Research on transformer-Kolmogorov–Arnold neural network open-pit mine slope deformation prediction model based on temporal interferometric synthetic aperture radar

Abstract

Slope stability is critical to ensuring safe mining operations; yet, traditional monitoring techniques often suffer from insufficient spatial resolution, limited coverage, and inadequate predictive accuracy. To address these limitations, this study proposes a collaborative monitoring and prediction framework that integrates temporal interferometric synthetic aperture radar (InSAR) with deep learning, using an open-pit mine in Guizhou as a case study. Small baseline subset InSAR (SBAS-InSAR) results reveal maximum subsidence rates of − 40 mm·yr⁻¹ in the phosphate quarry and − 70 mm·yr⁻¹ in the waste dump, with cumulative settlements of 106 and 236 mm, respectively. Comparison with ground-based leveling measurements from 22 monitoring benchmarks across two field campaigns yielded a correlation coefficient R of 0.92 between the SBAS-InSAR and leveling datasets. Based on the SBAS-InSAR-derived deformation time series, four predictive models—Long Short-Term Memory (LSTM), Transformer, LSTM–Kolmogorov–Arnold Neural Network (KAN), and Transformer–KAN—were constructed and evaluated. Among them, the Transformer–KAN model exhibited the best performance. On average across the six evaluated feature points, the Transformer–KAN model achieved a mean absolute error (MAE) of 6.12 (± 0.42) mm and a root mean square error (RMSE) of 7.98 (± 0.37) mm. At the best-performing point (Point F), the MAE and RMSE were 5.26 (± 0.27) mm and 6.81 (± 0.33) mm, respectively, representing a 42% improvement over the LSTM model at that location. The KAN module was used as a predictive component in the present study; its interpretability was not quantitatively evaluated. Further validation across additional sites is needed.