Permafrost cannot be directly detected from space, but many surface features of permafrost terrains and typical periglacial landforms are observable with a variety of EO sensors ranging from very high to medium resolution at various wavelengths. In addition, landscape dynamics associated with permafrost changes and geophysical variables relevant for characterizing the state of permafrost, such as land surface temperature or freeze-thaw state can be observed with space-based Earth Observation. Suitable regions to examine environmental gradients across the Arctic have been defined in a community white paper. These transects have been revised and adjusted within the ESA DUE GlobPermafrost project.

The ESA DUE GlobPermafrost project (2016-2019) developed, validated and implemented Earth Observation (EO) products to support research communities and international organizations in their work on better understanding permafrost characteristics and dynamics. Prototype product cases covered different aspects of permafrost by integrating in situ measurements of subsurface properties and surface properties, Earth Observation, and modelling to provide a better understanding of permafrost today. Both lowland (latitudinal) and mountain (altitudinal) permafrost issues were addressed.
A Permafrost Information System has been setup through which GlobPermafrost datasets can be accessed. This includes a data catalogue as well as a WebGIS. Both are linked to the Pangaea repository for easy data access.

http://apgc.awi.de/group/persys
http://maps.awi.de/map/map.html?cu=Globpermafrost_Overview#layers
https://www.pangaea.de/?q=project%3Alabel%3ADUE-GlobPermafrost

The thematic products include InSAR-based land surface deformation maps, rock glacier velocity fields, spatially distributed permafrost model outputs, land surface properties and changes, and ground-fast lake ice. Extended permafrost modelling (time series) is foreseen in the new ESA CCI+ Permafrost project (2018-2021). This will provide the key for our understanding of the changes of surface features over time identified in GlobPermafrost. Special emphasis will be on the evaluation and development of land surface models, to gain better understanding of the impact of climate change on permafrost and land-atmosphere exchange.

We will present an overview on technical developments made within GlobPermafrost and application examples of the services.

Local observations indicate that climate change and shifting disturbance regimes are causing near-surface permafrost degradation in Arctic and Boreal regions. Here, we quantify the abundance and distribution of key permafrost region disturbances (PRDs: lakes and lake dynamics, wildfires, and retrogressive thaw slumps) using dense time series analysis of 30-m resolution Landsat satellite imagery from 1999 to 2014 in conjunction with machine-learning classification, object-based image analysis and auxiliary data sources. Our dataset spans four latitudinal, continental-scale transects in North America (Alaska, Eastern Canada) and Eurasia (Western Siberia, Eastern Siberia), covering 2.4 million km² or ~10% of the northern permafrost region.
For the more than 640,000 lakes covering 5.04% of the study area a net decreasing lake area (-1.45%) was detected. We observed major lake area decrease in north-western Siberia and parts of western Alaska, while continental eastern Siberia (Central Yakutia) was affected by a significant increase in lake area. The Alaskan and Siberian Transects showed complex spatial dynamics with several influencing factors, such as permafrost extent, geomorphology or climate. The formerly glaciated Eastern Canadian Transect with predominantly bedrock geology exhibited a gradual transition from stable lakes in the south towards increasingly shrinking lakes in the north.
Fires were the most extensive PRD across boreal regions (6.62%), with major abundance in dry continental regions, such as interior Alaska (8.89%), Eastern Siberia (8.15%) and boreal Eastern Canada (5.06%). Western Siberia, with high abundance of wetlands, was less affected by wildfires (2.43%). Within our study regions, tundra fires were limited to northern and western Alaska (1.07%).
Active retrogressive thaw slumps, while dramatic on local scales, were spatially restricted to around 5 km² within the analyzed region. Due to their small size likely below the detection limit, many RTS may have not been identified safely. However, their clustered occurrence, mainly along former glacial margins (ice-sheets and mountain glaciers) and very ice-rich permafrost in Siberia and Alaska can help to target their identification with higher spatial resolution data.
Our analysis demonstrates the global-scale vulnerability of permafrost terrain and carbon pools to disturbances and potentially rapid future thaw across very large regions. Our findings highlight that these PRDs need to be included in next generation land surface models to correctly capture the permafrost carbon feedback.
New operational sensors, such as Planet or ESA’s Sentinel-2, which exceed Landsat’s spatial and temporal resolution, will shift the focus from a backward looking approach to potential near-real-time monitoring. Near-real time information allows for a much more complete understanding of diverse, rapid, and widespread disturbances in the large and remote Arctic permafrost region and their impacts on ecosystems and infrastructure.

