The Arctic-Boreal domain is currently undergoing extensive environmental changes. Mean annual air temperatures increased by up to 1.5–2
C over the 1960–2009 period [1
], and precipitation increased, with climate models showing further precipitation increases with additional warming [2
]. Both the increase in air temperatures and precipitation also affect the thermal state of permafrost, which is defined as ground that remains at or below a temperature of 0
C for at least two consecutive years and is very widespread in this region [3
]. Considering that about a quarter of the Northern Hemisphere is underlain by permafrost, which is 65% of the land north of 60 degrees [4
], these changes are of great importance to northern high latitude ecosystems [5
]. Continued warming of permafrost across the pan-Arctic region [6
] ultimately will result in widespread permafrost thaw and eventually cause near-surface permafrost loss on local to regional scales [7
]. Thawing permafrost has direct implications for hydrological systems [8
], ecosystems [5
], soil carbon decomposition and accumulation [9
], greenhouse gas emissions [10
], and infrastructure stability in permafrost areas [11
]. A multitude of landscape-scale changes, such as active layer deepening, thermokarst and talik formation, soil erosion, and changes in vegetation composition and surface hydrology can be linked to permafrost thaw [12
While permafrost itself cannot be observed directly with remote sensing, detectable surface disturbance can be linked to permafrost degradation. They can be distinguished between press and pulse disturbances, which are gradual and abrupt disturbance regimes respectively and alter the ecosystem and environment substantially [16
]. Press disturbances are driven by impacts and change factors unfolding over decades to centuries and occur gradually and continuously. Press disturbances include processes such as permafrost top-down thaw, as well as vegetation structure and composition changes. They usually occur on extensive regional scales. In contrast to that, pulse disturbances unfold within very short time periods from a few days to years. They are usually extreme in nature and affect the system rapidly as one-time or episodic short-term events, but arise locally in distinct areas. Pulse disturbances include soil erosion, thermokarst lake expansion, lake drainage, and wildfires. Press and pulse disturbances interact with one another as well; for example, press disturbances can trigger pulse disturbances. Exemplary is permafrost top-down thaw initiating rapid thermokarst development or soil erosion [17
]. To understand the complex feedbacks of permafrost degradation processes and resulting landscape and ecosystem dynamics in space and time, it is critical to track and quantify these not only at the field site scale, but at high spatial resolution across decadal time scales and extensive regions.
Remote sensing provides excellent tools to understand and quantify many of these dynamics linked to surface expressions and land cover changes associated with permafrost degradation across large regions [18
]. For example, Nitze et al. [21
] mapped widespread permafrost disturbances based on trend analysis using linear regression on Landsat multispectral index time series, highlighting the abundance and distribution of disturbances such as thermokarst lake changes, wildfires, and thaw slumps. Similarly, Pastick et al. [22
] analyzed the diversity of landscape and ecosystem dynamics for Alaska using Landsat-based time series analysis in combination with spectral metrics, climatic data, and topographical and soil information. Using statistical models, they predicted and understood the drivers of change and differentiated between gradual and abrupt disturbances. Apart from understanding the spatial coverage, magnitude, patterns, and dimension of disturbances, the temporal dynamics of disturbances such as initiation timing, duration and persistence, recovery time, and recurrence frequency are highly important parameters for monitoring and projecting Arctic-Boreal ecosystem and landscape trajectories in a changing climate. Space-borne remote sensing with a high temporal and spatial resolution allows detailed monitoring of both gradual and abrupt changes, such as press and pulse disturbances, but requires more detailed insights into the timing and magnitudes of spectral changes of observation targets.
Algorithms developed for monitoring forest disturbances, such as LandTrendr, manage to identify both abrupt and gradual changes with temporally-dense Landsat time series [23
]. Likewise, Sulla-Menashe et al. [24
] used Normalized Difference Vegetation Index (NDVI) Landsat time series for characterizing vegetation greening and browning patterns in Canadian boreal forests and identified the major driver of the signal to be vegetation recovery following disturbance. Implementing such algorithms developed for boreal ecosystems for extensive monitoring and assessing of disturbance types and dynamics, such as differentiating between press and pulse disturbances, across Arctic-Boreal permafrost regions will improve understanding of climate change and anthropogenic impacts in high latitudes on multiple temporal scales. Most of the schemes so far rely on the extensive Landsat record, which features the longest continuous data archive of optical Earth surface imagery, covering more than 45 years. However, the low temporal resolution of Landsat images and some environmental factors, such as a short growing season length, a challenging solar geometry, frequent snow and ice cover, and especially persistent cloud cover, negatively affect the availability of images for detailed studies [25
]. Therefore, often, only a rather low number of adequate Landsat images are available for time series analyses in high latitudes. In many regions, these are not sufficient for creating temporally dense time series, which are needed for extracting detailed trends of disturbances and especially their short-term dynamics.
