―A satellite-derived database for stand-replacing windthrow events in boreal forests of European Russia in 1986–2017‖

. Severe winds are among the main causes of natural disturbances in boreal and temperate forests. Here, we present a new GIS database of stand-replacing windthrowswindthrow events in the forest zone of the European RussiaEuropean Russia (ER) for the 1986–2017 period. Delineation of windthrowswindthrow areas was based on the full Landsat archive of Landsat images and two Landsat-derived products on forest cover change, namely the Global Forest Change and the Eastern‘ Europe Forest Cover Change datasets. Subsequent verification and analysis of each windthrow was carried out manually to 15 determine a the type of related storm event, its date or date range, and geometrical characteristics. The database contains 102,747 elementary areas of damaged forest that were combined into 700 windthrowswindthrow areasevents caused by 486 convective or non-convective storms. The database includes stand-replacing windthrowswindthrow only, which with an area > 0.05 km 2 5 ha and > 0.25 km 2 25 ha for events caused by tornadoes and other storms, respectively. Additional information such as contained weather station reports and event description from media sources is also provided. The total area of stand- 20 replacing windthrow amounts to 2966 km 2 , that is 0.19% of the forested area of the study region. Convective windstorms contribute 82.5% to total wind-damaged area, while tornadoes and non-convective windstorms are responsible for 12.9% and 4.6% of this area, respectively. Most of windthrowswindthrow events in the ERER happened

(1) The assessment includes a lot of subjective calls from the interpreter and there is no formal validation. I fully understand that validating the database is very challenging, but I would have at least expected some assessment on how accurate the data is. There are many steps involved in collecting the data, and some of them seem to be highly dependent on local knowledge. From this I assume that the accuracy will be higher in some regions (where there is ample local knowledge), but lower in others. In particular, I was wondering whether you searched the forest area of ER systematically or whether you applied any other "sampling strategy" to ensure you don't miss a storm event. Moreover, the accuracy will depend on the availability of HRI. Was HRI available for each year after 2001, or only for selected years? The authors address some of these issues in the discussion, which I appreciated, but maybe they should make clear from the beginning on that the database is a subjective collection and probably far from complete and/or consistent.
We agree that due to several limitations of the method and satellite data the database is spatially and temporally inhomogeneous and hence incomplete. We have added corresponding information into the Abstract. Sections 2, 4, and 6 were also amended.
We devoted the whole Section 6 to discussing of factors influencing the data accuracy including such factors as percentage of forest-covered area, forests species composition and forest management practices. In general, our data has highest accuracy for low-populated northern and eastern part of the ER, where forests cover 70-90% of the territory and dark-coniferous forests are widespread. In turn, the data may be less accurate for southern part of the study area, where some windthrow areas probably could be missed. The relevant information has been added to Section 6.
Each windthrow from our dataset was validated (that is, it was clarified that it is actually a windthrow), based on pre-and post-event Landsat/Sentinel-2 images, high-resolution images and additional information. So, the probability that any forest disturbance was mistakenly referred to a windthrow is minimal. However, we agree with the reviewer, that some thresholds used in the data collection process, in particular the thresholds for the minimum area of EDAs and windthrow, as well as the threshold for the minimum distance for separating successive windthrow, are somewhat subjective.
We should stress, that the searching of windthrow areas based on GFC/EEFCC was systematic. The GFC/EEFCC-based collection of forest loss areas with windthrow-like signatures was carried out separately for each region of the ER. A grid with 50 km cell size was built inside the region, which helped to organize the searching of windthrow-like forest disturbances. The relevant information has been added to Section 4.1.1.
(2) In the same line as the previous comment, many of the decision to group patches to windthrows, etc. are based on arbitrary thresholds. I wonder whether the authors have tested the sensitivity of their results to those thresholds. The allocation of patches to windthrow events might look very different with slight changes in threshold values.
Indeed, in our method, we used a number of thresholds that have some subjectivity but are based on previous studies. For instance, the 10-km threshold for separate successive windthrow from each other is based on study of Doswell and Burgess (1988), who proposed the 5-10 miles (8-16 km) threshold for the gap to separate one skipping tornado from two successive tornadoes. The threshold for minimum area of EDAs was chosen based on study of Koroleva and Ershov (2012) who showed high uncertainty of estimated geometrical characteristics of small-scale windthrow (less than 1800 m 2 , i.e., two GFC pixels). In addition, we decided to filter out such small-scale disturbances since it is virtually impossible to confirm their wind-related origin.
We performed sensitivity test for the latter threshold and found that the absence of minimum accepted area for EDAs will increase area of windthrow by 2-3% on average (up to 6%). The optimization of other threshold values can be evaluated in further studies that should involves ground-based data.
We have added this information to Sections 4.1.1 and 4.1.4.
(3) You do not explain how the Landsat data was processes. Please give some details on the processing.
A new figure 3 has been added that shows an example with all three hierarchical levels of the database, explains the determination of geometric characteristics of storm event, and shows the examples of parallel and successive windthrow. Correspondingly, numbers of all subsequent figures have been changed.

line 208-210: the choice of thresholds is not clearly defined
The paragraph has been rewritten for clarity.
We estimated threshold value from the statistics of ∆NDII raster. Firstly, we obtained the mean value and standard deviation of ∆NDII within the entire forest-covered area on image. Standreplacing forest disturbance inherently has ∆NDII values substantially higher than the image average. To separate stand-replacing forest disturbance from other forest-covered area, we used the threshold value of two standard deviations, which was previously tested by Koroleva and Ershov (2012). However, in some cases the ∆NDII distribution within the entire image was skewed (e.g., due to the presence of cloud decks or haze on the post-event image). In such cases, we lowered the threshold value of ∆NDII iteratively by comparing the detected changes with results of visual identification of windthrow on a post-event image (using several examples located in different parts of windthrow). As a result, actual threshold values ranged from 1.5 to 2 standard deviations. Then, a binary raster of detected changes (i.e., forest losses) has been created (see fig. 4d) and converted to a shapefile. This information has been added to Section 4.1.3.

lines 220-240: is not clear how the thresholds used in this section have been derived
We used the 10-km threshold as the slightly modified value proposed by Doswell and Burgess (1988) who suggested to use the 5-10 miles (8-16 km) threshold for a gap to discriminate between one skipping tornado and two successive tornadoes. The reference to (Doswell and Burgess, 1988) has been added to the revised version of the manuscript.
line 261: the ratio of length to width of 10:1 is based on previous studies conducted for European Russia?
Yes. This is typical ratio for tornado length and width for US (Schaefer and Edwards, 1999) and for Northern Eurasia .
The clarification has been added to the revised version of the manuscript.   Corrected.

line 472: is not clear what the authors meant by "contain fewer plots"
A typical tornado-induced windthrow event on average contains less EDAs than non-tornado induced one. The word -Plots‖ has been replaced by -EDAs‖.

Interactive comment from Barry Gardiner
Main Points 1. Language needs some help. I would recommend the paper being checked by a native English paper because in some places the meaning was not totally clear.
Language has been improved.
2. Unlike the earlier reviewer I was really impressed at how much effort was put into assessing the potential errors in the assessment. I also think the authors were completely honest about limitations in the data and their assessments. Such honesty will give really help users of the data.
In the revised version, we made minor changes according to the first reviewer suggestions.

Minor Points
1. I would recommend referencing the paper Ulanova, N.G., 2000. The effects of windthrow on forests at different spatial scales: a review. For. Ecol. Manage.135, 155-167. https://doi.org/10.1016/S0378-1127  This paper deals with the effects of wind in the Russian boreal forest. It is a brilliant and seminal paper.
The reference to this paper has been added in the Introduction section.
3. I think early on you need to tell the reader that your analysis was done manually and not automated. That is important to understand the process. Also refer whenever possible to Fig. 2 which is super useful.
We have added the word ‗manually' into the Abstract. Additionally, the short sentence on the manual character of the data analysis has been added to the beginning of Section 4. The additional references to figure 2 have been added throughout the text.
4. Page 6. I think some more justification of the limits you set would be helpful, for example the lower limits on size of EDAs.
We removed all EDAs with an area ≤ 1800 m 2 that equals to area of two GFC pixels. We filtered out such small-scale disturbances since it is virtually impossible to confirm their wind-related origin. Moreover, the area of local windthrow can be almost three times overestimated by Landsat images (Koroleva and Ershov, 2012). Thus, this is the balance between slight underestimation of the total area of windthrow (by 2-3%) in the study region and substantial overestimation of the number of EDAs mistakenly referred to windthrow. We have decided to choose the first option.
We have added this rationale to Section 4.1.1.
5. Page 7. What did you do in highly populated areas? Did you just accept that you will have missed wind damage on some occasions?
Indeed, the windthrow data obtained for the period before 2000 (using the EEFCC dataset) may be incomplete for highly-populated regions of the ER due to assignment of forest losses to broad periods, i.e., 1986-1988 and 1989-2000. To partially avoid missing of windthrows, using Landsat images, we performed additional verification of all large-scale forest loss areas (with area more than 5 km 2 ) in these regions independently of their geometry, since windthrow areas can be totally masked out by logged areas. Thus, we were able to find three large-scale windthrow events in highly-populated regions of the ER. However, some windthrow events can still be missed.
The clarification has been added to Section 4.1.2.
The paragraph has been completely rewritten according to the suggestions of the first reviewer.
7. Page 8, line 222. Again justify the figure of 5-10 km. Do you use "rules" or is this judgement left to the observer?
We used the 10 km threshold, which is in the range of 8-16 km (5-10 miles) proposed by Doswell and Burgess (1988) to discriminate between one skipping tornado and two successive tornadoes. Therefore, if the nearest EDAs were located at a distance more than 10 km from each other, they belonged to different windthrow. If the distance between them was less or equal to 10 km, they belong to one windthrow except for several cases. The exceptions were associated with changes of windthrow direction, transformations of one windthrow type to another identified by the HRI, and abrupt change of forest damage degree.
8. Page 9, line 261. I think "predominately" is better than "prevalently". It is of note, that we cannot determine whether the trees were felled or broken by the wind based on satellite images, even having very high resolution. Therefore, we use a single term -windthrow‖ for all types of wind-induced forest damage. This clarification has been added to the beginning of Section 5.
Successive windthrow areas induced by one storm event follow downwind one after another and approximately fall on one straight line (the angle of deviation from this line does not exceed 10-20°). Such windthrow are presumably induced by one convective cell generating a sequence of squalls or tornadoes. In contrast, parallel windthrow areas that located within one storm event are situated parallel to each other (with an angle less than 30°). They are presumably associated with two or more different convective cells or mesocyclones, generating squalls or tornadoes, often embedded into one mesoscale convective system. The examples of parallel and successive windthrow are shown at new Fig. 3 Dark-coniferous forests in the ER consist of three dominating tree species such as Picea abies, Picea obovata and Ábies sibírica. We have added this information to section 2.1.

