Geographically weighted regression-based carbon stock estimation and spatiotemporal evolution in terrestrial ecosystems within the Ebinur Lake Basin

The carbon stock of terrestrial ecosystems serves as an important indicator for regional ecological product valuation, as well as carbon emission reduction and carbon sink enhancement. Remote sensing-based estimation and spatial inversion of the carbon stock in terrestrial ecosystems can provide significant references for evaluating the carbon sequestration potential of ecosystems and achieving the "dual carbon" goals. Based on the remote sensing variables extracted from the Landsat8 OLI images, combined with single-band data, vegetation indices, texture factors, and topographic factors, this study determined the optimal bandwidth using the Akaike information criterion corrected (AICc) and cross-validation (CV). Then, it constructed a geographically weighted regression (GWR) model using the Gaussian, Bi-square, and Exponential kernel functions to estimate the carbon stock of the Ebinur Lake Basin. Through comparison between the GWR and the multiple linear regression (MLR) models, this study selected the optimal one to estimate the spatial distribution of carbon stock. The results indicate that the GWR model exhibited higher accuracy than the MLR model. The optimal GWR model with CV-determined bandwidth and Exponential kernel function achieved an accuracy improvement of 16.31%~66.69%, significantly reflecting spatial heterogeneity. The carbon storage of the basin in 2023 was estimated to be approximately 426.28×10⁶ t, with aboveground, underground, and soil carbon stocks accounting for 31.83%, 24.33%, and 43.28%, respectively. From 2014 to 2023, the carbon stock showed a decreasing trend and a spatial distribution pattern characterized by low stock encircling the Ebinur Lake and high stock in its periphery. Among all terrestrial ecosystems, grassland witnessed the most significant decline in carbon stock.