1 Introduction

Changing climate represents the most important threat for mountain forests in the Mediterranean region and affect forests vitality and productivity (Lindner et al. 2014). Although increasing temperatures, atmospheric CO2 concentration. and nitrogen fertilization suggest an enhancement of plant growth rates at the global scale (Frank et al. 2015; Pan et al. 2011; Xie et al. 2015), the increased frequency of extreme weather events such as spring frosts, heat waves, prolonged drought, and storms (Diffenbaugh et al. 2018; Garcia et al. 2014; Kramer et al. 2000; Williams 2014; Williams et al. 2014) represent the main limitation to forest growth rate especially in the Mediterranean region. Under these ongoing climate changing conditions, Mediterranean forest ecosystems are facing increasing threats related to wildfire risks, pest outbreaks, and forest dieback.

Despite the increasing trend in mean temperatures forecast for all the climate change scenarios (IPCC 2014), extreme events such as spring frosts represent a relevant climatic stress factor for mountain vegetation in the Mediterranean region, due to increases in daily temperature variance characterizing the ongoing climatic change (Gu et al. 2008; Nicholls and Alexander 2007; Pagter and Arora 2013; Rigby and Porporato 2008; Vitasse et al. 2018; Wang et al. 2014).

Warmer winters and spring temperatures affect plant phenology because the predicted opening of vegetation buds and leaf expansion increases plant vulnerability to late spring frosts (Augspurger 2013). According to Jeong et al. (2011), European forests showed an early start of the growing season by 3.5 days decade−1 and a total increase of the growing season by 6.5 days decade−1.

These extreme events according to their magnitude, frequency, and spatial scale may affect not only forest productivity but also species composition and distribution with important consequences at an ecosystem level. Nevertheless, most of these events are not taken into account because of their small spatial scale, but their frequency combined with their distribution at a local scale and their effects distributed over several growing seasons, represent a key driver of forest productivity loss and a carbon uptake reduction unaccounted for the ecosystem level (Bascietto et al. 2018; Hufkens et al. 2012; Kim et al. 2014; Príncipe et al. 2017).

The main impacts of spring frosts are related to foliar mortality due to ice formation within buds and leaf cells and tissues. Frost damages at a higher intensity may affect apical meristems causing changes in the canopy architecture with a generalized reduction in tree productivity due to late refoliation from branch and epicormic dormant or adventitious buds (Augspurger 2009; Dittmar et al. 2006; Kramer and Kozlowski 1979; Vitasse et al. 2014).

Mountain forests in the Mediterranean region are particularly exposed to frost disturbance due to the combination of anticipation of phenological phases induced by warmer late winter temperatures and spring frost events. In particular, beech forests (Fagus sylvatica L.), widely distributed in Europe from the southern latitudinal limit in Sicily up to the northern latitudinal limit in south Norway, show a higher sensitivity to late spring frost than other forest species (Kreyling et al. 2012; Príncipe et al. 2017). In the Mediterranean region, beech forests represent the climax forests, growing at higher altitudes from 900 to 1000 m asl up to the vegetation limit (Houston et al. 2016). In the last two decades, beech forests at the southern limit are experiencing long-term drought stress related to climate change, limiting forest productivity, and at the same time shaping the lower altitudinal population limit (Jump et al. 2006). Moreover, the phenological phase anticipation induced by warmer late winter temperatures of southern beech populations, with respect to Central and South-East European populations, as observed by Visnjic and Dohrenbusch (2004), makes Mediterranean populations more exposed to spring frost events.