References:
Doswell, C.A. and Burgess, D.W.: On some issues of United States tornado climatology. Monthly Weather Review, 116, 495-501, 1988. Foga, S., Scaramuzza, P.L., Guo, S., Zhu, Z., Dilley, R.D., Beckmann, T., Schmidt, G.L., Dwyer, J.L., Hughes, M.J., and Laue, B. Cloud detection algorithm comparison and validation for operational Landsat data products. Remote Sensing of Environment, 194, 379-390. http://doi.org/10.1016/j.rse.2017.03.026, 2017: European extra-tropical storm damage risk from a multi-model ensemble of dynamicallydownscaled global climate models. Natural Hazards and Earth System Sciences, 11, 2847-2857, doi:10.5194/nhess-11-2847 Estimation of error in determining the forest windfall disturbances area on high spatial resolution space images of LANDSAT-TM. In: Current Problems in Remote Sensing of the Earth From Space, 9, 80-86, 2012. ( Abstract. Severe winds are among the main causes of natural disturbances in boreal and temperate forests. Here, we present a new GIS database of stand-replacing windthrowswindthrow events in the forest zone of the European RussiaEuropean Russia (ER) for the 1986-2017 period. Delineation of windthrowswindthrow areas was based on the full Landsat archive of Landsat images and two Landsat-derived products on forest cover change, namely the Global Forest Change and the Eastern' Europe Forest Cover Change datasets. Subsequent verification and analysis of each windthrow was carried out manually to 15 determine a the type of related storm event, its date or date range, and geometrical characteristics. The database contains 102,747 elementary areas of damaged forest that were combined into 700 windthrowswindthrow areasevents caused by 486 convective or non-convective storms. The database includes stand-replacing windthrowswindthrow only, which with an area > 0.05 km 2 5 ha and > 0.25 km 2 25 ha for events caused by tornadoes and other storms, respectively. Additional information such as contained weather station reports and event description from media sources is also provided. The total area of stand-20 replacing windthrow amounts to 2966 km 2 , that is 0.19% of the forested area of the study region. Convective windstorms contribute 82.5% to total wind-damaged area, while tornadoes and non-convective windstorms are responsible for 12.9% and 4.6% of this area, respectively. Most of windthrowswindthrow events in the ERER happened in to the summer, which that is in contrast to Western and Central Europe, where windthrowswindthrowthey mainly occur in autumn and winter. Due to numerousseveral data and method limitations, of the data sources and the use of expert knowledge at several stages of the 25 data collection workflow,the compiled database isf probably far from completespatially and temporally inhomogeneous and hence incomplete. But despite this incompleteness, the compiledpresented database provides a valuable source of spatial and temporal information on windthrow in ER and can be used by both science and management.The compiled database provides a valuable source of spatial and temporal information on windthrowswindthrow in the ERER and can be used by both science and managementcan be successfully used both in forest science and severe storm studies. The database is 30 available at https://doi.org/10.6084/m9.figshare.12073278.v3 (Shikhov et al., 2020). Forests are is a valuable natural resource that is important for economy, society and sustainable development. Forests ecosystems are regularly exposed by to natural disturbance agents such as fires, droughts, insect outbreaks, and or windstorms. Being an intrinsic part of forest ecosystem dynamics (Attiwill, 1994;Seidl et al., 2017), natural disturbancess cause substantial environmental and economic damage (Schelhaas et al., 2003;Gardiner et al., 2010;van Lierop et al., 2015). In boreal and temperate forests, windstorms constitutes one of the main drivers of natural disturbancess (Ulanova, 40 2000;Forzieri et al., 20192020). In Europe, windthrowswindthrow contribute more than a half to the total area of natural disturbancess, including abiotic and biotic causes (Schelhaas, 2003;Gardiner et al., 2010).
Recently, disturbance regimes have changed considerably in many forest ecosystems worldwide (Seidl et al., 2011(Seidl et al., , 2017Senf et al., 2018). Particularly, both the occurrence and severity of disturbances both has increased in different regions, including those related to forest fires (Westerling, 2016;Kukavskaya et al., 2016), insect outbreaks (Kautz et al., 2017), and 45 droughts (Millar et al., 2015). Researchers have revealed a statistically significant increase of wind-related forest disturbances in the Western, Central, and Northern Europe (Seidl et al., 2014;Gregow et al., 2017), and in the European part of Russia (Potapov et al., 2015).
The observed increase in the frequency and severity of windthrowswindthrow events is associated with changes in forest structure like like increasing growing stock a rise of growing stock and median age, primarily in coniferous forests 50 (Schelhaas et al., 2003;Senf et al., 2018), and with climatic changes as well (Overpeck et al., 1990;Lassig and Moĉalov, 2000;Seidl et al., 2011;). An intensification of winter windstorms (Gardiner et al., 2010;Usbeck et al., 2010;Gregow et al., 2017) and an increase in the frequency and intensity of severe convective storms in a the warm season (Overpeck et al., 1994;Diffenbaugh et al., 2013;Chernokulsky et al., 2017;Radler et al., 2019) can be considered as the main climatic drivers for increasing of wind-related damage in boreal and temperate forests. 55 For correct attribution of forest windthrowswindthrow to particular causes, it is important to obtain corresponding data on such events. Recently, several long-term databases of windthrowswindthrow events in boreal and temperate forests, often together with other types of disturbances, have been collected at a national and macro-regionalinternational scale. The longest windthrowswindthrow data series have been compiled in Sweden (Nilsson et al., 2003) and Switzerland (Usbeck et al., 2010) based on literature reviews and forestry services reports. European Forest Institute presented compiled the 60 database of destructive storms in European forests for 1951−2010 (Gardiner et al., 2010). A new GIS database of wind disturbances in European forests has been compiled in 2019 by aggregating multiple datasets collected by 26 research institutes and forestry services across Europe (Forzieri et al, (20192020). It comprises more than 80.000 forest areas that were disturbed by wind in 2000-2018. Compare to other European countries, windthrowswindthrow events in Russia remain substantially understudied. Long-term databases of windthrowswindthrow events have been collected only for individual 4 regions, for example, for the Middle Ural (Lassig and Moĉalov, 2000) and the Central Forest Reserve in the Tver' region (Ulanova, 2000).
The main data sources of exist windthrow databases in Russia were the literature reviews, reports of forestry services, aerial observations and field investigations (Skvortsova et al., 1983;Lassig and Moĉalov, 2000;Ulanova, 2000). Meanwhile, satellite images have become the important data source for windthrowswindthrow monitoring in Russian forests in recent 70 decades (Krylov et al., 2012). Indeed, satellite data can be especially informative for studying Russian low-populated boreal forests, known in Russia as the taiga, which represent the largest forested region globallyon the Earth. They cover approximately 7.63 million km 2 , which is 22% of the world's forest areas (WWF Russia's boreal forests, 2007).
Use of satellite images for obtaining information on windthrow was proposed back in 1975 (Sayn-Wittgenstein and Wightman, 1975). However, the widespread utilising of satellite data to estimate the inter-annual variability of 75 windthrowswindthrowwind-related forest damage (e.g. Fraser et al., 2005;Baumann et al., 2014) became feasible after the openingpublication of the free-available Landsat archive (Wulder et al., 2012), and two Landsat-based products, namely the Global Forest Change (GFC) map (Hansen et al., 2013) and the Eastern' Europe Forest Cover Change (EEFCC) (Potapov et al., 2015). Thus, GIS databases of windthrowswindthrow events have been collected for some Russian regions based on Landsat archive and GFC data, i.e., for the Ural and north-eastern part of the ERER (Shikhov and Zaripov, 2018;Shikhov et 80 al., 2019), Kostroma region and adjacent areas (Petukhov and Nemchinova, 2014), and South Sakhalin (Korznikov et al., 2019). Shikhov and Chernokulsky (2018) found 110 previously unknown tornado-induced windthrowswindthrow areas in the ERER based on satellite images. However, for the entire ER, there are only rough estimates of storm-related forest damage (Potapov et al., 2015). The contribution of various weather phenomena like (convective and non-convective windstorms, snowstorms, or and tornadoes) to the total wind-induced forest damage, geometrical characteristics of 85 windthjrows and alsoas well as inter-annual and seasonal distribution of windthrow events remains unknown for the territory of ER. SuchThe appearance of such data can be helpful both for forest science and management andas well as for severe storms investigations.
In this study, we present a detailed GIS database of relatively large and stand-replacing windthrow events in the forest zone of the the ERER for the period 1986-2017. UnlikeThe database contains the above-described studies, this database will 90 contain the stand-replacing windthrow areas with indication of storm event types and dates, geometrical characteristics of windthrows areas, and additional information. To compile the data and determine theirthese attributescharacteristics, Wwe use the Landsat archivethe archive of Landsat images, the Landsat-based forest loss data products GFC and EEFCC, highresolution satellite images from the public map services, supplementary information including weather stations observations, databases on hazardous weather events, damage reports in the media sources, and reanalysis data. We describe the used data 95 and the study region in section Section 2, and explain the database structure in Section 3. Section 4 describes the windthrow delineation process and assessment of the geometrical parameters of windthrowswindthrow areas. Section 5 presents spatiotemporal variability of windthrowswindthrowwind-damaged areas and distributions of their geometrical characteristics. Section 5 6 discusses the main limitations of the method and the compiled dataset, while Section 6 draws the main conclusions of the paper. 100 2 Region and data

The study region
The study region includes the forest zone of the the ERER (Fig. 1) between the forest-steppe transition zone on the south and forest-tundra transition zone on the north. The availability of the EEFCC dataset determines the eastern boundary of the study region that broadly coincides with the Ural Ridge. 105 We used the 250-m resolution map of the vegetation cover of Russia (Bartalev et al., 2016) to estimate forest-covered area and dominant forest species (Fig. 1). Forests cover 54.6% of the study regions, while individual forested area within this region is typically > 100 km 2 . The most widespread dominant forest species are dark-coniferous (Picea abies, Picea obovata, Pinus sylvestrisÁbies sibírica), light-coniferous (Pinus sylvestris), small-leaved (Betula pendula, Betula pubescens, Pópulus trémula) and some broadleaved species (Tília cordáta, Quercus robur et al.) (Kalyakin et al., 2004). Secondary (re-grown 110 after logging or wildfires) small-leaved and mixed forests cover approximately 61% of the total forested area. Old-growth dark-coniferous forests are widespread on the western slope of the Northern Ural and the adjacent plain, and pine forests cover the largest area (>100 thousand. km 2 ) on the northwest of the the ERER (Fig. 1).

