Terrestrial Carbon Storage Estimation in Guangdong Province (2000–2021)

Abstract

(1) Terrestrial ecosystems are critical carbon sinks, and the accurate assessment of their carbon storage is vital for understanding global carbon cycles and formulating climate change mitigation strategies. (2) This study integrated vegetation indices, meteorological factors, land use data, soil/vegetation types, field sampling, and a convolutional neural network (CNN) model to estimate the carbon storage of terrestrial ecosystems in Guangdong Province. (3) Total carbon storage increased by 0.11 Pg from 2000 to 2021, with vegetation carbon gains (+0.19 Pg) offsetting soil carbon losses (−0.08 Pg), with the latter primarily being driven by reduced soil carbon in forest ecosystems. (4) Northern and eastern Guangdong exhibit high potential for enhancing carbon storage capacity, which is crucial for achieving regional carbon peaking and neutrality targets.

Dataset: https://doi.org/10.5281/zenodo.15013461

Dataset License: Creative Commons Attribution 4.0 International

1. Summary

This dataset comprises estimates of total terrestrial ecosystem carbon storage, vegetation carbon storage, soil carbon storage, total carbon density, vegetation carbon density, and soil carbon density in Guangdong Province, China. The data were estimated using a convolutional neural network (CNN) model integrating multisource data, including remote sensing data, meteorological data, land use/cover data, vegetation and soil types, and field sampling data (soil sampling data and vegetation field survey data). The sampling campaign was supported by the National Key R&D Program of China [Grant No. 2023YFD1900100]. Public access to this dataset will contribute to advancing regional carbon cycling research and further enhance the accuracy of terrestrial ecosystem carbon storage estimation.

2. Data Description

2.1. Land Use/Cover Data

• Source: GLC_FS30D (Global Land Cover Fine Classification Product, Table 1);
• Resolution: 30 m;
• Temporal Span: 2000–2021;
• Processing Steps: Reclassified into 10 forest, 9 shrub/grassland, and 4 cropland subtypes;
• Access: DOI: https://doi.org/10.12237/casearth.64d094d1819aec27a589a856.

Table 1. Data sources.

Data Type	Temporal Coverage	Spatial Resolution	Source
Field data	2018, 2021	--	Field sampling and surveys
LUC	2000–2021	30 m	GLC_FS30D
VEG	--	1 km	Resource and Environment Science Data Center
SOIL	--	1 km	HWSD2.0
TEMP/PRE	2000–2021	1 km	Resource and Environment Science Data Center
RESI, NPP, NDVI, EVI	2000–2021		MODIS data processed via Google Earth Engine (GEE)
DEM		30 m	ASTER GDEM V3

Table 1. Data sources.

Data Type	Temporal Coverage	Spatial Resolution	Source
Field data	2018, 2021	--	Field sampling and surveys
LUC	2000–2021	30 m	GLC_FS30D
VEG	--	1 km	Resource and Environment Science Data Center
SOIL	--	1 km	HWSD2.0
TEMP/PRE	2000–2021	1 km	Resource and Environment Science Data Center
RESI, NPP, NDVI, EVI	2000–2021		MODIS data processed via Google Earth Engine (GEE)
DEM		30 m	ASTER GDEM V3

2.2. Remote Sensing Indices

• Variables: NDVI, EVI, RESI (Remote Sensing Ecological Index), and NPP (Net Primary Productivity);
• Source: MODIS products (MOD13Q1, MOD17A3H) via GEE (Table 1);
• Resolution: 250 m (NDVI/EVI); 500 m (NPP);
• Processing: Vegetation growing season (April–October) maximum value compositing masked for cloud cover using QA bands;
• Access: NASA Earthdata (requires GEE API access).

2.3. Meteorological Data

• Variables: Mean annual temperature (TEMP) and total annual precipitation (PRE);
• Source: Resource and Environment Science Data Center (Table 1);
• Resolution: 1 km (spatially interpolated from station data);
• Method: Thin-plate spline interpolation with elevation correction;
• Access: DOI:10.12078/2022082501.