An approach to increase the data base in high latitudes is to combine Landsat time series, which relies on a single platform only, with images from additional optical sensors. A combined data base allows creating temporally much denser time series for the Arctic-Boreal domain, thereby increasing opportunities for monitoring permafrost region disturbances and ecosystem changes. The European Space Agency (ESA) Copernicus Sentinel-2 mission started in 2015 and consists of two satellites (S-2A, S-2B), which have a combined revisit time of five days [27
]. Landsat-8 and Sentinel-2’s sensors are both multispectral optical systems, have Sun-synchronous polar orbits, and obtain data with a comparable resolution. Both sensors record multispectral data with bands that cover similar wavelength ranges (Table 1
). Sentinel-2 sensors’ similar characteristics to Landsat and shortened revisit time are now leveraging the combined multispectral observation power of three platforms observing the Earth’s surface at 10–30 m resolution, significantly increasing the chance for capturing cloud-free observations at high latitudes. Li and Roy [28
] quantified the combined revisit time of Landsat 8, Sentinel-2A, and Sentinel-2B, showing that the global average revisit interval was 2.9 days. With the increased poleward orbit overlap and the enhanced spatial overlap between acquisition swaths, the high latitudes (>70
N) feature the highest number of acquired images within a year (>400 images) and the shortest revisit interval (<1 day) [28
]. This dramatic increase in image numbers and shortened revisit time for a combined Landsat-8 and Sentinel-2 time series could facilitate tracking permafrost region disturbances in much higher detail.
In terms of time series analyses, we do not just require cloud-free images, but rather depend on clear pixels for every location in one study area. A high number of clear pixels available in one observation period (summer months) indicates a good overall data coverage. Figure 1
shows the average number of clear pixels in the Lena Delta (see Section 2.1
) for the summer months of every year. It becomes obvious that the average number of clear pixels for every location in the Lena Delta increases dramatically, when adding Sentinel-2 to the analysis. This underlines that relying not only on one sensor platform, but on multiple ones enhances the database drastically.
For a combined time series, it is necessary to align the image products of the different sensors geometrically and radiometrically, ensuring that the same point on the ground is captured with similar spectral reflectance properties in both products. This issue was already addressed for the Landsat sensor series, defining adjustment schemes for slight differences in reflectance measurements, ensuring data continuity and that the Landsat image products from the different sensors (Landsat-5 TM, Landsat-7 ETM+, and Landsat-8 OLI) can be used in a combined time series analysis [29
]. To extend the data continuity incentive beyond the Landsat sensors and also to be able to include Sentinel-2, previous studies compared Landsat-8 with Sentinel-2 data and defined bandpass adjustments between corresponding Landsat-8 and Sentinel-2 bands for a better fit. Despite similar wavelength ranges covered by the corresponding bands, the spectral response functions per sensor differ, and this impact has to be assessed before a combined time series is applied [33
]. Previous comparison studies varied in application region, scale, input datasets, ranging from simulated to acquired image products, and covered temporally close to same-day acquisitions [33
]. Regardless of these differences, the comparison approaches are always alike, examining corresponding Landsat-8 and Sentinel-2 bands, comparing pixel-pairs, and deriving regression equations for bandpass alignments. Claverie et al. [38
] conducted a comparison between Landsat-8 and Sentinel-2 data for 91 test sites worldwide and created a globally-applicable Harmonized Landsat and Sentinel-2 surface reflectance dataset (HLS). While a few sample sites are situated in Alaska, Canada, and Northern Scandinavia, sites from the Siberian high latitudes are not included. Flood [39
] was first to compare same-day Landsat-8 with Sentinel-2 acquisition pixel-pairs of corresponding bands for Australia and defined band-wise adjustment equations. Furthermore, Flood [39
] applied the global HLS bandpass correction to the same dataset, but concluded that the HLS adjustment did not sufficiently adjust the Sentinel-2 data to Landsat-8 data. As a consequence, their region-specific derived bandpass adjustments showed a better fit than the HLS adjustment. An overall result of the previous studies was that Landsat-8 and Sentinel-2 data were comparable, but for a combined analysis, the datasets have to be adjusted to avoid detecting change because of a bias between sensors and not because of changes on the ground, which is particularly important for studying trends. Despite the effort of creating a globally-applicable harmonized product [38
], local to regional studies showed that locally-derived bandpass adjustments are superior to the HLS product [34
]. Therefore, it is recommendable to create region-specific adjustments between Landsat-8 and Sentinel-2 for regional assessments and applications.
