In order to study tundra–forest transition, we selected the monitoring sites satisfying two predefined criteria: a.

The sites are covered by historical high-resolution satellite imagery – Corona archive .

Based on these criteria, we identified three main study areas (Table , Fig. ). Area 1 (the largest area) is located on the east coast of the river Nadym; area 2 is between the south coast of the Gulf of Ob and the river Yarudey, and area 3 is to the west of the city of Nadym (Fig. ). Detailed maps of the study sites can be found on the geoportal “Nadym. Changes in 50 years (1968–2018) (NC50)” (https://ageoportal.ipos-tmn.ru/nadym/, last access: 3 May 2020). A description of the geoportal is given in Appendix A.

The northern part of study area 1 is in the continuous permafrost zone, while its southern part and study areas 2 and 3 are in the discontinuous permafrost zone . The vegetation zones include southern tundra, the forest–tundra ecotone and northern taiga , with the treeline crossing the study sites . The southern tundra zone is represented by dwarf-shrub, moss and lichen tundra; polygonal peatlands; and dwarf-shrub, lichen woodlands. The vegetation community of the forest–tundra consists of larch and spruce–larch, dwarf-shrub moss woodlands, and dwarf-shrub lichen tundra. The northern taiga is represented by larch and spruce–larch dwarf-shrub and moss–lichen forests and shrub–moss–lichen wetlands in flat or hilly terrain. The boundaries of these vegetation zones in the study areas, according to , are shown in Fig. . Further on, we use the definition and boundaries of vegetation zones following topographic maps. A description of plant species and the permafrost state at several sites obtained during the field campaign in 2019 is provided in Appendix B.

In addition, Fig.  shows the forest mask derived from the topographic maps by Ilyina et al. (1985). Forested areas include dense forests, shrubs in the floodplains and river deltas, and sprout forest. Overall, 3570 km2 or 17.5 % of the total study area was covered by forests, out of which 15.5 % was dense forest, 1.2 % was sprout forest and less than 1 % was shrubs. Dry tundra (“moss and lichen tundra” in the topographic maps) covered 11 370 km2 or 55.7 % of the total study area.

Anthropogenic activity in the study areas includes oil and gas mining (fields Medvezhye and Yarudeyskoe, major gas pipelines Yamburg-Tula, Yamburg-Yelets, SRTO-Ural, Urengoy-Center, Urengoy–Uzhgorod). The first geological surveys started in the 1950s, and industrial exploration of the areas started in 1967. A new railway road, the Northern Latitudinal Railway (Salekhard–Novy Urengoy), is currently under construction within study area 1. Environmental impact due to anthropogenic activities and climate change has been monitored since the 1970s .

In order to assess the climatic conditions in the region, we used meteorological data (http://meteo.ru, last access: 10 July 2019) from three stations, Novy Port, Nyda and Nadym (Supplement), located along a 200 km latitudinal transect from the north (Novy Port) to the south (Nadym). Novy Port station (67∘41′ N, 72∘52′ E; 12 ma.s.l.) is located to the north of the study areas. This is the closest station to the northern boundaries of the study areas. Nyda station (66∘37′ N, 72∘57′ E; 10 ma.s.l.) is located within study area 1. Nadym station (65∘32′ N, 72∘32′ E; 7 ma.s.l.) is located near the southern borders of the study areas (50 km from study area 3). The data set contained 52 years of quality-checked meteorological observations (from 1966 to 2018), including air temperature (time resolution 3 h), precipitation (time resolution 12 h), and cloudiness 0 to 10 (time resolution 3 h). Based on this data set, we calculated the mean monthly temperatures (the sum of daily mean temperatures divided by the number of days in a month), monthly cumulative precipitation (the sum of daily precipitation measurements for a given month), the daily minimum and maximum temperatures for all the years, and the mean annual temperature and annual precipitation.

In order to assess climatic factors influencing vegetation dynamics, we studied the length and mean temperature of the growing season, which we define as the period with daily mean temperatures above +5 ∘C . We calculated growing degree days following : 1GDD5=∑T‾≥5(T‾-5), where T‾ is the daily mean temperature [∘C]. Another important abiotic factor for vegetation is humidity. We calculated potential evapotranspiration (PET) from the temperature and cloudiness measurements using Bonan's modification of the Taylor–Priestley model . We used PET and precipitation to calculate a dryness index (DI) , i.e., the ratio between annual PET and precipitation (P): 2DI=PETP. Based on GDD5 and DI, we classified the vegetation zones in Novy Port, Nyda and Nadym using the thresholds introduced in the bioclimatic model SiBCliM (see Table  based on the data from ). Due to the fact that the stations are located along the latitudinal transect close to the study sites, they should be representative of sites' vegetation state.

