https://orcid.org/0009-0007-9274-4988
Correction
2 Feb 2026: Cheng L, Yang Zl, Xia Yy, Guo Ww, Ren Rq, et al. (2026) Correction: Mapping spatiotemporal distribution of forest carbon density in Xizang, China. PLOS ONE 21(2): e0342180. https://doi.org/10.1371/journal.pone.0342180 View correction
Figures
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.
Citation: Cheng L, Yang Zl, Xia Yy, Guo Ww, Ren Rq, Fang Jp (2025) Mapping spatiotemporal distribution of forest carbon density in Xizang, China. PLoS One 20(9): e0332890. https://doi.org/10.1371/journal.pone.0332890
Editor: Laxmi Kant Sharma, Central University of Rajasthan, INDIA
Received: March 19, 2025; Accepted: September 5, 2025; Published: September 25, 2025
Copyright: © 2025 Cheng et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Data Availability Statement: The data underlying the results presented in the study are available from Google Earth Engine (https://developers.google.cn/earth-engine), which provides access to Landsat 8 satellite data. The data is available under the CC BY 4.0 license. For more information, please visit the Google Earth Engine website: https://developers.google.cn/earth-engine.
Funding: Authors: LC, ZLY, YYX, WWG, RQC, JPF. Project title and grant number: Spatial Patterns of Carbon Sinks in Tibetan Forest Ecosystems and the Factors Affecting Them (XZ202401ZY0090); Xizang University Graduate High-level Talent Cultivation Research Fund Project (2022-GSP-B011). Funding unit: Xizang Science and Technology Agency; Xizang University.
Competing interests: The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Introduction
Climate change has become one of the global community’s most pressing issues [1–4]. As the most significant carbon reservoir in terrestrial ecosystems, forests are irreplaceable in mitigating climate change [5,6]. Trees are the main component of forest ecosystems [7]. They provide a variety of ecological services and can sequester and store carbon, thereby significantly affecting the carbon balance of the ecosystem [8,9]. From 2015 to 2020, forest cover decreased by 10 million hectares per year due to urbanization, human resettlement, and agricultural encroachment [10,11]. Xizang is considered one of China’s most important forest reserves [12,13], and its importance to the carbon cycle is self-evident. Therefore, monitoring the forest carbon density of the Tibetan Plateau is crucial to improving the global forest carbon sink capacity and ensuring a safe and livable environment for human beings in the future.
Traditional methods for estimating forest carbon stocks include directly or indirectly measuring forest biomass and then calculating based on the carbon content in the biomass [14]. The main types of forest carbon stock estimation methods in recent years can be summarized into three categories [15,16]: Inventory-based estimates, satellite-based estimates, and process-based estimates [17]. Inventory-based methods use regional forest inventory data to estimate forest carbon stocks, including forest type, stand age, stand density, stand volume, average tree height, and diameter at breast height [18,19]. Satellite-based methods estimate forest carbon stocks by building relationships between band combinations and stand volumes, using optical remote sensing data and synthetic aperture radar satellite data (SAR) [20–22]. There are two main types of process-based approaches: geostatistical modeling and mechanistic modeling. Geostatistical modeling combines forest inventory data with environmental factors such as topography, elevation, slope, and aspect to construct statistical models to estimate forest biomass at a regional scale [23,24]. Mechanistic modeling quantitatively estimates forest carbon storage through mass-energy exchange or biogeochemical processes [25]. Google Earth Engine (GEE) [26] is an open-access remote sensing tool and cloud database that contains trillions of satellite images of different resolutions captured by multispectral satellites. Therefore, using multispectral satellite data to estimate forest carbon density has the advantages of open access, high temporal resolution, and comprehensive coverage. However, due to limited spectral information, accurately estimating forest carbon density using multispectral data still faces challenges [27].
Factors such as forest age, climate, and soil have significant impacts on forest carbon density [28]. In addition to these factors, topography (such as slope and aspect) and soil properties are also key factors affecting the distribution of forest carbon stocks [29]. Topography and soil properties affect canopy density and forest age indirectly or directly by changing vegetation growth factors, which are the most critical drivers of vegetation carbon storage [30]. Therefore, topography and soil properties may be crucial in revealing carbon storage at the regional scale. In addition, differences in altitude directly affect forest carbon density [31], which is mainly due to the increase in organic matter content and decreases in soil bulk density caused by suitable temperature [32–34], rich biodiversity [35,36], and careful management measures such as fertilization and soil loosening [37,38], which promote the increase of organic matter content and the reduction of soil bulk density, making trees grow more luxuriantly and improving carbon density. Therefore, the growth differences of different species along altitude and latitudinal gradients may provide valuable information for estimating forest carbon density [39]. However, the combination of geographic and environmental factors (GEF, geographic information and environment auxiliary variables) in estimating forest carbon density shows a gap. Therefore, the study of forest carbon density estimation using GEF is critical and urgent.
To fill this gap, this study combines Landsat 8 satellite images with machine learning algorithms to estimate forest carbon density. To this end, we used 919 samples from Xizang, China. Considering the adequate characterization of environmental factors by band information, spectral indices, and environmental auxiliary variables in Landsat 8 data, this study will use these variables and the geographic information of the sample points to characterize GEFs. The research results are expected to provide valuable reference and technical support for global forest carbon density monitoring and enhance the practicality of multispectral data in environmental science applications.