Initial data
We used multiple data sources to collect information on windthrowswindthrow events for the 1986-2017 period. 115 Particularly, we utilized satellite data to delineate windthrowswindthrow areas and determine a storm event type and used additional information to determine the dates of storm events.

Primary information for windthrow delineation and verification
The Landsat-based GFC data were utilised to search and delineate windthrowswindthrow that occurredtheforest areas affected by windthrow in 2001-2017. The data come as the integer raster with a 30 m cell size. It contains information on 120 stand-replacing forest disturbances at annual temporal resolutionforest disturbances that classified with a one-year step. In the boreal forest regions, the overall accuracy of the forest loss detection in the GFC is 99.3%, while user's and producer's accuracies are 93.9% and 88.0%, respectively (Hansen et al., 2013). Here, producer's accuracy is the ratio of correctly classified forest loss area to the actual forest loss area; user's accuracy is the ratio of correctly classified forest loss area to the same area according to the verified forest loss area. The data were downloaded from 125 http://earthenginepartners.appspot.com/google.com/gMG7KbLG.
The EEFCC dataset was used to search and delineate windthrowswindthrow the areas affected by windthrow occurred in 1986-2000. The data come as the integer raster with a 30 m cell size. It contains information on forest loss that classified into four broad periods : 1986-1988, 1989-2000, 2001-2006 and 2007-2012. This rough time determination is associated with 6 rareness of the Landsat images between 1989 and 1998. The detection of gross forest loss in the EEFCC has producer's and 130 user's accuracy of 88% and 89%, respectively (Potapov et al., 2015). The data were downloaded from https://glad.geog.umd.edu/dataset/eastern-europe-forset-cover-dynamics-1985-2012/. Landsat images (L1T processing level), i.e., images from the Landsat Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager (OLI), were used to confirm the wind-related nature of forest disturbances, 135 determine the storm types, dates (or ranges of dates) of windthrowswindthrow occurrence in 1986-2017. It addition, many windthrowswindthrow areas appeared occurred before 2001 were delineated with Landsat images (see Section 3.1.3 for details).
Sentinel-2 images were used to confirm the wind-related nature of forest disturbances, determine the storm types, dates (or ranges of dates) of windthrowswindthrow events occurrence for the 2016--2017 period. The data wereWe downloaded 140 Landsat and Sentinel-2 images from https://earthexplorer.usgs.gov/ and https://eos.com/landviewer.
High-resolution (0.5-2 m) satellite images, hereinafter HRI, were used to discriminate the type of a storm eventwindstorm or tornadocausing a windthrow. The HRI are available for several years from 2001 to the present (uUsually 2--8 high-resolution images are available for the entire period 2001--2017). No HRI are available before 2001. To view and 145 analyze HRI, we used mainly Google Earth Pro, while other The HRI images were downloaded obtained from Google Earth Pro and public map services (, i.e., Google Maps, Bing Maps, ESRI Imagery, and Here), were used to a lesser degreeESRI Imagery, and Google Earth Pro. We should highlight that tThe availability of HRI substantially varyied foramong different parts of ER. ThusIn particular, some areas in the northern part of the ER are not covered by HRI.

Additional information on storm events
Information of 3-hourly weather reports was used to determine storm event dates, and match the reported wind gusts, if any, with windthrowswindthrow events. We utilized information on observed wind speed, precipitation, hail and thunderstorm occurrence. The routine meteorological observations have been collected at 402 meteorological stations located within the studied area and have been initially processed at the All-Russian Research Institute of Hydrometeorological Information-155 World Data Center (RIHMI-WDC) from 1966 to the present (Bulygina et al., 2014).
Monthly reviews of hazardous weather events occurred in Russia, which are published in the Russian Meteorology and Hydrology journal (http://mig-journal.ru/en/archive-eng) but not translated, were also used to determine storm event dates for the 2001--2017 period. Additionally, these reviews contain the descriptions of hazardous weather events and damage reports. We included this information into our database. 160 The RIHMI-WDC database of hazardous weather events (Shamin et al., 2019) and information from regional departments of the Russian state weather service were also utilized to determine the dates of several storms that caused windthrowswindthrow events in 1986--2017.

7
Media news and witness reports in social networks, including photos and videos, were used for obtaining additional information on the type of event, i.e., tornadic or non-tornadic, for the 1986--2017 period. 165 Data from meteorological satellites Terra/Aqua MODIS (from 2001) and Meteosat-8 (from 2016) were used for obtaining additional information on storm events causing windthrow, especially to determine storm date and time. In particular, the Collection 6 MODIS Active Fire data (Giglio et al., 2016) were used to discriminate fire-and wind-related forest disturbances in 2001-2017. Data were downloaded from https://earthdata.nasa.gov/data/near-real-time-data/firms. Data from Russian weather radars (Dyaduchenko et al., 2014) were used only for several events occurred in 2012, 2014, and 170 2016 to determine the time of storm event causing a windthrow.

Structure of the GIS database
The compiled database of stand-replacing windthrowswindthrow events in the forest zone of the ERER in 1986-2017 is publicly available at https://doi.org/10.6084/m9.figshare.12073278.v3https://doi.org/10.6084/m9.figshare.12073278.v36 (Shikhov et al., 2020). We divided the spatial and attributive information on windthrowswindthrow events into three 175 hierarchical levels that correspond to three GIS layers, i.e., three shapefiles (.shp), in the database:  -Elementary damaged area‖ (EDA), that is a single-part polygon of wind-damaged forest;  -Windthrow‖, that represent a group of closely spaced forest disturbances, i.e., a multipart polygon, associated with one storm event;  -Storm event track‖, that is a cluster of windthrowswindthrow areas with having identical similar direction and 180 having the same date (or same date range) of occurrence, which were most likely induced by one convective or nonconvective storm.
GIS layers have WGS84 geographic coordinate system (EPSG:4326). The key fields ID and storm_ID associates each damaged area with the spatial features in the datasets of windthrowswindthrow and storm event tracks respectively using one-to-many relation. ID values of windthrow areas are set according to the date of occurrence of storm events. Within one 185 year, numbers are first set for windthrow areas with known dates, and then for ones with unknown dates. If two or more windthrow areas are caused by one storm event, their numbers are sequential according to storm movement directionly. The numbering of EDAs is organized according to the numbering of windthrow areas. EDAs related to one windthrow area are numbered from the lower left corner, that is, from southwest to northeast. The structure of the attribute tables of each shapefile (stored in .dbf files) is presented at Tables 1-3. The determination process of the presented characteristics is 190 described in Section 4 and schematically presented at Fig. 2. Figure 3 shows an example with all three hierarchical levels of the database.

Methods: windthrow delineation and parameters determination
The process of windthrowswindthrow identification and attribution to a particular type includes four stages ( Fig.2): (1) delineation of a windthrow using the Landsat-based GFS and EEFCC products or time series of Landsat or Sentinel satellite 195 images, (2) subsequent verification of a windthrow using the HRI and determination of the type of a storm event causing a windthrow, (3) estimation of geometrical characteristics of a windthrow, and (4) determination of storm date or range of dates by utilizing additional information. Most of data collection stages were performed manually using standard GIS tools, except for the data extraction from the GFC and EEFCC products and calculation of windthrows the geometrical characteristics of windthrow areas (that were automated with Python language). Due to numerousseveral limitations of the 200 data sources and the use of expert knowledge at severaldifferent stages of the data collection workflow (Fig. 2), the compiled database is spatially and temporally inhomogeneous and hence incompleteif probably far from complete. In particular, the database lacks small-scale forest disturbances with area below thresholds (Fig. 2). The main data and method limitations of the data collection workflow are discussed in the Section 6.

GFC-based delineation (2001-2017)
We systematically searched through the GFC dataset for forest loss areas that have characteristic windthrow-like signatures.
The searching of windthrow areas based on GFC was rather systematic. The GFC-based collection of forest loss areas with windthrow-like signaturessearch was carried out separatelyperformed for each regioncell of a supplemental of the Russian 210 Federation. A grid with 50 km cell size that was built inside each regionfor the ER, which helped to organize the searching of windthrow-like forest disturbancesthe searching was performed sequentially in each cell of the grid. (the searching was performed sequentially in each cell of the grid) However, it is highly likely that some windthrow areas were missed (see Section 6 for more details). Since we used the threshold values of windthrow area (0.05 km 2 for tornado-induced windthrow areas and 0.25 km 2 for other windthrow), we did not estimate the forest disturbance with a smaller area. The relevant 215 information has been added to the section 4.1.1.
In particular, we looked for windthrowswindthrow the forest disturbances with the shape that elongated stretched along the direction of storm or tornado movement. Wind-related forest disturbances rarely have quasi-circular/elliptic or regular shapes that are characteristic for fire-related disturbances and logged areas, respectively Chernokulsky, 2018, Shikhov et al., 2019). Windstorm-or snow/icestorm-caused windthrowswindthrow areas have amorphous spatial structure 220 and a varying degree of forest damage, whereas tornado-induced windthrowswindthrow areas have quasi-linear spatial structure and almost total removal of a canopy . After selecting an area affected by a windthrow, we extracted respective pixels from the GFC data and converted them from raster to multipart vector polygons, which consist of many singlepart polygons, so-called ‗elementary damaged areas' (EDAs, fig. 2, fig. 36). In addition, we 9 removed all EDAs with an area ≤ 0.00181800 km 2 , that is two GFC pixels. We removed all EDAs with an area ≤ 1800 m 2 t 225 that equals to area of two GFC pixels. We filtered out such small-scale disturbance since it is virtually impossible to confirm their wind-related origin. Moreover, the area of local windthrow can be almost three times overestimated by Landsat images (Koroleva and Ershov, 2012). We found, that the absence of minimum accepted area for EDAs will increase area of windthrow by 2-3% on average (up to 6% for several windthrow areas with amorphous spatial structure). However, the number of EDAs mistakenly referred to windthrow can be substantially overestimated. Thus, the use of minimum accepted 230 area for EDAs is the balance between slight underestimation of total area of windthrow in the study region and substantial overestimation of the number of EDAs mistakenly referred to windthrow. We have decided to choose the first option. The optimization of other threshold values can be evaluated in further studies that should involves ground-based data.
We filtered out such small-scale disturbances since it is often impossible to confirm their wind-related origin. Moreover, their area of local windthrows can be almost three times overestimated by Landsat images (Koroleva and Ershov, 2012). By 235 removing small-scale EDAs, we can slightly underestimate the area of windthrow in general, but the number of EDAs mistakenly referred to windthrow substantially reduced.
In total, we delineated 450 windthrowswindthrow areas using the GFC dataset, and clarified contours of 126 of them manually using the Landsat, Sentinel-2, or HRI images (see Section 4.2 for details).

