Objectives  The complex terrain makes it difficult to construct coal mining subsidence basins and extract horizontal movements in the Loess Plateau, Western China. The traditional surface subsidence monitoring methods are often used to collect data along a given profile, which is inefficient. Repeat-pass interferometric synthetic aperture radar (InSAR) technology is prone to incoherence in areas with large gradient surface displacements, making it difficult to meet the surface subsidence monitoring accuracy requirements in mining areas. We propose a new method to construct mining subsidence basins and extract horizontal movements based on unmanned aerial vehicle light detection and ranging point cloud data.

  Methods  A range of terrain factors are combined to build a deep neural network (DNN) model to extract stable areas with limited topographic effects during the construction of subsidence basins, and the optimal value interpolation algorithm is used to fit the stable areas to obtain the complete subsidence basins. The binary shape context feature operator is then integrated with a variety of terrain factors to improve the feature matching algorithm to extract horizontal movements in the mining subsidence area. Based on this, we design a scheme to extract horizontal movements of the main sections, and analyze the relationships between the horizontal movement extraction errors and point cloud density and terrain factors.

  Results  The results in Yushen Mining area show that the DNN model combined with terrain factors can effectively extract the stable areas with the modeling error of subsidence being reduced even under complex terrain, which provides a new method for the construction of coal mining subsidence basins.

  Conclusions  The horizontal movement curve extracted by the improved feature matching algorithm integrated with terrain features conforms to the basic law of the horizontal motion of coal mining subsidence, and the terrain factors that have strong correlations with horizontal movement deviations can be used as indicators to represent the performance of the improved feature matching algorithm.