So far, to the best of our knowledge, no comparison of Landsat-8 and Sentinel-2 data or an assessment of a potential combined use in time series analyses for the Siberian Arctic-Boreal region has been done. The overall objective of this paper is to compare the spectral characteristics of Landsat-8 and Sentinel-2 data across different types of landscapes and ecosystems in the study domain and assess their compatibility in Arctic-Boreal permafrost regions for the potential use in combined time series. To address this objective, we focus on three sub-objectives: (1) comparison of the spectral characteristics of corresponding bands from Landsat-8 and Sentinel-2 data of same-day acquisitions for three individual study sites in Eastern Siberia that broadly differ in ecosystem and land cover characteristics; (2) defining spectral band adjustments between Landsat-8 and Sentinel-2 data for each of the three study sites; and (3) determining a set of spectral band adjustments that is applicable to Eastern Siberia generally and comparing them to the application of the global Harmonized Landsat Sentinel-2 product.
Our overall comparison and adjustment results showed a good fit between Sentinel-2 and Landsat-8 in the high latitudes. Despite this, the Batagay study site showed the biggest differences between Landsat-8 and Sentinel-2 reflectance values in the visible wavelength bands. The Batagay same-day image pair, compared to the Lena Delta and Yakutsk example, also contained the highest cloud cover with 38.6% and 49.8% (Table 2
). While the cloud and cloud shadow detection algorithms still had high omission errors [57
], the Batagay images particularly were visibly contaminated with thin cirrus clouds that were difficult to mask. This decreased the quality of the images [67
] and likely was the main cause for a lower correlation coefficient between spectral bands, since the visible wavelength ranges were particularly affected by clouds and cirrus [68
]. Landsat-8 and Sentinel-2 were differently sensitive to cirrus clouds, and the acquisition delay of 25 min between the two sensors was sufficient for a change in cirrus conditions in terms of cirrus presence and intensity, which decreased the comparison quality. Claverie et al. [38
] observed this in images with a high cloud cover content as well. The geographic location of the Batagay site between two mountain ranges with very frequent cloud cover and highly variable cloud dynamics precludes better same-day image pair qualities. Arguably, the Batagay image pair should therefore be considered with additional caution; however, in spite of the slightly less favorable adjustment fit, the Batagay study site still had good overall adjustment results and illustrated the applicability of this approach also for challenging areas.
All scatter plots had a reasonable amount of scatter around the one-to-one relationship. Experiments with simulated Landsat-8 and Sentinel-2 data by Zhang et al. [33
] demonstrated that the scatter was emerging from the differing spectral response functions of the bands. Another source for value scatter likely was the different spatial resolutions of the sensors. Mandanici and Bitelli [34
] performed tests that showed that spatially-heterogeneous target surfaces created a lower Pearson correlation coefficient in band comparisons [34
]. The Sentinel-2 sensors’ high-resolution bands most likely capture the landscape heterogeneity to a greater extent compared to Landsat-8, which then is not averaged out completely in the resampling process to 60-m pixels. Flood [39
] and Roy et al. [69
] further looked at misregistration, different viewing geometries, and atmospheric path length as potential sources for errors between the sensors. All these points contribute to the scatter between Sentinel-2 and Landsat-8 reflectance values. Overall, we are confident with our adjustment results as the Pearson’s correlation coefficients were high throughout the analyses.
