Prediction of overburden layer thickness based on spatial heterogeneity analysis and machine learning models in hillslope regions

The spatial distribution of overburden layer thickness (OLT) is crucial for landslide susceptibility prediction and slope stability analysis. Due to OLT spatial heterogeneity in hillslope regions, combined with the difficulty and time consumption of OLT sample collection, accurately predicting OLT distribution remains a challenging. To address this, a novel framework has been developed. First, OLT samples are collected through field surveys, remote sensing, and geological drilling. Next, the heterogeneity of OLT’s spatial distribution is analyzed using the probability distribution of OLT samples and their horizontal and vertical distributions. The OLT samples are categorized and the small sample categories are expanded using the synthetic minority over-sampling technique (SMOTE). The slope position is selected as a key conditioning factor. Subsequently, 16 conditioning factors are applied to construct OLT prediction model using the random forest regression algorithm. Weights are assigned to each OLT sample category to balance the uneven distribution of sample sizes. Finally, the Pearson correlation coefficient, mean absolute error (MAE), root mean square error (RMSE), and Lin’s concordance correlation coefficient (Lin’s CCC) are employed to validate the OLT prediction results. The Huangtan town serves as the case study. Results show: (1) heterogeneity analysis, SMOTE-based OLT sample expansion strategy and slope position selection can significantly mitigate the effect of spatial heterogeneity on OLT prediction. (2) The Pearson correlation coefficient, RMSE, MAE and Lin’s CCC values are 0.84, 1.173, 1.378 and 0.804, respectively, indicating excellent prediction performance. This research provides an effective solution for predicting OLT distribution in hillslope regions.