Landscape changes in the terrestrial Arctic occur on different time scales. One can normally distinguish between gradual and abrupt changes. Permafrost top-down thaw, changes in pedological processes and vegetation structure as well as aridification and paludification are gradual changes. In contrast to that, wildfires, thermokarst formation, thermokarst lake drainage, flooding and rapid soil erosion are abrupt and sometimes even discrete change events. A high temporal resolution in observations is required to properly monitor both types of changes. Particularly discrete change events that happen abruptly can be identified and characterized more precisely with better temporal resolution. However, a better understanding of the nature of gradual changes can also be gained since short periods of increased change or stagnation can also be identified with higher temporal resolution.

LandTrendr is an algorithm developed by Kennedy et al. (2010) to capture a wide range of forest disturbance and recovery occurrences based on Landsat time series. The algorithm creates cloud-free mosaics from multiple images and extracts the temporal trajectory of spectral data on a pixel-by-pixel basis. Furthermore, it segments the temporal trajectories based on regressions and point-to-point fittings of spectral indices, depicting then both long-term trends and abrupt changes (Kennedy et al., 2010). As optical remote sensing in the Arctic is highly restricted by frequent cloud cover and low illumination angles, the number of useable images per summer season is low. However, combining data of multiple optical systems bypasses this limitation. Combining data from Landsat and the newly launched ESA Copernicus Sentinel-2 mission, enables the use of data from three satellites (L-8, S-2A, S-2B), which shortens the revisit time to less than five days in high latitudes, enhances the likelihood of cloud-free images considerably, and increase the number of images per summer season dramatically. Using both Landsat and Sentinel-2 images will lead to much denser time-series and therefore capturing changes more precisely.

In this study in progress, we apply the LandTrendr algorithm to three different areas in North Eastern Siberia: the central Lena Delta, Batagay in the Yana Highlands and Yukechi in central Yakutia. These three sites are located along an approximate longitudinal transect and exemplify the different climatic, geologic, geomorphologic and vegetational conditions of the heterogeneous North Eastern Siberian permafrost landscapes that experience different disturbances. In order to ensure a full coverage of the study sites and to obtain higher temporal resolutions, we create a homogenous time series combining both Landsat and Sentinel-2 data. We anticipate that the LandTrendr algorithm is transferable to Arctic study areas and that the results will benefit from the higher spatial coverage and especially temporal resolution when combining Landsat and Sentinel-2. We hypothesize that we can distinguish between different permafrost disturbances, both gradual and abrupt changes, like thermokarst, erosion and fires, by applying key spectral indices, such as NDVI, NDWI, NBR and tasselled cap indices and by adapting the temporal segmentation parameters to permafrost landscape changes.

Kennedy, R.E., Yang, Z. and Cohen, W.B., 2010. Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr — Temporal segmentation algorithms. Remote Sensing of Environment, 114(12): 2897-2910.

Permafrost is an Essential Climate Variable (ECV) within the Global Climate Observing System (GCOS), which is characterized by subsurface temperatures and the depth of the seasonal thaw layer. Complementing ground-based monitoring networks, the Permafrost CCI project funded by the European Space Agency (ESA) 2018-2021 will establish Earth Observation (EO) based products for the permafrost ECV spanning the last two decades. Since ground temperature and thaw depth cannot be directly observed from space-borne sensors, we will combine a variety of satellite and reanalysis data in a ground thermal model. The model output allows to characterize the changing ground temperatures and thaw depths in Arctic and High-Mountain permafrost environments.
As recently demonstrated for the Lena River Delta in Northern Siberia, the algorithm uses remotely sensed data sets of Land Surface Temperature (MODIS LST/ ESA LST CCI) and landcover (ESA Landcover CCI) to drive the transient permafrost model CryoGrid 2, which yields ground temperature at various depths, in addition to thaw depth. For the circumpolar CCI product, we aim for a spatial resolution between 10 and 1km, but ensemble runs will be performed for each pixel to represent the subgrid variability of snow and land cover. The correct representation of ground properties is critical for the performance of the transient algorithm, in particular for reproducing the depth of the thaw layer. Therefore, the project will compile a new subsurface stratigraphy product which also holds great potential for improving Earth System Model results in permafrost environments. Furthermore, the representation of the winter snow cover insulating the ground is of outstanding importance for the reliability of thermal modelling. We discuss the feasibility of using different EO snow products (ESA Snow CCI) to help constrain the model, in conjunction with the model representation of snow processes, such as wind redistribution and water infiltration, which strongly affect the ground thermal regime on a range of spatial and temporal scales.
We report on simulation runs for various permafrost regions and characterize the accuracy and ability to reproduce trends against ground-based data. Finally, we evaluate the feasibility of future “permafrost reanalysis” products, exploiting the information content of various satellite products to deliver the best possible estimate for the permafrost thermal state over a range of spatial scales.

