Strong winds may uproot and break trees and represent a major natural disturbance for European forests. Wind disturbances have intensified over the last decades globally and are expected to further
rise in view of the effects of climate change. Despite the importance of such natural disturbances, there are currently no spatially explicit databases of wind-related impact at a pan-European scale. Here, we present a new database of wind disturbances in European forests (FORWIND). FORWIND is comprised of more than 80 000 spatially delineated areas in Europe that were disturbed by wind in the period 2000–2018 and describes them in a harmonized and consistent geographical vector format. The database includes all major windstorms that occurred over the observational period (e.g. Gudrun, Kyrill, Klaus, Xynthia and Vaia) and represents approximately 30 % of the reported damaging wind events in Europe. Correlation analyses between the areas in FORWIND and land cover changes retrieved from the Landsat-based Global Forest Change dataset and the MODIS Global Disturbance Index corroborate the robustness of FORWIND. Spearman rank coefficients range between 0.27 and 0.48 (
Natural forest disturbances represent a serious peril for maintaining productive forests. Studies indicate that their occurrence can reduce primary production and partially offset carbon sinks or even turn forest ecosystems into carbon sources (Kurz et al., 2008; Yamanoi et al., 2015; Ziemblińska et al., 2018). This is particularly critical for windthrow and tree breakage due to strong winds, which represents a major natural disturbance for European forests (Schelhaas et al., 2003; Seidl et al., 2017). Such disturbances are intensifying globally, a trend that is expected to continue with further climate change (Bender et al., 2010; Knutson et al., 2010; Seidl et al., 2014).
European windstorms are associated with areas of low atmospheric pressure that typically occur in the autumn and winter months (Martínez-Alvarado et al., 2012). Deep low-pressure areas frequently track across the northern Atlantic Ocean towards Western Europe, passing the northern coasts of Great Britain and Ireland and into the Norwegian Sea. However, when they track further south, they can potentially hit any country in Europe. In 1999, windstorm Lothar damaged approximately 165 million cubic metres of timber mainly in France, Germany and Switzerland (Gardiner et al., 2010), which is equivalent to about 140 % of the average annual roundwood harvested in the countries affected (FAOSTAT, 2019). In 2005, 75 million cubic metres were damaged by windstorm Gudrun in Sweden (Gardiner et al., 2010), equivalent to about 1 year of cuttings in the same area (FAOSTAT, 2019). In 2007, windstorm Kyrill caused the loss of 49 millon cubic metres of timber in Germany and the Czech Republic. In 2009 and 2010, windstorms Klaus and Xynthia hit forests in France and Spain and caused timber losses totalling approximately 45 million cubic metres. In 2018, windstorm Vaia hit the northeastern regions of Italy causing a damaged growing stock volume of about 8.5 million cubic metres.
The socio-economic consequences of wind disturbances can be especially critical for local economies highly dependent on the forest sector. Countries in Northern Europe, Central Europe and Eastern Europe, where the forest sector may cover up to 6 % of the national GDP (Forest Europe, 2015), are, therefore, potentially more vulnerable to wind-related impacts.
Despite the risks they pose, spatially explicit databases of wind
disturbances across Europe currently do not exist. Recent assessments of
current and future forest damage due to windstorms at a European scale are
based on catalogues of disturbances collected at country level (Gregow et al., 2017; Schelhaas et al., 2003; Seidl et al., 2014). Such databases
(e.g. FORESTORM) are subject to multiple sources of bias and uncertainty
associated with the diversity of the underlying inventories. Furthermore,
estimates of forest damage aggregated at a national scale may only partially
represent the spatial variability of the phenomenon. In fact, the coarse
spatial resolution of such data hampers inferential analysis of potential
drivers of forest vulnerability and their use in spatially explicit models
to monitor or forecast wind-related impacts (Masek et al., 2015; Phiri and Morgenroth, 2017). Despite the lack of systematic mapping of wind
disturbances in European forests, a multitude of local, national and
transnational initiatives have accurately mapped forest areas affected by
wind over the last decades. These data represent highly informative
observational records for characterizing spatial patterns of forest damage.