Detecting the spatial distribution of these disturbances and the recovery patterns of forest vegetation cover, represent key information for forest management and forest productivity analysis and modeling in order to make reliable estimates of carbon uptake at the regional scale. In the last few decades, Landsat satellites (Landsat 2, 3, 4, 5, and 7) have provided near-global moderate-resolution imagery based on multispectral scanners. Landsat images have been used by the scientific community for ecosystem level monitoring, including the management and analysis of agriculture and forest, and land cover change assessment (Roy et al. 2014). Since 2013, the new Landsat 8 satellite, equipped with the new sensors Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS), provides imagery with refined bands designed to operate in continuity with Landsat TM and ETM+ and two bands in the region of the thermal infrared, for a better characterization of land and vegetation cover. Landsat TM, ETM+, and Landsat 8 time series have been used to detect on a small scale growing season phenology up to large-scale vegetation changes based on the normalized difference vegetation index (NDVI) detection change (Cohen and Fiorella 1998; Coppin et al. 2004; Hayes and Sader 2001; Kovalskyy et al. 2012; Lu et al. 2004; Mancino et al. 2014; Pasquarella et al. 2016). Being NDVI anomalies related to sudden changes of several biotic (e.g., pests) and abiotic (e.g., drought stress, flooding, and geomorphological processes) agents, the spatio-temporal analysis of NDVI pattern based on satellite image time series, may help to detect the impact of known short-term extreme weather events such as spring frosts.

In this study, we analyze the effects of a regional scale spring frost disturbance, which occurred during April 2016 in southern Italy, using 16-day Landsat 8 (OLI/TIRS) time series covering the 2016 growing season. In particular, we investigate the possibility of identifying frost disturbance legacies on forest canopy cover at the end of the growing season, and to analyze spatial variability of late spring frost damages on beech forests based on remote sensing analysis.

Thus, the main objectives of the present study are (i) to assess the spatial pattern of the damage to a beech forest canopy due to a late spring frost and correlated with the peak decrease of NDVI and (ii) to analyze the NDVI temporal recovery after this late spring frost.

2 Materials and methods

2.1 Study area and frost event

The study area is represented by the beech forests covering the Apennine Mountain range of the Basilicata Region in southern Italy. The geomorphology of the region is mostly characterized by mountainous terrain covering almost 90% of its area. The main mountain range is represented by the Apennine Mountains in the eastern part of the region (46.8%), (Online Resource1/OR1) containing the highest peaks of the southern Apennine range, reaching 2248 m asl at Monte Pollino. The western part of the region is mainly characterized by low hills (45.2%) with an altitude below 600 m, sloping towards flat plains on the south coast, along the Ionian Sea.

The climate is consistent with the topographic variability of the region, alternating between a Mediterranean humid type in the mountain area and a Mediterranean dry type in the western part of the region (Mancino et al. 2014). According to the Basilicata Region Forest Atlas (Costantini et al. 2006) (OR1), beech forests occupy 29,900 ha, representing 8.5% of the total regional forests and distributed up to the vegetation limit (1900–2000 m asl).

The late spring frost event studied in this paper was a fast cold air mass with north easterly winds, which occurred on the 25th and 26th of April 2016 (DOY 116/117). Sub-freezing temperatures occurred above 900–1000 m asl, with minimum temperatures reaching − 5.63 °C on DOY 116, as derived from the MODIS/Terra Land Surface Temperature and Emissivity (LST/E) Daily L3 Global 1-km Grid product (MOD11A1) (Fig. 1). Measured daily minimum temperatures were selected according to local meteorological stations closest to the study area (ALSIA Regional Agency, http://www.alsia.it/opencms/opencms/Servizio/SAL/rete.html). The Online Resources (OR2) provided in supplementary material clearly shows the drop of temperatures on the 25th and 26th of April 2016. The frost event severely damaged green leaves within an irregular altitudinal range along the beech forest distribution, depending on the combination between local geomorphic variables like slope and aspect and plant phenology. The low limit of the frost-affected area was determined by the altitude of sub-freezing temperatures, while the upper limit was determined by the plant phenology characterized by unfolded leaves on that date. As a consequence of this combination, the entire slope of the Apennine mountain range turned brown producing a “burned band” persistent during most of the vegetative season (Fig. 2).

