Search

We propose a simple modification to standard ResNet architectures during training--L2 normalization over feature space--that produces results competitive with state-of-the-art Out-of-Distribution (OoD) detection performance. When L2 normalization is removed at test time, the L2 norm of feature vectors becomes a surprisingly good proxy for network uncertainty, whereas this behaviour is not nearly as effective when training without L2 normalization. Intuitively, familiar images result in large magnitude vectors, while unfamiliar images result in small magnitudes. Notably, this is achievable with almost no additional cost during training, and no cost at test time.

We propose a simple modification to standard ResNet architectures--L2 normalization over feature space--that substantially improves out-of-distribution (OoD) performance on the previously proposed Deep Deterministic Uncertainty (DDU) benchmark. We show that this change also induces early Neural Collapse (NC), an effect linked to better OoD performance. Our method achieves comparable or superior OoD detection scores and classification accuracy in a small fraction of the training time of the benchmark. Additionally, it substantially improves worst case OoD performance over multiple, randomly initialized models. Though we do not suggest that NC is the sole mechanism or a comprehensive explanation for OoD behaviour in deep neural networks (DNN), we believe NC's simple mathematical and geometric structure can provide a framework for analysis of this complex phenomenon in future work., Comment: 19 pages, accepted by Transactions of Machine Learning Research, 2022

The timing of snowmelt is an important turning point in the seasonal cycle of small Arctic catchments. The TerraSAR-X (TSX) satellite mission is a synthetic aperture radar system (SAR) with high potential to measure the high spatiotemporal variability of snow cover extent (SCE) and fractional snow cover (FSC) on the small catchment scale. We investigate the performance of multi-polarized and multi-pass TSX X-Band SAR data in monitoring SCE and FSC in small Arctic tundra catchments of Qikiqtaruk (Herschel Island) off the Yukon Coast in the Western Canadian Arctic. We applied a threshold based segmentation on ratio images between TSX images with wet snow and a dry snow reference, and tested the performance of two different thresholds. We quantitatively compared TSX- and Landsat 8-derived SCE maps using confusion matrices and analyzed the spatiotemporal dynamics of snowmelt from 2015 to 2017 using TSX, Landsat 8 and in situ time lapse data. Our data showed that the quality of SCE maps from TSX X-Band data is strongly influenced by polarization and to a lesser degree by incidence angle. VH polarized TSX data performed best in deriving SCE when compared to Landsat 8. TSX derived SCE maps from VH polarization detected late lying snow patches that were not detected by Landsat 8. Results of a local assessment of TSX FSC against the in situ data showed that TSX FSC accurately captured the temporal dynamics of different snow melt regimes that were related to topographic characteristics of the studied catchments. Both in situ and TSX FSC showed a longer snowmelt period in a catchment with higher contributions of steep valleys and a shorter snowmelt period in a catchment with higher contributions of upland terrain. Landsat 8 had fundamental data gaps during the snowmelt period in all 3 years due to cloud cover. The results also revealed that by choosing a positive threshold of 1 dB, detection of ice layers due to diurnal temperature variations resulted in a more accurate estimation of snow cover than a negative threshold that detects wet snow alone. We find that TSX X-Band data in VH polarization performs at a comparable quality to Landsat 8 in deriving SCE maps when a positive threshold is used. We conclude that TSX data polarization can be used to accurately monitor snowmelt events at high temporal and spatial resolution, overcoming limitations of Landsat 8, which due to cloud related data gaps generally only indicated the onset and end of snowmelt.

We characterize and compare creep preceding and following the complex 2011 Pampahasi landslide (∼40 Mm3±50 %) in the city of La Paz, Bolivia, using spaceborne radar interferometry (InSAR) that combines displacement records from both distributed and point scatterers. The failure remobilized deposits of an ancient complex landslide in weakly cemented, predominantly fine-grained sediments and affected ∼1.5 km2 of suburban development. During the 30 months preceding failure, about half of the toe area was creeping at 3–8 cm a−1 and localized parts of the scarp area showed displacements of up to 14 cm a−1. Changes in deformation in the 10 months following the landslide demonstrate an increase in slope activity and indicate that stress redistribution resulting from the discrete failure decreased stability of parts of the slope. During that period, most of the landslide toe and areas near the head scarp accelerated, respectively, to 4–14 and 14 cm a−1. The extent of deformation increased to cover most, or probably all, of the 2011 landslide as well as adjacent parts of the slope and plateau above. The InSAR-measured displacement patterns, supplemented by field observations and optical satellite images, reveal complex slope activity; kinematically complex, steady-state creep along pre-existing sliding surfaces accelerated in response to heavy rainfall, after which slightly faster and expanded steady creeping was re-established. This case study demonstrates that high-quality ground-surface motion fields derived using spaceborne InSAR can help to characterize creep mechanisms, quantify spatial and temporal patterns of slope activity, and identify isolated small-scale instabilities; such details are especially useful where knowledge of landslide extent and activity is limited. Characterizing slope activity before, during, and after the 2011 Pampahasi landslide is particularly important for understanding landslide hazard in La Paz, half of which is underlain by similar large paleolandslides.

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.