However, they are collected by different institutes and are often difficult
to retrieve or poorly documented. Since 2012, the Copernicus Emergency
Management Service (
In this study, we try to fill the above-mentioned gap. To achieve this, we collected and harmonized 89 743 forest areas damaged by wind into a consistent geospatial dataset. The work was carried out through a unique joint effort of 27 research institutes and forestry services across Europe. This collaboration led to the first spatially explicit database of wind disturbances in European forests over the period 2000–2018, hereafter referred to as the FORWIND database. We believe that it provides essential spatial information for improving our understanding of forest damage from wind and can assist in large-scale systematic monitoring and modelling of forest disturbances and their effects on the land–atmosphere system. In the following sections, we describe the data collection, the harmonization process and the cross-comparison performed against satellite retrievals of changes in vegetation cover and data from national inventories of forest disturbances. We conclude the data description with some examples of the possible usage of the FORWIND database.
List of institutions responsible for wind disturbance mapping and the corresponding number of records collected and acquisition methods employed.
Conversion table for passing data from class of damage to degree of damage. Records of windstorms that occurred in Italy in 2015 (Toscana) and in 2018 (Veneto) are already expressed as damage degree in a consistent range between 0 (no damage) and 1 (full destruction of forest pattern).
We collected wind disturbance events caused by windstorms or tornadoes that
occurred in Europe between 2000 and 2018. A wind disturbance event is
represented by a geo-referenced polygon that delineates the damaged forest
stand, regardless of the degree of damage. The original acquisition of the
polygons was made by aerial and satellite photointerpretation or field survey
(Table 1). Therefore, the polygons are delineated when a reasonably
homogeneous patch of damaged forest is detected from the ground or remotely.
The data were managed mostly on the Google Earth Engine platform (Gorelick et al., 2017) to efficiently quantify the extent of disturbances over large scales and extract additional informative attributes (e.g. Hansen et al., 2013; McDowell et al., 2015). We structured the data collection process in four main phases, described below.
Attribute table of the FORWIND database. The name and description of the
attributes associated with each wind disturbance in FORWIND and listed in
the .dbf file are given. Missing data are reported as
The FORWIND database is the final output of the data collection procedure
and it is publicly available at
Spatial and temporal distribution of wind disturbances in the
FORWIND database.
Examples of wind disturbances recorded in the FORWIND database.
Overall, FORWIND includes 89 743 records, corresponding to a million hectares of forest area affected by wind disturbances during the 2000–2018 period. Each record should not be viewed as independent as a single storm may cause multiple, geographically disjunct disturbances. At a European level, the median forest disturbance patch caused by wind measures 1.07 ha (Table 4). However, there is substantial variability across disturbances and countries likely driven by the high heterogeneity of forest and landscape characteristics. Figure 1 shows the spatial and temporal variations of records in the FORWIND database. In order to better visualize the data, we summed the areas affected by wind disturbances in 0.5
Statistics of wind disturbance records collected in the FORWIND database aggregated at country level and for all of Europe.
Representativeness of FORWIND. The first estimate of representativeness at the European level accounts for damaging wind events that occurred during the 2000–2018 period in the countries currently included in FORWIND. The second estimate of representativeness at the European level accounts for all damaging events occurring during the 2000–2018 period, including those countries currently missing in FORWIND.
Continued.
The lack of alternative datasets with the same spatially explicit mapping of wind disturbances as in FORWIND does not allow for a standard validation exercise. Therefore, we evaluated the validity of FORWIND based on the plausibility of the collected spatial delineations of wind disturbances with respect to two satellite-based proxies of forest disturbances and estimates of forest damage reported in national inventories.
FORWIND was initially compared with satellite-based estimates of forest
cover loss derived from the Global Forest Change maps (Hansen et al., 2013) (GFC,
For each selected FORWIND record we computed the area of affected forest based on the spatial delineation of the polygon and the corresponding Landsat-derived forest cover loss and calculated the correlation between the two sets of estimates. In order to account for the spatial dependence structure of FORWIND data, correlation values were derived for 100 subsets of 1000 records randomly selected from the entire dataset. The final estimate of correlation was then quantified as the average of the correlation values derived from the 100 subsets.