Fig. 1
figure 1

MODIS/Terra Land Surface Temperature and Emissivity (LST/E) Daily L3 Global 1-km Grid product (MOD11A1) for DOY 116 (left panel) and 117 (right panel) 2016

Full size image
Fig. 2
figure 2

The “burned band effect” of late spring frost event over beech forest in the Pollino National Park area (a). Detailed picture of the event in the same area at 1600 m asl taken in June 2016 (b)

Full size image

2.2 Landsat 8 OLI/TIRS data and image pre-processing

To assess the extension of beech forest areas affected by the frost disturbance and analyze the NDVI variability across a geographic range, we used 16-day orthorectified and terrain-corrected collection 1 level 1 Landsat 8 OLI/TIRS images (Table 1). The images were downloaded from the USGS EarthExplorer website (https://earthexplorer.usgs.gov/), and registered with UTM coordinates, Datum WGS84, Zone 33 N (path = 188; row = 32). The quality of the images was assessed from Landsat 8 OLI quality assessment (QA) band, by masking contaminated pixels (clouds, cloud shadows, cirrus, and water) using a cloud masking Qgis plugin (Corredor 2018). All images were radiometrically calibrated in order to convert the quantified energy expressed in digital numbers (DN) into at-sensor radiance. Then, images were corrected to top-of-atmosphere (TOA) reflectance which is the planetary reflectance corrected by incorporating exo-atmospheric solar irradiance, Earth–Sun distance, and solar elevation angle at the image acquisition, provided by the associated metadata file (MTL). Finally, all the images were atmospherically corrected using the atmospheric correction algorithm Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH), based on MODTRAN4 radiative transfer code (Felde et al. 2003). The model was parameterized according to the FLAASH user’s manual for multispectral imagery (ENVI 2009). For all the corrected images, the NDVI was calculated using the following Eq. 1, where ρRED and ρNIR are band 4 (red, 0.64–0.67 μm) and band 5 (near-infrared, 0.85–0.88 μm) reflectance respectively.

$$ \mathrm{NDVI}=\frac{\rho_{\mathrm{NIR}}-{\rho}_{\mathrm{RED}}}{\rho_{\mathrm{NIR}}+{\rho}_{\mathrm{RED}}} $$
(1)
Table 1 16-day Landsat 8 scenes for 2015 and 2016 growing season
Full size table

Finally, NDVI images were clipped to the area covered by beech forests according to the forest class distribution of the Basilicata Region Forest Atlas (Costantini et al. 2006).

2.3 NDVI differencing

In order to identify frost disturbance legacies on forest canopy cover after seasonal NDVI recovery, and thus to identify the extension of beech forests affected by late spring frost, we implemented the NDVI differencing technique (Lunetta et al. 2006; Lyon et al. 1998; Nabuurs et al. 2007; Podeh et al. 2009; Schowengerdt 1997; Wilson and Sader 2002). This technique is based on the change detection analysis between two NDVI images and identifies the pixels representing significant positive and negative changes in the tails of a normal distribution of a NDVI difference image histogram. On the contrary, the pixels with no significant changes are clustered around the mean. The NDVI difference image (hereafter ΔNDVI) was then calculated as the difference between the Landsat 8 retrieved NDVI values on the 15th August 2016 (DOY 228) and the value on the 13th August 2015 (DOY 225) (Fig. 3). The ΔNDVI was then reclassified using a threshold value calculated as (μ – n × σ); where μ represents the ΔNDVI pixels mean, σ is the ΔNDVI pixels standard deviation, and n represents the threshold value. The threshold identifies the values of ΔNDVI < (μ – n × σ) representing the pixels characterized by a significant change in NDVI from 1 year to another due to the frost effects. In order to identify the best fit of n × σ to the average ΔNDVI, the n value was calibrated through the trial and test approach, by comparing different n values (n = 0.6, n = 1, and n = 1.5) against ground-truth control points. To this purpose, between May and June 2016, we walked through the forest and selected 21 sites where the canopy appeared heavily affected by frost damage, and 15 sites where the canopy appeared totally unaffected. For each site, we geo-referenced a control point and we used them for the assessment of the best fitting of n × σ of the NDVI difference image average. The complete list of ground-truth control points is now provided in the Online Resources (OR3). Finally, the best fit of n × σ thresholds was determined by Cohen’s Kappa inter-rater agreement coefficient, which identified n = 1 (OR4).

