Work

Terrain attributes and seasonal Sentinel-2 covariates for machine learning-based digital mapping of soil organic carbon fractions in a Himalayan watershed

Year: 2026

Type: article

Abstract: Accurate spatial prediction of soil organic carbon (SOC) fractions is crucial for carbon accounting, mitigating climate change and sustainable land management of montane ecosystems. This study used four machine-learning (ML) models viz., Cubist, Random Forest (RF), boosted regression trees (BRT) and weighted k-nearest neighbours (wkNN) as part of a digital soil-mapping (DSM) system in the prediction of SOC, particulate organic carbon (POC), and dissolved organic carbon (DOC) across the Chattergul watershed, Ganderbal district, Kashmir. 157 surface soil samples at a depth of 0–15 cm with three covariate ensembles were studied, the terrain attributes based on SRTM DEM (DEM only), the terrain attributes based on DEM + summer Sentinel-2 spectral indices (DEM + SUM) and DEM + autumn Sentinel-2 indices (DEM + AUT). The values of the SOC were between 1.0% and 4.74% (mean 2.59%, SD 0.96%), and the mean values of POC and DOC were 1,421.21 mg kg − 1 and 20.49 mg kg − 1 respectively. Elevation, topographic position index (TPI), and valley depth were the dominant terrain predictors across all models. Cubist with DEM-only covariates achieved the best SOC prediction (RMSE = 0.73%; rRMSE = 28.2%; R² = 0.58). wkNN produced the most accurate POC predictions across all covariate sets (RMSE = 422 mg kg − 1 ; rRMSE = 29.7%; R 2 = 0.12) and DOC (RMSE = 1.29 mg kg − 1 ; rRMSE = 6.3%; R² = 0.58) across all covariate configurations. Moran’s I tests confirmed no significant spatial autocorrelation of model residuals ( p > 0.05; n = 157). These results demonstrate that terrain covariates alone suffice for SOC fraction mapping in this forest-dominated Himalayan catchment, while multi-temporal satellite imagery complements prediction of soil physico-chemical and biological properties, providing a reproducible and scalable multi-temporal framework for SOC monitoring and carbon-sequestration policy in humid Himalayan catchments.

Source: Scientific Reports

DOI: 10.1038/s41598-026-66839-w