EEFCC-based delineation (1986-2000) 240
For the EEFCC data, we performed similarly to the GFC searching and delineation of windthrowswindthrow areas with however some limitations. The main limitation is related to the classification of forest losses into broad periods, i.e., 1986-1988 and 1989-2000. Thereby, windthrow area can be correctly delineated only if it lacks overlap with other forest disturbances, namely loggings and wildfires, occurred in the same period. For instance, in highly-populated areas, salvage loggings are usually performed in 1-2 years for most of wind-damaged forests. Such windthrowswindthrow areas were 245 delineated by the Landsat images with semi-automated NDII-based method (see Section 4.1.3). Based on the EEFCC, we were able to delineate windthrowswindthrow areas with high confidence mainly in the low-populated northern part of the the ERER (Fig. 43). To partially avoid missing of windthrow areas, using Landsat images, we performed additional verification of all large-scale forest loss areas (with area more than 5 km 2 ) in highly-populated regions independently of their geometry, since windthrow areas can be totally masked out by logged areas. We were able to find three large-scale windthrow events (≥ 250 10 km 2 ) in these regions. However, some windthrow events can still be missed.
To avoid missing large-scale windthrow in highly-populated areas, we additionally verified in such regions all large-scale forest disturbances (with an area ≥ 5 km 2 , that is not typical for industrial logging in the forest zone of ER). These forest loss areas were verified by multi-temporal Landsat images independently of their geometry, since windthrow areas can be totally masked out by logged areas. Thus, we were able to find three large-scale (≥ 10 km2) windthrow areas in highly-populated 255 regions of ER. However, some windthrow event can be missed.
In total, we delineated 153 windthrowswindthrow areas using the EEFCC dataset. Contours of the 32% of them were then substantially clarified manually with the Landsat images, obtained before and after the storm events. Another 22 windthrowswindthrow areas that occurred before 2001 were delineated manually using the Landsat images. As for the GFC, we removed all EDAs with an area < 0.0018 k1800 m 2 , since it is often impossible to confirm their wind-related origin. 260

NDII-based delineation (1987-2000)
We used Landsat TM/ETM+ images (Level 1T) obtained before and after the storm event in the growing season to delineate seven7 In total, seven large-scale windthrowswindthrow, occurred before 2001, were delineated by comparing Landsat TM/ETM+ images obtained before and after the storm event in the growing season. We used the difference of Normalized 265 Difference Infrared Index (NDII, Hardisky et al., 1983) to detect and delineate wind-related disturbances. High efficiency of the NDII using for windthrowswindthrow identification on Landsat images has been shown previously Wang and Xu, 2010;Chernokulsky and Shikhov, 19842018). NDII was formulated as follows: where TM4 and TM5 are the reflectance in the bands 4 (0.85 μm) and 5 (1.65 μm) of Landsat TM/ETM+ data, while the 270 difference was calculate as ∆NDII= NDII before -NDII after , where subscripts ‗before' and ‗after' denotes two closest to an event cloud-free images obtained, respectively, before and after the windthrow occurrence, but in the growing season only.
We applied none atmospheric correction algorithm for preprocessing Landsat images, since NDII is based on the nearinfrared (0.76 -0.90 nm) and middle-infrared (1.55 -1.75 nm) spectral bands that are almost insensitive to atmospheric impact. the near-infrared (0.76 -0.90 nm) and middle-infrared (1.55 -1.75 nm) spectral bands that are less sensitive to 275 atmospheric influences in comparison to visible spectrum bands. For NDII-based delineation process, we used only images with cloudiness less than 10% based on CFMask algorithm (Foga et al., 2017). For other purposes (verification, type and date determination), we visually inspected Landsat images for lacking clouds over the area of interest (i.e., a windthrow area)Cloud masking was also not required since we used thousands of images, and most of them were needed only to determine the dates of previously delineated windthrow. In its turn, the images that were used for windthrow delineation 280 were cloudless within the area of interest.
The masking of forested lands was performed on the ‗before' image with the use of Iterative Self-Organizing Data Analysis Technique Algorithm (Ball and Hall, 1965) unsupervised classification. Then, the NDII was calculated only within the mask of the forested area. The same technique was successfully applied previously to delineate windthrowswindthrow areas caused by the 1984 Ivanovo tornado outbreak (Chernokulsky and Shikhov, 19842018). 285 WindthrowsWindthrow and other forest disturbancess are characterized by a sharp decrease of the NDII. However, threshold values of ∆NDII for distinguishing between stand-replacing disturbances and moderate damaged or undamaged forests, differ for each pair of images. We estimated threshold value from the statistics of ∆NDII raster. Firstly, we obtained the mean value and standard deviation of ∆NDII within the entire forest-covered area on image. Stand-replacing forest disturbance inherently has ∆NDII values substantially higher than the image average. To separate stand-replacing forest 290 disturbance from other forest-covered area, we used the threshold value of two standard deviations, which was previously tested by Koroleva and Ershov (2012). However, in some cases the ∆NDII distribution within the entire image was skewed (e.g., due to the presence of cloud decks iness decks or haze on the post-event image). In such cases, we lowered the threshold value of ∆NDII iteratively by comparing the detected changes with results of visual identification of windthrow on a post-event image (using several examples located in different parts of windthrow). As a result, actual threshold values 295 ranged from 1.5 to 2 standard deviations. Then, a binary raster of detected changes (i.e., forest losses) has been created (see fig. 54d) and converted to a shapefile.We estimated threshold value from the statistics of ∆NDII raster. Firstly, we obtained the mean value and standard deviation of ∆NDII within forest-covered area. Stand-replacing forest disturbance has ∆NDII values substantially higher than the average for the image. To separate stand-replacing forest disturbance from other forestcovered area, we used threshold value of two standard deviations, which was previously tested by Koroleva and Ershov 300 (2012). However, in some cases the ∆NDII distribution for the entire image was asymmetric (e.g. due to the presence of cloudiness on the post-event image), In such cases, we corrected threshold value manually, to ensure the best fit with the results of the visual identification of windthrow. In most cases, the pixels with values of ∆NDII that exceed the average value for the forested areas of the entire image by more than two standard deviations, indicate the stand-replacing forest disturbances (Koroleva and Ershov, 2012). However, this value may be less if these disturbances hold the substantial part of 305 the image. We estimated the threshold values from a sample of ∆NDII, obtained within the forested area, and corrected it in several cases to ensure the best fit with the results of the visual identification of windthrows. The threshold values ranged from 1.5 to 2 standard deviations for different pair of images. On At the next step, windthrowswindthrow areas were separated from logged areas and other disturbances (see Section 3.2). The EDAs ≤ 0.0018 k1800 m 2 were removed.

Combining delineated polygons to a windthrow areas and windthrowswindthrow areas to a storm event tracks
In general, a group of closely spaced EDAs, caused by one storm event, was assigned to one windthrow. By the ‗close distance' we meant in most cases a distance of tens or hundreds of meters between the nearest EDAs. This distance is 315 determined manually by the proportion of stand-replacing damage (the distance decrease as its increasing), and the presence of treeless areas. Maximum distance between nearest EDAs combined to one windthrow area However, it may reach 5-10 km, if a windthrow crossed treeless areas ( fig. 7 may be an example).
Most of windthrowswindthrow areas were extracted from the GFC dataset (450 windthrowswindthrow), EEFCC dataset (153 windthrowswindthrow) or with NDII-based methods (7 windthrowswindthrow). For tThese windthrowswindthrow 320 areas were, we first automatically delineated a gross outline of a windthrow as a multi-part polygons, and then we specified exact contours of its their componentssingle-part polygons (EDAs);. aAfter that, we correctly merged them to a windthrow itself (Ffig. 2). We delineated other 90 windthrowswindthrow areas manually using the Landsat, Sentinel-2, or HRI images -30, 17, and 43 windthrowswindthrow areas, respectively. In this case, we first delineated EDAs and then merged them into a windthrow. 325 Many storms induced a series of successive windthrowswindthrow areas, which are separated from each other by tens or even hundreds of kilometres of undamaged forests, treeless areas or water bodies (Ffig. 63). In general, we divided the damaged areas into two separate windthrowswindthrow (two records in the dataset), if the gap between them exceeded 10 km. The This threshold is based on the study of Doswell and Burgess (1988), who proposed the 5-10 miles (8-16 km) threshold for the gap to separatediscriminate between one skipping tornado fromand two successive tornadoes. A similar 330 threshold value (8 km) was previously used proposed to separate one skipping tornado from two successive tornadoes (Doswell and Burgess, 1988;Shikhov and Chernokulsky, 2018). A few exceptions were associated with changes of windthrow direction, transformations of one windthrow type to another identified by the HRI, i.e., the tornado-induced to non-tornado-induced, and with abrupt change of forest damage degreefrom 60--80% to 5--10% of stand-replacing disturbances. In these cases, the distance between two distinct windthrowswindthrow areas was lessfor instance, the 335 minimum distance was about 1 km when a tornado-induced windthrow transformed to a squall-induced one.
If several successive or quasi-parallelclose windthrowswindthrow areas have similar direction, differing by no more than 30°, and the same date (or date range) of occurrence, we assigned them to one storm event ( fig. 6). We highlight successive and parallel windthrow areas (Fig. 3). Successive windthrow areas induced by one storm event follow downwind one after another and approximately fall on one straight line (the angle of deviation from this line does not exceed 10-20°). Such 340 windthrow are presumably induced by one convective cell generating a sequence of squalls or tornadoes. In contrast, parallel windthrow areas that located within one storm event are situated parallel to each other (with an angle less than 30°). They are presumably associated with two or more different convective cells or mesocyclones, generating squalls or tornadoes, often embedded into one mesoscale convective system. Successive windthrow areas follow downwind one after another and approximately fall at one straight line (the angle of deviation from this line does not exceed 10-20°). Such windthrow are 345 presumably induced by one convective cell generating a sequence of squalls or tornadoes. In contrast, parallel windthrow are located parallel to each other or at a large angle (over 45°). They are presumably associated with two or more different convective cells or mesocyclones, generating squalls or tornadoes, often embedded into one mesoscale convective system.
The expert-based process of windthrow areas combining to a storm event was based as well on various additional information including the storm dates and types (see next sections), information from weather station reports, eye-witness 350 and newspaper reports, data from meteorological satellites, and so on. Threshold value of a distance between different windthrow areas related to one storm event was not determined, and maximum distance between them reaches 150 km. In total, the dataset of storm event tracks contains 486 items.
The threshold values used in this section (maximum distance to combine EDAs to windthrow area and minimum distance to 355 separate two successive windthrow areas) have some subjectivity, and their modification may substantially change the number of allocated windthrow areas in the dataset. However, these thresholds determine only length and width of single 13 windthrow, and they does not affect the total area of windthrow, which is the main characteristic of wind-induced forest disturbances. The optimization of above-described threshold values can be considered as a further research direction.