2.4. Soil and Vegetation Data

• Soil Type: HWSD2.0 (Harmonized World Soil Database v2.0);
• Vegetation Type: Vegetation Atlas of China;
• Access: FAO HWSD (DOI: https://doi.org/10.4060/cc3823en); Plant Science Data Center (https://doi.org/10.12282/plantdata.0155, accessed on 8 February 2025).

2.5. Field Sampling Data

• Soil Samples: 2316 sites (soil sampling depths were set at 50 cm for croplands and 30 cm for forest/grassland ecosystems, with soil organic carbon content calculated based on the corresponding sampling depths; organic carbon was measured via dry combustion. The sampling was conducted from June to November in both 2018 and 2021 and was supplemented with soil survey data from agricultural land classification and grading to ensure an even distribution of sampling points).
• Vegetation Samples: 1264 sites (carbon storage was quantified through the integrative calculation of stem volume, wood density, and carbon concentration following the methodological protocols outlined in the IPCC Guidelines for National Greenhouse Gas Inventories. Similarly, sampling was also carried out from June to November in 2018 and 2021, with additional forest resource survey data used to achieve an even distribution of sampling points).
• Quality Control: Outliers removed using the ±3σ threshold.
• Spatial representativeness validated via Thiessen polygon analysis.
• Access: Restricted (available upon request for academic use).

2.6. Data Processing Workflow

A three-phase quality assurance protocol was systematically implemented across preprocessing, model training, and postprocessing stages. Data processing workflow is presented in Figure 1.

2.6.1. Preprocessing Quality Assurance

• Spatiotemporal Standardization:

All raster datasets, including LUC, VEG, SOIL, TEMP, PRE, RESI, NPP, NDVI, and EVI, were reprojected to the WGS_1984_UTM_Zone_49N (EPSG:32649) coordinate system using bilinear resampling. The spatial resolution was standardized to 500 m through pixel aggregation, with all outputs exported in GeoTIFF format compliant with ISO 19115-2 [1] geospatial metadata standards.

• Data Sanitization (

  Table 2. Data sanitization.

  Process	Parameters	Outcome
  Outlier removal	±3σ threshold + Moran’s I spatial autocorrelation filter	7.2% samples excluded
  Spatial representativeness	Thiessen polygon analysis	92.3% coverage of ecological zones

  ):

2.6.2. Model Training Verification

• Architectural Regularization:
• Spatial Representativeness Validation:

Voronoi tessellation generated 2357 non-overlapping polygons across a 179,800 km² study area;

Achieved 93% spatial coverage compliance (polygon spacing < 2 × mean NND [8.7 km]).

2.6.3. Postprocessing Uncertainty Quantification

• A triple-component error mitigation framework was developed (

  Table 3. Error propagation analysis.

  Source	Contribution (%)	Mitigation
  Input data noise	38.7	Gaussian process regression imputation
  Model structure bias	29.1	Ensemble learning with 3 CNN variants
  Spatial extrapolation	32.2	Geographically weighted error correction field

  ):

2.7. File Structure

Carbon storage and carbon density dataset:

scd_gd_00p.tif (soil carbon density_GuangDong_2000);

vcd_gd_00p.tif (vegetation carbon density_GuangDong_2000);

tcd_gd_00p.tif (total carbon density_GuangDong_2000);

scs_gd_00p.tif (soil carbon storage_GuangDong_2000);

vcs_gd_00p.tif (vegetation carbon storage_GuangDong_2000);

tcs_gd_00p.tif (total carbon storage_GuangDong_2000).

3. Methods

All variables for 2000–2021 were preprocessed to unify coordinate systems, clip study area boundaries, and assemble a time-series carbon storage factor database.

Field sampling data were filtered using ArcGIS geostatistical tools to remove outliers. A 500 m × 500 m grid was overlaid across the province, with factor values extracted at grid centroids to generate carbon storage base maps.

Convolutional neural networks (CNNs)—a deep learning architecture specialized in grid-structured data processing—automatically extract spatial features through hierarchical learning [2,3,4].