Abstract
Climate warming is a major global challenge, and forests, essential carbon sinks, are critical in mitigating its effects. Forest carbon density is a key parameter in assessing the carbon sinks. Traditional estimating methods of forest carbon density are time-consuming, labor-intensive, and difficult to apply on a large scale. Combining multispectral data with machine learning offers a promising solution, but accurately estimating forest carbon density remains challenging due to the band limitations of multi-spectral data. This study proposes a novel approach to address this limitation gap. We utilized Landsat 8 data and 919 samples from Xizang, China, simultaneously constructed geographic (GEO) and environmental factors (GEF) for estimating forest carbon density for the first time, and adopted three models to evaluate the effectiveness. The results indicate that the extreme gradient boosting (XGB) model is significantly better, the average R2 exceeds 0.77, especially in Rikaze exceeds 0.96. The total relative importance of GEF in the modelling exceeded 60%, Geo was the most critical variable, followed by CI. This study successfully used multi-spectral data to quantify the spatiotemporal distribution of forest carbon density and demonstrated that GEO and GEF are indispensable, which is expected to provide new perspectives and technical support for global carbon sink monitoring.