A novel geospatial machine learning approach to quantify non-linear effects of land use/land cover change (LULCC) on carbon dynamics

Land Use and Land Cover Change (LULCC) introduces considerable uncertainties in its relationship with carbon emissions, posing challenges to nature-based climate mitigation. The effective establishment and protection of carbon sinks through land use management lack clarity. To address this challenge, this study utilized top-down interannual grided CO2 data from satellite observations, revealing a 7 % decline in carbon sinks in mainland China from 2016 to 2019. Faced with this anomaly, we proposed a novel approach that combined machine learning with traditional regression analysis, to investigate the nonlinear relationship between annual spatiotemporal variations in net carbon exchange and LULCC. Sentinel-2 imagery was employed for high-resolution (10 m) LULC classification based on uniform rules. Particularly, time-series LULC class probabilities were considered and estimated using a deep learning framework via the Google Earth Engine (GEE) cloud platform, which allows us to access the effects of dynamic LULCC. GIS methods were applied to enhance machine learning interpretability, integrating multi-source remote sensing datasets, particularly for capturing nonlinear features in spatiotemporal aspects of LULCC and carbon emissions. Threshold effects revealed how LULCC transformed areas were associated with carbon sinks or sources. The results mapped carbon sink shrinkage locations, highlighting correlations with significant reductions in snow cover (−6.25 %), changing water patterns (−1.3 %), urban expansion (1 %), and mixed forest changes (regrowth of 4 % and deforestation of −1%). This research aims to advance understanding of carbon emissions through remote sensing, bridging different Earth observation data within a geostatistical context, and expects to provide a new validation method for the bottom-up approach.