A Data Mining Approach for Multivariate Outlier Detection in Postprocessing of Multitemporal InSAR Results

Displacement maps from multitemporal InSAR (MTI) are usually noisy and fragmented. Thresholding on ensemble coherence is a common practice for identifying radar scatterers that are less affected by decorrelation noise. Thresholding on coherence might, however, cause loss of information over the areas undergoing more complex deformation scenarios. If the discrepancies in the areas of moderate coherence share similar behavior, it appears important to take into account their spatial correlation for correct inference. The information over low-coherent areas might then be used in a similar way the coherence is used in thematic mapping applications such as change detection. We propose an approach based on data mining and statistical procedures for mitigating the impact of outliers in MTI results. Our approach allows for minimization of outliers in final results while preserving spatial and statistical dependence among observations. Tests from monitoring slope failures and undermined areas performed in this work have shown that this is beneficial: 1) for better evaluation of low coherent scatterers that are commonly discarded by the standard thresholding procedure, 2) for tackling outlying observations with extremes in any variable, 3) for improving spatial densities of standard persistent scatterers, 4) for the evaluation of areas undergoing more complex deformation scenarios, and 5) for the visualization purposes.