Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Spring phenology of oak stands in the Western Carpathians: validation of satellite metrics from MODIS using ground-based observations

Spring phenology of oak stands in the Western Carpathians: validation of satellite metrics from... The study focuses on the validation of the leaf unfolding (LU) onset of oak stands in the Western Carpathians in 2000–2021 derived from MODIS satellite data. LU onset was derived from the annual trajectories of the Normal- ised difference vegetation index (NDVI) fitted with a double sigmoid logistic function. The satellite metric Grow- ing speed day (GSD) corresponding to the LU onset is represented by the first derivative of the sigmoid function. Ground-based observations from 22 phenological stations of the Slovak Hydrometeorological Institute (SHMI) were used to validate the date of GSD. The results showed a good agreement between the medians of ground- and satellite-based LU onset dates. In addition to the median, the LU onset at the 5th and 95th percentiles were com- pared. For both percentiles, we found differences in the onset from MODIS and SHMI. The 5th percentile of the LU onset derived from MODIS was determined later than the one from SHMI data. With the 95th percentile, it was the opposite. As a result, the range determining the duration of LU onset from MODIS was significantly shorter than from SHMI observations. The trend analyses over the period 2000–2021 revealed a shift to the earlier onset of LU ~ −1 −1 0.33 day.year (p = 0.13) from satellite and ~ 0.32 day.year from ground-based observations (p = 0.08). The vali- dated LU onset and trends using the median allow analysing of the oak stands’ response to changing environmental conditions. However, the differences at the 5th and 95th percentiles, i.e. at the beginning and the end of the LU onset duration, remained unexplained. Key words: MOD09; MYD09; NDVI; leaf unfolding; validation Editor: Milan Barna of the course of phenological events. The AVHHR’s reso- 1. Introduction lution of 8 km was the main limiting factor in their use. Forest phenology is a traditional scientific discipline In Slovakia, satellite phenology has gained an impor- that examines the time course of important, periodi- tance following the launch of the Terra and Aqua satellites cally recurring life manifestations of forest trees, the so- (2000, 2002, NASA Earth Observation System Satel- called phenological phases. The onset of phenophases lites) with the Moderate Resolution Imaging Spectrora- depends mainly on the weather (Štefančík 1995) and diometer (MODIS). For the red and infrared bands, the is influenced by the changing climate. Therefore, the pixel size of 250 × 250 m is already sufc fi iently detailed to phenophases’onsets are considered indicators of cli- identify and select a sufficient number of homogeneous mate change impacts on forests (Richardson et al. 2013; pixels with one type of landscape cover for analyses, even Gray & Ewers 2021). A new dimension to study forest in the fragmented area of the Slovak forests (Bucha et al. ecosystem responses to changing environmental condi- 2011). Currently, a consistent database was created from tions at global and macro-regional levels was provided 2000 to 2021. It is based on MODIS products MOD09GQ by the satellite-based phenology using the Advanced and MOD09GA derived from the Terra satellite and Very High-Resolution Radiometer (AVHRR) (Stöckli & MYD09GQ and MYD09GA from the Aqua satellite. Vidale 2004; Piao et al. 2006; Heumann et al. 2007). The The products contain a quality layer that allows elimi- main benefit compared to sampling methods based on nating pixels with reflectivity affected by clouds, cloud ground-based observations lies in the possibility of quasi- shadows or high aerosol content. Several papers demon- continuous (in day intervals) and large-scale monitoring strated the suitability of using the Normalized Differen- *Corresponding author. Tomáš Bucha, e-mail: tomas.bucha@nlcsk.org, phone: +421 45 5314 156 © 2022 Authors. This is an open access article under the CC BY 4.0 license. T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 tial Vegetation Index (NDVI) derived from satellite data percentiles is important for understanding the response to monitor phenophases in forests (Zhang et al. 2003; of oak stands in the boundary conditions of their occur- Beck et al. 2006; Soudami et al. 2008). rence. Modelling methods are used to derive satellite-based The second goal of the work is to derive and compare phenological metrics. Under modelling, we understand temporal trends of the LU onset from ground surveys and a mathematical expression of the course of NDVI devel- satellite data. We assume that the trends derived from opment during the year and determine the onset of major both data sets will be identical. Due to the continuity phenological events. Berra & Gaulton (2021) report that of data acquisition with MODIS, after the validation it the logistic function (Zhang et al. 2003) and its modic fi a - will be possible to analyse LU changes more accurately tions (Fisher et al. 2006; Fisher & Mustard 2007) is the and use the new knowledge in proposing climate change most commonly used function in deriving vegetation phe- adaptation measures in forests. nology at local and regional levels. In Slovakia, the double logistic function was used for modelling (Bucha & Koreň 2017; Lukasová et al. 2014, 2019; Barka et al. 2019; Tholt 2. Material and methods 2021) through the PhenoProfile software application 2.1. Study area (© Milan Koreň). The main advantage of this application is the full user control of modelling parameters and the We carried out the research in the forest stands of Slo- possibility to refine the modelling procedures. Its main vakia with a dominant presence of pedunculate oak functionalities were described by Bucha & Koreň (2014). (Quercus robur L.) and/or sessile oak (Quercus petraea The rapid expansion in remote sensing and ground- Liebl.) (Fig. 1). These are economically important trees based phenology data acquisition has supported signifi- with a total share of 10.4% of forest area (Moravčík et al. cant advances in plant phenological research, as shown 2021). The species are not differentiated and are analysed in the recent review (Piao et al. 2019). However, some together as one genus in the study. research results point at methodological risks in deriving Geographically, most of the territory belongs to the and interpreting phenological models based on satellite province of the Western Carpathians, while a part of the data from various sensors (Norris et al. 2020; Younes et eastern Slovakia roughly from 21°15’ longitude belongs al. 2021). Therefore, the validation of derived regional to the province of the Eastern Carpathians (Kočický & products still appears to be very topical and necessary. Ivanič 2011). The occurrence of assessed oak stands is Data from terrestrial measurements are used for concentrated in the 1st to 3rd forest vegetation zones validation and parameterization of remote sensing from the 8-level scale (Zlatník 1976). It is an area with outputs (Hmimina et al. 2013). The onset of individual an average annual temperature above 5.5 °C, annual pre- phenophases derived from MODIS for beech stands cipitation totals up to 800 mm and a growing season of in Slovakia was validated by Pavlendová et al. (2014) 150 to 180 days. The dominant soil on which oak stands and Lukasová (2019) by comparison them with obser- occur are cambisols, followed by leptosols, luvisols and vations from the network of phenological stations of the planosols (Miklós 2002). Slovak Hydrometeorological Institute (SHMI). In addi- tion to SHMI data, they also used their own observations. A validation for oak stands has not been performed, yet. 2.2. Oak stands selection Therefore, our first goal is to validate a regional When selecting oak stands, we used the tree species map phenological model derived from NDVI from MODIS of Slovakia from Landsat (Bucha 1999) with a resolution satellite images for oak stands. The observations from resampled to 250 m in a combination with the current the SHMI regional phenological stations for the period database of the Forestry Information System (FIS). The 2000–2021 were used to validate the day of the onset database contains data at the compartment level (NFC of the leaf unfolding (LU) spring phenophase. SHMI Zvolen – https://gis.nlcsk.org/lgis/). Data from FIS