Assessing Potential of Multisource Satellite Data and Machine Learning Models for Cropland Soil Organic Carbon Prediction in Plateau Lake Basin

Accurate spatial quantification of cropland soil organic carbon (SOC) in plateau lake basins is crucial for assessing the carbon sequestration potential in ecologically fragile regions. This study developed a machine learning (ML) framework that integrates multi-source satellite-derived environmental covariates (topography, climate, vegetation, soil properties, and parent materials) to estimate SOC distribution in the Erhai Lake basin. Using 432 topsoil samples (0–20 cm), we systematically compared 15 models, including conventional ML approaches (e.g., random forest, support vector machine, and light gradient boosting machine) and deep learning (DL) models (e.g., long short-term memory, recurrent neural network, and multilayer perceptron). The results showed that DL models achieved higher predictive accuracy than conventional ML models, reducing RMSE by 0.1680 g kg⁻¹ and increasing R², RPIQ, and CCC by averages of 0.0225, 0.1143, and 0.0253, respectively, although conventional ML models exhibited greater robustness. Elevation and temperature were identified as dominant factors controlling SOC spatial patterns, with higher concentrations clustered in the western and northern subbasins. Greater prediction uncertainty in the northwestern and eastern margins was associated with complex terrain heterogeneity. Spatially explicit SOC mapping derived from the integration of multi-source satellite data and ML models offers innovative approaches for carbon management in ecologically fragile lacustrine agroecosystems.