A Global Soil Spectral Library and Soil Organic Carbon Estimation Dataset Based on Geographical Stratification
Description
This dataset comprises 784 topsoil samples selected from the Open Soil Spectral Library (OSSL), including soil organic carbon content, geographic coordinates, and visible–near-infrared spectral data covering 400–2400 nm at a standardized 10 nm interval. The dataset also integrates ten environmental covariates: clay content, silt content, cation exchange capacity, soil pH, elevation, slope, annual mean temperature, annual precipitation, NDVI, and land-cover type. Based on these covariates, the samples were classified into five geographical environmental zones using a Gaussian mixture model. The dataset supports comparisons between global and geographically stratified models for soil organic carbon estimation and facilitates the regional application of global soil spectral libraries.
Files
Steps to reproduce
The dataset was derived from the Open Soil Spectral Library (OSSL). Samples lacking geographic coordinates, soil organic carbon (SOC) measurements, or visible–near-infrared (Vis–NIR) spectra were excluded, after which only topsoil samples were retained. The final dataset contained 784 georeferenced samples, mainly originating from the ISRIC and USDA–NRCS Kellogg Soil Survey Laboratory spectral libraries. Before spectral measurement, soil samples were air-dried, ground, and passed through a 2 mm sieve. Vis–NIR reflectance spectra were measured over 350–2500 nm using ASD FieldSpec spectrometers with white-reference calibration. SOC was determined using the Walkley–Black wet-oxidation method or the LECO high-temperature combustion method, and measurements from different sources were harmonized. To reduce differences among instruments and remove noisy edge wavelengths, all spectra were resampled to 10 nm intervals and restricted to the 400–2400 nm range. Ten environmental covariates were extracted for each sample location, including clay content, silt content, cation exchange capacity, soil pH, elevation, slope, annual mean temperature, annual precipitation, NDVI, and land-cover type. Soil covariates were obtained from SoilGrids at 250 m resolution, whereas topographic, climatic, vegetation, and land-cover variables were prepared at approximately 1 km resolution. Geographical environmental zones were identified using a Gaussian mixture model based on the standardized environmental covariates. Candidate solutions containing two to seven zones were compared using the Akaike information criterion, Bayesian information criterion, negative log-likelihood, and silhouette score, and the optimal solution divided the samples into five zones. A fuzzy C-means classification based on spectral similarity was also generated for comparison. Random forest models were then fitted to the complete dataset and to individual zones to evaluate whether geographical stratification improved SOC prediction. Model performance was assessed using R², RMSE, Lin’s concordance correlation coefficient, and RPIQ.
Institutions
- Xinjiang UniversityXinjiang, Ürümqi
- Chinese Academy of SciencesBeijing, Beijing