Fig. 3
figure 3

Landsat 8 bands composition (4, 3, 2) (path = 188; row = 32) for the 13th of August 2015 (DOY 225) and the 15th of August 2016 (DOY 228)

Full size image

2.4 NDVI spatio-temporal variability analysis

In order to analyze the effects of the frost disturbance on beech forest NDVI variability on a regional scale, we analyzed the 2016 growing season using 16-day time step calculated NDVI related to local geographical variables. In particular, we used the information derived from the 20-m digital terrain model (Ministero dell’Ambiente e della Tutela del Territorio e del Mare 2012) to calculate elevation slope and aspect for the study area. The geographic and temporal relationships between the dependent (NDVI) and the independent variables (DOY as time, elevation, and slope) were investigated via generalized additive models (GAMs) (Hastie and Tibshirani 1990). GAMs are a non-parametric extension of the generalized linear model (GLM) allowing the inclusion of a non-parametric smoothing function and thus fitting a nonlinear relation among variables. Further, the “aspect” term (dummy coded as North 270° to 90° and South 90° to 270°) was included in the model as a grouping factor, and a tensor product function (i.e., an anisotropic interaction term between x and y) was used both for latitude and longitude and elevation and slope to account for the spatial and geomorphic interactions, respectively. The level of complexity (nonlinearity) of model terms was determined by the estimated degrees of freedom (e.d.f.) of the smoother by using the generalized cross-validation (GCV) criterion, which measures the predictive-squared error of the model (Wood 2006), with low GCV values representing the best compromise between model complexity and fit to the observed data. The contribution of each predictor was assessed by comparing simpler against complex models by the Chi-square likelihood ratio tests. The Wald statistics tested the significant difference of smooth terms from null functions. Finally, the implementation of a specific autocorrelation structure AR(1) in the model accounted for temporal residual autocorrelation.

Data availability

The dataset generated and analyzed during the current study is available in the Figshare repository, https://doi.org/30410.6084/m9.figshare.6409235.v1. (Nolè et al. 2018). Datasets not peer-reviewed.

3 Results

The implementation of the NDVI differencing technique provided a detailed map, at a 30-m spatial resolution, of the beech forest affected by the frost disturbance, based on the NDVI differences at the end of the growing season (Fig. 4). The area affected by frost disturbance, calibrated against ground-truth control points, was obtained using a threshold of 1.0 × σ (n = 1) that showed the best accuracy between the different n values tested (n = 0.6, n = 1, and n = 1.5) (OR4).

Fig. 4
figure 4

Distribution of the beech forest areas affected and unaffected by the frost event

Full size image

The frost affected areas, identified with ΔNDVI < (μ – 1 × σ), cover about 13% (3887.26 ha) of the total regional beech forests (29,902 ha), distributed along a 400-m altitudinal range with a mean altitude of 1440 m asl. Figure 5a shows the difference between the average NDVI values of frost affected and frost unaffected beech forests during the vegetative season. An almost full recovery of average NDVI values between the two forest areas is reached at mid-July (DOY 196). During the rest of the vegetative season (DOY 212, 228, 244, 260), the average NDVI seasonal pattern for frost affected and unaffected beech forests constantly shows a slight difference.