Materials and methods
Study area
As shown in Fig 1, Xizang is located in the southwest of China, with geographical coordinates of 26°50’ to 36°53’ north latitude and 78°25’ to 99°06’ east longitude, and an average altitude of more than 4,000 meters. Affected by topography, landforms, and atmospheric circulation, the climate in the region is characterized by severe cold and dryness in the northwest and warm and humid weather in the southeast. The average annual temperature is about 8°C, with the lowest monthly average temperature dropping to -16°C and the highest monthly average temperature exceeding 16°C. The annual precipitation is between 74.8 and 901.5 mm, with precipitation mainly concentrated from June to September, accounting for 80% to 90% of the annual precipitation. The permanent population is 3.65 million. The forest area is 14.9099 million hectares, and the forest stock is 2.283 billion cubic meters, ranking fifth and first in China respectively.
The figure shows two maps: the base map on the left shows the Xizang provincial boundary and is sourced from Natural Earth (public domain; https://www.naturalearthdata.com); the map on the right shows a digital elevation model (DEM) hillshade, sourced from USGS SRTMGL1 V003 (public domain; https://lpdaac.usgs.gov/products/srtmgl1v003/). Sampling areas are marked in blue, and black dots indicate the sampling points.
Sample acquisition and statistical properties
From May to the end of June 2016, the system collected a total of 919 samples, with a total sampling area of 836,640 square kilometers, and recorded the GPS coordinate information of each sample. The sampling standard is that the measurement is carried out at a height of 1.3 meters from the upslope root neck of the tree trunk. The measurer should stand in the upslope position of the sample tree, ensure that the circumference is perpendicular to the central axis of the trunk, and tighten it before reading the data. All trees’ diameter at breast height is measured using a steel circumference ruler, and the reading is accurate to 0.1 cm. For samples with a breast diameter of more than 60 cm, if it is impossible to measure with a steel circumference ruler, use a tape measure to measure the circumference at breast height and convert it into diameter. Before measuring the ruler, the vines, mosses, and other attachments on the trunk should be removed, and the fallen bark on the coniferous trees should also be removed before measurement. The forest carbon density was then calculated in the laboratory using the data obtained from the sampling. Due to the wide distribution of samples, we divided the study area into five regions (Shannan (28), Rikaze (19), Naqu (21), Linzhi (465), and Changdu (386)) according to the municipal administrative regions.
Landsat 8 multi-spectral data
Due to a malfunction of the scan line corrector of Landsat 7 ETM+, all data after May 31, 2003, are defective. Therefore, this study uses Landsat 8 data (Collection 2, Tier 1, and Level 2) obtained through Google Earth Engine [26], which has undergone geometric correction and atmospheric correction. It covers coastal B1, visible light (blue band (B2), green band (B3), and red band (B4)), near-infrared (B5), and shortwave infrared (B6, B7). The resolution of all bands is 30 m. Due to each region’s vast area, the data date span is relatively large to obtain complete images. However, we controlled each year to be consistent, which makes the years comparable. The specific dates of the images are shown in Table 1.
Estimating models
The nonlinear models include The Extreme Gradient Boosting (XGB) [40] and the Random Forest Regression (RFR) [41] models. The XGB is an ensemble learning technique founded on the principles of gradient-boosting machines. Hyperparameters include max-depth, min-child-weight, gamma, subsample, colsample-bytree, eta, and lambda. Moreover, the RFR, a well-established ensemble learning approach, enhances prediction models’ accuracy and robustness by incorporating multiple decision trees. Hyperparameters include trees, leaf, and fboot.The linear models include the Partial Least Squares Regression (PLSR) [42]; it serves effectively in addressing the linear relationship between the predictor variable X and the response variable Y.
Formulas and explanations
Evaluation metrics for model accuracy.
In evaluating model accuracy, the Coefficient of Determination (R2) is a critical metric for assessing model quality. The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are essential indicators for gauging model error. Consequently, a model demonstrating superiority should manifest high R2 values concurrently with low RMSE and MAE values.
In the equations above: yi represents the actual measured values, denotes the predicted values,
is the mean of all measured values, n is the number of samples.Additionally, STDEV refers to the standard deviation of the actual values in the testing dataset.
Spectral indexes and environmental auxiliary variables.
Forest carbon density is generally believed to be closely related to vegetation growth, and soil quality degradation will directly affect vegetation growth, water retention capacity, and the content of various components in the soil (such as organic matter, soil carbon, and clay). Therefore, this study uses spectral information, indices, and environmental auxiliary variables in Landsat images as a feature candidate set and improves forest carbon density estimation accuracy through feature selection.
Specifically, the spectral indexes employed include the Soil Adjusted Vegetation Index (SAVI), its modified version (Modified Soil Adjusted Vegetation Index (MSAVI)), and the Normalized Difference Soil Organic Carbon (NDSOC) Index [43]. Their formulas are as follows:
In the above formulas, R represents the red band, NIR represents the near-infrared band, SWIRa represents Band 6 of the short-wave infrared, and L is set to 0.5. The calculation of the NDSOC index is determined based on the maximum correlation coefficient between soil organic matter (SOM) and each band.
