Estimating soil organic carbon content in reclaimed mining areas using unmanned aerial vehicle multispectral remote...

Soil organic carbon (SOC) is a key indicator for assessing soil quality. Rapid and accurate acquisition of the spatial distribution of SOC content can guide production management in reclaimed farmland in mining areas. This study takes the Xinglongzhuang Coal Mine in Yanzhou District, Jining City, as the study area, with reclaimed farmland SOC as the research subject. Based on unmanned aerial vehicle (UAV) multispectral remote sensing images and ground sampling point data, three sets of feature variables were constructed: a full variable set, a variance inflation factor (VIF) screened variable set, and a genetic algorithm (GA) screened variable set. Four deep learning algorithms, namely convolutional neural network (CNN), graph neural network (GNN), recurrent neural network (RNN), and transformer, are employed to construct a total of 12 SOC content estimation models. Based on the optimal estimation model CNN-GA, the model was optimized and improved by generating pseudo-reflectance samples through Kernel-SMOTE algorithm oversampling. The results indicate that (1) following variable selection through the use of GA, the accuracy of the majority of estimation models saw a significant improvement. (2) Among the four deep learning algorithms, the model constructed using the CNN algorithm achieved the highest accuracy. (3) CNN-GA-SMOTE was identified as the optimal estimation model, achieving an R² of 0.864 and RMSE of 0.827 g kg^-1 for the modeling set, with R² of 0.857 and RMSE of 0.796 g kg^-1 for the validation set. The SOC content of reclaimed farmland was estimated using this model, with an overall range of 7.53~13.40 g kg^-1. This study provides technical support for soil quality assessment in reclaimed mining areas.