was observations were used to verify the hypothesis that the used to calculate the proportion of oak in 250 m pixels. local extreme of the first derivative of Fisher’s sigmoidal We excluded marginal pixels from the tree species map. function (Fisher & Mustard 2007) is related to the leaf The reason is that these pixels in the MOD09 products unfolding onset. Based on the knowledge of the Fisher’s may be spectrally contaminated with other land cover function course and its validation in beech stands (Pav- categories (Townshend et al. 2000) and thus distort the lendová et al. 2014), we assume that the LU phenophase derived NDVI values. We also excluded pixels where the of oak will correspond to the GSD (Growth Speed Day) satellite metric. The LU onset phenophase was chosen oak proportion was less than 70%. Thus, created masks of because it is used as the start of the growing season. homogeneous oak stands contained 981 pixels. Of these, In contrast to published studies, we validate the onset of we selected 413 pixels by stratified random sampling, on which the LU onset was analysed (Fig. 1). The number LU derived from MODIS over the entire duration of the of points ensured the representatives of the sample. At phenophase using 5-percentile, median (50-percentile) and 95-percentile. The validation of the 5th and 95th the same time, the iterative calculation of 6 parameters 192 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Fig. 1. Forests with a predominant occurrence of oak (brown colour), the spatial distribution of SHMI phenological stations (red circles) and analysed pixels from MODIS with the occurrence of oak ≥ 70% (black crosses). The forests are shown in grey. The numbering of the phenological stations is in accordance with Table 2. of Fisher’s function modelling the NDVI profile during the relationship: the year was manageable in about 6 hours on a PC with ܫܴܰ Ȃ  ܴ݁݀ ଼ସଵ Ȃ଼ ଻଺ ௡ ௠ ଺ଶ ଴Ȃ ଺଻ ଴ ௡௠ a six-core processor. The altitudinal distribution of the ܰܦܸܫ ൌ ൬ ൰ [1] ܫܴܰ ൅ ܴ݁݀ ଼ସ ଵȂ ଼଻଺ ௡ ௠ ଶ଺ ଴Ȃ ଻଺ ଴ analysed oak pixels is shown in Fig. 2a. Fig. 2b depicts the altitudinal distribution of SHMI stations. The layer of quality was taken from the MOD/ MYD09GA products. Using the layer, we excluded pixels affected by cloud cover, cloud shadows and high aerosol 2.3. Modelling of the annual NDVI course content from the analyses. By overlaying the NDVI layer The solution was based on MODIS images, specifically adjusted in this way with a layer of selected oak stands, on products MOD/MYD09GQ and MOD/MYD09GA a database suitable for modelling oak phenology was collections 5 and 6. The products represent the spec- created. The procedure for choosing and pre-processing tral reflectance on the earth’s surface, i.e. the effect of MODIS images and deriving NDVI is described in detail absorption and scattering of radiation in the atmosphere in Bucha & Koreň (2014 and 2017). was eliminated. From the MOD09GQ and MYD09GQ The subject of the analysis was the NDVI time series products, we derived the NDVI using the red and infrared from March to July and the years 2000 to 2021. The bands with a pixel resolution of 250 × 250 according to annual development of NDVI was modelled by a double Fig. 2. a) Frequency histogram of oak stands (with oak share of 70% and more) along the altitudinal range. b) Frequency his- togram of altitude distribution of SHMI stations. X–axis: altitude (m a.s.l.). Y–axis: a) number of pixels (MODIS); b) number of SHMI stations. 193 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 logistic function (Fisher & Mustard 2007): most leaves are likely to emerge and that deciduous for- ests with different crown canopies or even small amounts Y W Y  Y [2] of conifers can be compared at the date of onset. The GSD PL Q DPS P  Q W P  Q W H   H date was calculated for each of the 413 analysed pixels for each year from 2000–2021. where v(t) is NDVI observed at time t; v and v are param- min amp eters corresponding to the minimum value of the NDVI and its amplitude; m and n parameters control the shape and slope of 1 1 the curve of the ascending (spring) phase, and m and n param- 2 2 2.4. Data validation and statistical evaluation eters control the descending (autumn) phase. We used ground-based phenological observations to validate and reveal possible errors in modelling LU onset The advantage of using this type of the sigmoid func- from MODIS. From the SHMI network, we used 22 forest tion is that it is differentiable. The extremes of the first phenological stations focused on oak evaluation (Fig. 1; and second derivatives of equation 2 were used to derivate Table 2), where more than 95% of the data on the LU satellite metrics and corresponding phenological events. onset were available in individual years from 2000–2021. The connection of spring phenophases from ground Observations were made individually using a telescope. observations and satellite metrics is shown in Table 1 10 pre-dominant or dominant trees were evaluated at based on Pavlendová et al. (2014). each plot. In the next text, where it is necessary to dis- tinguish the LU onset between ground and satellite Table 1. Satellite-based phenological metrics and correspond- observations, we use the abbreviations LU_SHMI and ing vegetative phenophases by Pavlendová et al. (2014). LU_MODIS. Satellite The description of satellite phenological metrics derived from [2] – metrics Corresponding vegetation phenophases GAD Maximum of the second derivative in a spring phase ~ Bud break onset Table 2. Basic characteristics of selected SHMI forest pheno- Extreme of the first derivative in a spring phase (spring inflection point) logical stations. GSD ~ Leaf unfolding (LU) onset Station number Lati- Longi- Altitude Minimum of the second derivative in a spring phase ~ End of leaf un- Station name GDD – see Fig. 1. tude tude [m a.s.l.] folding GAD – Growing Acceleration Day; GSD – Growing Speed Day; GDD – Growing Deceleration 1 Riadok 48° 30’ 17° 10’ 183 Day; DOY – Day of Year. 2 Hajnáčka 48° 13’ 19° 57’ 250 3 Lučenec 48° 20’ 19° 41’ 250 4 Levice 48° 13’ 18° 37’ 280 5 Stakčín - CHOTINA 49° 00’ 22° 13’ 300 6 Železná studienka 48° 10’ 17° 07’ 320 7 Stupava 48° 16’ 17° 02’ 320 8 Hvezdáreň 48° 22’ 17° 17’ 350 9 Čebovská Bukovina 48° 14’ 19° 18’ 360 10 Kokava nad Rimavicou 48° 34’ 19° 51’ 370 11 Kšinná 48° 48’ 18° 21’ 374 12 Prešov-Cemjata 48° 58’ 21° 10’ 375 13 Zborov 49° 22’ 21° 18’ 400 14 Myjava 48° 46’ 17° 35’ 450 15 Jasov Lesy-Premonstr. 48° 40’ 20° 58’ 450 16 Bytča-Starovec 49° 14’ 18° 33’ 460 17 Nové Mesto nad Váhom 48° 46’ 17° 50’ 500 18 Ždaňa 48° 34’ 18° 45’ 500 19 Kecerovce 48° 49’ 21° 24’ 550 20 Zvolen 48° 34’ 19° 10’ 718 Fig. 3. The course of the NDVI curve in the spring phase. 21 LS Veľké Kapušany 48° 32’ 22° 05’ 108 The black dots represent the NDVI in a given pixel on a given 22 LS Svinica 48° 44’ 21° 28’ 350 DOY. The SHMI stations are evenly distributed across Slo- In this work, we analyse only the phenophase of leaf vakia so that the highest possible number of tree species unfolding. LU onset was evaluated according to the can be observed simultaneously in each station (Fig. scale of the manual for phenological observations for the 1). Therefore, it was impossible to match stations with pan-European monitoring system (Preuhsler 1999) and analysed oak pixels from MODIS. For this reason, we according to the scale developed by SHMI (Braslavská & used the two-sample Kolmogorov-Smirnov test to test Kamenský 1996). The LU onset is considered the DOY the agreement of the altitude distribution of two sample when the first light green leaves, smaller than in adult - sets, i.e. 413 pixels from MODIS and 22 SHMI stations. hood, appeared on at least half of the individuals of the The hypothesis H0: The two samples follow the same given observed group. altitudinal distribution was tested against H1: The alti- The GSD satellite metric corresponds to the LU onset tudinal distributions of the two samples are different. The phenophase. GSD is defined as the date when the sig - reason for testing the equality of altitudinal distributions moid reaches its half-maximum value or peak of the first is the strong dependence of LU onset to the altitude. Any derivative of the equation [2]. Fisher & Mustard (2007) differences in sample sets would bias the comparison of state that the half-maximum represents the date when differences between LU_SHMI and LU_MODIS, as well 194 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 as validation of LU onset modelling from MODIS data 3. results in individual years 2000–2021. 