Results for the whole dataset are shown in Fig. 3a. Overall, we found a
modest but significant Spearman rank correlation coefficient (
Validation of the FORWIND database.
FORWIND was also compared with an independent dataset of satellite-based
estimates of forest disturbance as expressed by the MODIS-based Global
Disturbance Index (Mildrexler et al., 2009) (MGDI,
Overall, we found a low but significant correlation coefficient (
FORWIND data were finally compared with estimates of damaged growing stock
volume (GSV) that are recorded at country level in the FORESTORM database
for five windstorm events: Slovakia in 2004, Sweden in 2005 (windstorm Gudrun),
Germany in 2007 (windstorm Kyrill), the Czech Republic in 2007 (windstorm Kyrill)
and France in 2009 (windstorm Klaus). We derived the damaged GSV by multiplying
the estimated GSV by the percentage damaged, both of which are reported in
FORESTORM. An analogous metric was derived from FORWIND data by first
calculating, for each FORWIND record, the amount of GSV lost by multiplying
the areal average GSV by the damage level reported for the record. As the
damage level was only reported for windstorm Klaus, for the other events we assumed a
damage level equal to the average level reported for windstorm Klaus weighted on the
spatial extent of each record. The GSV was retrieved from the GlobBiomass
dataset (Santoro et al., 2018;
Overall, results show that the magnitude of damage estimated from FORWIND and FORESTORM are largely different, except for windstorm Klaus in 2009 in France, for which we found a very good agreement (Fig. 3e). For most of the events, however, FORESTORM tends to systematically give higher forest damage estimates than FORWIND, with differences exceeding 90 %. We note that such differences persist when we derive FORWIND estimates of damaged GSV assuming a 100 % damage degree for all records (not shown). Therefore, the uncertainty in the damage degree in FORWIND does not substantially affect the difference between FORWIND and FORESTORM. We recognize that estimates of forest damage based on FORWIND are fully dependent on the GSV derived from GlobBiomass. Indeed, any deviations of the mapped GSV from the true forest state are inherently translated into our damaged GSV estimates. In particular, the GSV map refers to the year 2010; therefore, it is very likely that it largely reflects the biomass conditions following, rather than preceding, the windstorm events (all five events considered in this validation exercise occurred before 2010).
In order to disentangle such source of bias, we derived country-scale
estimates of average GSVs for the year 2000 (pre-event conditions) from the
State of Europe's Forest dataset (Forest Europe, 2015) (
Similar to the previous results, we found higher values of damaged GSVs in FORESTORM than in our estimates based on the integration of FORWIND and country values of GSVs (Fig. 3f), except for windstorm Klaus. We recognize that FORWIND could miss some wind damage occurrences, for instance due to incomplete detection of wind disturbance from aerial photointerpretation or difficulties in mapping inaccessible areas by ground surveys. However, according to the institutions responsible for the data acquisition, the forest areas affected by the windstorm events considered in this validation exercise were exhaustively mapped. Therefore, possible residual omissions are expected to only marginally affect our results. We, therefore, argue that a possible source of error may be associated with the FORESTORM database. Estimates of forest damage from FORESTORM originate from different sources and are collected by multiple actors. Hence, the loss figures should be viewed in light of their potential biases, including a possible overestimation of the true impacts.
For demonstration purposes, we show a series of possible applications of the FORWIND database. We recognize that the examples described in the following sections are an oversimplification of the relationships observed in nature and of the biomechanical processes that may cause wind disturbances or that can be triggered by wind disturbances. More sophisticated approaches could be employed to better explore and predict the forest response functions to wind disturbances. For example, multiple variables, susceptibility factors and drivers (e.g. tree species, tree dimension, management regimes, planting patterns, soil depth and snow cover) contribute concurrently to modulate the forest response to wind disturbances (Hart et al., 2019; Klaus et al., 2011; Mitchell, 2013), and their contribution should be analysed in a multidimensional space (e.g. Sect. 5.1 and 5.2). Therefore, the approaches described here should not be considered a reference methodology but only as informative applications for exploring the usefulness of the FORWIND database.