The environmental auxiliary variables encompass the Normalized Difference Vegetation Index (NDVI), which is associated with vegetation health [44], along with the Enhanced Vegetation Index (EVI) [45]. For indexes indicating surface characteristics like Land Surface Brightness (LSB), Land Surface Greenness (LSG), and Land Surface Wetness (LSW) [20,46,47], the formulas are as follows:
In the calculation of LSB, to fully reflect the surface brightness, we included all non-thermal infrared bands (B1–B7). We calculated LSG using a simplified vegetation indicator, which shows the difference between the near-infrared and green light bands. This difference calculation can effectively reveal the presence of vegetation because the reflectance of vegetation in the near-infrared band is generally higher and lower in the green light band. For the humidity index (LSW), we used the difference between the near-infrared and short-wave infrared bands as an indicator because these two bands are more sensitive to moisture, reflecting the surface humidity conditions. Therefore, NIR is the near-infrared band, B is the blue band, R is the red band, Bandmean represents the average reflectance of the B1–B7 bands, G is the green band, and SWIRa is Band 6 of the short-wave infrared.
Regarding soil properties, the Soil Brightness Index (BI) and the Normalized Difference Moisture Index (NDMI) are selected to evaluate soil erosion and moisture conditions [48,49]; the formulas are as follows:
Where NIR is the near-infrared band, R is the red band, and SWIR is the average reflectance of the B6 and B7 short-wave infrared bands.
Additionally, to monitor soil salinization, the Soil Salinity Index (SI) and the Soil Salinity Remote Sensing Monitoring Index (SDI) are utilized, effectively reflecting soil salinity [43,50,51]; the formulas are as follows:
Where B is the blue band, R is the red band, and NIR is the near-infrared band.
Finally, for assessing soil parameters, the Clay Index (CI) is chosen [52,53]. These indexes, together with the original spectral data, form a comprehensive feature set for modelling; the formula is as follows:
Where SWIRa and SWIRb represent B6 and B7 of the short-wave infrared, respectively.
Explanations for normalization of metrics.
After standardization, the maximum value in the raw data is mapped to value b, while the minimum value is mapped to value a. The standardization formula is as follows:
Where represents the normalized value, x is the original data value, and
and
are the minimum and maximum values in the dataset, respectively. This formula ensures that the data are redistributed within the interval [a,b], facilitating more uniform data processing for subsequent analytical procedures. This means that the influence of different scales is effectively eliminated, enhancing the rationality and efficacy of data comparisons.
In this formula, the maximum value of the dataset is mapped to a, and the minimum value is mapped to b. Here, denotes the normalized value, x represents the original data value,
is the maximum value in the dataset, and
is the minimum value. In this study, a is set to 0.2, and b is set to 1.
The calculation processes in this study were carried out in MATLAB R2022b and Python 3.9 open-source libraries.
Methods of feature selection
Given the high variability and potential nonlinear characteristics of forest carbon density distribution, we chose nonlinear Spearman correlation analysis [54–56] (accessed on 13 Aug 2024) instead of Pearson correlation analysis. For feature selection, we will comprehensively consider the correlation of each feature in the five regions, sort the features in descending order, and finally select the first few features with the highest correlation.
The experiment used three models to establish carbon density estimation models and subsequent evaluation and analysis to prevent the randomness of experimental results, using 5-fold cross-validation. Divide all datasets into five parts, taking one part as the test set and the other as the training set without repetition. Then, take the average of the metric values from the validation sets of the model as the final evaluation metric. Model performance evaluation adopts three metrics: R2, RMSE, and MAE. Evaluate the estimation performance of the model based on larger R2, as well as smaller RMSE and MAE.
All algorithms and models in this study were implemented using Python 3.8 and MATLAB R2022b.
Results
Sample statistical properties
We conducted a statistical analysis of the forest carbon density of samples in each region Fig 2. The analysis results showed that the coefficient of variation of samples in each region was more significant than 1, showing spatial heterogeneity and the data distribution deviated significantly from the normal distribution.
This figure’s horizontal axis represents the Forest carbon density content, and the vertical axis represents the number of samples. The histogram displays the number of samples within different Forest carbon density content ranges. The red dashed line marks the average Forest carbon density content in the samples, the green dashed line represents the range of SD near the mean value, and the blue line describes the distribution characteristics of Forest carbon density content in the samples. Additionally, the legend includes further statistical information about the Forest carbon density content in the samples.
Results of feature selection
The correlation analysis results in Fig 3 show differences in the correlation between spectral index, band information, environmental auxiliary variables, and forest carbon density in each region. Given the previous method description, we comprehensively considered the feature correlations of the five regions, sorted the features in descending order, and finally selected the top six most relevant features. These features include Band3, Band4, Band7, NDSOC, LSB, NDMI, and CI. In addition, we are incorporating geographic information (Geo) as an additional standard feature.
Subfigures (a–e) show the correlation matrices for five regions: (a) Shannan, (b) Rikaze, (c) Naqu, (d) Linzhi, and (e) Changdu. The direction of the ellipse indicates the sign of the relationship, and its width and opacity reflect the mag-nitude of the relationship. Significance levels are denoted as ** for p<0.01, * for p<0.05, and × for p>0.05.
