Our study highlights the usefulness of very high resolution (VHR) images to detect various types of disturbances over permafrost areas using three example regions in different permafrost zones. The study focuses on detecting subtle changes in land cover classes, thermokarst water bodies, river dynamics, retrogressive thaw slumps (RTS) and infrastructure in the Yamal Peninsula, Urengoy and Pechora regions. Very high-resolution optical imagery (sub-meter) derived from WorldView, QuickBird and GeoEye in conjunction with declassified Corona images were involved in the analyses. The comparison of very high-resolution images acquired in 2003/2004 and 2016/2017 indicates a pronounced increase in the extent of tundra and a slight increase of land covered by water. The number of water bodies increased in all three regions, especially in discontinuous permafrost, where 14.86% of new lakes and ponds were initiated between 2003 and 2017. The analysis of the evolution of two river channels in Yamal and Urengoy indicates the dominance of erosion during the last two decades. An increase of both rivers’ lengths and a significant widening of the river channels were also observed. The number and total surface of RTS in the Yamal Peninsula strongly increased between 2004 and 2016. A mean annual headwall retreat rate of 1.86 m/year was calculated. Extensive networks of infrastructure occurred in the Yamal Peninsula in the last two decades, stimulating the initiation of new thermokarst features. The significant warming and seasonal variations of the hydrologic cycle, in particular, increased snow water equivalent acted in favor of deepening of the active layer; thus, an increasing number of thermokarst lake formations.

This work aims at evaluating the sensitivity of landslide susceptibility mapping (LSM) to sampling design in lithologically-heterogeneous areas. We hypothesize that random sampling of the landslide absence data in such areas can be biased by statistical aggregation of the explanatory variables, which impact the model outputs. To test this hypothesis, we train a Random Forest (RF) model in two different domains, as follows: (1) in lithologically heterogeneous areas, and (2) in lithologically homogeneous domains of the respective areas. Two heterogeneous areas are selected in Japan (125 km²) and Romania (497 km²), based on existing landslide inventories that include 371 and 577 scarps, respectively. These areas are divided into two, respectively three domains, defined by lithological units that reflect relatively homogeneous topographies. Fourteen terrain attributes are derived from a 30 m SRTM digital elevation model and employed as explanatory variables. Results show that LSM is sensitive to a random sampling of the absence data in lithologically heterogeneous areas. Accuracy measures improve significantly when sampling and LSM are conducted in lithologically homogeneous domains, as compared to heterogeneous areas, reaching an increase of 9% in AUC and 17% in the Kappa index.

Sampling efforts are constrained by limited availability of resources. Therefore, methods to reduce the number of samples, while still achieving reasonable accuracy are needed. Land-surface segmentation (LSS) has proven a powerful technique to partition digital elevation models (DEMs) and their derivatives into relatively homogeneous areas, which can be further employed as support in soil sampling. Though topography is one of the main soil forming factors, a robust assessment of the potential of this technique to digital soil mapping (DSM) is still missing. In this study, we aimed at evaluating the potential of LSS in stratifying a landscape into relatively homogeneous areas, which can be used as strata for guiding the selection of sampling points in DSM. The experiments were carried out in two study areas where soil samples were available. Land-surface derivatives were derived from DEMs and segmented with a tool based on the multiresolution segmentation algorithm, into objects considered as homogeneous soil-landscape divisions. Thus, one sample was randomly selected within each segment from the existing sample data, based on which predictions of soil classes/sub-orders and properties, i.e. soil texture and A-horizon thickness, were made. Results were compared with predictions based on simple random sampling (SRS) and conditioned Latin hypercube (cLHS). The segmentation-based sampling (SBS) scheme performed better than SRS and cLHS schemes in predicting the A-horizon thickness, soil texture fractions and soil classes, showing a high potential of LSS in stratifying a landscape for the purposes of DSM. The novelty of this study is in the way strata are constructed, rather than in the sampling design itself. Further research is needed to demonstrate the value of a SBS design for practical use. The analyses presented here further highlight the importance of considering locally adaptive techniques in optimization of sampling schemes and predictions of soil properties.