Fig. 2: Performance on synthetic test data.
From: Autonomous extraction of millimeter-scale deformation in InSAR time series using deep learning
Top: performance of the reconstruction of fault deformation by our deep autoencoder, on synthetic noisy time series, as measured by structural similarity index (SSIM) between model output and deformation ground truth, as a function of signal-to-noise ratio (SNR, see “Methods”). Shades of blue show the distribution of SSIM as a function of SNR (counts per bins for 105 test samples). The black and gray lines show the median and 25th and 75th percentile of the SSIM in SNR bins, respectively. Bottom: examples of the data showing input time series, ground truth, and its reconstruction, for different signal-to-noise ratios, shown with matching numbers in the plot above. Note that the model outperforms the eye, recovering with reasonable fidelity deformation signals with SNRs down to a few percent.
