GSDEv2 Soil Dataset Reduces Surface Soil Moisture Error in CoLM2024 Experiment

Researchers have developed a new global soil dataset, designated GSDEv2, designed to serve as a high-resolution input for Earth system models. The dataset maps twenty distinct physical and chemical soil properties across six standard depth intervals on a nominal 90-meter grid. This level of spatial and vertical detail addresses significant gaps in existing global soil products, which often vary widely in property coverage, resolution, and the integration of regional soil information.

The development of GSDEv2 was led by a team of scientists including Gaosong Shi, Anne Gobin, Zhongwang Wei, Hua Yuan, Qingliang Li, Wei Shangguan, and Yongjiu Dai. The dataset is currently available as a preprint under review for the journal Earth System Science Data. The authors describe the framework as a synergy of observation-informed spatial prediction, soil-domain-specific modelling, and a stacking-based incorporation of regional soil products.

A key feature of GSDEv2 is its integration of compatible regional soil products from China, the United States, Africa, Canada, and Australia. These regional products were incorporated as auxiliary predictors through a stacking-based fusion method. By combining these diverse sources, the team aimed to create a vertically harmonized resource that is both spatially detailed and globally consistent.

The evaluation of the dataset’s global baseline performance utilized a profile-level random hold-out design. In this testing framework, 90 percent of the soil profiles were used for training the prediction models, while the remaining 10 percent were reserved for testing. This design characterizes the interpolation performance within the sampled observational domain. The results indicated that GSDEv2 achieved improved average prediction accuracy compared to previous versions and other widely used global soil datasets.

When comparing GSDEv2 against GSDEv1, SoilGrids 2.0, and HWSD 2.0 across 54 property-depth combinations, the new dataset demonstrated superior statistical performance. GSDEv2 showed the highest mean modelling efficiency coefficient and the lowest mean root-mean-square error among the evaluated combinations. Specifically, the median root-mean-square error reductions relative to GSDEv1, SoilGrids 2.0, and HWSD 2.0 were 22.2 percent, 5.8 percent, and 24.6 percent, respectively.

To further assess the value of regional fusion, the team conducted a spatially blocked case study in China focusing on surface soil organic carbon. This nested 60-kilometer spatial-block validation revealed that regional stacking reduced the root-mean-square error by 31.8 percent. Additionally, the mean 90 percent prediction-interval width was reduced by 38.7 percent while maintaining approximately nominal coverage.

The practical utility of GSDEv2 was demonstrated through a downstream sensitivity experiment using the CoLM2024 land-surface model. In this experiment, the GSDEv2-driven simulation reduced the annual-mean spatial root-mean-square error of surface soil moisture from 0.109 to 0.089 cubic meters per cubic meter. This improvement was observed relative to a simulation driven by the SoilGrids dataset. The experiment underscores the importance of high-quality soil data for the accuracy of land-surface simulations.

Furthermore, the GSDEv2-driven simulation increased the spatial correlation with adjusted ERA5-Land data from 0.741 to 0.802.

The dataset provides not only the 90-meter grid predictions but also spatially aggregated 1-kilometer mean and conditional-quantile products.

GSDEv2 is freely available to the scientific community through designated data repositories. The dataset is hosted at the Terre.tpdc and scidb.cn platforms, ensuring open access for researchers worldwide. The core Python processing workflow for GSDEv2 is also available, providing transparency and reproducibility for the methods used to generate the dataset.

The study acknowledges that existing global soil datasets have significant limitations in terms of uncertainty characterization and the integration of regional information.

The preprint status of the article indicates that it is currently undergoing peer review for publication in Earth System Science Data. The authors have provided a supplement and model code to support the reproducibility of their work.

The development of the Global Soil Dataset for Earth System Modeling version 2 represents a collaborative effort across multiple research institutions. Gaosong Shi carries affiliations with the School of Atmospheric Sciences at Sun Yat-sen University in Zhuhai, China, and the Department of Earth and Environmental Sciences at KU Leuven in Belgium. The author team also includes Anne Gobin from KU Leuven, corresponding author Wei Shangguan, Zhongwang Wei, Hua Yuan, and Yongjiu Dai from Sun Yat-sen University, alongside Qingliang Li from the School of Artificial Intelligence at Changchun Normal University. Together, the investigators established the framework to resolve persistent discrepancies across existing global soil datasets, which differ substantially in property coverage, spatial and vertical resolution, uncertainty characterization, and the integration of regional soil information.

To establish a globally consistent baseline while drawing on localized expertise, the system incorporates compatible regional soil products from China, the United States, Africa, Canada, and Australia as auxiliary predictors through stacking-based fusion. The resulting resource produces spatial predictions for twenty physical and chemical soil properties across six standard depth intervals on a nominal 90-meter grid. Alongside these high-resolution outputs, the methodology generates spatially aggregated 1-kilometer products containing both mean values and conditional-quantile distributions. This multi-tiered structure links observation-informed spatial prediction directly with soil-domain-specific modeling, providing modelers with a vertically harmonized foundation designed to accommodate varied spatial requirements and robust uncertainty analyses in regional and planetary environmental applications.

Evaluation of the baseline model relied on a profile-level random hold-out design where ninety percent of profiles served for training and ten percent were reserved for testing, effectively characterizing interpolation performance across the sampled observational domain. Benchmarking across fifty-four property-depth combinations jointly supported by GSDEv2, GSDEv1, SoilGrids 2.0, and HWSD 2.0 indicated that GSDEv2 achieved the highest mean modeling efficiency coefficient and lowest mean root-mean-square error. Compared to GSDEv1, SoilGrids 2.0, and HWSD 2.0, median root-mean-square error reductions reached 22.2 percent, 5.8 percent, and 24.6 percent, respectively. Regional fusion was further appraised through a nested 60-kilometer spatial-block validation for surface soil organic carbon in China, which yielded a 31.8 percent reduction in root-mean-square error and a 38.7 percent reduction in mean 90 percent prediction-interval width while maintaining nominal coverage.

Downstream utility was established in a land-surface sensitivity experiment using the CoLM2024 model, where driving simulations with GSDEv2 reduced the annual-mean spatial root-mean-square error of surface soil moisture from 0.109 to 0.089 cubic meters per cubic meter compared to SoilGrids, while elevating spatial correlation with adjusted ERA5-Land from 0.741 to 0.802. For open verification, the core Python processing workflow developed by Gaosong Shi is archived under Zenodo, and the complete dataset is published across the Terre.tpdc and scidb.cn repositories. The manuscript, submitted on 20 September 2026 and released for discussion on 12 November 2026 under a Creative Commons Attribution 4.0 License, is accompanied by a 9,816-kilobyte supplement alongside the 2,902-kilobyte preprint document during its ongoing review for Earth System Science Data.

Subject of Research: A Global Soil Dataset for Earth System Modeling (Version 2, GSDEv2)

Article Title: A Global Soil Dataset for Earth System Modeling (Version 2, GSDEv2)

Article References: Shi, G., Gobin, A., Wei, Z., Yuan, H., Li, Q., Shangguan, W., & Dai, Y. (2026). A Global Soil Dataset for Earth System Modeling (Version 2, GSDEv2). https://doi.org/10.5194/essd-2026-770

Image Credits: AI Generated

DOI: 10.5194/essd-2026-770

Keywords: Global, Soil, Dataset, Earth, System, Modeling, Version, GSDEv2