The initial state of the study areas was assessed using Corona imagery. Corona is a US program running from 1958 to 1972, which used satellite surveillance systems to obtain high-resolution photographic coverage from USSR–China and some other territories . The ground resolution of the imagery for subsequent KH-1 to KH-4 missions was continuously improved from 12 m (KH-1, 1960) to 2 m (KH-4, 1967). We identified 21 KH-4b frames under clear-sky conditions from 21 August 1968. Each frame consisted of four scanned fragments. The cropped fragments without color correction were georeferenced to the chosen orthomosaic (SPOT layer, Sect. ) using a third-degree polynomial method in software ArcGIS (v. 10.4.1). The RMSE in georegistration estimated using control points did not exceed 10–12 m. Then the fragments were organized into separate paths, and finally a mosaic was formed using the mosaic operator Last in ArcGIS. This mosaic characterized the state of study areas at the start of the active industrial exploration of the region (1967).

Further, we used Landsat Level-1 data to quantify burned areas for particular years within 1968–2018 . The data providing the best coverage of the study areas were available from the following years: 1988, 2001, 2016 and 2018 (Table ). The images were synthesized using near- and mid-infrared channels (Landsat 5 and 7 – 0.63–0.69, 0.76–0.90 and 1.55–1.75 µm; Landsat 8 – 0.64–0.67, 0.85–0.88 and 1.57–1.65 µm) as burned areas are identifiable in the infrared range of wavelengths. Landsat mosaics for all years were formed after color correction using the mosaic operator Blend in ArcGIS.

Mapping and quantification of the burned areas were performed by means of an object-based image analysis, successfully used for studies of landscape dynamics . In the first stage, we performed segmentation of mosaics using the algorithm “Multiresolution Segmentation” in eCognition software (v. 9.0). The segmentation was performed using parameters 40 for scale and 0.5 for color. The second stage, classification, was different for Corona and Landsat mosaics. In Corona mosaics, the spectral information was missing and we had to rely on the contrast of colors between non-disturbed tundra and burned areas. For visual classification we applied two criteria. First, non-disturbed tundra is lighter due to the presence of lichen in the vegetation community, whereas recently burned areas are dark. Second, burned areas are characterized by well-defined boundaries often coinciding with river coastlines. An example illustrating segmentation and the visual choice of burned areas is shown in Fig. . For Landsat mosaics, we used an unsupervised classification iterative self-organizing data analysis technique (ISODATA) (15 classes, a change threshold of 5 %). Further, we identified visually one or two classes corresponding to burned areas. The segments containing more than 90 % of pixels within these classes were identified as burned areas. The segments with 40 %–90 % of pixels within these classes, located at the periphery of large fires, were added to burned areas based on visual imagery check. Calculation of the areas of segments classified as burned areas was performed using the standard instrument Calculate Geometry in ArcGIS.

We studied the percentage and distribution of the burned sites and calculated the frequency of fire return. Corona and Landsat images showed that some years were characterized by particularly large-scale fires in the study areas (see an example for 1990 in the Supplement, Fig. S1). These years are referred to as the years of major fires. The burned areas can be detected in the satellite images for a few years after the fire. Burned areas in the Corona mosaic from 1968 were partially dated back to the fires in the period 1953–1964 based on geological surveys and early Corona images (the sources are listed below Table ). Landsat mosaics from 1988, 2001, 2016 and 2018 largely reflect the state of the study areas after the major fire years 1976, 1990, 2012 and 2016 (Table ). For a given territory, the fire interval was calculated as the difference in years between the major fires.

We have studied the vegetation dynamics using two methods. First, we used normalized difference vegetation index (NDVI) distributions to assess the state of vegetation in the sites burned at different times. Second, we studied the tundra–forest shift of dry tundra during the period of ca. 60 years visually comparing historical imagery to modern satellite data of high spatial resolution. A similar visual method was used by, e.g.,  to study transition from grassland to shrubs in northern Eurasia.