Regarding the production of the spatiotemporal distribution map of forest carbon density content, we used ArcGIS Pro software to classify the prepared Landsat 8 data, extract the pixel information of vegetation in the image, and finally use the optimal model for each region to estimate the forest carbon density content in each region.
Modelling results
Table 3 displays the metric values of each model under various methods. Fig 4 compares the accuracy of each model on the test set (Formulas 17 and 18).
Subfigures illustrate: (a) Changdu, (b) Linzhi, (c) Naqu, (d) Rikaze, and (e) Shannan. In this figure, R2 is positively normalized (Eq 17), while RMSE and MAE undergo inverse normalization (Eq 18). The distance from the center reflects model accuracy for each metric, facilitating comparison across models and regions.
The following conclusions can be drawn from Table 3 and Fig 4: (a) The nonlinear model occupies a dominant position in the prediction accuracy of each region, among which the performance of the XGB model is generally better than that of RFR and PLSR. However, in the Naqu region, RFR outperforms XGB, which may be related to the region-specific data characteristics. (b) Due to differences in samples from different regions, the model’s pre-diction errors in Naqu and Rikaze are relatively small, which shows that the data charac-teristics in these regions are more adaptable to the model, making the prediction results more accurate. (c) The model accuracy in the Rikaze and Shannan regions has the best performance. The R2 value of the Rikaze region is as high as 0.96, which is significantly better than other regions, indicating that the model fitting effect in this region is particularly excellent.
Specifically, nonlinear models such as XGB and RFR have outstanding estimation performance and can effectively capture the complex nonlinear relationship between car-bon density and spectral reflectance. In comparison, most linear models have limited esti-mation accuracy, which can be explained by the advantages of nonlinear models in handling complex relationships. In-depth analysis shows that incorporating GEF and multispectral data as environmental auxiliary variables into feature inputs can significantly improve the accuracy of forest carbon density estimation. This shows that by fusing these environmental variables and geographical information, the model can effectively explain the complex nonlinear relationship between GEF and forest carbon density, thereby achieving high ac-curacy in estimating forest carbon density.
Response and importance of each feature
Fig 5 shows the importance of different features in the optimal model for carbon density estimation in each region. Among them, Geo is identified as the most indispensable feature, always occupying the top two positions, especially in the Naqu region, where the importance of Geo accounts for more than 50%—followed by CI, ranking among the top three in all regions except Linzhi. The pie chart in the figure details the relative importance of GEF and Spectrum. Among them, GEF accounts for more than 60% of the overall importance, especially in the Shannan region, where the proportion of GEF exceeds 75%, highlighting the synergistic relationship between GEF and forest carbon density. This result shows that Geo features have significant effectiveness in carbon density estimation and are an indispensable key factor. At the same time, the relative importance of CI exceeds 10% on average across regions, which may be related to the direct impact of vegetation growth on forest carbon density. CI represents the clay content in the soil, and clay contributes to the absorption and assimilation of plant nutrients, resulting in a clear correlation between carbon density and CI, thereby improving the model’s learning ability. These findings further validate the critical role of environmental auxiliary variables constructed from Geo and Landsat data in carbon density estimation and highlight their critical impact and effectiveness on forest carbon density prediction.
Including (a) the XGB model of Changdu, (b) the XGB model of Linzhi, (c) the RFR model of Naqu, (d) the XGB model of Rikaze, and (e) XGB model of Shannan. In the figure presented, the column depicts the relative importance of individual features. At the same time, the pie chart illustrates the contribution rate (proportions of the relative importance) across various categories of features.
Dynamic spatial distribution of forest carbon density from 2016 to 2022
According to Fig 6, the forest coverage rate in various regions of Xizang is generally high, except in the Naqu region. Between 2016 and 2022, the carbon density of each region changed slightly. It is worth noting that the Linzhi and Shannan regions have the highest carbon densities, located in their respective middle and lower right parts. In addition, Shannan and Linzhi are adjacent to each other. Specifically, the right side of Shannan is Linzhi. Therefore, the higher carbon density in these two regions can be explained. The Naqu region not only has the lowest forest coverage, but its carbon density is also significantly lower usually less than 1.5 t/ha. These analysis results highlight critical areas in Xizang that require forest management and can provide valuable reference information for relevant government departments.
Each submap shows the spatial distribution of forest carbon density for a specific year and region. Due to uncontrollable factors such as atmospheric conditions and cloud cover, the image coverage in some areas may slightly differ, such as in Changdu (2019), and Rikaze and Shannan (2022). Base map and provincial boundaries are from Natural Earth (public domain; https://www.naturalearthdata.com). Remote sensing imagery was acquired from Landsat OLI/TIRS data (public domain; https://landsat.visibleearth.nasa.gov/) for each corresponding year.