The exploration of the relations between forest dynamics and scale can
reveal important information on ecosystem spatial organization by addressing
preservation of information integrity in upscaling and downscaling procedures of
land surface parameterization for ecological modelling applications (Forzieri and Catani, 2011). Here, we explore – in a simplified approach – the scaling relations of the degree of damage of wind disturbances collected in FORWIND. To achieve this we estimated, for each record, the cover fractions of different plant functional types (PFTs) including broadleaf deciduous (BrDe), broadleaf evergreen (BrEv), needleleaf deciduous (NeDe) and needleleaf evergreen (NeEv). Cover fractions were retrieved from the annual land cover maps of the European Space Agency's Climate Change Initiative (ESA-CCI,
Results show that all considered PFTs generally have a higher degree of
damage for wind disturbances with small spatial extent (Fig. 4a). This may
reflect a better delineation of small affected areas when the damage is
typically higher and homogeneous. Furthermore, the declining scaling
relations suggest potential spatially varying dampening effects of wind
severity due to landscape heterogeneity over large areas compared to more
homogeneous patterns in small forest patches. Model fitting shows reasonably
good performances, with
Scaling relations of the degree of damage.
Parameters and performance of fitting regression models expressing
the degree of damage as a function of the area affected. The relationship
between the degree of damage (
The vulnerability of forests to natural disturbances is a key determinant of
risk and reflects the propensity of a forest to be adversely affected when
exposed to hazardous events (IPCC, 2014). Vulnerability is largely controlled by local environmental conditions, such as climate and forest characteristics, which regulate the sensitivity of ecological processes to
disturbance agents (Lindenmayer et al., 2011; Seidl et al., 2016; Turner, 2010). Here, we employ FORWIND records to quantify the forest vulnerability as a function of the fraction of evergreen needleleaf forest and annual maximum wind speed. The fraction of NeEv was derived from the ESA-CCI product aggregated at 0.5
Susceptibility factors and drivers of forest vulnerability to wind
disturbances.
Wind disturbance areas manifest a substantial variability, as evident from
the generally high values of the coefficient of variation. However, when
data are spatially averaged at bin level, simple linear regression models
show a reasonably good fit, with
Natural disturbances are accelerating globally, but their full impact is not
quantified because we lack an adequate monitoring system. Remote sensing
offers a means to quantify the frequency and extent of disturbances over
landscape-to-global scales (McDowell et al., 2015). For instance, some pioneering studies have begun producing classification maps of various forest disturbance agents based on remote sensing data (Cohen et al., 2016; Hermosilla et al., 2015; Potapov et al., 2015; White et al., 2017). However, the attribution of forest change to windstorms remains challenging. Previous systematic monitoring has been performed only over limited areal extents and showed considerable uncertainty (Baumann et al., 2014; Schroeder et al., 2017) mostly due to the limited number of sampled wind-affected areas available for training and testing classification algorithms (Schroeder et al., 2017). In this respect, FORWIND data can be used to enhance our capability to detect and attribute forest damage due to windstorms from remote sensing data. Here, we tested different types of classification trees in combination with a Sentinel-2 imagery and the FORWIND database to automatically map wind disturbances that occurred following windstorm Vaia in October 2018 in the Dolomites in northern Italy (Pirotti et al., 2016). Google Earth Engine was used to create a single image composite from a stack of cloud-free pixels (11 and 28 images acquired before and after the windstorm event, respectively). The median was used as a reducer over the vector of pixel values derived from each image, after masking cloudy pixels using the cloud probability raster delivered from atmospheric, terrain and cirrus correction of the Sen2Cor processor (Louis et al., 2018). Further masking was applied to process only pixels covered by forest, using the 2018 estimated forest cover map from the Global Forest Change 2000–2018 dataset (Hansen et al., 2013). Binary classification, i.e. damaged versus non-damaged, was applied over a set of 1000 completely damaged areas retrieved from FORWIND and 1000 non-damaged areas. Half of these were used for training and validation, the other half for unbiased testing of the model performance. The feature vector used for predictors included reflectance values recorded by Sentinel-2 after radiometric and atmospheric correction (i.e. bottom of atmosphere) and a
tasselled cap (TC) transform of reflectance bands to the brightness,
greenness and wetness domain. The TC was added as it is reasonable that
wind-affected areas will provide a higher degree of brightness and lower
degree of greenness with respect to undisturbed areas (Baumann et al., 2014). Several machine learning algorithms were employed, including Random Forest,
Extremely Randomized Forest, Gradient Boosting Machines, Deep Neural Networks and Stacked Ensemble, all trained and cross-validated based on
Results, based on the best performing classification model (Random Forest), provided very promising accuracy, with an F1 score of 0.97, 27 false positives and 1 false negative over 915 pixels used for testing (507 undamaged and 408 damaged). Figure 6 shows mapped probability of wind occurrence – with blue to red, respectively, representing zero to one probability of a heavily hit area in the Veneto region. Based on visual comparison with ground data, the automatic classification is able to capture the spatial patterns of wind damage. It is worth noting that damage in forest–non-forest nexus is less accurate due to pixel mixing. Another point worth further investigation is that data may be defined as false positives from binary classification but could actually be true positives that were not mapped due to human error. On the other hand, false negatives may be true negatives in the sense that small patches of standing trees might be present in mapped areas due to the understandable minimum level of detail that must be adopted.
