Optimizing Soil Properties Mapping: A Comparative Study of UAV and Satellite Imagery With Ensemble Learning Framework

Accurate mapping of multiple soil properties is essential for enhancing soil quality and optimizing agricultural management practices. Given the complex interactions among soil properties, generating distribution maps for multiple soil properties provide richer informationto support agricultural decision-making. This study proposes an effective digital soil mapping (DSM) framework that integrates multisource data [PlanetScope imagery, uncrewedaerial vehicle (UAV) data] with feature selection methods and ensemble learning to predict key soil properties: soil organic carbon (SOC), total nitrogen, pH, available phosphorus, and available potassium. Results demonstrated that data integration markedly improved prediction accuracy. Multitemporal PlanetScope imagery outperformed UAV imagery, improving $\mathbf {R^{2}}$ by around 0.20 and reducing its standard deviation by around 0.10. Integrating pedotransfer functions with DSM techniques enhanced SOC modeling ($\mathbf {R^{2}}$ increased from 0.70 to 0.82), yielding an SOC map that better reflects the carbon–nitrogen ratio and the underlying physical mechanisms. Among feature selection methods, the variable importance method outperformed genetic algorithms and the correlation-based methods, improving $\mathbf {R^{2}}$ by 0.07–0.24. A metamodel based on gradient boosting decision tree (GBDT) was developed to adaptively stack light gradient boosting machine (LightGBM), support vector regression (SVR), and random forest (RF). The LightGBM model achieved a 5% reduction in normalized root-mean-square error compared to RF and SVR. The EM_GBDT model delivered the highest performance, with $\mathbf {R^{2}}$ exceeding 0.75 and concordance correlation coefficient exceeding 0.80 for all soil properties. These results emphasize the capability of the proposed approach for field-scale DSM and provide valuable insights for supporting sustainable soil management and enhancing agricultural production.