Discussion
Reasons for the response of the environmental factors to Forest carbon density
Forest carbon density is affected by a complex interaction of multiple environmental factors, which together determine the forest’s carbon storage capacity. First, climatic factors such as temperature, precipitation, and humidity [57] directly affect plant growth. Suitable temperatures and sufficient precipitation often significantly promote photosynthesis of plants [58], thereby enhancing carbon fixation and storage capabilities. Temperature changes affect plant physiological activities [59], while changes in precipitation and humidity directly affect soil moisture supply [60]. The combined effect of these factors determines the carbon storage potential of forests. Second, soil properties also play an essential role in forest carbon density. The texture, fertility, and pH of the soil directly affect the growing environment of plants. Fertile soil can provide abundant nutrients to support healthy plant growth [61], thereby effectively increasing the carbon storage capacity of forests. At the same time, soil texture and pH significantly affect carbon accumulation and release [62], and the combined effect of these factors determines the overall carbon density of the forest. Finally, the impact of different vegetation types on carbon density is also apparent [63]. Biodiverse forests often exhibit higher carbon densities [64]. These forests possess dense canopies and enhance carbon sequestration and storage through complex ecosystem interactions [65]. Therefore, the diversity and coverage of vegetation types play a crucial role in carbon storage. In summary, the interaction between climatic conditions, soil characteristics, and vegetation types has a crucial impact on forest carbon density. This is why environmental auxiliary variables such as CI, NDMI, and NDSOC play a crucial role in this study. Therefore, understanding these related environmental factors will help promote the study of global carbon density with multispectral images and play an essential role in the balance of ecological cycles.
Limitations and caveats
Although this study makes significant progress in forest carbon density monitoring, some limitations still deserve further discussion. First, the general applicability of the model has not been thoroughly verified. The data used in the current study mainly come from Xizang, China, which makes us have some doubts about the effectiveness of this method in other geographical areas. Therefore, future research should focus on testing and validating the adaptability and robustness of this method in different geographical landscapes to ensure its widespread application on a global scale. Second, while incorporating GEF data does provide a valid estimate of carbon density, the availability and quality of this data can vary significantly across regions. This data inhomogeneity may limit the method’s effectiveness in regions where spectral data are scarce. Future work needs to consider how to optimize this approach in areas where data are scarce or explore how to supplement and enhance data sources to improve the accuracy and reliability of estimates. Finally, although the model has integrated multiple variables in its estimates, it has not taken into account some unmeasured factors that may affect the results, such as microclimate conditions and specific land use practices. These factors may have a significant impact on estimates of forest carbon density. Future research should consider incorporating these unmeasured factors into models to improve the accuracy and comprehensiveness of estimates further. In summary, although this study provides a powerful tool, its applicability, data dependence, and potential missing factors need to be overcome and optimized in future research.
Implications for future research and applications
This study successfully developed a novel and highly versatile forest carbon density estimation method that combines GEF data with advanced machine learning algorithms. Through the application of this method, researchers and practitioners can estimate forest carbon density with optimal accuracy, providing crucial technical support for forest management and environmental protection. In addition, the successful implementation of this method has opened up new avenues for research and practical applications, especially in addressing human sustainability challenges. Specifically, similar techniques can be extended to other critical areas, such as soil carbon estimation, to further promote a deeper understanding and protection of the environment. The open-source tool developed based on this method will facilitate researchers and practitioners around the world to apply this method more conveniently for related research. The promotion and application of this tool will improve the research level in environmental protection and sustainability and strengthen international cooperation to jointly address global environmental challenges. Through this innovative technology and tool, we are expected to achieve more significant results in promoting scientific research and practical applications.
Conclusions
Our findings from this study can be summarized as follows:
Effectiveness of nonlinear techniques: This study covers an area of 836,640 square kilometers. Based on multispectral image data, two nonlinear models (XGB and RFR) and one linear model (PLSR) were used to estimate forest carbon density by comprehensively considering geographic information, environmental variables, and spectral information. The results show that the XGB model performs significantly better than other models, although it is slightly less effective in Naqu. Moreover, especially in Rikaze, the R2 of the model exceeds 0.96, while the comprehensive R2 across regions exceeds 0.77. This result, combined with the vast study area, proves the importance of nonlinear models and shows that forest carbon density has a highly nonlinear relationship with geographic information, environmental variables, and spectral information, confirming the effectiveness of this study’s experimental design.
Effectiveness and importance of GEF: Geo is the most critical factor, consistently ranking in the top two, especially in the Naqu area, where its importance exceeds 50%. Second, CI ranks in the top three regions except Linzhi. Overall, the relative importance of GEF exceeds 60%, and in the Shannan region, it exceeds 75%, emphasizing the close relationship between GEF and forest carbon density. The relative importance of CIs averages over 10% across regions. These results verify the critical role of environmental variables constructed from Geo and Landsat data in carbon density estimates, highlighting their significant impact on prediction models.
Changes in forest carbon density in Xizang: Most areas in Xizang have high forest coverage, except Naqu. Between 2016 and 2022, the carbon density of each region did not change much. Linzhi and Shannan had the highest carbon density, while Naqu had the lowest forest coverage and carbon density below 1.5 t/ha. This highlights the critical areas for forest management in Xizang and provides a reference for relevant government departments to formulate policies.