3.1. Comparison of altitudinal distribution of For the nationwide evaluation of the LU onset (LU_SHMI and LU_MODIS) of oak stands, we calcu- SHMI and MODIS data sets lated the median value in individual years of the period The results of the Kolmogorov-Smirnov test, with which 2000–2021. We expressed the intra-annual variability we tested the equality of the altitudinal distribution of the with the 5th and 95th percentiles. These represent the LU two sample sets, are presented in Table 3. The overlap onset day values below which 5 (95) % of the distribution of the cumulative distribution functions of SHMI and lies. This means that the 5th percentile represents the day MODIS samples documents the equality of their alti- of LU onset at lower altitudes or in southern areas. The tudinal range in Fig. 4. Therefore, we consider that the 95th percentile represents the day of LU onset at higher samples follow the same altitudinal distribution and are altitudes or in northern regions of oak occurrence. We suitable for: note that the median represents the 50th percentile. The uncertainty in the LU onset assessment from (i) comparing the onset of LU from ground-based and MODIS was analysed with statistical metrics such as MODIS observations, a Coefficient of determination (R ), Mean bias error (ii) identification of modelling inaccuracies in individ- (MBE) and Mean absolute error (MAE). The statistical ual years 2000–2021, metrics were used to compare the model’s performance (iii) validation of LU onset from MODIS. in deriving the GSD using equation [2] with SHMI obser- vations. Table 3. Two-sample Kolmogorov-Smirnov test for compar- ing the equality of the altitudinal distributions of ground- Coefficient of determination (R ): based and MODIS phenological plots. ۍ ې ത ത σ ሺ ሻ ൫ܻ െܻ ൯ ܻ െܻ ௙ ௙ ௠ ௠ ଶ N u m b e r Significance ێ ۑ ܴ ൌ [3] Source p – value Interpretation ێ ଶ ۑ of plots level – alpha ට ത ത σ ൫ܻ െܻ ൯ σሺ ܻ െܻ ሻ ௙ ௙ ௠ ௠ ۏ ے As the computed p–value > alpha, one SHMI 22 cannot reject the null hypothesis H0: 0.883 0.05 Mean bias error (MBE): The two samples follow the same dis- MODIS 413 σ ሺܻ െܻ ሻ ௠ ௙ tribution. ܯܤ ܧ ൌ [4] Mean absolute error (MAE): σ หܻ െܻ ห ௠ ௙ ܯܣ ܧ ൌ   [5] where Y = ground-based value (LU onset from SHMI), Y = f m modelled value (LU onset from MODIS), Ῡ = average of Y f f values, Ῡ = average of Y values, N = number of years in the m m analysed period 2000–2021. We used a simple linear correlation and regression analysis to derive temporal trends in LU onset from SHMI and MODIS data. The input was the time series of the median values of the LU onset calculated for indi- vidual years of the period 2000–2021. We assessed the signic fi ance of the trends by testing the signic fi ance of the sample correlation coefficients. Under hypothesis H0, Fig. 4. Comparison of the equality of the cumulative altitu- the correlation coefficient in the population is equal to 0. dinal distribution of ground-based phenological plots of the It is proven that testing the signic fi ance of the correlation SHMI stations and MODIS pixels. coefficient is identical to testing the significance of the regression coefc fi ient (Šmelko 1991). The test character - istic t was used. For t > t , the hypothesis H0 is rejected at α,f 3.2. Validation of LU onset from MODIS α % level of significance, and with P = 1 – α% confidence using ground-based observations it is considered that both the correlation and the regres- sion sample coefficients are different from 0. Input data on the LU onset from ground-based SHMI We verie fi d whether the trends derived from MODIS observations (LU_SHMI) and MODIS (LU_MODIS) and SHMI were significantly different by testing the are presented in Table 4. The course of both onsets is hypothesis about the equality of the two regression coef- graphically compared in Fig. 5a–c, where the medians, ficients. 5th and 95th percentiles express the day of LU onset for individual years of the period 2000–2021. 195 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Table 4. Median, 5th and 95th percentiles of leaf unfolding from MODIS (413 pixels) and from phenological observations in the SHMI network (22 stations). Year LU_MODIS median 5th percentile 95th percentile LU_SHMI median 5th percentile 95th percentile 2000 111 109 113 114 105 119 2001 119 116 122 117 101 125 2002 119 111 121 114 104 125 2003 117 115 121 118 111 125 2004 116 112 122 119 110 128 2005 115 110 119 122 110 134 2006 115 112 121 120 110 129 2007 105 101 111 109 99.2 120 2008 115 110 122 115 103 126 2009 104 102 111 108 103 119 2010 115 111 119 118 105 125 2011 108 105 113 111 105 120 2012 115 110 119 116 102 125 2013 117 115 121 117 110 123 2014 96 91 102 103 89 119 2015 111 108 117 113 90.3 123 2016 105 102 109 108 90.5 121 2017 106 97 119 107 93.1 123 2018 107 105 108 108 96.3 117 2019 106 102 114 110 93.1 119 2020 109 106 116 110 97.1 121 2021 124 119 129 123 105 135 Average 111.59 107.68 116.77 113.64 101.48 123.68 Table 5 shows the differences (∆) in days between the with the original parameters. Therefore we consider the LU onset from ground observations and MODIS using LU_MODIS onset DOY as confirmed. MBE and MAE statistics derived according to relations The regression dependence between the onset of [4] and [5]. We found the smallest differences in the LU LU_SHMI and LU_MODIS is expressed in Figures 5d–f. onsets between the medians and the largest for the 95th The highest correlation coefc fi ient and the concentration percentile. The difference between the 5th percentiles of the points around the line y = x was observed for the was also substantial. The median MBE of −2.05 days median LU values (Fig. 5d). indicates an earlier determination of LU onset from The calculated differences between the 95th and 5th MODIS compared to ground observations. Similarly, percentiles, which determine the duration of the LU MBE of −6.91 days for the 95th percentile represents an onset phase, are shown in Fig. 6a for the entire monitored earlier determination of LU onset from MODIS. Con- period 2000–2021. The graph shows that the duration versely, the MBE of 6.20 days for the 5th percentile rep- of LU onset from MODIS was systematically under- resents the later onset from MODIS versus SHMI. estimated compared to ground observations of SHMI. The significant difference results from the two already described opposite tendencies of LU determination at the Table 5. Mean bias error (MBE) and Mean absolute error (MAE) in days between SHMI and MODIS leaf unfolding on- 5th and 95th percentiles. set for median, 5th percentile and 95th percentile. ∆ LU ∆ LU ∆ LU median 5th_percentile 95th_percentile MBE −2.05 6.20 −6.91 3.3. Temporal trend of leaf unfolding onset MAE 2.77 6.29 6.91 Median values of the day of LU onset, intra-annual vari- A close relationship between the median LU onset ability expressed by the 5th and 95th percentiles for the values from MODIS and SHMU is evident from the individual years of the period 2000–2021 and the calcu- course of DOY during the entire observed period (Fig. lated linear trends of the LU onset are shown in Fig. 7a 5a). It confirms the suitability of the GSD metric for and 7b for SHMI and MODIS selection of oak stands, determining LU onset from MODIS. At the same time, respectively. it was possible to identify the years with greater differ- Fig. 7a and 7b show considerable intra- and inter- ences between satellite and ground-based DOYs that annual variability of DOY of leaf unfolding onset. In may indicate possible errors in determined values that the Carpathian conditions, the intra-annual variability need to be checked. These are mainly the years 2002, is mainly conditioned by the altitude and the related air 2005, 2006 and 2014. For these years, we re-derived temperature. The inter-annual variability is determined the LU_MODIS onset day (GSD metric) by changing by the spring weather. Both SHMI and MODIS obser- the modelling parameters. We expanded the parameter vations show a tendency towards an earlier LU onset ranges of function [2] and doubled the subspaces within over the period 2000 to 2021. The regression coefficient which the iteration process took place. The calculation determines the magnitude of the change per year. For extended to more than three days, but in no case did the ground-based SHMI observations, this change was −1 difference exceeds 0.2 days compared to the estimate −0.32 days.year (p = 0.08) and for MODIS it was 196 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Fig. 5. Left: Comparison of the leaf unfolding onset from SHMI and MODIS observations for the period 2000–2021: a) median, b) 5th percentile, c) 95th percentile. X–axis: year; Y–axis: Day of the year (DOY). Right: Dependence between the leaf unfolding onset from SHMI and MODIS observations for d) median, e) 5th percentile, f) 95th percentile. X– and Y–axes: Day of the year (DOY). The dashed line corresponds to the line y = x. Fig. 6. Course (a) and dependence (b) between the duration of the LU onset on 22 SHMI phenological stations and 413 MODIS pixels. Duration is expressed in days and is calculated as the difference between the 95th and 5th percentiles for each year of the period 2000–2021. 