Verification of windthrowswindthrow events and determination of its type
At the second stage, we performed expert-based verification for each forest disturbanceverified each windthrow area in the database using pre-and post-event Landsat/Sentinel-2 images, high-resolution images and additional informationthe HRI or, in the lack of the HRI, the Landsat/Sentinel-2 images. This verification was performed to ensure the forest disturbance was caused by wind and to determine a type of a storm caused a windthrow. In total, we verified 54% of windthrowswindthrow 365 areas with the HRI, mainly for the 2001-2015 period. Other windthrowswindthrowevents were verified using the Landsat images (22% of windthrowswindthrow), the Sentinel-2 images (9%) and additional data sources like weather station and eye-witness reports (15%). As a result, the probability that any forest disturbance was mistakenly referred to a windthrow is minimal.
In addition, we used the last cloud-free Landsat or Sentinel-2 image obtained before a storm and first image obtained after to 370 separate windthrowswindthrow areas from other disturbances, mainly from logged areas. We removed forest disturbances that were not related to a storm event (Fig. 65). During the verification, we also found and delineated several storm-damaged areas that were missed in the GFC/EEFCC data. Such areas are located mainly in small-leaved or broadleaved forests. After the verification, we determined the type of a windthrow depending on a weather phenomenon induced this windthrow. We selected tornado-induced and non-tornado-induced windthrowswindthrow areas, the latter were subdivided into induced by 375 convective and by non-convective storms. In turn, non-convective storms include also snowstorms, which are indicated in the database but not analyzed separately further in the paper. By convective storms we mean squalls and downbursts; however, this more detailed division lacks in the database.
To distinguish tornado-induced windthrowswindthrow areas from other wind-related disturbances, we determined the direction of fallen trees using the HRI. Indeed, the main signature of tornado-induced windthrowswindthrow is the 380 counterclockwise, or infrequently clockwise, rotation of the fallen trees (Beck and Dotzek, 2010;Shikhov and Chernokulsky, 2018). In the lack of the HRI, we considered three additional signatures of tornado-induced windthrowswindthrow, namely (1) quasi-linear structure of a windthrow with a ratio of length and width ≥ 10:1, (2) a gradual turn of a storm track, and (3) predominatelyprevalently total removal of forest stands . Note, that the ratio of length and width of tornado track ≥ 10:1 is also typical for U.S. (Schaefer 385 and Edwards, 1999). Based on these three signatures and additional information from weather station reports, witness reports, photos and videos, we assigned the high or medium degree of certainty of storm type determination for each windthrow (Table 4).
WindthrowsWindthrow areas, caused by non-convective windstorms or snowstorms, have as well specific geometrical features that seen at satellite images. Specifically, windthrowswindthrow areas related to non-convective windstorms 390 typically have enormous length and width of the damage track, up to 200 and 45 km respectively, with however slightly or moderate damaged forests. Caused by non-convective windstorms sStand-replacing disturbances caused by non-convective windstorms are usually occur in dark coniferous forests only (Dobbertin et al., 2002;Schmoeckel and Kottmeier, 2008).
Since non-convective storms affect large areas and last for relatively long period, they are typically well-reported by weather stations, which simplify the attribution of related windthrowswindthrow. In its turn, snowstorm-induced 395 windthrowswindthrow areas are distinguishable from other disturbances primarily based on the dates of occurrencethey happen usually in autumn; although, one severe snowstorm occurred in early summer. It is of note, that we found none of snowstorm-induced stand-replacing windthrow happen in winter.
After the determining of a storm event type, we excluded from the database the tornado-induced windthrowswindthrow with an area ≤ 0.05 km 2 and non-tornado-induced windthrowswindthrow with an area ≤ 0.25 km 2 . We took into account the 400 following reasons during exclusion of such small-scale windthrowswindthrow areas: 1. Difficulty to prove that these disturbances are were actually were caused by wind, especially in the lack of the HRI.
2. Difficulty to determine wind storm event dates with the Landsat images for these windthrowswindthrow areas.
3. High uncertainty of estimated geometrical characteristics of small-scale windthrowswindthrow (Koroleva and Ershov, 2012;Shikhov and Chernokulsky, 2018). 405 Only five squall-induced windthrowswindthrow with an area < 0.25 km 2 were stored in the database, since they are associated with severe weather outbreaks with proven dates. It is of note, that a typical tornado-induced windthrow consist of a relatively small number of EDAs with total removal of forest stands that are well-detected by the Landsat images. In its turn, a typical non-tornado-induced windthrow include larger number of small-scale (i.e., 2-4 Landsat pixels) areas of standreplacing disturbances, that are worse detected by satellite images. This difference results in the necessity of using two 410 distinct thresholds for tornado-and non-tornado-induced windthrowswindthrow areas.
The threshold values used in Sections 4.1-4.2 have some subjectivity, and their modification may substantially change the number of allocated windthrow areas in the dataset. The optimization of above-described threshold values can be evaluated in further studies that should involves ground-based data.

Estimation of geometrical parameters of windthrowswindthrow areas and storm tracks and its accuracy
We used Landsat data and the Landsat-based products GFC and EEFCC to estimate geometrical parameters of windthrowswindthrow areas. We determined the path length (L), mean and maximum widths (W mean and W max ), and damaged area (A) for each windthrow using the technique that had been successfully implemented for tornado-induced windthrowswindthrow areas . The calculation of these parameters was performed in the 420 Lambert Equal Area and Equidistant projection for North Asia to avoid possible projection-related distortions.
We calculated A in the ArcGIS 10.4 as the sum of area of forest damaged plots, which are attributed to one windthrow. We determined L as a length of the central line drawn through a damaged area, i.e. distance between two farthest points of a windthrow. The central line was created automatically (using a Python tool) as a distance between two farthest points of a windthrow. It is insensitive to the allocation of patches to windthrow area. 425 We calculated W mean as the mean length of several transects that are perpendicular to a storm track with a 200 m step; this step had been found optimal in terms of quality and counting efficiency . Only standreplacing windthrowswindthrow areas were taken into account in this calculation. In comparison to , where W max were calculated manually using the HRI data, in this study, we assigned the length of the largest transect to W max because the lack of the HRI for many windthrowswindthrow areas. 430 In addition to windthrow characteristics, we estimated geometric characteristics of EDAs and those of storm tracks.
Particularly, for EDAs, we calculated their area A EDA . For storm tracks, we estimated maximum and mean width (W TRmean and W TRmax ), path length (L TR ), and damaged area (A TR ). We calculated W TRmean based on the same transects that were used to calculate W mean but without excluding undamaged forests and treeless areas. Similarly, length of the largest transect that includes undamaged forests and treeless areas was assigned to W TRmax (Fig. 67). If a track consists of two (or more) parallel 435 windthrowswindthrow areas, then its width was calculated within the outermost boundaries of these windthrowswindthrow areas (Fig. 3). The same calculation was performed for L TR in case of two (or more) subsequent windthrowswindthrow areas ( Fig. 3). Thus, the used 10-km threshold (see Section 4.1.4) may influence geometrical characteristics of single windthrow area, but do not affect those of a storm event.
We assessed the accuracy of GFC-based estimates of windthrow geometrical parameters by comparing them with the same 440 parameters calculated manually with the use of HRI using. We performed such procedure for ten windthrowswindthrow areas caused by squalls, whose area ranges from 0.26 to 6.09 km 2 ( Table 5). Distribution of their A is close to the one for the full dataset.
We delineated manually all EDAs within these ten windthrowswindthrow areas using the HRI. In total, we found 837 and 947 EDAs, according to the GFC and the HRI data respectively. Owing to relatively correct georeference of the Landsat data 445 (Landsat Collection 1, 2019), we found no systematic spatial bias between contours of GFC-based and HRI-based windthrowswindthrow areas. Despite their general matching, there is no complete overlap due to different spatial resolution of the GFS and HRI (Fig. 78). For example, one GFC-based EDA may intersect with several HRI-based ones, and vice versa. We found, that only 66.5% of the total area is attributed to windthrowswindthrow in both GFC and HRI, while EDAs with small area can be missed. In particular, 263 HRI-based EDAs with the total area of 0.97 km 2 were completely missed in 450 the GFC, while 146 GFC-based EDAs with the total area of 0.52 km 2 were missed in the HRI. For overlapped EDAs, we found the mean absolute error and root mean square error of A EDA estimates amounted to 27.6% and 13.1%, respectively. We found that the relative error decreases for large EDAs and for those having a simple shape, i.e., quasi-circular. The user's and producer's accuracies increase from 20-25% for EDAs with A EDA < 0.01 km 2 to 70-75% for EDAs with A EDA > 0.1 km 2 . In general, for the overlapped EDAs, the GFC overestimates their A EDA (by 4% on average) primarily in coniferous forests. 455 Mutual effect of more frequent omission of small EDAs in the GFC compare to the HRI and overestimation of overlapped EDAs results in approximate equality of total area of delineated windthrowswindthrow -17.11 km 2 and 17.13 km 2 based on the GFC and HRI, respectively.
For entire windthrowswindthrow area, we as well calculated as well an accuracy of their geometrical characteristics estimatingestimatesion. In particular, we calculated the user's and producer's accuracies of the GFC-based delineation for 460 each of ten selected windthrow. These accuracies are mainly determined by the complexity of windthrow shapes and composition. In particular, the accuracy is higher for a windthrow consisting of relatively small number of simple-shape EDAs. Otherwise, the accuracy decreases down to 50% for a windthrow with ishaving very amorphous spatial structure. In our sample, the GFC data tends to overestimate the area of windthrowswindthroweight cases out of ten were overestimated. The mean absolute percent error (MAPE) for A is 14.6%. The major overestimation of A by the GFC data, as 465 well as W mean and W max , was revealed for relatively small windthrowswindthrow areas. This is in line with the previous findings by Koroleva and Ershov (2012). They who showed that the reliable estimate (with 15% accuracy) of the damaged area using the Landsat images is possible only for windthrowswindthrow areas exceeding 0.026 km 2 . It is of note, that for tornado-induced windthrowswindthrow areas, Shikhov and Chernokulsky (2018) found, that the GFC data generally tends to underestimate A, with MAPE amounted to 17.9%. 470 The assessment of geometrical parameters of windthrowswindthrow areas appeared occurred before 2000 and found by the EEFCC is challenging due to the low availability of the HRI or other independent data sources, e.g. the data of forestry services. WindthrowsWindthrow areas induced by storm events that occurred >20 years ago can be delineated by the HRI only if athey storm passed through old-growth forests that have not been affected by other disturbances, i.e., timber harvesting or wildfires, in subsequent years. Such forests are widespread only in the northeastern part of the the ERER 475 (Pakhuchiy, 1997). We found five EEFCC-based windthrowswindthrow appeared occurred between 1998 and 2000 that were most well-detected by the HRIfour tornado-induced and one non-tornado induced. We delineated them with the EEFCC and the HRI and compare their characteristics (Table 6). We found general overestimation of A, W mean and W max in the EEFCC, that was larger than in the GFC. It may be related to the inclusion into a windthrow not only real wind-damaged pixels but also surrounding pixels where tree had died after a windthrow appearance mostly because of bark beetles (Köster 480 et al., 2009). Intensity of this mortality is highest at a second year after a storm event (Köster et al., 2009).