Our study successfully estimated the spatial distribution of carbon density over an extensive range of years based on multispectral data and proposed an effective carbon density estimation method (incorporating GEF information). At the same time, we determined the optimal model (XGB) and verified that the combination of the two has the best effect. This method effectively expands the application scope of multispectral data. It strongly supports national policy formulation related to carbon sinks and critical technical support for the global forest carbon sink monitoring challenge.
References
- 1. Adnan M, Xiao B, Bibi S, Xiao P, Zhao P, Wang H. Addressing current climate issues in Pakistan: An opportunity for a sustainable future. Environ Challenges. 2024;15:100887.
- 2. Guedes WP, Branchi BA, Bicudo da Silva RF. Climate-induced migration: The need to address human rights to water and sanitation for a growing displaced population. Environ Sci Policy. 2024;158:103799.
- 3. Dubois T, Hadi BAR, Vermeulen S, Ballantyne P, Dobermann A, Fan S, et al. Climate change and plant health: Impact, implications and the role of research for mitigation and adaptation. Global Food Secur. 2024;41:100750.
- 4. Hayashi N, Delaney AE. Climate change, community well-being, and consumption: Reconsidering human-environment relationships in Greenland under global change. Polar Sci. 2024;41:101102.
- 5. Litton CM, Raich JW, Ryan MG. Carbon allocation in forest ecosystems. Global Change Biol. 2007;13(10):2089–109.
- 6. Pan Y, Birdsey RA, Fang J, Houghton R, Kauppi PE, Kurz WA, et al. A large and persistent carbon sink in the world’s forests. Science. 2011;333(6045):988–93. pmid:21764754
- 7. Köbel M, Príncipe A, Soares C, Pinho P, Nunes A, Branquinho C. More than trees: Stand management can be used to improve ecosystem diversity, structure and functioning 20 years after forest restoration in drylands. Sci Total Environ. 2023;902:166107.
- 8. Ibrahim MM, Lin Y, Guo Z, Guo C, Rao X, Liu S, et al. Increasing tree diversity reduces spatial heterogeneity of soil organic carbon and promotes carbon storage in subtropical forests. Agric Ecosyst Environ. 2024;371:109077.
- 9. Wang S, Roland B, Adhikari K, Zhuang Q, Jin X, Han C, et al. Spatial-temporal variations and driving factors of soil organic carbon in forest ecosystems of Northeast China. Forest Ecosyst. 2023;10:100101.
- 10. Rajapakshe R, Karthigesu J, Thavananthan S, Sivachandiran S, Navaneetham V, Sinnamani S. Restoring a dry tropical forest through assisted natural regeneration: Enhancing tree diversity, structure, and carbon stock. Trees Forests People. 2024;17:100616.
- 11.
Kaviani Rad A, Zarei M, Pourghasemi HR, Tiefenbacher JP. The COVID-19 crisis and its consequences for global warming and climate change. Computers in Earth and Environmental Sciences. Elsevier; 2022. p. 377–85. https://doi.org/10.1016/b978-0-323-89861-4.00006-3
- 12. Kang Y, Chen C. Case investigation on the construction mode of rural landscapes in the Qinghai Tibet Plateau, Mengda National Nature Reserve. Heliyon. 2024;10(17):e37033. pmid:39296138
- 13. Su X, Shen Y, Zhou W, Liu Y, Cheng H, Liu G. Mapping forest disturbance and recovery for ecological security improvement on the Qinghai-Tibet Plateau: A case study from Three Parallel Rivers Region. Global Ecol and Conserv. 2022;33:e01983.
- 14.
Li HK, Lei YC. Estimation and evaluation of forest biomass carbon storage in China. Beijing: Chinese Forestry Press; 2010.
- 15. Sun W, Liu X. Review on carbon storage estimation of forest ecosystem and applications in China. For Ecosyst. 2019;7(1).
- 16. Sun T, Qi J, Huang H. Discovering forest height changes based on spaceborne lidar data of ICESat-1 in 2005 and ICESat-2 in 2019: A case study in the Beijing-Tianjin-Hebei region of China. For Ecosyst. 2020;7(1).
- 17. Piao S, Fang J, Ciais P, Peylin P, Huang Y, Sitch S, et al. The carbon balance of terrestrial ecosystems in China. Nature. 2009;458(7241):1009–13. pmid:19396142
- 18. Tang X, Zhao X, Bai Y, Tang Z, Wang W, Zhao Y, et al. Carbon pools in China’s terrestrial ecosystems: New estimates based on an intensive field survey. Proc Natl Acad Sci U S A. 2018;115(16):4021–6. pmid:29666314
- 19. Zhang L, Lai G, Zeng W, Zou W, Yi S. Effect of climate on carbon storage growth models for three major coniferous plantations in china based on national forest inventory data. Forests. 2022;13(6):882.
- 20. Zhang X, Jia W, Sun Y, Wang F, Miu Y. Simulation of spatial and temporal distribution of forest carbon stocks in long time series—Based on remote sensing and deep learning. Forests. 2023;14(3):483.
- 21. Vatandaşlar C, Abdikan S. Carbon stock estimation by dual-polarized synthetic aperture radar (SAR) and forest inventory data in a Mediterranean forest landscape. J For Res. 2021;33(3):827–38.