197 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Fig. 7. Leaf unfolding onset day (median), duration (5th and 95th percentile) and onset trends of oak stands for the period 2000–2021 according to a) SHMI observations and b) MODIS. X–axis: year; Y–axis: Day of the year (DOY). −1 −0.33 days.year (p = 0.13). The test of the hypoth- 111.6 DOY in MODIS and 113.6 DOY in SHMI obser- esis about the equality of regression coefficients proved vations. that the slope of both regression lines was identical In both data sets, we recorded the earliest onset of (p < 0.001). LU in 2014, when the median DOY for Slovakia was 96 (MODIS), or 103 (SHMI). The year 2014 was historically the first year in Slovakia when the cumulative value of the average daily air temperature exceeded all previous 4. Discussion observations since 1951 (Bochníček 2014 – available 4.1. Notes on phenological modelling online). and validation of results In 2021, we observed the atypical late onset of LU DOY equal to 124 DOY (MODIS) and 123 DOY (SHMI). The day of the leaf unfolding onset was derived from According to the SHMI meteorological report, the tem- NDVI annual trajectories fitted with the double logis - perature in March 2021 was normal, but the tempera- tic function according to formula [2]. Only the spring ture in April was below average in the entire territory of period from about 70 to 200 DOY was analysed. The Slovakia and the weather was unstable and cold. This parameters m , n of equation [2] are decisive for autumn 2 2 type of weather had a contrary character compared to the phenophases and do not affect the calculation of spring above average air temperatures in April in the previous parameters in eq. [2]. Therefore, the NDVI images from decade. It was related to atypical circulation conditions DOY above 200 were not included in the modelling. in the northern hemisphere resulting from the global cli- The LU onset determined from 413 MODIS pixels mate change (SHMÚ, 2021). and 22 SHMI phenological stations and expressed by the median had a comparable course during the monitored period 2000–2021 (Fig. 5a). We checked the four more substantial differences in 2002, 2005 and 2006 and 2014 4.3. Intra-annual variability of leaf unfolding by recalculating the DOY of LU onset from MODIS. The onset refined DOY derivation did not differ substantially from The intra-annual variability of the LU onset is mainly the original one. We did not n fi d an exact explanation for conditioned by climatic conditions resulting from the the identified differences, as significant differences did altitude and the current development of the weather not occur in other years. It is also not possible to rule out in a given year. The air temperature plays a decisive the error of the ground-based phenological observations role (Škvareninová 2008). Gafenco et al. (2022) found due to the subjectivity of the observers. a longitudinal shift in the budburst of Quercus petraea in Romania. Besides, the variation in leaf flush timing involves some ecological trade-offs in temperate regions. 4.2. Inter-annual variability of the leaf For example, the phenological form of pedunculate oak unfolding onset (Quercus robur var. praecox) in Eastern Poland flushes early in the season to avoid summer droughts but con- The days of LU onset expressed by the median during sequently suffers from later spring frost and insect her- 2000–2021 naturally varied depending on the spring bivory, whereas the other oak form (Quercus robur var. weather. Annual median values ranged from 96 to 124 tardiflora) flushes up to five weeks later in the season, DOY for MODIS and from 103 to 123 DOY for SHMI (Fig. 5a). The mean value for the period 2000–2021 was avoiding insect herbivory but suffering from summer 198 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 droughts (Wesołowski & Rowiński 2008; Puchałka et longitude has not yet been quantified in the oak forests. al. 2017). Considering the mask with the oak share of at least We expressed the range of oak stands LU onset in 70%, it is not possible to exclude the influence of other individual years using the 5th and 95th percentiles (Fig. tree species occurring in a given pixel on the LU onset. 5b and 5c). For both percentiles, we found differences in The tree species with the highest representation in Slo- determining the LU onset between MODIS and SHMI. vakia could have the most significant impact: beech at Moreover, the nature of the difference was contradic- the 95th percentile, hornbeam and Turkey oak, especially tory. The 5th percentile of the onset determined from at the 5th percentile due to their altitudinal occurrence. MODIS occurred later than from ground observations. Schieber et al. (2009) showed an earlier beech and horn- For the 95th percentile, we observed the opposite. When beam LU onset of up to 6 days (average for 1995–2007) determining the duration of the leaf unfolding onset in compared to Sessile oak in central Slovakia. On the con- Slovakia from MODIS, the phenophase was the shortest trary, a later LU onset is reported in the case of Turkey in 2000 and 2018, lasting only 5 or 4 days. The longest oak. When analysing LU onset with an 85% oak mask phenophase duration was found in 2017, equal to 23 (136 pixels), we found no shift for the 5th percentile com- days. In other years, it varied from 7 to 13 days. Accord- pared to the 70% mask. For the 95th percentile, there was ing to ground-based phenological observations of SHMI, a 1-day shift to a later LU onset and thus a prolongation the shortest phenophase duration was in 2013, 2000 of the phenophase. and 2003, namely 14 and 15 days. The most extended Both effects (a shift due to the earlier LU onset of duration of the phenophase was found in 2015, namely beech in analysed pixels and possible latitudinal shift) 33 days. In other years, it varied from 16 to 31 days. could extend the average 10.1 days-long duration of the Due to the opposite nature of determining the 5th and LU onset, found in our work, by 2–2.5 days. Despite this, 95th percentiles of LU onset from MODIS compared to most of the differences remain unexplained. SHMI, the resulting range of the phenophase duration was even more different than for the individual percen- tiles. This was manifested by a substantially shorter range 4.4. Statistics to determine the leaf unfolding of the LU onset in MODIS than in SHMI observations. onset The average duration of the phenophase from MODIS for From the analysis above of the intra- and inter-annual var- the entire monitored period was 10.1 days, while accord- iability of the LU onset and from the regression depend- ing to SHMI it was up to 23.2 days. ence between LU_MODIS and LU_SHMI expressed in A partial explanation for the significant difference Fig. 5d–f, it follows that for determining the onset, it is is the uneven geographical distribution of SHMI sta- most appropriate to use the median value. tions and pixels from MODIS. From Fig. 1 we can see The coefficient of determination between satellite that some SHMI stations are located further south (sta- and ground observations was R = 0.82 (Fig. 5d), and tions no. 2, 4, 6–9, 21) but also further north (stations the Mean bias error (MBE) was −2.05 days, indicating no. 5, 13, 16) than most MODIS pixels. Considering the the earlier detection of the LU onset from MODIS (Table south-north and partly the west-east dependence of the 5). The 5th and 95th percentiles are inappropriate met- onset of spring phenophases (with increasing latitude rics if the presented methodical approach is used. In the and continentality, the phenophase is delayed) and the case of the 5th percentile (R = 0.53, MBE = 6.20 days), smaller number of SHMI stations, it