.1007/s41748-025-00656-5 · 被引用次数：10 · 研究领域：Land Use and Ecosystem Services、Coastal wetland ecosystem dynamics、Peatlands and Wetlands Ecology

Abstract Wetlands are vital for global carbon storage, yet face significant pressures. This study quantifies wetland landscape pattern changes and their impact on carbon storage in the transboundary Irtysh River Basin (IRB) from 2000 to 2020, identifies key landscape drivers, and projects future carbon storage under distinct scenarios for 2030. We utilized multi-temporal land cover data (GWL_FCS30), landscape metrics (Fragstats), the InVEST model for carbon storage estimation, interpretable machine learning (NGBoost coupled with SHAP analysis) to link landscape patterns to carbon dynamics, sensitivity analysis, and the PLUS model for scenario-based future projections (Natural Scenario - S1, Wetland Protection - S2, Wetland Degradation - S3). From 2000 to 2020, total wetland area increased by 10,417 km², primarily driven by marsh and swamp expansion, resulting in a net carbon storage increase from 2.827 × 10⁸ tC to 2.885 × 10⁸ tC (net gain: 5.8 × 10⁶ tC). Sensitivity analysis revealed high responsiveness (Sensitivity Index = 10.812) of carbon storage to wetland area change. The NGBoost model accurately predicted carbon storage based on landscape metrics (MSE = 0.682, RMSE = 0.8259, MAE = 0.7811). SHAP analysis identified the aggregation index (AI), largest patch index (LPI), and number of patches (NP) as the most critical landscape predictors influencing carbon storage. Future projections for 2030 estimate total carbon storage at 3.229 × 10⁸ tC under S1 (stabilization), increa...