- 22. Silveira EMO, Radeloff VC, Martinuzzi S, Martinez Pastur GJ, Bono J, Politi N, et al. Nationwide native forest structure maps for Argentina based on forest inventory data, SAR Sentinel-1 and vegetation metrics from Sentinel-2 imagery. Remote Sens Environ. 2023;285:113391.
- 23. Destan S, Yilmaz O, Sahin A. Making objective forest stand maps of mixed managed forest with spatial interpolation and multi-criteria decision analysis. iForest. 2013;6(5):268–77.
- 24. Dai W, Fu W, Jiang P, Zhao K, Li Y, Tao J. Spatial pattern of carbon stocks in forest ecosystems of a typical subtropical region of southeastern China. For Ecol Manag. 2018;409:288–97.
- 25. Wanlong S, Yowhan S, Baishuo H, Xuehua L. An individual tree-based model for estimating regional and temporal carbon storage of Abies chensiensis forest ecosystem in the Qinling Mountains, China. Ecol Model. 2023;479:110305.
- 26. Gorelick N, Hancher M, Dixon M, Ilyushchenko S, Thau D, Moore R. Google earth engine: Planetary-scale geospatial analysis for everyone. Remote Sens Environ. 2017;202:18–27.
- 27. Guo L, Sun X, Fu P, Shi T, Dang L, Chen Y, et al. Mapping soil organic carbon stock by hyperspectral and time-series multispectral remote sensing images in low-relief agricultural areas. Geoderma. 2021;398:115118.
- 28. Guo Y, Peng C, Trancoso R, Zhu Q, Zhou X. Stand carbon density drivers and changes under future climate scenarios across global forests. For Ecol Manag. 2019;449:117463.
- 29. Li D, Ning Z, Chen G, Li Y, Cui B, Wang Q, et al. The effect of land use and land cover on soil carbon storage in the Yellow River Delta, China: Implications for wetland restoration and adaptive management. J Environ Manage. 2024;367:122097. pmid:39102784
- 30. Xu L, Shi Y, Fang H, Zhou G, Xu X, Zhou Y, et al. Vegetation carbon stocks driven by canopy density and forest age in subtropical forest ecosystems. Sci Total Environ. 2018;631–632:619–26. pmid:29533798
- 31. Sasanifar S, Alijanpour A, Shafiei AB, Rad JE, Molaei M, Álvarez-Álvarez P. Understanding how forest ecosystem services are affected by conservation practices and differences in elevation: A study in the Arasbaran biosphere reserve, Iran. Ecol Eng. 2024;203:107230.
- 32. Li J, Ding J, Yang S, Zhao L, Li J, Huo H, et al. Depth-dependent driver of global soil carbon turnover times. Soil Biol Biochem. 2023;185:109149.
- 33. Wang M, Guo X, Zhang S, Xiao L, Mishra U, Yang Y, et al. Global soil profiles indicate depth-dependent soil carbon losses under a warmer climate. Nat Commun. 2022;13(1):5514. pmid:36127349
- 34. Saimun MdSR, Karim MdR, Sultana F, Arfin-Khan MAS. Multiple drivers of tree and soil carbon stock in the tropical forest ecosystems of Bangladesh. Trees For People. 2021;5:100108.
- 35. Ralhan D, Rodrigo R, Keith H, Stegehuis AI, Pavlin J, Jiang Y, et al. Tree structure and diversity shape the biomass of primary temperate mountain forests. For Ecosyst. 2024;11:100215.
- 36. Wu A, Zhou G, He H, Hautier Y, Tang X, Liu J, et al. Tree diversity depending on environmental gradients promotes biomass stability via species asynchrony in China’s forest ecosystems. Ecol Indicat. 2022;140:109021.
- 37. Grüneberg E, Schöning I, Hessenmöller D, Schulze E-D, Weisser WW. Organic layer and clay content control soil organic carbon stocks in density fractions of differently managed German beech forests. For Ecol Manag. 2013;303:1–10.
- 38. Toriyama J, Hak M, Imaya A, Hirai K, Kiyono Y. Effects of forest type and environmental factors on the soil organic carbon pool and its density fractions in a seasonally dry tropical forest. For Ecol Manag. 2015;335:147–55.
- 39. Singh G, Chauhan V, Thakur CL, Verma ML, Bishist R, Prakash P, et al. Studies on natural regeneration, floristic composition, biomass, carbon density and soil properties along an altitudinal gradient in the north-western Himalayas, India. For Ecol Manag. 2023;548:121391.
- 40.
Chen T, Guestrin C. XGBoost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining; 2016. p. 785–94. https://doi.org/10.1145/2939672.2939785
- 41. Breiman L. Random forests. Mach Learn. 2001;45(1):5–32.
- 42. Wold S, Ruhe A, Wold H, Dunn WJ III. The collinearity problem in linear regression. The partial least squares (PLS) approach to generalized inverses. SIAM J Sci and Stat Comput. 1984;5(3):735–43.
- 43. Zhang S, Tian J, Lu X, Tian Q. Temporal, spatial dynamics distribution of organic carbon content of surface soil in coastal wetlands of Yancheng and China from 2000 to 2022 based on Landsat images. CATENA. 2023;223:106961.
- 44. Fensholt R, Proud SR. Evaluation of Earth Observation based global long term vegetation trends—Comparing GIMMS and MODIS global NDVI time series. Remote Sens Environ. 2012;119:131–47.
