大速率、不均匀的地面沉降已经威胁到人类的生产活动，高精度的沉降预测结果对于地质灾害的精准防控具有重要意义。为掌握地面沉降的演化规律，利用现场观测数据或InSAR数据开展了多项预测研究。然而，由于空间异质性的存在，大范围地面沉降的准确预测仍然是一项挑战。在这项研究中，从数据驱动的角度提出了一种顾及空间异质性的大范围地面沉降时空预测方法STLSTM（Spatio-temporal Long Short-Term Memory）。首先，通过聚类识别地理空间中的均质子区；然后，在每个子区中，一个特别的长短期记忆LSTM（Long Short-Term Memory）网络被用来捕捉局部位置的非线性特征；最后，利用预训练的网络对未来时刻的地面沉降进行定量预测。在实验部分，哨兵1号影像数据被用来比较STLSTM与其他8种基准方法的性能，利用空间统计指标分析了模型的有效性。结果表明，STLSTM在152 s内达到了最高的预测精度（71.4%），且能够有效弱化空间异质性对大区域沉降预测任务的影响。总之，这项研究将空间异质性处理策略融合到深度学习模型中，实现了高精度、高时效的大范围地面沉降时空预测。

The rapid and uneven ground subsidence has threatened human production activities, and high-precision subsidence prediction results are of great significance for the precise prevention and control of geological disasters. In order to grasp the evolution law of ground subsidence, a number of prediction studies have been carried out using field observation data or InSAR data. However, due to the existence of spatial heterogeneity, accurate prediction of large-scale ground subsidence is still a challenge.In this study, a spatio-temporal prediction method considering spatial heterogeneity for large-scale ground subsidence STLSTM (Spatio-temporal Long Short-Term Memory) is proposed from a data-driven perspective. First, clustering is used to identify homogenous subregions in geographic space; then, in each subregion, a special Long Short-Term Memory (LSTM) networks are used to capture the nonlinearity features of local locations; Finally, the pre-trained network is used to quantitatively predict the ground subsidence at the future time.In the experimental part, the sentinel-1 image data was used to compare the performance of STLSTM with the other 8 benchmark methods, and the effectiveness of STLSTM was analyzed using spatial statistical indicators. The results show that STLSTM achieves the highest prediction accuracy (71.4%) within 152 secs, and can effectively weaken the effect of spatial heterogeneity on large-scale subsidence prediction tasks.In conclusion, this paper integrates the spatial heterogeneity processing strategy into the deep learning model, and large-scale subsidence prediction is realized with high precision and time efficiency.