Automated INSAR time-series analysis tool for geological interpretations in near-real time

Introduction: Large Earth Observation (EO) datasets are becoming increasingly challenging to analyse due to their size, a problem that will continue to worsen with longer satellite operating times. The ESA Sentintel-1 satellites have been operating together for 8 years, collecting InSAR data of the Earth’s surface to monitor geological and man-made changes to the surface. The wealth of data obtained from these satellites means that their time series can be analysed to monitor ground motions due to various hazards around key areas such as infrastructure, high-risk fault lines, and homes. Here, we present an automatic tool which analyses these large InSAR datasets and interprets the cause or process behind deformations. This algorithm will aid in the task of understanding substantial ground motions in vulnerable areas, as well as monitoring longterm movements, which is becoming increasingly important with the worsening effects of climate change. Methodology: The algorithm developed for this study was tested using a sample set of European Ground Motion Service (EGMS) data [1] for the Midlands region of England, UK, containing time series data for almost 290,000 measurement points (MPs). For each one of them, the algorithm first uses a Seasonal-Trend decomposition using LOESS (STL) [2] to extract the trend and seasonality components of the signal. Then, it uses an automatic piecewise linear regression (PLR) algorithm (also developed for this study) to fit the trend with a variable number of linear segments depending on the trends in the displacement time series. To determine the optimal number of segments whilst preventing overfitting, this algorithm gradually increases the complexity of the model by adding one more segment. If the increase in the adjusted R2 is larger than 3%, the more complex model is accepted and a new segment is added. This process continues until this increase becomes equal to or smaller than 3%, at which point additional segments will provide little improvement. MPs for which the PLR algorithm required only one segment to fit their trend are then classified as “Linear”, while those needing two or more segments are “Nonlinear”. Linear trends are further grouped into “stable”, “subsidence” and “uplift”, based on a velocity threshold. MPs with linear velocities within the range (±threshold) are considered to be stable and assigned to group 0, while those with linear velocities above or below this threshold are labelled as uplift and subsidence and assigned groups 1 and -1 respectively. Results: Figure 1 shows the results for three MPs in our sample dataset. The plots also contain the trend component obtained from the STL decomposition, the results of the PLR fit of the trend, and the ground velocities obtained for each segment. The adjusted R2 for these cases ranges from 0.68 to 0.99, highlighting the ability of our automatic processing tool to successfully capture the characteristics of, and changes in, the original signal. Additionally, the algorithm contains no geological or process-related information, nor does it assume any kind of spatial correlation, however the resulting data is coherent in space. Fig. 1: Ground displacement time series data for three MPs in the test data, presenting a) linear stable, b) linear subsiding and c) non-linear trend behavior. Red lines are the trend component from STL, and blue lines are the PLR results. Black lines are the original signal of deformation. Conclusions: We have outlined a process independent data analysis tool for the automatic interpretation of InSAR-derived deformation time series data. Overall, the preliminary results for our test dataset indicate our algorithm successfully captures the main characteristics of the original signals and provides valuable information about spatiotemporal patterns present in the time series. This tool is currently applicable to EGMS data and is being adapted for different datasets to enable interpretations for locations beyond Europe. There is scope to incorporate climate models or geological maps to monitor areas vulnerable to hazards with additional context in near-real time.