Fig. 5
figure 5

Seasonal patterns of average NDVI (disturbed and undisturbed) and NDVI differences for growing seasons 2016 (DOY 148, 164, 180, 196, 212, 228, 244, 260) (left panel). Altitudinal patterns of average NDVI at 25-m altitudinal steps during the 2016 growing season (right panel)

Full size image

Figure 5 shows the seasonal patterns of average NDVI for each Landsat 8 scene (DOY 148 to 228) calculated at 25-m altitudinal steps. The NDVI vs elevation seasonal patterns clearly show the NDVI drop and recovery for each DOY and highlight that most of the canopy recovery occurs between the end of June and the end of July (DOY 180–212). Furthermore, the NDVI vs elevation patterns at DOY 212 and 228, show lower NDVI values between 1000 m asl and 1750 m asl, than the values at higher altitudes above 1750 m, which are usually lower during an undisturbed growing season. This seasonal pattern shows that the frost event actually affected a wide altitudinal range and thus there was a lower altitudinal limit of the frost affected area compared to that detected by the NDVI differencing technique referring to the mid-August NDVI. This difference in the results is due to the NDVI differencing calibration based on visual observations using ground-truth control points only in stands with canopies heavily affected by frost damages.

The spatial distribution of frost affected area, based on terrain aspect, shows a higher distribution of the forest areas facing North-West to North-East (Fig. 6), accounting for approximately 63% of the total frost affected beech forests. Based on this distribution, a generalized additive model was applied to evaluate the spatio-temporal NDVI variability according to northerly and southerly expositions (Table 2 and Fig. 7) which accounts for more than 47% of the explained deviance. The most appreciable relationship detected by the model shows an NDVI increasing trend during the vegetative season up to the maximum values reached in mid-August. Interestingly, the NDVI values are significantly higher on the northerly slope than on the southerly one, and thus suggesting a higher impact of the frost disturbance over the beech forests facing south. The same differences between both expositions are also evident by examining the slope influence on NDVI pattern. A rather distinct picture emerges from GAMs highlighting a clear relationship between NDVI and elevation, where the frost disturbance “footprint” can be observed along the altitudinal range through an evident decrease in NDVI values. In particular, the main NDVI decrease occurs from about 1250 m and peaks at 1500 m asl, representing the altitudinal range affected by frost disturbance, with constant positive differences between north and south facing forest NDVI. However, the relation showed a different pattern between northerly and southerly expositions, with the southerly NDVI presenting a first drop at lower altitudes than the northerly one. Although the North and South facing patterns appear quite similar, the forest NDVI estimates differences (Fig. 7f) below 1000 m asl and between 1700 and 1900 m asl, show NDVI values following their natural pattern with the southerly forest NDVI higher than the northerly one.

Fig. 6
figure 6

Percent distribution of frost affected beech forest compared with site aspect

Full size image
Table 2 Model summary for the generalized additive models fitted to the NDVI
Full size table
Fig. 7
figure 7

Predicted smoothing curve (left panels) of NDVI and DOY, slope, and elevation variables assessed using generalized additive model (GAM). Full and dashed lines represent northerly and southerly exposition classes. The shaded areas indicate the 95% confidence interval. In these plots, a positive slope of the line shows a positive effect of the x variable, and a negative slope of the line indicates a negative effect. Right panels shows the estimated difference between the two conditions, i.e., north and south exposition. Bounding vertical dotted lines identify statistical differences detected by performing simultaneous 95% confidence intervals around the estimates

Full size image

4 Discussion

The study of the effects of the regional scale late spring frost on Mediterranean beech forests showed an effective approach for the spatio-temporal variability detection of frost disturbance and a detailed description of the remote sensing retrieved NDVI patterns (White et al. 2014). The NDVI differencing technique has been mainly used for the detection of forest vegetation changes and disturbances (Zhang et al. 2003, 2004). In this case study, we used the NDVI differencing technique in order to identify the area affected by frost damages and presenting at the end of the growing season the disturbance legacies on forest canopy NDVI.