cannot be ruled out we revealed a substantial shift to a later onset derived that the mentioned factors affected the comparability of from MODIS compared to SHMI. On the contrary, for SHMI and MODIS data despite the agreement of their the 95th percentile (R = 0.63, MBE = −6.91 days), there altitudinal distribution. For example, Stakčín and Zborov was a substantially earlier determination of the LU onset stations, which are the most northern and at the same from MODIS compared to SHMI. time the most eastern, but also Bytča-Starovec, which The magnitude of the deviation in the median values is located in the northwest of the country, regularly had of LU onset between satellite and ground assessments the late LU onset. This affectes the 95th percentile of LU (MBE −2.05 days and MAE 2.77 days) confirms our onset. The southerly located SHMI stations, Hvezdáreň assumption that the GSD satellite metric represents the in the west, Hajnáčka in the middle and Veľké Kapušany phenophase LU onset and is a suitable metric for satellite- in eastern Slovakia, usually had the earliest onsets of the based observations of the mentioned phenophase. The phenophase. This was manifested in the 5th percentile determination of the LU onset at marginal conditions towards the earlier LU onset compared to MODIS. Barka of oak stand occurrence, expressed by the 5th and 95th et al. (2019) quantified a delay in LU onset towards the percentiles, and the duration of the LU onset were not North by 1.05 days per 100 km from MODIS data for sufficiently validated. The differences between ground beech forests in the Pannonian-Carpathian macro- and satellite observations should be the subject of further region. The longitudinal delay was observed as well but research. was not significant. The LU onset shift due to latitude/ 199 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 to an earlier onset of LU in 22 years. The differences 4.5. Temporal trend of the leaf unfolding onset between ground and satellite observations at the 5th The length of the monitored period (2000–2021) is suf- and 95th percentiles, i.e. at the beginning and the end of ficient for a statistically relevant derivation of temporal the duration of the LU onset in Slovakia, were not suf- trends in the LU onset. The statistical evaluation results ficiently explained. at the national level showed a tendency towards an ear- Based on the achieved results, it will be possible to lier LU onset of oak stands. The rate of change derived expand the scope of modelling the spring phenophase −1 from ground observations was −0.32 days.year , and leaf unfolding onset using continuous satellite data −1 −0.33 days.year in MODIS observations (see regres- from MODIS. However, the derivation of more accu- sion coefficients in Fig. 7a and 7b). Over the entire rate results on the reaction of trees to the manifestations monitored period of 22 years, this represented a shift of of climate change along the altitudinal gradients requires ~7 days to an earlier start of leaf unfolding. This result a clarification of the differences in the LU onset between is in line with Škvareninová’s (2014) finding based on ground-based and satellite data at the peripheral occur- ground observations in Slovakia from 1988–2013, which rence of oak stands. Only subsequently it will be possi- showed ane earlier LU onset of Quercus robur by 7 days ble to apply the acquired knowledge, for example, in the −1 (−0.27 days.year ). The values of the regression coeffi- forestry decision-making about changes in the transfer cients (LU_SHMI and LU_MODIS) are almost identical, of forest reproductive material between forest seed the difference between them was statistically insignifi - regions or about changes in the regeneration and target cant. This conr fi ms our assumption that MODIS satellite tree species composition in management models appli- data and the GSD metric derived from them are suitable cable in the adaptive forestry. for temporal analyses and derivation of trends in the LU onset of oak stands. This finding is important because the shift to an earlier onset of the phenophase is a consid- Acknowledgement ered response of woody plants to climate change (Berra This publication is the result of the project implementation Cen- & Gaulton 2021). Thus, the use of satellite data makes it tre of Excellence of Forest-based Industry, ITMS: 313011S735, possible to reliably analyse the reaction of oak stands even supported by the Operational Programme Integrated Infrastruc- at regional or local levels due to the significantly larger ture funded by the ERDF. Study was partially supported by the MODIS dataset compared to ground surveys. Slovak Research and Development Agency under the contract No. APVV–20–0365. 5. Conclusion In the presented study, we validated the accuracy of the references satellite-based modelling of the leaf unfolding onset Barka, I., Bucha, T., Molnár, T., Móricz, T., Somogyi, Z., spring phenophase. LU onset of a large homogeneous Koreň, M., 2019: Suitability of MODIS-based NDVI set of oak stands (413 pixels) from the MODIS was com- index for forest monitoring and its seasonal applica- pared with data from 22 ground-based phenological sta- tions in Central Europe. Central European Forestry tions in the Slovak part of the Carpathians territory. In Journal, 65:206–217. the case of satellite data, the LU onset corresponds to the Beck, P. S. A, Atzberger, T. C., Hogda, K. A., Johansen, extreme of the first derivative of the double logistic func- B., Skidmore, A. K., 2006: Improved monitoring of tion [2] in its spring phase, represented by the GSD met- vegetation dynamics at very high latitudes: a new ric. From the verie fi d statistics, median, 5-percentile and method using MODIS NDVI. Remote Sensing of 95-percentile, we found the closest relationship between Environment, 100:321–334. ground-based and satellite determination of LU onset at Berra, E., Gaulton, R., 2021: Remote sensing of temper- the median calculated from onset days in individual pixels ate and boreal forest phenology: A review of progress, and SHMI stations. The coefc fi ient of determination was challenges and opportunities in the intercomparison R = 0.82 with 2.05 days earlier detection of the LU onset of in-situ and satellite phenological metrics. Forest by MODIS. We can therefore consider the LU onset from Ecology and Management, 480:118663. MODIS expressed by the median as validated and reliable Braslavská, O., Kamenský, L., 1996: Fenologické pozo- for further analyses. rovania lesných drevín. Metodický predpis Sloven- The temporal analysis of the trend over the period ského hydrometeorologického ústavu. SHMÚ, 22 s. 2000–2021 also confirmed the consistency of determin- Bucha, T., 1999: Classification of tree species compo - ing LU onset using the median values. The statistical test sition in Slovakia from satellite images as a part of confirmed that the regression coefficients determining monitoring forest ecosystems biodiversity. Zvolen, the rate of the temporal change derived from SHMI and LVÚ Zvolen, Acta Instituti Forestalis Zvolen, Tomus MODIS data were identical. We found a shift of ~7 days 9:65–84. 200 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Bucha, T., Priwitzer, T., Koreň, M., 2011: Modelovanie Norris, J. R., Walker, J. J., 2020: Solar and sensor geom- fenologického vývoja lesných porastov pomocou etry, not vegetation response, drive satellite NDVI vegetačného indexu NDVI odvodeného zo satelitných phenology in widespread ecosystems of the western snímok MODIS. Lesnícky časopis – Forestry Journal, United States. Remote Sensing of Environment, 249: 57:187–196. 