Remote sensing classification of windthrows. Probability of windthrow obtained from random forest classification of Sentinel-2 reflectance bands and their tasselled cap transformation in a sampled area of the Dolomites in northern Italy affected by windstorm Vaia in October 2018. Black polygons show the actual wind disturbances.
Land surface models (LSM) are key components of Earth System Models that are widely applied to support policy-relevant assessments on the impact of climate change on terrestrial ecosystems (Quéré et al., 2018). Recently, windstorm effects have been incorporated in LSMs (Chen et al., 2018). However, these models are hampered by the lack of harmonized spatially explicit information on windstorms required as input for robust model parameterization and large-scale representation of wind disturbance. In such contexts, the FORWIND database represents a valuable source of harmonized wind-affected forest areas for improving model calibration and/or evaluation. To illustrate such possible applications, FORWIND was used as an independent data source to evaluate the LSM ORCHIDEE (revision r4262), which simulates windthrow damage and was parameterized with observations prior to the FORWIND time frame.
ORCHIDEE r4262 was parameterized to the extent possible with observed
parameter values. Nevertheless, tuning windthrow parameters remained
necessary for gustiness, maximum damage rate (which is a parameter to
account for the large simulations units, i.e. 2500 km
The model simulated a total annual damage of 30 Mm
Differences in spatial and temporal definitions between ORCHIDEE and FORWIND
were partly accounted for by extracting storm damage estimates from ORCHIDEE
only when the storm was included in FORWIND. Following this, the ORCHIDEE
model appears to overestimate the damage rate in years with small storms but
failed to estimate the damage rate of windstorm Klaus in 2009 (Fig. 7). This suggests
that the tuned relaxation factor for the damage function (
Observed and simulated cumulated forest area damaged by windstorms
between 2000 and 2015 over Europe. The observed damage area was extracted
from the FORWIND dataset (shown in blue), whereas the simulated area comes
from ORCHIDEE r4262 with
These results show that evaluating the capacity of land surface models to project storm damage hinges on our ability to precisely define the storm events recorded in the databases and our ability to use this information to estimate key model parameters such as the relaxation factor and the maximum damage rate.
FORWIND may also be employed to improve the predictive performances of slope stability models that rely on water–soil interactions and soil mechanics. Vegetation affects terrain properties in a variety of way, including the modification of hydraulic conductivity, the regulation of evapotranspiration and the increase in soil strength by apparent root cohesion (Amundson et al., 2015; De Baets et al., 2008). This, in turn, may strongly condition terrain response to external forcing, such as intense rainfall and seismic shaking, leading to mass wasting in the form of shallow landslides and soil erosion (Moos et al., 2016; Ruiz-Colmenero et al., 2013).
