Objective The existing spatiotemporal prediction methods of land subsidence have some problems, such as poor ability to capture time-series features and ignoring spatial neighborhood features, which lead to poor reliability of spatiotemporal prediction of land subsidence.

Method This paper proposes a spatiotemporal prediction of land subsidence based on convolutional long short-term memory (ConvLSTM) neural network, which can capture time-series features and spatial neighborhood features. First, based on small baseline subset interferometric synthetic aperture radar (SBAS-InSAR), Sentinel-1A images were used to obtain the spatiotemporal InSAR data of land subsidence at Beijing Capital International Airport (BCIA). Then, a ConvLSTM model is constructed to predict the land subsidence in the next year. Permanent scatters InSAR, SBAS-InSAR and benchmark results are used to cross-verify the reliability of time-series InSAR results. Time-series InSAR land subsidence data are segmented by sliding windows to form a many-to-one dataset model. Combined with wavelet transform and evaluation indexes to determine the optimal time step of the prediction model, the ConvLSTM spatiotemporal prediction model of time-series InSAR land subsidence is established.

Results The R² of the predicted and real results reaches 0.997. Meanwhile, the performance of the proposed model is further evaluated based on the image evaluation index structural similarity (SSIM) and multi-scale structural similarity (MS-SSIM), which reach 0.914 and 0.975, respectively. In addition, support vector regression, multilayer perceptron, convolutional neural network and the long-term memory neural network model are compared and analyzed, and each index reveales that the proposed model is optimal.

Conclusion The ConvLSTM spatiotemporal prediction model predicts that the maximum cumulative subsidence at BCIA will reach 157 mm by November 2022. This study can provide key technical support for the early prevention of urban land subsidence.