12013. Pavlendová, H., Snopková, Z., Priwitzer, T., Bucha, T., Bucha, T., Koreň, M., 2014: Fenológia dubových a buko- 2014: Using of long-term phenological observations vých porastov v období 2000–2014. In: Bucha, T. et of SHMI and NFC for validation of regional phenol- al. (ed.): Satelity v službách lesa. Bratislava, SAP, ogy model for European beech. In: Roznovsky, J., p. 101–115. Litschmann, T. (eds.): Proceedings of the Inter- Bucha, T., Koreň, M., 2017: Phenology of the beech national Conference Mendel and Bioclimatology. forests in western Carpathians from MODIS for Masaryk University (Brno, Czech Republic) 3–5 Sept 2000–2015. iForest - Biogeosciences and Forestry, 2014, p. 294–311. Available at: http://www.better- 10:537–546. cities.eu/bulletin/wp-content/uploads/2014/09/ Fisher, J. I., Mustard J. F., 2007: Cross-scalar satellite mendel_ConferenceProceedingsMendel.pdf phenology from ground, Landsat and MODIS data. Piao, S., Fang, J., Zhou, L., Ciais, P., Zhu, B., 2006: Remote Sensing of Environment, 109:261–273. Variations in satelite-derived phenology in China’s Gafenco (Pleşca), I. M., Pleşca, B. I., Apostol, E. N., temperate vegetations. Global Change Biology, Şofletea, N., 2022: Spring and Autumn Phenology 12:672–685. in Sessile Oak (Quercus petraea) Near the Eastern Piao, S. L., Liu, Q., Chen, A. P., Janssens, I. A., Fu, Y. Limit of Its Distribution Range. Forests, 13:1125. S., Dai, J. H. et al., 2019: Plant phenology and global Gray, R. E. J., Ewers, R. M., 2021: Monitoring Forest climate change: Current progresses and challenges. Phenology in a ChangingWorld. Forests, 12:297. Global Change Biology, 25:1922–1940. Heumann, B.W., Seaquist, J.W., Eklundh, L., Jönsson, Preushler, T., 1999: Manual on methods and criteria for P., 2007: AVHRR derived phenological change in the harmonized sampling, assessment, monitoring and Sahel and Soudan, Africa, 1982–2005. Remote Sens- analysis of the effects of air pollution on forest. Part ing of Environment, 108:385–392. IX, Phenological observation. UN-ECE, 35 p. Hmimina, G., Dufrêne, E., Pontailler, J. Y., Delpierre, N., Puchałka, R., Koprowski, M., Gričar, J., Przybylak, R., Aubinet, M., Caquet, B. et al., 2013: Evaluation of the 2017: Does tree-ring formation follow leaf phenol- potential of MODIS satellite data to predict vegeta- ogy in Pedunculate oak (Quercus robur L.)? European tion phenology in different biomes: An investigation Journal of Forest Research, 136:259–268. using ground-based NDVI measurements. Remote Richardson, A. D., Keenan, T. F., Migliavacca, M., Sensing of Environment, 132:145–158. Ryu, Y., Sonnentag, O., Toomey, M., 2013: Climate Kočický, D., Ivanič, B., 2011: Geomorfologické členenie change, phenology, and phenological control of veg- Slovenska. Štátny geologický ústav Dionýza Štúra, etation feedbacks to the climate system. Agricultural Bratislava, Available at: https://www.geology.sk/ and Forest Meteorology, 169:156–173. new/sites/default/files/media/geois/PrehladneM - Schieber, B., Janík, R., Snopková, Z., 2009: Phenology of apy/GM_mapa.pdf four broad-leaved forest trees in a submountain beech Lukasová, V., Lang, M., Škvarenina, J., 2014: Seasonal forest. Journal of Forest Science, 55:15–22. changes in NDVI in relation to phenological phases, Soudami, K., Maire, G. M., Dufrene, E., Francois, Ch., LAI and PAI of beech forests. Baltic Forestry, 20:248– Delpierre, N., Ulrich, E. et al., 2008: Evaluation of the onset of green-up in temperate deciduos broadleaf Lukasová, V., Bucha, T., Škvareninová, J., Škvarenina, forests derived from Moderate Resolution Imaging J., 2019: Validation and Application of European Spectroradiometer (MODIS) data. Remote Sensing beech Phenological Metrics Derived from MODIS of Environment, 122:2643–2655. data along an Altitudinal Gradient. Forests, 10:60. Stöckli, R., Vidale, P. L., 2004: European plant phenol- doi:10.3390/f10010060 ogy and climate as seen in a 20-year AVHRR land- Miklós, L. 2002: Atlas krajiny Slovenskej republiky : surface parameter dataset. International Journal of Landscape atlas of the Slovak Republic (1. vyd.). Remote Sensing, 25:303–3330. Bratislava, Ministerstvo životného prostredia SR. Škvareninová, J., 2008: Vyhodnotenie nástupu jarných Moravčík, M. et al., 2021: Report on the forest sector of fenofáz duba letného (Quercus robur L.) v Zvolenskej the Slovak republic 2020 – Green report. Ministry of kotline vo vzťahu k teplotným sumám. Meteorolo- Agriculture and Rural Development of the Slovak gický časopis, 11:15–20. Republic, Bratislava and National Forest Centre, Škvareninová, J., 2014: Phenological signs of pedun- Zvolen, 69 p. culate oak (Quercus robur L.) in Slovakia as a bio- indicator of forest ecosystems, weather extremes and climate change. Zprávy lesnického výzkumu, 59:250–255. 201 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Štefančík, I., 1995: Fenológia v lesníctve: 1. Začiatok Zlatník, A., 1976: Přehled skupin typů geobiocénů vegetačnej činnosti. Lesnícky časopis – Forestry původně lesních a křovinných ČSSR. Zpravodaj Journal, 41:131–139. Geograc fi kého ústavu Československé akademie věd, Šmelko, Š., 1991: Štatistické metódy v lesníctve. VŠLD Brno, 13:55–64. Zvolen, skriptá, 276 s. Zhang, X., Friedl, H. A., Schaaf, B. S., Strahler A. H., Tholt, U., 2021: Modelovanie lesníckej fe n o l ó g i e Hodges, J. C. F., Gao, F. et al., 2003: Monitoring veg- s použitím vegetačných indexov spektrorádiometra etation phenology using MODIS. Remote Sensing of MODIS. Diplomová práca. Zvolen, Lesnícka fakulta Environment, 84:471–475. TU vo Zvolene. Internet Sources Townshend, J. R. G., Huang, S. N., Kalluri, V., Defries, SHMÚ, 2021: Bulletin – Meteorológia a klimatológia. R. S., Liang S., 2000: Beware of the per-pixel char- Slovenský hydrometeorologický ústav, 27:66. Availa- acterization of land cover. International Journal of ble at: https://www.shmu.sk/File/ExtraFiles/KMIS/ Remote Sensing, 21:839–843. publikacie/BMaK_0421.pdf Wesołowski, T., Rowiński, P., 2008: Late leaf develop- NFC Zvolen: Available at: http:/ / gis. nlcsk. org/ lgis/ ment in pedunculate oak (Quercus robur): An anti- Bochníček, O., 2015:Available at: https://www.shmu. herbivore defence? Scandinavian Journal of Forest sk/sk/?page=2049&id=592 Research, 23:386–394. Younes, N., Joyce, K. E., Maier, S. W., 2021: All models of satellite-derived phenology are wrong, but some are useful: A case study from northern Australia. International Journal of Applied Earth Observa- tion and Geoinformation, 97:102285. Available at: https://doi.org/10.1016/j.jag.2020.102285 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Forestry Journal de Gruyter

Spring phenology of oak stands in the Western Carpathians: validation of satellite metrics from MODIS using ground-based observations

Download PDF
12 pages

Loading next page...
 
/lp/de-gruyter/spring-phenology-of-oak-stands-in-the-western-carpathians-validation-HsMveJmpVn
Publisher
de Gruyter
Copyright
© 2022 Tomáš Bucha et al., published by Sciendo
ISSN
0323-1046
eISSN
2454-0358
DOI
10.2478/forj-2022-0014
Publisher site
See Article on Publisher Site

Abstract

The study focuses on the validation of the leaf unfolding (LU) onset of oak stands in the Western Carpathians in 2000–2021 derived from MODIS satellite data. LU onset was derived from the annual trajectories of the Normal- ised difference vegetation index (NDVI) fitted with a double sigmoid logistic function. The satellite metric Grow- ing speed day (GSD) corresponding to the LU onset is represented by the first derivative of the sigmoid function. Ground-based observations from 22 phenological stations of the Slovak Hydrometeorological Institute (SHMI) were used to validate the date of GSD. The results showed a good agreement between the medians of ground- and satellite-based LU onset dates. In addition to the median, the LU onset at the 5th and 95th percentiles were com- pared. For both percentiles, we found differences in the onset from MODIS and SHMI. The 5th percentile of the LU onset derived from MODIS was determined later than the one from SHMI data. With the 95th percentile, it was the opposite. As a result, the range determining the duration of LU onset from MODIS was significantly shorter than from SHMI observations. The trend analyses over the period 2000–2021 revealed a shift to the earlier onset of LU ~ −1 −1 0.33 day.year (p = 0.13) from satellite and ~ 0.32 day.year from ground-based observations (p = 0.08). The vali- dated LU onset and trends using the median allow analysing of the oak stands’ response to changing environmental conditions. However, the differences at the 5th and 95th percentiles, i.e. at the beginning and the end of the LU onset duration, remained unexplained. Key words: MOD09; MYD09; NDVI; leaf unfolding; validation Editor: Milan Barna of the course of phenological events. The AVHHR’s reso- 1. Introduction lution of 8 km was the main limiting factor in their use. Forest phenology is a traditional scientific discipline In Slovakia, satellite phenology has gained an impor- that examines the time course of important, periodi- tance following the launch of the Terra and Aqua satellites cally recurring life manifestations of forest trees, the so- (2000, 2002, NASA Earth Observation System Satel- called phenological phases. The onset of phenophases lites) with the Moderate Resolution Imaging