Land surface state, frozen versus unfrozen conditions, can be captured globally with satellite data obtained by microwave sensors. The number of frozen days can be derived in case of daily availability of measurements. Such a sampling density can be achieved in high latitudes from coarse resolution sensors such as scatterometer and radiometers. It has been hypothesized that the number of frozen days can give an estimate for mean annual ground temperature, although it does not account for crucial factors such as soil and snow properties. The aim of the study was to quantify the accuracy of such an approach and discuss the uncertainties with respect to different sensor types and environmental conditions contributing to the ESA DUE GlobPermafrost project (2016-2019).
Coarse spatial resolution microwave satellite data (Metop Advanced Scatterometer - ASCAT - and Special Sensor Microwave Imager - SSM/I; 12.5km and 25km nominal resolution; 2007 - 2012) which provide the necessary temporal sampling were used. The ASCAT surface state product has been developed as part of the ESA DUE Permafrost project (2009-2012, Naeimi et al. 2012). An SSMI surface state product is available from Paulik et al. (2014). The MAGT from GTN-P (Global Terrestrial Network - Permafrost) borehole records at coldest sensor depth (MAGTc) was derived and used for calibration and validation (separated into two distinct time periods).
MAGTc could be obtained with an RMSE of 2.2°C from ASCAT and 2.5°C from SSM/I surface state records using a linear model. Comparison to Snow Water Equivalent data (GlobSnow) and consideration of melt day information from ASCAT demonstrated the discrepancies of the simple approach. Difference between years and datasets showed that especially central Siberia and southern Yakutia are regions with higher complexity regarding MAGT retrieval.
For further evaluation, within the framework of the ESA CCI+ Permafrost project, the modelled MAGT was compared to the temperature at top of permafrost (TTOP) from Obu et al. (2018) derived by the Cryogrid-1-model showing an overall RMSE of 2.18. This surprisingly low difference was closer investigated by comparing the in-situ MAGT-datasets used for both studies and the modelled MAGT-values regionally.
A linear regression using MAGTTTOP from Obu et al. (2018) resulted in a lower R² (0.49) than for MAGTc (R² 0.62) and amplifies non-linear aspects in the regression model. Nevertheless, the RSE of the new Model was 1.9°C compared to 2.5°C of MAGTc and 2.18°C for Cryogrid-1. The model shows a similar extent of the area in which MAGTc is lower than -1 °C, although its mean ground temperature was mainly higher than for the ASCAT-model. Especially the Canadian Arctic and mountainous areas show temperature differences larger than 2 °C (negative difference). Both models show an underestimation of MAGT in the southern Alaskan mountainous areas compared to in-situ data, an error depending on the linear regression approach, but there are also inconsistencies in the Cryogrid-approach not appearing for MAGTc. This strengthens the impression that there is still a predicament regarding input-data accuracy which alters modelling quality. Maybe surface-state-information can help to refine some of the datasets used in physical models.

The Sentinel-1A/1B constellation provides observations of microwave backscatter almost on a daily basis at relatively high spatial resolution for the northern boreal forest zone. The C-band observations are useful in estimating the soil conditions of the top soil layer, such as the Frost/Thaw (F/T) state of the soil. Yet, due to the high sensitivity of the C-band signal to vegetation, estimating the soil F/T state directly from the measured backscatter is not feasible over dense vegetation, such as boreal forests.

