Net ecosystem productivity (NEP) serves as a key indicator for assessing regional carbon sink potential, with its dynamics regulated by nonlinear interactions among multiple factors. However, its driving factors and their coupling processes remain insufficiently characterized. This study investigated terrestrial ecosystems in Yunnan Province, China, to elucidate the drivers of NEP using 14 environmental factors (including topography, meteorology, soil texture, and human activities) and 21 remote sensing features. We developed a research framework based on “Feature Selection–Machine Learning–Mechanism Interpretation.” The results demonstrated that the Variable Selection Using Random Forests (VSURF) feature selection method effectively reduced model complexity. The selected features achieved high estimation accuracy across three machine learning models, with the eXtreme Gradient Boosting Regression (XGBR) model performing optimally (R² = 0.94, RMSE = 76.82 gC/(m²·a), MAE = 55.11 gC/(m²·a)). Interpretation analysis using the SHAP (SHapley Additive exPlanations) method revealed the following: (1) The Enhanced Vegetation Index (EVI), soil pH, solar radiation, air temperature, clay content, precipitation, sand content, and vegetation type were the primary drivers of NEP in Yunnan. Notably, EVI’s importance exceeded that of other factors by approximately 3 to 10 times. (2) Significant interactions existed between soil texture and temperature: Under low-temperature conditions (−5 °C to 12.15 °C), moderate clay content (13–25%) combined with high sand content (40–55%) suppressed NEP. Conversely, within the medium to high temperature range (5 °C to 23.79 °C), high clay content (25–40%) coupled with low sand content (25–43%) enhanced NEP. These findings elucidate the complex driving mechanisms of NEP in subtropical ecosystems, confirming the dominant role of EVI in carbon sequestration and revealing nonlinear regulatory patterns in soil–temperature interactions. This study provides not only a robust “Feature Selection–Machine Learning–Mechanism Interpretation” modeling framework for assessing carbon budgets in mountainous regions but also a scientific basis for formulating regional carbon management policies.

Accurate estimation of forest aboveground carbon stock (AGC) is essential for understanding carbon accounting and climate change. In previous studies, the extracted factors, such as spectral textures, vegetation indices, and textural features, were used to estimate the AGC. However, few studies examined how different factors affect estimation accuracy in detail. Meanwhile, there are also many uncertainties in the collection and processing of the field data. To quantify the various uncertainties in the process of AGC estimation, we used the random forest (RF) to establish estimation models based on field data and Sentinel-1/2 images in Shangri-La. The models included the band information model (BIM), the vegetation index model (VIM), the texture information model (TIM), the Sentinel-2 factor model (S-2M), and the Sentinel-1/2 factor model (S-1/2M). Then, uncertainties resulting from the plot scale and estimation models were calculated using error equations. Our goal is to analyze the influence of different factors on AGC estimation and to assess the uncertainty of plot scale and estimation models quantitatively. The results showed that (1) the uncertainty of the measurement was 3.02%, while the error of the monocarbon stock model was the main uncertainty at the plot scale, which was 9.09%; (2) the BIM had the lowest accuracy (R² = 0.551) and the highest total uncertainty (22.29%); by gradually introducing different factors in the process of modeling, the accuracies improved significantly (VIM: R² = 0.688, TIM: R² = 0.715, S-2M: R² = 0.826), and the total uncertainty decreased to some extent (VIM: 14.12%, TIM: 12.56%, S-2M: 10.79%); (3) the S-1/2M with the introduction of Sentinel-1 synthetic aperture radar (SAR) data has the highest accuracy (R² = 0.872) and the lowest total uncertainty (8.43%). The inaccuracy of spectral features is highest, followed by vegetation indices, while textural features have the lowest inaccuracy. Uncertainty in the remote-sensing-based estimation model remains a significant source of uncertainty compared to the plot scale. Even though the uncertainty at the plot scale is relatively small, this error should not be ignored. The uncertainty in the estimation process could be further reduced by improving the precision of the measurement and the fitting of the monocarbon stock estimation model.