Spectrora- depends mainly on the weather (Štefančík 1995) and diometer (MODIS). For the red and infrared bands, the is influenced by the changing climate. Therefore, the pixel size of 250 × 250 m is already sufc fi iently detailed to phenophases’onsets are considered indicators of cli- identify and select a sufficient number of homogeneous mate change impacts on forests (Richardson et al. 2013; pixels with one type of landscape cover for analyses, even Gray & Ewers 2021). A new dimension to study forest in the fragmented area of the Slovak forests (Bucha et al. ecosystem responses to changing environmental condi- 2011). Currently, a consistent database was created from tions at global and macro-regional levels was provided 2000 to 2021. It is based on MODIS products MOD09GQ by the satellite-based phenology using the Advanced and MOD09GA derived from the Terra satellite and Very High-Resolution Radiometer (AVHRR) (Stöckli & MYD09GQ and MYD09GA from the Aqua satellite. Vidale 2004; Piao et al. 2006; Heumann et al. 2007). The The products contain a quality layer that allows elimi- main benefit compared to sampling methods based on nating pixels with reflectivity affected by clouds, cloud ground-based observations lies in the possibility of quasi- shadows or high aerosol content. Several papers demon- continuous (in day intervals) and large-scale monitoring strated the suitability of using the Normalized Differen- *Corresponding author. Tomáš Bucha, e-mail: tomas.bucha@nlcsk.org, phone: +421 45 5314 156 © 2022 Authors. This is an open access article under the CC BY 4.0 license. T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 tial Vegetation Index (NDVI) derived from satellite data percentiles is important for understanding the response to monitor phenophases in forests (Zhang et al. 2003; of oak stands in the boundary conditions of their occur- Beck et al. 2006; Soudami et al. 2008). rence. Modelling methods are used to derive satellite-based The second goal of the work is to derive and compare phenological metrics. Under modelling, we understand temporal trends of the LU onset from ground surveys and a mathematical expression of the course of NDVI devel- satellite data. We assume that the trends derived from opment during the year and determine the onset of major both data sets will be identical. Due to the continuity phenological events. Berra & Gaulton (2021) report that of data acquisition with MODIS, after the validation it the logistic function (Zhang et al. 2003) and its modic fi a - will be possible to analyse LU changes more accurately tions (Fisher et al. 2006; Fisher & Mustard 2007) is the and use the new knowledge in proposing climate change most commonly used function in deriving vegetation phe- adaptation measures in forests. nology at local and regional levels. In Slovakia, the double logistic function was used for modelling (Bucha & Koreň 2017; Lukasová et al. 2014, 2019; Barka et al. 2019; Tholt 2. Material and methods 2021) through the PhenoProfile software application 2.1. Study area (© Milan Koreň). The main advantage of this application is the full user control of modelling parameters and the We carried out the research in the forest stands of Slo- possibility to refine the modelling procedures. Its main vakia with a dominant presence of pedunculate oak functionalities were described by Bucha & Koreň (2014). (Quercus robur L.) and/or sessile oak (Quercus petraea The rapid expansion in remote sensing and ground- Liebl.) (Fig. 1). These are economically important trees based phenology data acquisition has supported signifi- with a total share of 10.4% of forest area (Moravčík et al. cant advances in plant phenological research, as shown 2021). The species are not differentiated and are analysed in the recent review (Piao et al. 2019). However, some together as one genus in the study. research results point at methodological risks in deriving Geographically, most of the territory belongs to the and interpreting phenological models based on satellite province of the Western Carpathians, while a part of the data from various sensors (Norris et al. 2020; Younes et eastern Slovakia roughly from 21°15’ longitude belongs al. 2021). Therefore, the validation of derived regional to the province of the Eastern Carpathians (Kočický & products still appears to be very topical and necessary. Ivanič 2011). The occurrence of assessed oak stands is Data from terrestrial measurements are used for concentrated in the 1st to 3rd forest vegetation zones validation and parameterization of remote sensing from the 8-level scale (Zlatník 1976). It is an area with outputs (Hmimina et al. 2013). The onset of individual an average annual temperature above 5.5 °C, annual pre- phenophases derived from MODIS for beech stands cipitation totals up to 800 mm and a growing season of in Slovakia was validated by Pavlendová et al. (2014) 150 to 180 days. The dominant soil on which oak stands and Lukasová (2019) by comparison them with obser- occur are cambisols, followed by leptosols, luvisols and vations from the network of phenological stations of the planosols (Miklós 2002). Slovak Hydrometeorological Institute (SHMI). In addi- tion to SHMI data, they also used their own observations. A validation for oak stands has not been performed, yet. 2.2. Oak stands selection Therefore, our first goal is to validate a regional When selecting oak stands, we used the tree species map phenological model derived from NDVI from MODIS of Slovakia from Landsat (Bucha 1999) with a resolution satellite images for oak stands. The observations from resampled to 250 m in a combination with the current the SHMI regional phenological stations for the period database of the Forestry Information System (FIS). The 2000–2021 were used to validate the day of the onset database contains data at the compartment level (NFC of the leaf unfolding (LU) spring phenophase. SHMI Zvolen – https://gis.nlcsk.org/lgis/). Data from FIS was observations were used to verify the hypothesis that the used to calculate the proportion of oak in 250 m pixels. local extreme of the first derivative of Fisher’s sigmoidal We excluded marginal pixels from the tree species map. function (Fisher & Mustard 2007) is related to the leaf The reason is that these pixels in the MOD09 products unfolding onset. Based on the knowledge of the Fisher’s may be spectrally contaminated with other land cover function course and its validation in beech stands (Pav- categories (Townshend et al. 2000) and thus distort the lendová et al. 2014), we assume that the LU phenophase derived NDVI values. We also excluded pixels where the of oak will correspond to the GSD (Growth Speed Day) satellite metric. The LU onset phenophase was chosen oak proportion was less than 70%. Thus, created masks of because it is used as the start of the growing season. homogeneous oak stands contained 981 pixels. Of these, In contrast to published studies, we validate the onset of we selected 413 pixels by stratified random sampling, on which the LU onset was analysed (Fig. 1). The number LU derived from MODIS over the entire duration of the of points ensured the representatives of the sample. At phenophase using 5-percentile, median (50-percentile) and 95-percentile. The validation of the 5th and 95th the same time, the iterative calculation of 6 parameters 192 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Fig. 1. Forests with a predominant occurrence of oak (brown colour), the spatial distribution of SHMI phenological stations (red circles) and analysed pixels from MODIS with the occurrence of oak ≥ 70% (black crosses). The forests are shown in grey. The numbering of the phenological stations is in accordance with Table 2. of Fisher’s function modelling the NDVI profile during the relationship: the year was manageable in about 6 hours on a PC with ܫܴܰ Ȃ  ܴ݁݀ ଼ସଵ Ȃ଼ ଻଺ ௡ ௠ ଺ଶ ଴Ȃ ଺଻ ଴ ௡௠ a six-core processor. The altitudinal distribution of the ܰܦܸܫ ൌ ൬ ൰ [1] ܫܴܰ ൅ ܴ݁݀ ଼ସ ଵȂ ଼଻଺ ௡ ௠ ଶ଺ ଴Ȃ ଻଺ ଴ analysed oak pixels is shown in Fig. 2a. Fig. 2b depicts the altitudinal distribution of SHMI stations. The layer of quality was taken from the MOD/ MYD09GA products. Using the layer, we excluded pixels affected by cloud cover, cloud shadows and high aerosol 2.3. Modelling of the annual NDVI course content from the analyses. By overlaying the NDVI layer The solution was based on MODIS images, specifically adjusted in this way with a layer of selected oak stands, on products MOD/MYD09GQ and MOD/MYD09GA a database suitable for modelling oak phenology was collections 5 and 6. The products represent the spec- created. The procedure for choosing and pre-processing tral reflectance on the earth’s surface, i.e. the effect of MODIS images and deriving NDVI is described in detail absorption and scattering of radiation in the atmosphere in Bucha & Koreň (2014 and 2017). was eliminated. From the MOD09GQ and MYD09GQ The subject of the analysis was the NDVI time series products, we derived the NDVI using the red and infrared from March to July and the years 2000 to 2021. The bands with a pixel resolution of 250 × 250 according to annual development of NDVI was modelled by a double Fig. 2. a) Frequency histogram of oak stands (with oak share of 70% and more) along the altitudinal range. b) Frequency his- togram of altitude distribution of SHMI stations. X–axis: altitude (m a.s.l.). Y–axis: a) number of pixels (MODIS); b) number of SHMI stations. 