When inspecting the scatter plots, a noticeable number of pixels displayed data discrepancies in the NIR (B8A/B5, B8/B5) and SWIR (B11/B6, B12/B7) band comparisons of Yakutsk and the Lena Delta. In the Lena Delta, Landsat-8 captured little reflectance variability in bands B5 and B6 (between 0.08 and 0.13), while Sentinel-2 recorded reflectance values from 0–0.35 in bands B8A, B8, and B11 for the corresponding pixels. In Yakutsk, the relationship was reversed, where Sentinel-2 recorded nearly no variation in values (0–0.05) in bands B8A, B8, and B11, while Landsat-8 indicated values ranging from 0–0.2 in bands B5 and B6 for the same pixels. This data illustrated the different sensitivities of the two sensors in signal reception and translation to reflectance values. We identified these pixels by building masks to only select the pixels associated with these discrepancies. The majority of these pixels were water or mixed water-land pixels. In the Lena Delta, those pixels were mainly from sandbanks. Despite applying a water mask, these pixels were not masked as they did not fully comply with water characteristics only. The sandbanks have a high bottom reflectance through the thin water column, resulting in mixed reflectance signals, and the extent of water coverage changes strongly with water level changes. Therefore, the water mask failed to mask these pixels properly, and evidently, the Landsat-8 and Sentinel-2 NIR and SWIR bands picked up and converted the mixed signals differently. Likewise, the striking pixels in the Yakutsk NIR and SWIR bands were water body pixels, as well. The Yakutsk area contains many thermokarst lakes and ponds that vary in size and depth, which fosters the likelihood of pixels containing a mixed water and ground signal. Vuolo et al. [35
] made similar observations in pixel differences and accredited the bigger differences to three error sources: heterogeneous and complex terrains, the quality of cloud and cloud shadow masking algorithms, and the challenging and differing atmospheric correction results. Our analyses and study site characteristics confirmed these points and challenges, which underlines that a spectral band comparison can only be based on clear, high-quality images, as reflectance artifacts decrease the quality of the comparison and the adjustment scheme results. Further attention has to be given to the cloud, cloud shadow, and water masking. Under challenging conditions, a manual selection and masking of clear areas for analysis may be inevitable. The NDVI comparison results showed that both Landsat-8 and Sentinel-2 NDVI captured similar vegetation traits and that green vegetation estimations derived from either system would be very similar. However, in this case, it was inevitable to adjust the Sentinel-2 bands spectrally to Landsat-8 since a moderate amount of non-adjusted Sentinel-2 NDVI values were negative in contrast to the Landsat-8 NDVI values.
The overall results demonstrated that Sentinel-2- and Landsat-8-acquired reflectance values for corresponding pixels correlated well in Arctic-Boreal permafrost regions in Eastern Siberia. This was observed for all three study sites separately and also in the combined adjustment for the Eastern Siberian region. This underlines the expected sensor compatibility and agrees with results from similar studies [33
]. However, our study went slightly beyond some of the previous ones, which partly based their evaluation on simulated Landsat-8 and/or Sentinel-2 image products [33
], and not on acquired images. We showed that it is possible to reduce the persistent difference in spectral responses between the measured reflectance values of the spectral bands by deriving and then applying an ordinary least squares regression model, adjusting Sentinel-2 to resemble Landsat-8 reflectance values in Arctic-Boreal permafrost regions. It is possible to create a seemingly perfect fit between spectrally corresponding bands at the individual study site. The best adjustment results were obtained in the local assessments (LDA, BatA, YukA), but closely followed by the regional approach (esA). The ES analysis shows that combining regional data, and deriving overarching adjustment coefficients will lead to a very good fit between Landsat-8 and ES adjusted Sentinel-2 reflectance values in Eastern Siberia as well. We therefore considered our derived adjustment coefficients to be representative for Eastern Siberia and that they were applicable for a Sentinel-2 to Landsat-8 adjustment throughout Eastern Siberia. The implementation of the globally-applicable HLS coefficients did not improve the spectral comparability of Landsat-8 and Sentinel-2 reflectance values at any of our three Arctic-Boreal study sites or on the regional ES level. This might be because only a few Arctic study sites and merely one Siberian site were included in the HLS product, and therefore, the Eastern Siberian landscape and vegetation forms were not represented enough in the HLS study. This finding is similar to the adjustment results from Flood [39
] for Australian landscapes, and thus, we do not recommend using globally-derived adjustment coefficients for local or regional scale studies, but instead determine region-specific coefficients for a Sentinel-2 to Landsat-8 adjustment.