Landsat time-series (LTS) archived the multi-temporal hyperspectral images, providing freely accessible and long-term optical data for estimating forest aboveground carbon stock (ACS). Due to LTS carrying noise, there were such issues as bias, outliers and missing values in ACS estimation. Hence, a new filtering method named Terrain-Perceive Spatiotemporal Filtering (TP-STF) was developed to improve the estimation accuracy. In TP-STF, landforms were classified based on the terrain data. A computer recognizable identifier was generated by perceiving each terrain unit. Combining the discriminative criteria with the spatiotemporal information, the TP-STF adaptively selected performant filtering to reconstruct LTS. Then, the random forests regression (RFR) was employed to estimate ACS of Pinus densata in Shangri-La, Yunnan, China. Compared with the other filtering, the TP-STF method's reconstructed LTS had the best modeling accuracy and the highest prediction accuracy, with R² = 0.903, RMSE = 17.049 t/hm², P = 81.080% and rRMSE = 19.691%. The ACS results using TP-STF and RFR were: 6.56 million tons in 1987, 6.44 million tons in 1992, 6.33 million tons in 1997, 6.35 million tons in 2002, 6.72 million tons in 2007, 6.70 million tons in 2012, and 7.04 million tons in 2017. The TP-STF could effectively denoise the LTS images in high-altitude regions, providing a new approach to improve the accuracy of remote sensing-based forest ACS estimation.

In forest resource surveys, using sampling methods to estimate aboveground carbon stock (ACS) can significantly reduce survey costs. This study improves the accuracy of ACS estimation by optimizing the stratified sampling design. The sampling process was divided into two stages: stratification and intra-stratum sampling. For stratification, remote sensing features were used as stratification variables, and a spatial clustering stratification method was introduced. For intra-stratum sampling, a composite method, Spatially Correlated Poisson Disk Sampling (SCPDS), was proposed. Using Random Forest (RF) and the sample points selected by SCPDS, the ACS was estimated and compared with traditional sampling methods for Pinus densata in Shangri-La, Yunnan, China. The results showed that (1) by selecting effective stratification variables (e.g., texture features), the required sample size was reduced by up to 19.35% compared to that of simple random sampling; (2) the Ward clustering method greatly improved stratification heterogeneity; (3) for intra-stratum sampling, the SCPDS method ensured spatial independence within strata, particularly at low sampling rates (1%–5%), where its error was significantly lower than that of other methods, indicating greater stability and improved accuracy; (4) the SCPDS-based model achieved the best fitting accuracy, with R² = 0.886. The total carbon stock of Pinus densata using RF was 7,872,787.5 t, closely matching forest management inventory (FMI) data. Through sampling, even with a relatively small sample size, the representative plots can still accurately reflect ACS estimates that are consistent with those derived from large-scale plot surveys. Thus, the optimized stratified sampling method effectively reduced sampling costs while significantly enhancing the stability and accuracy of the results.

Accurate estimation of forest aboveground carbon stock (AGC) is essential for understanding carbon accounting and climate change. In previous studies, the extracted factors such as spectral textures, vegetation indices, and textural features were used to estimate the AGC. However, few studies examined how different factors affect estimation accuracy in detail. Meanwhile, uncertainties in measurement and the fitting of mono-carbon stock model lack calculation. To quantify the various uncertainties in the process of AGC estimation, we used the Random Forest (RF) to establish estimation models based on field plots and Sentinel-1/2 images in Shangri-La. The models included the Band Information Model (BIM), Vegetation Index Model (VIM), Texture Information Model (TIM), Sentinel-2 factor Model (S-2M), and Sentinel-1/2 factor Model (S-1/2M). Then, uncertainties resulted from the plot scale and estimation models were calculated using error equations. Our goal is to analyze the influence of different factors on AGC estimation and to assess the uncertainty of plot scale and estimation models quantitatively. The results showed that: (1) the uncertainty of measurement was 3.02%, while the error of mono-carbon stock model was the main uncertainty at the plot scale, which was 9.09 %; (2) the BIM has the lowest accuracy (R2 = 0.551) and the highest total uncertainty (22.29%); by gradually introducing different factors in the process of modeling, the accuracies improved significantly (VIM: R2 = 0.688, TIM: R2 = 0.715, S-2M: R2 = 0.826) and the total uncertainty decreased to some extent (VIM: 14.12%, TIM: 12.56%, S-2M: 10.79%); (3) the S-1/2M with the introduction of Sentinel-1 Synthetic Aperture Radar (SAR) data has the highest accuracy (R2 = 0.872) and the lowest total uncertainty (8.43%). The uncertainty of spectral features is highest, following by vegetation indices, while textural features have the lowest uncertainty. Uncertainty in the remote sensing-based estimation model remains a significant source of uncertainty compared with the plot scale. Although texture features of SAR data are effective in reducing uncertainty, but uncertainty is lower at the plot scale. The total estimation uncertainty could be reduced by improving the precision of measurement and the fitting of the mono-carbon stock estimation model.