Below we describe the method based on the NDVI. This index reflects the state of biomass, i.e., vegetation greenness, density and development . The NDVI was calculated from the red (R) and near-infrared (NIR) radiation reflected by vegetation: 3NDVI=NIR-RNIR+R. We chose two scenes from one path of Landsat 8 from 30 June 2018 (Table ). We used Level-2 data (Committee on Earth Observation Satellites – CEOS) after atmospheric correction by the standard Landsat 8 Operational Land Imager (OLI) atmospheric correction algorithm . The NDVI was calculated in ArcGIS using standard tools.

We analyzed NDVI distributions for study area 1, separately for dry tundra burned in different years (as detected in satellite imagery mosaics from 1968, 1988, 2001, 2018) and for non-disturbed dry tundra. We considered the territories covered by one satellite overpass. Moreover, analysis of burned tundra included only the sites burned once within 60 years. In order to make the study areas more balanced by size, we merged the data sets from the burned territories detected in 1968 and 1988, for which we might expect that vegetation had recovered after the fires, and omitted 2016 due to insufficient size of the fire-affected territory. As a result, all study areas were larger than 600 km2 (1968 + 1988 – 600 (405 + 195), 2001 – 1565, 2018 – 867 km2, non-disturbed – 937 km2). Using the dates of major fires (Table ) and the date of the Landsat mosaic (2018), we can assume that the NDVI within burned tundra sites detected in 1968 and 1988 reflects the state of vegetation after fires more than 42 years ago, that in 2001 reflects the state of vegetation after fires 28 years ago, and that in 2018 reflects the state of vegetation after fires 2 years ago. The NDVI in non-disturbed tundra refers to tundra not affected by fires during the whole study period.

The visual method can be described as follows. The data sets used in the analysis of vegetation dynamics are summarized in Table . The initial state of vegetation was assessed using the Corona mosaic (Sect. ) and topographic maps. The topographic maps were used to develop forest and dry tundra masks using automatic tracing in Easy Trace software (v. 8.7). The resulting vector layer was checked and corrected using the Corona mosaic.

Assessment of the current state of vegetation was based on SPOT and Resurs-P data (Table ). The study areas were almost fully covered by the mosaic of SPOT-6 and SPOT-7 imagery. Ca. 10 % of the study areas were covered by three paths of Resurs-P (product level 1C; channels blue, green, red and NIR). Resurs-P data were co-registered to the SPOT mosaic in ArcGIS. The data were synthesized in the visible-radiation range without color adjustment. The SPOT mosaic was used as a pluggable web map layer without additional processing.

Firstly, based on the topographic map and Corona mosaic, we identified two types of dry tundra sites: those burned before 1968 (Sect. ) and those not affected by fires during the whole study period. The area of burned tundra was 1090 km2, and the area of non-disturbed tundra was 5300 km2. Then we introduced random sample circles with 100 m diameters into the subsets. The samples were spaced at intervals of 2 km within burned tundra (157 samples) and at intervals of 5 km within non-disturbed tundra (231 samples). We included only the circles for which we could confidently state that there were no trees in 1968. In the modern imagery, we distinguished between the following types of vegetation: tundra (no trees), woodland (trees covered less than 50 % of the area inside the circle) and forest (trees covered more than 50 % of the area inside the circle). We visually compared the areas inside the circles in the historical and modern images (see examples in Fig. ). We introduced three change classes: “no change”, when tundra was identified in both the old and the modern images; “to woodlands”, when tundra turned to woodlands; and “to forest”, when tundra turned to forest.

To quantify the topographic slopes, we used a digital elevation model ArcticDEM (Table ). The model is a mosaic with a high spatial resolution (2 m). The stereo couples for this version of the model were prepared in 2010–2017. It is currently the most detailed source of information on the relief. Co-registration of the mosaic elements, calculation of slopes and preparation of the raster layer for the geoportal NC50 (Appendix A) was carried out in ArcGIS.

We calculated topographic slopes within the circles where we visually assessed the vegetation shift. For calculation, we used circles with a 50 m diameter located in the centers of larger sample circles (Sect. ). For each circle, we calculated the mean slope. We separated data corresponding to burned tundra and non-disturbed tundra and further stratified them by vegetation shift (no shift, to woodlands, to forest), thus splitting the data set into six classes. Finally, we calculated the median values as well as the 25th and 75th percentiles of mean slopes for each class.