The calibration of the threshold through the trial and test approach, based on ground-truth control points, confirmed the reliability of the methodology to identify significant differences between the NDVI calculated for mid-August of the 2 years 2016 and 2015 (Pu et al. 2008).

The frost-affected NDVI patterns observed during the 2016 vegetative season suggest a second leaf unfolding starting from the end of June (DOY 180) after almost a 2-month delay, and then completed at the end of July (DOY 212). The second leaf unfolding required about 30 days for the complete NDVI recovery. As shown by Menzel et al. (2015), this period of time is consistent with the observations found in central European beech forests However, it is worth noting that the delay between the frost event and the second leaf unfolding is double that of Menzel et al. (2015). This delay is probably due to the severity of the frost event (Hufkens et al. 2012) reaching at higher elevations temperatures far below − 3 °C which is considered the threshold for permanent leaf damage (Dittmar et al. 2006; Sakai and Larcher 1987).

Results from GAM highlighted significant influence of slope, altitude, and aspect on NDVI values where a clear distinct pattern between northerly and southerly exposure emerged. The low NDVI values in forest stands facing south, distributed from about 1000 to 1600 m asl, might be considered as a consequence of the high phenotypic plasticity characterizing the phenological shift of beech populations, according to both elevation and exposition (Vitasse et al. 2010). Moreover, the north-facing beech forest canopies might be characterized by an incomplete bud flush at the moment of the frost event, preserving more buds from frost damage.

A complex NDVI pattern related to elevation represents an interesting result of our analysis, showing independent patterns between northerly and southerly forest exposure. The double drop in the southerly NDVI curve can be interpreted according to the level of canopy browning. The first one, which occurred around 1000 m asl, may suggest a partial canopy frost damage. The second drop, with north/south NDVI curves overlapping each other, may correspond to the forest altitudinal belt which was completely affected by frost disturbance as also highlighted by the NDVI differencing. The lower altitude beech population experienced non-lethal temperatures that contributed to reduced damages at the canopy level but still gave slightly lower NDVI values than the unaffected beech forest observed at full canopy recovery.

Although beech forests are characterized by a high resilience in growth capacity in response to late spring frost (Príncipe et al. 2017), the combination of subsequent extreme events may dramatically reduce the species resilience. Considering the forest NDVI as a proxy of forest productivity, the effects of spring frost may have a negative impact on forest carbon uptake due to defoliation. The observed reduction of NDVI and the shortening of the growing season may reduce forest productivity by 7–14%, as recently reported by Bascietto et al. (2018) and by Príncipe et al. (2017). Furthermore, a secondary bud flush during the same growing season causes a depletion of plant carbon reserve. Thus, in the context of the mountainous Mediterranean region, the increasing frequency and duration of extreme events, as was the case with the spring frost of 2016 followed by the summer heat wave of 2017 (Di Giuseppe et al. 2017), can further reduce forest productivity and affect the species resilience at the same time.

5 Conclusion

The 2016 late spring frost event provided an opportunity to explore Mediterranean mountain beech forest sensitivity and recovery patterns to frost disturbance through the spatio-temporal analysis of forest NDVI. The implementation of the NDVI differencing technique based on Landsat 8 images allowed the identification of the spatial distribution of the frost disturbance on the forest canopy by the detection of a persistent reduction of forest canopy NDVI at the end of the growing season. About 13% of the total regional beech forests, distributed across a 400 m altitudinal range, had canopies damaged by this late spring frost. The main results also highlighted the significant influence of topographic variability on NDVI values with a distinct pattern between northerly and southerly exposure emerging due to the marked differences in the phenological stage of these trees when the frost occurred. Then, after a 2-month lag in leaf unfolding, the full canopy recovery of the affected areas took about 30 days based on the NDVI pattern analysis. The results of this study represent a contribution to the better understanding of Mediterranean beech forest sensitivity to the increasing threats of extreme events under changing climatic conditions.