Abstract: The rapid and accurate monitoring of soil organic carbon (SOC) is of great significance for evaluating the quality of reclaimed soils in mining areas. With reclaimed soils from the Xinglongzhuang Coal Mine as the object, we constructed five types of inversion variables based on drone multispectral imagery and sample SOC content: band mathematical transformation groups, traditional spectral index groups, improved spectral index groups, full-variable groups, and groups selected by the improved greedy algorithm (IGA). We further built SOC inversion models with six machine learning algorithms, AdaBoost, backpropagation neural network (BPNN), CatBoost, LightGBM, random forest (RF), and XGBoost. The results showed that: 1) When the improved spectral index group was used as the variable, the accuracy of inversion model was higher than that of the traditional spectral index group. 2) When the IGA-selected variable group was used, the accuracy and stability of the model significantly improved. 3) The LightGBM model using the IGA-selected variable group was the optimal SOC inversion model for reclaimed soils, with a modeling set coefficient of determination (R²) of 0.825, root mean square error (RMSE) of 0.914, a validation set R² of 0.793, and RMSE of 0.945. 4) The inverted SOC content ranged from 7.75 to 13.60 g·kg^-1, with an average of 10.48 g·kg^-1, which was consistent with the sample SOC measurements. These fin-dings could provide technical support for the planning and implementation of land reclamation in mining areas.

Key words: soil organic carbon, reclamation, unmanned aerial vehicle, multispectral remote sensing, machine learning, inversion model