The thickness of the black soil horizon in sloping farmland within China's black soil region is primarily affected by various elements, including terrain, climate, and anthropogenic activity. While conventional studies on soil thickness prediction primarily rely on topographic variables and vegetation indices, they often overlook the potential importance of spectral data. This study employs Sentinel-2 optical imagery data from May 2023, during the bare soil phase, in conjunction with topographic features and vegetation indices, to investigate the efficacy of various input variables in predicting soil thickness in sloping farms within the watersheds of the black soil region. The model was trained and validated using 157 sampling points of black soil horizon thickness (BSHT) through three machine learning techniques: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANNs), to assess the influence of various variable combinations on soil thickness prediction. The findings show that models incorporating spectral information (R² ranging from 0.62 to 0.71) have better explanatory power for predicting BSHT than models without (R² ranging from 0.68 to 0.75). While topographic factors were strong predictors, including spectral information significantly enhanced prediction accuracy. The findings indicated that RF exhibited superior prediction accuracy compared to XGBoost and ANNs among the three methodologies. This study's findings yield novel insights for accurately predicting soil thickness on sloping farmland within the black soil region and furnish scientific support for soil conservation and sustainable agricultural growth.