Fig 1.
MRF-GCN consists of five parts: Graph generation module, Encoding module, GCN module, Training module, and Prediction module.
Fig 2.
Study area and point layout map.
(a) is the InSAR-based control network, (b) is the corresponding adjacency matrix; (c) is the manual-based control network, and (d) is the corresponding adjacency matrix. InSAR images reprinted from Earth Resources Observatory and Science (EROS) Center under a CC BY license. InSAR image obtained from the Earth Resources Observatory and Science (EROS) Center (http://eros.usgs.gov).
Fig 3.
Schematic diagram of the study area.
InSAR image reprinted from Earth Resources Observatory and Science (EROS) Center under a CC BY license. Map source: Anhui Provincial Department of Natural Resources, Bozhou Natural Resources and Planning Bureau. Insar image of the study area obtained from the Earth Resources Observatory and Science (EROS) Center (http://eros.usgs.gov).
Fig 4.
The model converges in the region of 180 to 220 epochs.
Table 1.
Comparison of experimental results of different prediction models.
PCC is the Pearson correlation coefficient, and MSE means mean square error.
Fig 5.
Comparison between predicted and observed values.
Time series of subsidence of points p1, p6 and p19 selected from the In-SAR control network.
Fig 6.
Comparison between predicted and observed values.
Time series of subsidence of points p5, p7 and p11 selected from the manual level observation network.
Fig 7.
The red, green, and blue scatter points correspond to the error distribution of MRF-GCN, LSTM, and ARIMA, respectively.
Fig 8.
MRF-GCN and LSTM time series comparison.
Fig 9.
Heatmap of Pearson correlation coefficient.
The Pearson correlation coefficient between any two points is indicated by different color shades.