Determination of windthrow dates
We aimed to establish the exact date or even the exact time for each windthrow appearance. However, due to data constraints, dates of some windthrowswindthrow events were determined with accuracy ∼< 6 months accuracy. We iteratively refined date, or a date range, by using different data. The process, related to the determination of date of tornado-485 induced windthrowswindthrow only, had been described previously in .
First, the year of a windthrow can be obtained directly from the Landsat products but with some limitations. In the GFC, forest disturbances are accompanied with information on the year of event occurrence. However, the exact year is determined correctly only for 75.2% of eventsforest loss pixels; for 2124.58% of eventsthem, the date can be either a 1--2 year earlier or a year later (Hansen et al., 2013). In the EEFCC, a year of windthrow occurrence is not explicitly determined 490 and came within the ranges 1986-1988 and 1989-2000 years. Next, we refined a range of dates based on all available images from the Landsat and Sentinel-2 satellites. The accuracy of such refinements depends on a frequency of observations and cloudiness. The availability of cloudless Landsat images varied from year to yearstrongly fluctuated from year to year. The lowest number of cloud-free images (2--4 images a year on average) is available for 2003--2006 data are available (Potapov et al., 2015). 495 Hence, the worst accuracy of windthrow date determination is typical for these years. The lowest frequency of satellite observations in the study area, namely 2-4 cloud-free images per year, took place in 2003-2006, 2008, and 2012 years when only Landsat-7 data were available (Potapov et al., 2015). On average, 8-10 images per year can be used for windthrow identification and dates determination. In turn, Due to Sentinel-2A satellite launching, number of images per year had an abrupt increase after the summer of 2016tThe highest frequency of satellite imagery, namely ten images per month for a 500 location, was achieved in 2016-2017 after the start of the Sentinel-2 mission. We used images taken throughout a year.
Despite the frequency of cloudless images in autumn and winter was lower than in summer season, it was sufficient for analysis. Thus, It is of note also that winter images (of land covered with snow) were successfully used for windthrow identification, especially if a storm occurred at the end of summer season, and autumn season lacked cloud-free images.
wintertime images (of landobtained when land surface is covered with snow) were widely used for windthrow identification, 505 especially if a storm occurred at the end of summer season, and autumn season lackedwas without cloud-free images. The frequency of obtaining of cloudless images in autumn and winter was lower than in summer season, but is it is sufficient for analysis, especially considering that we used all images cloud-free over the area of interest (i.e. over a windthrow area).
Further, given the satellite-derived range of event possible dates, we made the subsequent analysis using additional data such 510 as weather station observations, various databases and reviews on hazardous weather events, damage reports, photos and videos in the media and social networks, and reanalysis data (see  for details). This analysis allowed to establish the exact dates for 48.4% of all windthrowswindthrow events including 39.2% and 59.7% of tornado-and non-tornado-induced windthrowswindthrow events, respectively.
The dates of storm-induced windthrowswindthrow events were defined more successful than those for tornado-induced 515 windthrowswindthrowones due to the local nature of convective storms, especially of tornadoes, and a relatively large distance between Russian weather stations. Specifically, the average and median distance between nearest weather stations within the study area amounted to 53.7 and 49.9 km, respectively. Wherein, many storm events were reported by weather stations located on a storm path at a distance of 50-100 km from a windthrow, while the closest stations did not reported strong wind gusts since they were away from a storm path. In total, we matched storm reports of weather stations, namely 520 reports with wind gusts ranges from 15 m/s to 34 m/s, only with 34.5% of windthrowswindthrow events with known date.
Another reason for more successful determination of dates of appearance for large-scale windthrowswindthrow areas than for small-scale windthrowswindthrowones, e.g., tornado-induced, is an increase of probability that a corresponding storm passes through a settlement(s) and this is covered in the media. In total, we used media reports, information from regional weather services, witness photos and videos, existed scientific literature (e.g., Dmitrieva and Peskov, 2013;Petukhov and 525 Nemchinova, 2014;Shikhov and Chernokulsky, 2018; to specify the date and time of 29.7% of windthrowswindthrow events. Dates and time of some cases (7.8% of all cases) were established using images from meteorological satellites Terra/Aqua MODIS and METEOSAT-8, and Russian weather radar data (Dyaduchenko et al. 2014). However, the routine usage of these data is time-consuming and limited due to some access restrictions. Subsequent clarification of windthrow exact time can be 530 carried out in further studies.

Windthrow type
The compiled database includes three shapefiles (.shp), corresponding to three hierarchical levels such as elementary damaged areas, windthrowswindthrow, and storm events. The database includes 102747, 700, and 486 objects for each level, 535 respectively ( fig. 2). The total area of the spatial features is equal 2966.1 km 2 . It is of note, that we cannot determine whether the trees were felled or broken by the wind based on satellite images, even having very high resolution. Therefore, we use a single term -windthrow‖ for all types of wind-induced forest damage.
The overwhelming majority of found stand-replacing windthrowswindthrow in the ERER, namely 97.4% of windthrowswindthrowthe events and 95.3% of wind-damaged area, are associated with convective storms and tornadoes 540 (Table 7). More than a half of all windthrowswindthrow areasevents are tornado-induced with however relatively small damaged area (less than 13% of the total wind-damaged area). Non-convective storms and snowstorms are responsible for less than 5% of the area of stand-replacing windthrowswindthrow in the ERER. This is somewhat in contrast to Western and Central Europe, where most of windthrowswindthrowforest damage are is induced by non-convective wind events, namely winter storms, caused by strong extratropical cyclones (Gardiner et al., 2010;Gregow et al., 2017). Indeed, winter 545 windstorms affects less Eastern Europe compare to Western and Central Europe (Haylock, 2011). In addition, in the ERER and Northern Europe, ground is usually frozen during winter and prevents trees from falling because of windstorms (Suvanto et al., 2016).
Among 486 storm events that caused windthrowswindthrow, 381 yielded only one windthrow area (Fig. 9), primarily tornado-induced. The rest 105 storms resulted in a smaller number of windthrowswindthrow events (319) but larger 550 damaged area -2276.6 km 2 , namely 76.8% of all damaged area. Most of these storms induced two or three successive or parallel located windthrowswindthrow areas, and only 14 of them caused ≥ 5 windthrowswindthrowones. We found maximum of 17 separate windthrowswindthrow areas that related to one storm. We found 71 storm events result in two or more successive windthrowswindthrow areas, while 12 storm events lead to formation of two or more parallel windthrowswindthrow areas, and 22 storm events include a family of both parallel and successive 555 windthrowswindthrowones ( fig. 36). The maximum distance between two nearest successive and two parallel windthrowswindthrow areas amounts to 150 and 26 km, respectively.
It should be noted, that a single storm may cause both tornado-and non-tornado induced windthrowswindthrow, e.g. a supercell can lead to formation of a tornado and a rear-flank downdraft (Karstens et al., 2013) both causing forest damage. In total, we found 30 storms that resulted in formation of two types of windthrowswindthrow. 560 We managed to match several storm events with reports at weather stations, in particular the database contains 89 such cases. Among these 89 station reports, we found eight reports with wind gusts ≥ 30 m/s, 14 reports with wind gusts 25-29 m/s, and 30 reports with wind gusts 20-24 m/s. This information have been included in the database, and can be used in further studies to estimate the critical wind speed causing windthrowswindthrow and to analyse the role of other accompanying weather phenomena, e.g. with snow, heavy rainfall, large hail, etc. 565