- 45. Kunkel VR, Wells T, Hancock GR. Modelling soil organic carbon using vegetation indices across large catchments in eastern Australia. Sci Total Environ. 2022;817:152690. pmid:34974006
- 46. Hamzehpour N, Shafizadeh-Moghadam H, Valavi R. Exploring the driving forces and digital mapping of soil organic carbon using remote sensing and soil texture. CATENA. 2019;182:104141.
- 47. Healey S, Cohen W, zhiqiang Y, Krankina O. Comparison of tasseled cap-based landsat data structures for use in forest disturbance detection. Remote Sens Environ. 2005;97(3):301–10.
- 48. Vieira AS, do Valle Junior RF, Rodrigues VS, da Silva Quinaia TL, Mendes RG, Valera CA, et al. Estimating water erosion from the brightness index of orbital images: A framework for the prognosis of degraded pastures. Sci Total Environ. 2021;776:146019. pmid:33652307
- 49. Sedaghat A, Shahrestani MS, Noroozi AA, Fallah Nosratabad A, Bayat H. Developing pedotransfer functions using Sentinel-2 satellite spectral indices and machine learning for estimating the surface soil moisture. J Hydrol. 2022;606:127423.
- 50. Wang Z, Zhang F, Zhang X, Chan NW, Kung H-T, Ariken M, et al. Regional suitability prediction of soil salinization based on remote-sensing derivatives and optimal spectral index. Sci Total Environ. 2021;775:145807. pmid:33618298
- 51. Yang J, Zhao J, Zhu G, Wang Y, Ma X, Wang J, et al. Soil salinization in the oasis areas of downstream inland rivers—Case study: Minqin oasis. Quarter Int. 2020;537:69–78.
- 52. Sabetizade M, Gorji M, Roudier P, Zolfaghari AA, Keshavarzi A. Combination of MIR spectroscopy and environmental covariates to predict soil organic carbon in a semi-arid region. CATENA. 2021;196:104844.
- 53. Taghizadeh-Mehrjardi R, Nabiollahi K, Kerry R. Digital mapping of soil organic carbon at multiple depths using different data mining techniques in Baneh region, Iran. Geoderma. 2016;266:98–110.
- 54. Xu Y, Hao Z, Li Y, Li H, Wang L, Zang Z, et al. Distribution of selenium and zinc in soil-crop system and their relationship with environmental factors. Chemosphere. 2020;242:125289. pmid:31896204
- 55. Radeloff VC, Dubinin M, Coops NC, Allen AM, Brooks TM, Clayton MK, et al. The dynamic habitat indices (DHIs) from MODIS and global biodiversity. Remote Sens Environ. 2019;222:204–14.
- 56. Akkajit P, Tongcumpou C. Fractionation of metals in cadmium contaminated soil: Relation and effect on bioavailable cadmium. Geoderma. 2010;156(3–4):126–32.
- 57. Wei X, Chen D, Hou P, Hu X, Qiu L, Zhang J, et al. Daily trunk radial growth patterns in relation to precipitation in orange trees in the dry-hot valley region of Southwest China. J Hydrol. 2024;641:131856.
- 58. Wu Z, Dijkstra P, Koch GW, Peñuelas J, Hungate BA. Responses of terrestrial ecosystems to temperature and precipitation change: A meta-analysis of experimental manipulation. Global Change Biology. 2011;17(2):927–42.
- 59. Lahti M, Aphalo PJ, Finér L, Ryyppö A, Lehto T, Mannerkoski H. Effects of soil temperature on shoot and root growth and nutrient uptake of 5-year-old Norway spruce seedlings. Tree Physiol. 2005;25(1):115–22. pmid:15519993
- 60. Xu S, Wang J, Sayer EJ, Lam SK, Lai DYF. Precipitation change affects forest soil carbon inputs and pools: A global meta-analysis. Sci Total Environ. 2024;908:168171. pmid:37923258
- 61. Vicca S, Luyssaert S, Peñuelas J, Campioli M, Chapin FS 3rd, Ciais P, et al. Fertile forests produce biomass more efficiently. Ecol Lett. 2012;15(6):520–6. pmid:22472207
- 62. Zhao Z, Dong P, Fu B, Wu D, Zhao Z. Soil organic carbon distribution and factors affecting carbon accumulation in natural and plantation forests in tropical China. Ecol Indic. 2023;148:110127.
- 63. Guo M, Yang L, Zhang L, Shen F, Meadows ME, Zhou C. Hydrology, vegetation, and soil properties as key drivers of soil organic carbon in coastal wetlands: A high-resolution study. Environ Sci Ecotechnol. 2024;23:100482. pmid:39318543
- 64. Poorter L, van der Sande MT, Thompson J, Arets EJMM, Alarcón A, Álvarez-Sánchez J, et al. Diversity enhances carbon storage in tropical forests. Global Ecol Biogeogr. 2015;24(11):1314–28.
- 65. Bastin J-F, Finegold Y, Garcia C, Mollicone D, Rezende M, Routh D, et al. The global tree restoration potential. Science. 2019;365(6448):76–9. pmid:31273120