Mission-critical applications that rely on deep learning (DL) for automation suffer because DL models struggle to provide reliable indicators of failure. Reliable failure prediction can greatly improve the efficiency of a system, because it becomes easier to predict when human intervention is required. DL-based systems thus stand to benefit greatly from robust measures of uncertainty over model predictions. Monte Carlo dropout (MCD), a Bayesian method, and deep ensembles (DE) have emerged as two of the most popular and competitive ways to perform uncertainty estimation. Although literature exploring the usefulness of these approaches exists in medical imaging, robotics and autonomous driving domains, it is scarce to non-existent for remote sensing, and in particular, synthetic aperture radar (SAR) applications. To close this gap, we have created a deep learning model for road extraction (hereafter referred to as segmentation) in SAR and use it to compare standard model outputs against the aforementioned most popular methods for uncertainty estimation, MCD and DE. We demonstrate that these methods are not effective as an indicator of segmentation quality when measuring uncertainty (as indicated by model softmax outputs) across an entire image but are effective when uncertainty is measured from the set of road predictions only. Furthermore, we show a marked improvement in the correlation between prediction uncertainty and segmentation quality when we increase the set of road predictions by including predictions with lower softmax scores. We demonstrate the efficacy of our application of MCD and DE methods with an experimental design that measures performance in real-world quality assessment using in-distribution (ID) and out-of-distribution (OOD) data. These results inform the development of mission-critical deep learning systems in remote sensing. Tasks in medical image analysis that have a similar morphology to road structures, such as blood vessel segmentation, can also benefit from our findings.

Landslides can have complex, spatially strongly inhomogeneous surface displacement fields with discontinuities from multiple active lobes that are deforming while failing on nested slip surfaces at different depths. For synthetic aperture radar interferometry (InSAR), particularly at lower resolutions, these characteristics can cause significant aliasing of the wrapped phase. In combination with steep terrain and seasonal snow cover, causing layover and temporal decorrelation, respectively, traditional phase unwrapping can become unfeasible, even after topographic phase contributions have been removed with an external high-resolution digital surface model (DSM). We present a novel method: warp demodulation that reduces the complexity of the phase unwrapping problem for noisy and/or aliased, low-resolution interferograms of discontinuous landslide displacement. The key input to our warp demodulation method is a single (or several) reference interferogram(s) from a high-resolution sensor mode such as TerraSAR-X Staring Spotlight with short temporal baseline and good coherence to allow localization of phase discontinuities and accurate unwrapping. The task of constructing suitable phase surfaces to approximate individual to-be-demodulated interferograms from the reference interferogram is made difficult by strong and spatially inhomogeneous temporal, seasonal, and interannual variations of the landslide with individual lobes accelerating or decelerating at different rates. This prevents using simple global scaling of the reference. Instead, our method uses an irregular grid of small patches straddling strong spatial gradients and phase discontinuities in the reference to find optimum local scaling factors that minimize the residual phase gradients across the discontinuities after demodulation. Next, for each to-be-demodulated interferogram, from these measurements we interpolate a spatially smooth global scaling function, which is then used to scale the (discontinuous) reference. Demodulation with the scaled reference leads to a residual phase that is also spatially smooth, allowing it to be unwrapped robustly after low-pass filtering. A key assumption of warp demodulation is that the locations of the phase discontinuities can be mapped in the reference and that they are stationary in time at the scale of the image resolution. We carry out extensive tests with simulated data to establish the accuracy, robustness, and limitations of the new method with respect to relevant parameters, such as decorrelation noise and aliasing along phase discontinuities. A realistic parameterization of the method is demonstrated for the example of the Fels Glacier Slide in Alaska using a recent late-summer high-resolution staring spotlight interferometric image pair from TerraSAR-X to demodulate. We show warp demodulation results for also recent but early-summer, partially incoherent interferograms of the same sensor, as well as for older and coarser aliased interferograms from RADARSAT-2, ALOS-1, and ERS.

We discuss enhanced processing methods for high resolution Synthetic Aperture Radar (SAR) interferometry (InSAR) to monitor small landslides with difficult spatial characteristics, such as very steep and rugged terrain, strong spatially heterogeneous surface motion, and coherence-compromising factors, including vegetation and seasonal snow cover. The enhanced methods mitigate phase bias induced by atmospheric effects, as well as topographic phase errors in coherent regions of layover, and due to inaccurate blending of high resolution discontinuous with lower resolution background Digital Surface Models (DSM). We demonstrate the proposed methods using TerraSAR-X (TSX) Staring Spotlight InSAR data for three test sites reflecting diverse challenging landslide-prone mountain terrains in British Columbia, Canada. Comparisons with corresponding standard processing methods show significant improvements with resulting displacement residuals that reveal additional movement hotspots and unprecedented spatial detail for active landslides/rockfalls at the investigated sites.

We analyze the sensitivity of a large (area extent ∼3 km²), deep-seated gravitational slope deformation (Fels slide, Alaska Range) to three specific drivers: (i) liquid surface water input from ERA-5 reanalysis snow melt and rainfall; (ii) locally projected seismic activity of Alaskan earthquakes; and (iii) lowering of Fels Glacier at the slide toe estimated from topographic data. A surface displacement map-series is derived from 1991 to 2016 spaceborne multi-sensor InSAR data (ERS, RADARSAT-1/2, ALOS, TerraSAR-X) using adaptive demodulation to unwrap interferograms of variable spatial resolution and quality. On this series we use independent component analysis (ICA) to uncover five displacement patterns that map to independently moving domains of the slide and then correlate the corresponding temporal pattern intensities with the suspected drivers. We find significant sub-annual correlation between displacement pattern intensities and seasonal water input variations. The correlation can be optimized, for each ICA pattern, by choosing appropriate values of temporal smoothing and lag to create depth-propagated versions of the water input driver. Lag time results ranging from one to 3 weeks relate to shallower and deeper propagations of water input, driving the different deformation patterns. For two of the deformation patterns, seasonal sensitivity to water input was strongly amplified by the 2002 M_w7.9 Denali earthquake. Sensitivity of these patterns remained high for 4 years until abruptly dropping to below pre-earthquake values, which suggests a highly non-linear modulation by the seismic driver. Other deformation patterns show a steady intensity increase that appears linked to the deglaciation driver. Despite these observations, the inter-annual variations in ICA pattern intensities show no clear predictability by individual drivers or driver combinations. This suggests that the mechanical and hydraulic evolution of the slide, especially after damaging events such as earthquakes or heavy rainfall, is a crucial factor not adequately modeled in our approach. Despite this limitation, our analysis provides the first direct evidence that the Fels slide comprises several independently moving domains that respond differently to the suspected drivers as is suggestive of a complex slope deformation.