Spatial distribution of windthrowswindthrow areas
WindthrowsWindthrow eventsareas occur in the entire forest zone of the the ERER (Fig. 910). However, the highest density is observed near the 60° N and somewhat coincides with the highest percentage of forest-covered area (see Fig. 1). It is of note, that two windthrowswindthrow areas are located north of 66° N and one of them is even north of the Arctic Circle. The dominant direction of both tornado-induced and other windthrowswindthrow is SW-NE (Fig.14b15b), which is in line with 570 the previous studies on tornado climatology in Northern Eurasia Chernokulsky et al., 2020).
Three areasregions, where windthrowswindthrow have has affected more than 0.75% of forests, can be highlighted (Fig.   10a11a). Two of them are related to the catastrophic storms which occurred on 27 June 2010 and 29 July 2010. In total, these two storms have damaged 1140 km 2 of forests, which is 38.4% of the total area of stand-replacing windthrowswindthrow in 575 the ERER in 1986-2017. The third area is located on the western slope of the Northern Ural and coincides with the largest massive of dark-coniferous forests in the the ERER (Pakhuchiy, 1997). The largest most important windthrowswindthrow events occurred here in June 1993, July 2012 and October 2016. The latter was induced by snowstorm. The relatively high frequency of windthrowswindthrow in this region was emphasized previously (Lassig and Mocalov, 2000;Shikhov and Chernokulsky, 2018;. It was hypothesized that it may be related to the combination of several factors, 580 namely widespread old-growth forests, a high precipitation rate, and large soil wetness, which all contribute to the forests wind susceptibility (Dobbertin, 2002).
The highest density of tornado-induced windthrow is found between 59° and 62° N, 48° and 56° E (Fig. 110, b), which is in a good agreement with the previous estimates . However, when the percentage ratio of tornado-damaged area of to the total forested area is considered is higher in the western part of the the ERER , then the 585 western part of the ERER becomes the most affected by tornadoes (Fig. 10b11b). It is of note, that higher values of so-called convective instability indices are also observed in this region (Taszarek et al., 2018).

20
The species composition and age of forest stands have substantial influence on the spatial distribution of windthrowswindthrow (Dobbertin, 2002, Suvanto et al., 2016Gregow et al., 2017). However, the available data on the forests species composition for the entire ER (Fig. 1) have too coarse spatial resolution (i.e., 250 m) to overlap them with 590 windthrows data based on 30-m Landsat images. CorrectUsing the presented dataset, estimates the of relationships between windthrowswindthrow area and forest stands characteristics can be carried out in future studies at a regional scale.

Temporal variability of windthrowswindthrow and storm events
We successfully determined the year of occurrence for all windthrowswindthrow events and the month of occurrence for 263 (67.9%) tornado-induced and 224 (71.5%) non-tornado-induced windthrowswindthrow events. We established the dates of 595 occurrence for 339 windthrowswindthrow events, including 149 tornado-induced (39.2%) and 187 (59.7%) non-tornadoinduced windthrowswindthrowones. It is of note, that the dates of most impacted impactful large-scale windthrowswindthrow with damaged area > 10 km 2 were determined for 44 out of 49 cases (90%). WindthrowsWindthrow events with known dates have a total area of 2599 km 2 , i.e., 87.7% of the total wind-damaged area.
The storm-damaged area has a relatively high inter-annual variability (Fig. 1112). The largest area of windthrowswindthrow, 600 i.e. >1200 km 2 , is found in 2010, when two exceptional storm events were occurred. An extremely high number of tornadoinduced windthrowswindthrow events occurred in 2009 and 2017. Storm events causing windthrowswindthrow are have been observed every year and ranges from 2 to 36, with the maximum in 2012 and minimum in 2001. In general, annual number of windthrowswindthrow and storm events was lower before 2001 when the EEFCC data were used to identify windthrowswindthrow, and higher after 2001, when the GFC data were utilized. Annual number of windthrowswindthrow 605 events for these periods amount to 12.1 and 30.5, respectively; in its turn, annual number of storm events amounts to 8.3 and 20.9. This temporal inhomogeneity, related to different initial data used, should be taken into account when inter-annual variability is analyzed. More details on the dataset limitations are provided in Discussion the Ssection 6.
WindthrowsWindthrow events occur in the the ERER from May to October (Fig. 1213). No winter windthrowswindthrow were was found. The seasonal maximum of the number of windthrowswindthrow events is found in Juneboth for 610 tornadoes and for other storm events. This is in concordance with the previous estimates on the tornado climatology Chernokulsky et al., 2020). Maximum frequency of the occurrence of storm events causing windthrowswindthrow is also observed in June. Moreover, more than 90% of storm events with known dates occur in summer. It is important to note, that we failed to establish the month of occurrence appearance for 127 tornado-induced windthrowswindthrow areas and 98 non-tornado induced windthrowswindthrowones, with thewhich have total area of 245 615 km 2 .
Sometimes, two or more storm events causing windthrowswindthrow occurred in ER on the same day. In total, we found seven outbreaks with more than ten windthrowswindthrow eventsareas per day. The most remarkable outbreaks occurred on 18 July 2012 when nine storms resulted in 25 windthrowswindthrow eventsareas, and on 7 June 2009 when five storms resulted in 24 windthrowswindthrowwindthrow areasones. However, the largest forest damage is associated with a single 620 21 storm, namely the long-lived convective storm -Asta‖ (Suvanto et al., 2016). This storm has passed over the northwestern part of the ERER and Finland on 29 July 2010 and has damaged 639 km 2 of forests in Russia.
No winter windthrow was found. It is of note, that both Landsat-based products GFC and EEFCC reveals stand-replacing windthrow area regardless of the season of its appearance. In particular, if windthrow happened in winter it would be clearly seen on image taken in subsequent vegetation period because of rather slow forest recovery process. Therefore, the revealed 625 lack of winter windthrow is feasible due to the climatic conditions of the study area and does not associated with data limitations. In particular, winter storms from Western Europe reach the territory of Russia already weakened (Haylock, 2011), In addition, in ER and Northern Europe, while low temperatures and soil freezing also prevents trees from falling because of windstorms during winter season (Suvanto et al., 2016). According to (Suvanto et al., 2016), winter windthrow are not typical for Finland as well. 630 We restored the time of occurrence with 6-h accuracy for 216 windthrowswindthrow events -136 among them using weather station reports and 80 using other data sources. We found 122 windthrowswindthrow events (56.4%) occurred between 15.00 and 21.00 of local time (LT), which coincides with the afternoon maximum of the development of deep convection. However, several most impactful storms, including for instance the ‗Asta' storm, occurred around midnight at LTlocal time. No windthrowswindthrow found between 06.00 and 10.00 local timeLT during the morning minimum of the 635 convection diurnal cycle. The similar diurnal cycle was found for tornado events in the Northern Eurasia (Chernokulsky et al., 2020).

Geometrical parameters of windthrowswindthrow areas, elementary damaged areas, and storm tracks
Area of EDAs varies between 0.0018 to 30.9 km 2 . Most of EDAs are less than 0.01 km 2 (Fig. 13a14a), but their total area is 640 less than 10%. In turn, 1% of the largest EDAs account for 36.8% of the total area of windthrowswindthrow. Using Kolmogorov-Smirnov (K-S) test, we found that at 0.01 significant level we can reject the null hypothesis that two samples of A EDA within each pair of windthrow types are drawn from the same distribution (at 0.01 level). Because of small sample size of windthrowswindthrow areas induced by non-convective storms, later in the article we will not discuss the results of K-S test to compare distributions of characteristics of this type with those of other types. 645 Tornado-induced windthrowswindthrow areas contain fewer plotsEDAs, than other windthrowswindthrowwindthrow areas induced by strong wind (Fig. 13b14b). Particularly, most of tornado-induced windthrowswindthrow areas include 10-25 EDAs, and only 2.5% of them consists of more than 100 EDAs. In contrast, about 43% of non-tornado induced windthrowswindthrow areas includes more than 100 EDAs, while 5.5% of them consists of more than 1000 EDAs. Based on K-S test, we found that samples of number of EDAs in tornado-and convective storm induced windthrowswindthrow areas 650 are from different distributions.
A relatively small number of severe storm events are responsible for most of the area of windthrowswindthrow (Fig.   14a15a). Indeed, the ten most destructive storm events occurred in the ERER over 1986-2017 damaged 1758 km 2 of forests, namely 59.2% of the total area of windthrowswindthrow in the database. This peculiarity is less pronounced for tornadoinduced windthrowswindthrow areas, since their area usually is less than 10 km 2 . Particularly, ten tornadoes with the largest 655 area damaged 96.6 km 2 of forests -25.5% of the total tornado-damaged area. Thus, the distribution of tornado-damaged area is less skewed to high values, than the distribution of other windthrowswindthrow areas. The K-S test shows that samples of A for tornado-and convective storm induced windthrowswindthrow areas are from different distributions.
Length of windthrowswindthrow ranges from 0.8 km to 283.6 km (Fig. 14b15b). More than 44% of tornado-induced windthrowswindthrow areas have path length < 5 km, while path length 5-15 km is most frequent for non-tornado-induced 660 ones. Based on K-S test, we found that samples of number of L for tornado-and convective storm induced windthrowswindthrow areas are from different distributions. The maximum length of storm track, consisting of several subsequent windthrowswindthrow areas, reaches 544 km. This damage track is caused by the storm on 27 June 2010. In addition, another nine storm tracks have a length exceeding 250 kmmost of them are among the most destructive in terms of forest-damaged area. Such series of windthrowswindthrow with an exceptionally long path length were likely caused by 665 derechos. Derechos are, i.e. long-lived mesoscale convective systems producing widespread damaging winds and causing large-scale forest damage in U.S. (Johns and Hirt, 1987;Peterson, 2000), and Europe . A few derecho events occur each year in Europe, and some of them induced catastrophic forest damage (Taszarek et al., 2019), and South America (Negrón-Juárez et al., 2010). Although, not a single derecho events have been reported previously in Russia. A more detailed further analysis of these storm events should be carried out to confirm their nature. 670 Most of tornado-induced windthrowswindthrow areas have W max and W mean less than 200 m (Fig. 14 15 c,d). Instead, the distribution of W max of non-tornado induced windthrowswindthrow areas shifted toward larger W max . In particular, 103 windthrowswindthrow areas (32.9%) have W max > 1000 m. The K-S test shows that samples of both W max and W mean for tornado-and convective storm induced windthrowswindthrow areas are from different distributions. Width of storm tracks is several times higher than the width of windthrowswindthrow areas. Moreover, the W TRmax of windthrowswindthrow areas 675 caused by non-tornadic storms is several times higher than the W TRmean . W TRmax exceeds 30 km for three widest convective stormstwo derechos occurred on 27 June 2010 and 29 July 2010, and one non-convective storm occurred on 7-8 August 1987.