Analysis of the indirect effects of wind damage on slope instability. Changes in NDVI and probability of landsliding following windstorm Vaia of October 2018 in the Dolomites in northern Italy, shown in
We have tested the capability of FORWIND to provide data for assimilation in shallow landslide hazard models and for model validation by selecting the dataset relative to windstorm Vaia in October 2018 in the Dolomites in northern Italy and using it to model indirect effects of wind disturbance on slope stability. A multivariate machine learning model for shallow landslide susceptibility has been trained and applied on pre-storm terrain attributes to reveal relative probability of occurrence and then applied again to post-storm conditions to measure the effects of forest disturbance on the hazard. The terrain attributes considered in the analysis include elevation, slope angle, slope curvature variability, local rainfall patterns, geo-mechanical classes, potential soil saturation, contributing area and pre- and post-storm Normalized Difference Vegetation Index (NDVI) maps from Landsat 8 level-2 imagery. The dataset was trained by a Random Undersampling (RUS) Boosted Random Forest regressor (Catani et al., 2013) on a validated shallow-landslide dataset derived from the Italian National catalogue IFFI (Trigila et al., 2013). The training process highlights that NDVI, typically considered a good proxy of biomass density, is ranked second in terms of explained variance and seems to strongly condition landslide susceptibility in the Dolomites. The FORWIND database collects dated and graded information on wind damage to forests that directly correlates to marked changes in NDVI values, as can be observed in Fig. 8a. The effects of the damage recorded in the FORWIND dataset are measurable by comparing the levels of susceptibility before and after the occurrence of windstorm Vaia (Fig. 8b). As can be appreciated in the map, the red areas, which reveal a marked increase in the probability of landslides, match the FORWIND polygons very well and clearly indicate the usefulness of the wind damage geographical databases in slope hazard prediction and modelling. In Fig. 8b, we also note some omission and commission errors. They, however, can be easily explained by noting that vegetation stripping (or vegetation scantiness) is only one of the factors contributing to landslides. Therefore, wherever Vaia has damaged forests but slopes are very gentle, no shallow landslides can be generated. On the other hand, outside FORWIND polygons landslides may still develop, due to the prevailing action of other factors, such as unfavourable geological conditions or strong concentrated rainfall.
The use of FORWIND data in landslide modelling is not limited to the cross-validation of biomass volume changes but can also be extended to the usage of the dataset as an additional predictor in multi-variate statistics. We noted that the overlapping of FORWIND polygons and NDVI stress (brown) areas shows few exceptions. In such areas, the two factors seem to behave independently. In particular, locations where wind damage does not correspond to a NDVI change might reveal cases where the possible storm effects on soil stability are not captured by satellite-based variations in biomass content and must be accounted for by a different metric. That, in turn, opens the way to important future developments in the usage of wind-driven damage datasets in slope stability forecasting.
The data used in this study are freely available at (
Modern and forthcoming Earth observation systems (McDowell et al., 2015), new generation of land surface models (Bonan and Doney, 2018), recent developments of cloud computing platforms (Gorelick et al., 2017) and machine learning approaches (Reichstein et al., 2019) are offering unprecedented opportunities to explore and predict ecosystem dynamics at an increasing spatio-temporal resolution and sophistication level. In light of such progress, it is of paramount importance to implement robust calibration and validation procedures based on reliable ground observations. In order to capture the variability of ecosystem response across wide environmental gradients, reference ground truth needs to be collected over large spatial scales. In this context, FORWIND represents an essential dataset to improve our capacity to detect, understand and predict wind disturbances and quantify their impact on forest ecosystems and the land–atmosphere system. The FORWIND database is the first Pan-European collection of spatially delineated forest areas affected by wind disturbances and includes all major events that occurred over the 2000–2018 period. Future research should aim to further populate FORWIND with missing damaging wind events.
GF designed the study. MP performed the data harmonization. MG assisted in data integration tasks. MM, CNik, MR, JT, DS, CNis, DJ, BG, FG, RC, AW, FP, FM, SI, WLS, KS, KZK, PSJ, MM, FS, LK, IH, MN, PW,and GC collected forest disturbance data. FP ran the classification models. FC ran the slope instability model. YYC and SL ran the ORCHIDEE model. GF analysed the data and wrote the manuscript with contributions from all co-authors.
The authors declare that they have no conflict of interest.
This research has been supported by the European Commission, Joint Research Centre (project FOREST@RISK).
This paper was edited by Scott Stevens and reviewed by Rupert Seidl and one anonymous referee.