Being able to adjust Sentinel-2 to Landsat-8 gives the possibility of combining Landsat and Sentinel-2 data in analysis applications. Here, we relied on the already existing data continuity strategy for the Landsat sensors and would like to extend the range of sensors by adjusting Sentinel-2 to Landsat-8 [29
]. Considering the substantial increase in available multispectral images for the high latitudes with Sentinel-2 (Figure 1
], the creation of dense time series is possible. A combined Landsat and Sentinel-2 time series will improve land cover and vegetation mapping and monitoring approaches in detecting shifts in vegetation composition and structure, as well as permafrost region disturbances in a warming climate. A key advantage of dense time series analysis is the capability to detect landscape dynamics and to differentiate between rapid and gradual changes, therefore describing permafrost region disturbances better and being able to conduct detailed trend analysis in Arctic-Boreal permafrost regions.
One of the main motivation to combine Landsat-8 and Sentinel-2 in high latitudes is the frequent cloud cover and therefore the low data availability of cloud-free images. To avoid this disadvantage of optical remote sensing altogether, one can consider using Synthetic Aperture Radar (SAR) sensors that penetrate the cloud cover and acquire data during bad weather conditions, as well. SAR data were applied to a variety of studies focusing on different aspects and applications in northern high latitudes, such as land cover mapping, successfully already [20
]. Nonetheless, one general conclusion is that the best results can be obtained from polarimetric SAR data together with optical remote sensing [71
]. Despite the advantages of active sensors, in particular SAR, towards frequent cloud cover, we trust assessing permafrost landscape changes with optical remote sensing. On the one hand, the optical remote sensing data archive, mainly Landsat, is the longest archive spanning more than 45 years, providing extensive records of the past. On the other hand, we see high potential in change detection and time series segmentation algorithms, such as LandTrendr, which we would like to apply to permafrost regions and assess disturbance regimes. However, the availability and advantages of SAR data will be an asset when assessing permafrost disturbances, as SAR data can potentially fill data gaps and help as supplementary material to identify and evaluate permafrost landscape disturbances well.
The overall objective of our study was to compare the spectral characteristics of Landsat-8 and Sentinel-2 across Arctic-Boreal permafrost regions and to assess their compatibility and potential use in a combined time series. Our spectral comparison of Landsat-8 and Sentinel-2 corresponding bands showed that these two sensors correlate well and depict similar trends with only minor differences in Eastern Siberian permafrost regions. Depending on the application purpose, one has to assess whether these differences are negligible or of relevance. With an ordinary least squares regression based on band comparisons of corresponding Landsat-8 and Sentinel-2 bands, we can successfully adjust the Sentinel-2 reflectance values to resemble Landsat-8 more and homogenize the two data sets. To avoid geolocation errors and account for the sensors’ different spatial resolutions, we downscaled our images to 60 m.
The comparison method was valid for adjustments on the local level for sites with different ecological and landscape characteristics, but also when merging data across Eastern Siberia and conducting the spectral comparison and adjustment regionally. This underlines the comparability and compatibility of the two sensors. In contrast, we found that the application of the Harmonized Landsat Sentinel-2 product generated no improvement in spectral adjustment. Relying on global adjustment schemes between Landsat-8 and Sentinel-2 is therefore not advisable for Eastern Siberia or for local to regional time series studies in general. Therefore, we recommend at least regionally-fitted spectral adjustments when jointly using Landsat-8 and Sentinel-2 data products. This minimizes the potentially introduced error from different spectral band responses when combining data from multiple sensors. Our data enable the combined use of Landsat and Sentinel-2 in future time series analysis and other landscape change approaches in Arctic-Boreal permafrost regions of Eastern Siberia. Despite the overall good adjustment results, small differences between the sensors remained, which can be attributed to several factors including heterogeneous terrain, poor cloud and cloud shadow masking, and mixed pixels.
Our assessment was solely based on the use of open source software (GEE, SNAP, Jupyter Notebook) and freely-available data sets (Landsat-8, Sentinel-2). This makes the methodological approach reproducible and allows wide usage also for other study areas. As our results were in line with previous local and regional studies, we acknowledge that these services and softwares are a valuable and useful tool appropriate for such comparison studies.