Discussion: Data and method limitations
Due to several data and method limitations, the presented database is spatially and temporally inhomogeneous and hence 680 incomplete. The presented database likely lacks many windthrowswindthrow that occurred in ER in 1986-2017.
Specifically, since most of windthrowswindthrow were delineated from the GFC and EEFCC datasets, forest loss areas which are initially missed or underestimated in these datasets, could be as well missed in our database. The verification performed with the Landsat images and the HRI allows to reduce these omissions. In particular, we found several windthrowswindthrow in small-leaved or broadleaved forests that were significantly substantially underestimated in the GFC 685 dataset.
The efficiency of the method depends on the percentage of forest-covered area. In general, our data has highest accuracyis more complete for low-populated northern and eastern part of the the ERER, where forests cover 70-90% of the territory and dark-coniferous forests are widespread (Bartalev et al., 2016). However, some regions in the northern part of the ER are not covered by HRI which prevents from thorough verification of some windthrow areas. 690 In turn, the data may be less accurate In thefor southern part of the study area, the dataset is likely less complete wheresince some windthrow areas probably could becan be overlooked missed. The most reliable estimates of wind-damaged area can be obtained for low-populated northern and eastern regions of the ERER, where forests cover 70-90% of the territory (Bartalev et al., 2016) (Fig.1). In the southern part, the probability of the windthrows omission is higher .In particular, 695 iIt is possible to miss overlook amiss windthrow if a storm or tornado passed through areas of intensive timber harvesting or agricultural lands . Salvage logging performed shortly after a storm event also complicates the identification of windthrowswindthrowwind-related forest damage (Baumann et al., 2014). However, in most cases, the time interval between storm event and salvage logging in the the ERER was quite long, i.e., more than a year, except for more populated southern regions. 700 Currently, the proposed method requires expert verification at almost all stages, which prevents to switch it into the automatic mode. The possibility of automated searching throughout the GFC and EEFCC datasets is limited by a wide variety of geometrical shapes of windthrowswindthrow shapes and their overlapping with other forest disturbances. The data collection process requires the use of numerous and diverse sources such as the HRI from various public web-services, weather station reports, eye-witness and media reports, etc. 705 While the algorithms for automated forest disturbances detection based on satellite data are well-developed and applied at the regional-to-global scale (Huo et al., 2019), automated attribution of forest disturbances to their causes, namely windstorms, logging, wildfires, insect outbreaks, and others, remain a critical challenge for remote sensingsatellite-based forest monitoring. The spectral characteristics of various types of disturbances, e.g., windthrowswindthrow and logged areas, are often similar (Baumann et al., 2014) that complicates the automated attribution automatization of attribution. The 710 promising approaches in this process is the complex use of spectral, temporal, and topography-related metrics (Oeser et al., 2017) as well as implementing of advanced image classification/segmentation methods (Oeser et al., 2017;Liu et al., 2018;Huo et al., 2019). In future studies, such approaches can be applied to automate delineation of windthrowswindthrow areas in the ERER using satellite data of with various spatial resolution.
We have to stress Ttemporal inhomogeneity of our database, especially for small-scale windthrowswindthrow areas, comes 715 fromdue to the following causes: 1. The use of two different Landsat-based products to search windthrow-like disturbancesthe EEFCC before 2001 and the GFC after. The GFC data have higher accuracy of forest loss detection and of initial time assigning, than the 24 EEFCC (see Section 4.1 for details), which allows to detect more windthrowswindthrow areas. Thus, the annual number of windthrowswindthrow events 2.5 times higher in the GFC period compare to the EEFCC period. 720 2. After 2002-2003, the HRI had become available, which made it possible to confirm the tornadic nature of windthrowswindthrow. The observed increase in the number of tornado-induced windthrowswindthrow events after 2003 is very likely related to the appearance of the HRI.
3. The start of the Sentinel-2 mission in 2015 providing the images with a 10 m spatial resolution (Drusch et al., 2012) had also increased the possibility for windthrow identification. 725 4. A strong decrease in the volume of timber harvesting occurred in the ERER, especially in its northeastern past, after the Soviet Union dissolution (Potapov et al., 2015). This could lead to more omission of windthrowswindthrow areas in the late 1980s compare to the subsequent period because of their masking outoverlapping with logginglogged areas.
5. The number of windthrow areas and storm events has been determined with the use of arbitrary threshold values. It 730 can be substantially change due to modification of these thresholds (see Sections 4.1.4. and 4.3. for more details).
So, the data on the number of windthrow areas may be more inhomogeneous than assessment of wind-affected area.
The optimization of other threshold values can be evaluated in further studies that should involves ground-based data.Thus, the presented database should be used for assessing interannual variability with caution. Special assumptions should be made to estimate linear trends. For instance, they can be obtained for particular regions, e.g. for those with little changes of 735 forestry practices, and for relatively large windthrowswindthrow areas, that are well-detected from both the EEFCC and the GFC data. For instance, linear trend of in number of windthrowswindthrow with area ≥ 1 km 2 amounts to 0.27 year -1 and is statistically significant at 0.05 level 1 . This increase of wind-related windthrowswindthrowforest disturbances is in line with observed increase of such characteristics as convective precipitation (Ye et al., 2017;Chernokulsky et al., 2019), convective cloudiness (Sun et al., 2001;Chernokulsky et al., 2011), convective instability indices (Riemann-Campe et al., 2009;740 Chernokulsky et al., 2017) in the the ERER in the last decades.
Currently, the proposed method requires expert verification at almost all stages, which prevents to switch it into the automatic mode. The possibility of automated searching throughout the GFC and EEFCC datasets is limited by a wide variety of windthrow shapes and their overlapping with other forest disturbance. The data collection process requires the use of numerous and diverse sources such as the HRI from various public web-services, weather station reports, eye-witness and 745 media reports, etc.
While the algorithms for automated forest disturbance detection based on satellite data are well-developed and applied at the regional-to-global scale (Huo et al., 2019), automated attribution of forest disturbance to their causes, namely windstorms, logging, wildfires, insect outbreaks, and others, remain a critical challenge for satellite-based forest monitoring. The spectral characteristics of various types of disturbance, e.g., windthrow and logged areas, are often similar (Baumann et al., 2014) 750 25 that complicates the automated attribution. The promising approaches in this process is the complex use of spectral, temporal, and topography-related metrics (Oeser et al., 2017) as well as implementing of advanced image classification/segmentation methods (Oeser et al., 2017;Liu et al., 2018;Huo et al., 2019). In future studies, such approaches can be applied to automate delineation of windthrow areas in the ER using satellite data with various spatial resolution.

Conclusions 755
The compiled GIS database contains the most complete information on a relatively large stand-replacing windthrowswindthrow areas in the forest zone of the ER in 1986--2017. The database contains 102747 elementary damaged areas, combined into 700 windthrowswindthrow areas, which were caused by 486 storm events. For each windthrow, we determined its type with degree of certainty, dates or date ranges, and geometrical characteristics. Database also contains weather station reports and links to additional information on storm events from the media. We included into the database 760 only the stand-replacing windthrowswindthrow with an area > 0.05 km 2 and > 0.25 km 2 for the tornado-and non-tornadoinduced windthrowswindthrow, respectively.
The total area of windthrowswindthrow area amounts to 2966 km 2 , namely 0.19% of the forested area within the study region. Most of windthrowswindthrow in the the ERER, i.e., 82.5% of the total wind-damaged area, are related to convective squalls and downbursts, which occur mainly in June and July. The ten most impactful storms are responsible for 59.2% of 765 the total forest damage. More than 55% of windthrowswindthrow events in the database are tornado-induced, but their contribution to total damaged area is much lowerit is less than 13%. Non-convective windstorms and snowstorms caused only 4.6% of storm-damaged area. The presented method has several limitations which that results in spatial and temporal inhomogeneity of the compiled database specifically for small-scale windthrowswindthrow areas and hence determine the dataset incompletentss. Because of influence of forest area percentage and forestry practice, such windthrowswindthrow areas can be rather missed in the southern part of the the ERER compare to the northern part. Because of coarser resolution of the EEFCC data and lack of the 775 HRI, such windthrowswindthrow areas can be rather missed before 2001. The obtained increases in number of windthrowswindthrow events and their affected area are mainly artificial. However, the positive trend is likely real for largescale windthrows, namely for ones with the area ≥ 1 km 2 .
Despite the incompleteness, tThe compiled database provides a valuable source of spatial and temporal information on windthrowswindthrow events in the the ERER, which previously has been incomplete. On the one hand, the database allows 780 estimate the role of wind-related disturbances in comparison with other natural disturbances in forests and improve our understanding of different forest species susceptibility to windstorms. On the other hand, the database presents a unique 26 source of information on storm and tornado events causing windthrowswindthrowforest damage in the the low-populated forest zone of the ERER. It includes numerous of previously unknown storms and tornadoes, which caused forest damage, and also clarifies information on known storm events. Thus, the database significantly substantially contributes to the 785 climatology of severe storms and tornadoes in the the ERER. Based on the compiled database, further studies may be carried out to determine the contribution of climate variability to the inter-annual variability of wind-related forest damage, and to quantify the risk of windthrows events in forests of the entire ER.

Author contributions.
ASh and ACh designed the study. ASh and ASe performed windthrow identification using satellite data. ASh and IA carried out an analysis of additional information to determine storm event types and dates. Ash and ACh, with contributions by IA, wrote the initial draft of the paper and produced the maps and figures.
Competing interests. The authors declare that they have no conflict of interest. 795    The HRI are unavailable or do not allow to determine the direction of the fallen trees and only two out of three additional signature are confirmed.
HRI are unavailable or do not allow to determine the direction of the fallen trees; quasi-linear structure of a windthrow without turns of a track, and a ratio of length and width < 10:1 The date of a storm event indicate a low probability of a convective storm (e.g., autumn season) and lack of elongation along the wind direction (especially for windthrowswindthrow induced by snowstorms)