During the last 50 years, the mean annual temperature at the meteorological stations has increased by 2.6–3.0 ∘C (Fig. ): from -9.5 to -6.5 ∘C in the north and from -6.5 to -4 ∘C in the south. In addition, the daily minimum temperature has increased by 4–5 ∘C (from -44 or -45 to -40 ∘C, in the Supplement, Figs. S2–S4).

The growing season typically lasts 3 months in the north and 3–4 months in the south (in the Supplement, Figs. S2–S4). The length of the growing season has increased by 24 d (in the Supplement, Figs. S1–S3): from 72 to 97 d in the north (Novy Port) and from 97 to 121 d in the south (Nadym). The mean temperature of the growing season has increased by 1.0–1.4 ∘C. In general, both the length and the mean temperature of the growing season exhibit great variability particularly during recent years (in the Supplement, Figs. S2–S4). Following the mean temperature and the length of the growing season, the growing degree days, GDD5, have increased over the last 50 years (Fig. ). The strongest increase in GDD5 (ca. 200 degree days) was observed during the 20 years from 1966 to 1985, after which GDD5 leveled off. The period from 1980 to 1995 seemed to be the most favorable period for new-forest establishment due to the relative stability of mean temperature and GDD5. On the contrary, the last decade was characterized by a strong variability in both mean annual temperature (the peak-to-peak value was ca. 3 ∘C) and GDD5 (the peak-to-peak value was larger than 400 degree days), which could make seedlings' survival less probable.

Mean annual precipitation increases from the north to the south: from 330–340 mm in Novy Port to 510 mm in Nadym (in the Supplement, Fig. S5). Precipitation in the northern part (Fig. ) is characteristic of the arctic regime with the annual value below 400 mm, whereas precipitation in the southern part (Fig. ) is characteristic of the temperate continental regime . Monthly variability in precipitation follows a well-pronounced seasonal cycle (in the Supplement, Fig. S5) with a maximum in August (July–August in the south). While the seasonal cycle of precipitation was similar in three stations, the inter-annual variability was not. Precipitation time series in Nyda and Novy Port closely followed each other between 1965 and 2005 (Fig. ). Since 2005, the annual precipitation has drastically decreased in Novy Port and it has increased in Nyda. Potential evapotranspiration increased at all three sites. Accordingly, the dryness index (DI) increased from 0.4 to 0.5–0.6 everywhere (in the Supplement, Fig. S6). During a few recent years, the DI in Novy Port exceeded unity several times.

Based on the observed growing degree days and DI, we assessed the change in the vegetation class at three stations (Table ). The vegetation class has changed at two sites (Nadym and Novy Port). This is a result of an increase in GDD5, while the DI was not a limiting factor for the vegetation class at any of the three stations. Due to the fact that GDD5 is a function of air temperature and the latter depends mainly on latitude, GDD5 from the three stations should be representative of our study areas. Therefore, we can conclude that (1) our study sites should have experienced vegetation change during 60 years, specifically from tundra to forest in the north, and (2) temperature and moisture regimes throughout our study areas are theoretically suitable for forests.

Weather conditions in northern Siberia can support large-scale wildfires during summertime. The fires occur mainly in July . Burned areas after five major fire events (Table ) are shown in Fig. a. The burned areas in Fig. a can overlap if the same site experienced more than one fire. Separate maps of the burned sites for each of the five cases (Table ) are available on the geoportal NC50 (Appendix A).

The overall area of burned sites was ca. 8300 km2, constituting more than 40 % of the total study area. Figure b shows the distribution of the fraction of burned area over years. The largest fires, contributing more than 30 % to the total burned area, occurred between 1989 and 2001 (mainly in 1990, Fig. S1). The fires in each of the periods 1953–1968, 1968–1988 and 2016–2018 contributed 10 %–20 % of the total burned area (Fig. b). In addition, ca. 10 % of that area was recognized in both 1988 and 2001 imagery. According to Fig. b, ca. 80 % of the burned area experienced fire once and 20 % experienced fires more than once during 60 years. Approximately 17 % of the fire-affected territory (1400 km2) burned twice; 2.5 % of the territory (200 km2) burned three times, and 0.2 % (20 km2) burned four times. These sites were characterized by a remarkably small period between consequent fires, which was estimated as 15–60 years if only major fires were taken into account (Table ).