A new approach for monitoring soil frost in the boreal forest region using Sentinel-1 co-polarized synthetic aperture radar (SAR) data has been proposed (Cohen et al., in press). The method is based on applying a simplified zeroth-order model of the radiative transfer equation, called the HUT (Helsinki University of Technology) semi-empirical forest backscattering model, to estimate the backscatter contribution of the ground and the forest canopy on the observed total backscatter. With the proposed approach, the soil F/T state is estimated using two retrieved values: 1) The backscatter contribution of the ground surface beneath the canopy and 2) the backscatter contribution of the forest canopy. These parameters are retrieved from high-resolution Sentinel-1 data, aggregating estimates to a grid with a cell spatial resolution of 1000 x 1000 m. The aggregation is performed by minimizing a cost function of model estimates and observations, using all high-resolution Sentinel-1 SAR image pixels within a given grid cell. Forest cover information at an equivalent resolution is required to determine forest properties (canopy cover or stem volume) for each Sentinel-1 SAR image pixel, allowing to fit the vegetation backscatter model over a range of forest properties. Finally, in order to separate frozen and thawed conditions, an optimal linear classification line in the 2-D plane formed by the retrieved ground and canopy backscatter values against in-situ measured freezing conditions was determined using a least sum of square errors estimation method. This was performed separately for different soil types.

On the other hand, the use of passive microwave observations at L-band has become increasingly relevant in Earth observation based soil F/T state estimation with the launch of the ESA SMOS and NASA SMAP missions. These sensors provide daily observation data at a global scale. It is relatively straightforward to measure the F/T state of soils using passive microwave sensors operating at low frequencies, due to the high dielectric contrast between frozen and free phases of water. Methods for monitoring soil F/T from both SMOS and SMAP have been developed (Rautiainen et al., 2016; Derksen et al., 2017). While appealing for large scale observations with daily temporal resolution and hemispheric to global spatial coverage, the applicability of passive microwave observations over heterogeneous terrain are limited by their inherent coarse spatial resolution.

Sentinel-1 based soil F/T estimates were compared with spatially and temporally corresponding SMOS based F/T estimates. The experiment was performed over a test area near Sodankylä, northern Finland, which is a typical northern boreal forest environment. The total size of the test area was 75 x 75 km, which is covered by 9 pixels of the SMOS F/T product, where each SMOS pixel contained 25 x 25 Sentinel-1 F/T estimation pixels. The general similarity of the S1 and SMOS estimates were tested for the freezing seasons of 2014-2017. The similarity of Sentinel-1 based soil F/T estimates with the SMOS estimates was also tested separately for the main individual soil classes; Fine- and Coarse Haplic Podzol, Umbric Gleysol, and Haplic Arenosol. Preliminary results show good general agreement between the active and the passive retrievals. The Sentinel-1 based soil F/T estimates are more sensitive to air temperature variations in the beginning of the freezing period, and react faster to the cooling and warming of weather, while the SMOS estimate shows smoother and more stable soil state changes during the freezing period. Soil freezing therefore occurs somewhat earlier according to the Sentinel-1 estimate. This is reasonable, since the C-band Sentinel-1 observes mainly the top of the soil and surface vegetation layers, while L-band emission observed by SMOS originates from deeper in the soil (5-10 cm depth) and thus shows slower changes of the soil state.

The ultimate aim for the future is to combine the Sentinel-1 soil F/T estimates with existing passive microwave L-band F/T detection methods, which have a better temporal but weaker spatial resolution. As a result, a synergistic product which applies C-band SAR to downscale passive microwave estimates could be implemented.


J. Cohen, K. Rautiainen, J. Ikonen, J. Lemmetyinen, T. Smolander, J. Vehviläinen, J. Pulliainen. "A Modeling-Based Approach for Soil Frost Detection in the Northern Boreal Forest Region With C-Band SAR," in IEEE Transactions on Geoscience and Remote Sensing. doi: 10.1109/TGRS.2018.2864635

K. Rautiainen, T. Parkkinen, J. Lemmetyinen, M. Schwank, A. Wiesmann, J. Ikonen, C. Derksen, S. Davydov, A. Davydova, J. Boike, M. Langer, M. Drusch, J. Pulliainen, SMOS prototype algorithm for detecting autumn soil freezing, Remote Sensing of Environment, Volume 180, 2016, Pages 346-360, ISSN 0034-4257, https://doi.org/10.1016/j.rse.2016.01.012.

C. Derksen, X. Xu, R. S. Dunbar, A. Colliander, Y. Kim, J. S. Kimball, T. A. Black, E. Euskirchen, A. Langlois, M. M. Loranty, P. Marsh, K. Rautiainen, A. Roy, A. Royer, J. Stephens, Retrieving landscape freeze/thaw state from Soil Moisture Active Passive (SMAP) radar and radiometer measurements, Remote Sensing of Environment, Volume 194, 2017, Pages 48-62, ISSN 0034-4257, https://doi.org/10.1016/j.rse.2017.03.007.