193 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 logistic function (Fisher & Mustard 2007): most leaves are likely to emerge and that deciduous for- ests with different crown canopies or even small amounts Y W Y  Y [2] of conifers can be compared at the date of onset. The GSD PL Q DPS P  Q W P  Q W H   H date was calculated for each of the 413 analysed pixels for each year from 2000–2021. where v(t) is NDVI observed at time t; v and v are param- min amp eters corresponding to the minimum value of the NDVI and its amplitude; m and n parameters control the shape and slope of 1 1 the curve of the ascending (spring) phase, and m and n param- 2 2 2.4. Data validation and statistical evaluation eters control the descending (autumn) phase. We used ground-based phenological observations to validate and reveal possible errors in modelling LU onset The advantage of using this type of the sigmoid func- from MODIS. From the SHMI network, we used 22 forest tion is that it is differentiable. The extremes of the first phenological stations focused on oak evaluation (Fig. 1; and second derivatives of equation 2 were used to derivate Table 2), where more than 95% of the data on the LU satellite metrics and corresponding phenological events. onset were available in individual years from 2000–2021. The connection of spring phenophases from ground Observations were made individually using a telescope. observations and satellite metrics is shown in Table 1 10 pre-dominant or dominant trees were evaluated at based on Pavlendová et al. (2014). each plot. In the next text, where it is necessary to dis- tinguish the LU onset between ground and satellite Table 1. Satellite-based phenological metrics and correspond- observations, we use the abbreviations LU_SHMI and ing vegetative phenophases by Pavlendová et al. (2014). LU_MODIS. Satellite The description of satellite phenological metrics derived from [2] – metrics Corresponding vegetation phenophases GAD Maximum of the second derivative in a spring phase ~ Bud break onset Table 2. Basic characteristics of selected SHMI forest pheno- Extreme of the first derivative in a spring phase (spring inflection point) logical stations. GSD ~ Leaf unfolding (LU) onset Station number Lati- Longi- Altitude Minimum of the second derivative in a spring phase ~ End of leaf un- Station name GDD – see Fig. 1. tude tude [m a.s.l.] folding GAD – Growing Acceleration Day; GSD – Growing Speed Day; GDD – Growing Deceleration 1 Riadok 48° 30’ 17° 10’ 183 Day; DOY – Day of Year. 2 Hajnáčka 48° 13’ 19° 57’ 250 3 Lučenec 48° 20’ 19° 41’ 250 4 Levice 48° 13’ 18° 37’ 280 5 Stakčín - CHOTINA 49° 00’ 22° 13’ 300 6 Železná studienka 48° 10’ 17° 07’ 320 7 Stupava 48° 16’ 17° 02’ 320 8 Hvezdáreň 48° 22’ 17° 17’ 350 9 Čebovská Bukovina 48° 14’ 19° 18’ 360 10 Kokava nad Rimavicou 48° 34’ 19° 51’ 370 11 Kšinná 48° 48’ 18° 21’ 374 12 Prešov-Cemjata 48° 58’ 21° 10’ 375 13 Zborov 49° 22’ 21° 18’ 400 14 Myjava 48° 46’ 17° 35’ 450 15 Jasov Lesy-Premonstr. 48° 40’ 20° 58’ 450 16 Bytča-Starovec 49° 14’ 18° 33’ 460 17 Nové Mesto nad Váhom 48° 46’ 17° 50’ 500 18 Ždaňa 48° 34’ 18° 45’ 500 19 Kecerovce 48° 49’ 21° 24’ 550 20 Zvolen 48° 34’ 19° 10’ 718 Fig. 3. The course of the NDVI curve in the spring phase. 21 LS Veľké Kapušany 48° 32’ 22° 05’ 108 The black dots represent the NDVI in a given pixel on a given 22 LS Svinica 48° 44’ 21° 28’ 350 DOY. The SHMI stations are evenly distributed across Slo- In this work, we analyse only the phenophase of leaf vakia so that the highest possible number of tree species unfolding. LU onset was evaluated according to the can be observed simultaneously in each station (Fig. scale of the manual for phenological observations for the 1). Therefore, it was impossible to match stations with pan-European monitoring system (Preuhsler 1999) and analysed oak pixels from MODIS. For this reason, we according to the scale developed by SHMI (Braslavská & used the two-sample Kolmogorov-Smirnov test to test Kamenský 1996). The LU onset is considered the DOY the agreement of the altitude distribution of two sample when the first light green leaves, smaller than in adult - sets, i.e. 413 pixels from MODIS and 22 SHMI stations. hood, appeared on at least half of the individuals of the The hypothesis H0: The two samples follow the same given observed group. altitudinal distribution was tested against H1: The alti- The GSD satellite metric corresponds to the LU onset tudinal distributions of the two samples are different. The phenophase. GSD is defined as the date when the sig - reason for testing the equality of altitudinal distributions moid reaches its half-maximum value or peak of the first is the strong dependence of LU onset to the altitude. Any derivative of the equation [2]. Fisher & Mustard (2007) differences in sample sets would bias the comparison of state that the half-maximum represents the date when differences between LU_SHMI and LU_MODIS, as well 194 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 as validation of LU onset modelling from MODIS data 3. results in individual years 2000–2021. 3.1. Comparison of altitudinal distribution of For the nationwide evaluation of the LU onset (LU_SHMI and LU_MODIS) of oak stands, we calcu- SHMI and MODIS data sets lated the median value in individual years of the period The results of the Kolmogorov-Smirnov test, with which 2000–2021. We expressed the intra-annual variability we tested the equality of the altitudinal distribution of the with the 5th and 95th percentiles. These represent the LU two sample sets, are presented in Table 3. The overlap onset day values below which 5 (95) % of the distribution of the cumulative distribution functions of SHMI and lies. This means that the 5th percentile represents the day MODIS samples documents the equality of their alti- of LU onset at lower altitudes or in southern areas. The tudinal range in Fig. 4. Therefore, we consider that the 95th percentile represents the day of LU onset at higher samples follow the same altitudinal distribution and are altitudes or in northern regions of oak occurrence. We suitable for: note that the median represents the 50th percentile. The uncertainty in the LU onset assessment from (i) comparing the onset of LU from ground-based and MODIS was analysed with statistical metrics such as MODIS observations, a Coefficient of determination (R ), Mean bias error (ii) identification of modelling inaccuracies in individ- (MBE) and Mean absolute error (MAE). The statistical ual years 2000–2021, metrics were used to compare the model’s performance (iii) validation of LU onset from MODIS. in deriving the GSD using equation [2] with SHMI obser- vations. Table 3. Two-sample Kolmogorov-Smirnov test for compar- ing the equality of the altitudinal distributions of ground- Coefficient of determination (R ): based and MODIS phenological plots. ۍ ې ത ത σ ሺ ሻ ൫ܻ െܻ ൯ ܻ െܻ ௙ ௙ ௠ ௠ ଶ N u m b e r Significance ێ ۑ ܴ ൌ [3] Source p – value Interpretation ێ ଶ ۑ of plots level – alpha ට ത ത σ ൫ܻ െܻ ൯ σሺ ܻ െܻ ሻ ௙ ௙ ௠ ௠ ۏ ے As the computed p–value > alpha, one SHMI 22 cannot reject the null hypothesis H0: 0.883 0.05 Mean bias error (MBE): The two samples follow the same dis- MODIS 413 σ ሺܻ െܻ ሻ ௠ ௙ tribution. ܯܤ ܧ ൌ [4] Mean absolute error (MAE): σ หܻ െܻ ห ௠ ௙ ܯܣ ܧ ൌ   [5] where Y = ground-based value (LU onset from SHMI), Y = f m modelled value (LU onset from MODIS), Ῡ = average of Y f f values, Ῡ = average of Y values, N = number of years in the m m analysed period 2000–2021. We used a simple linear correlation and regression analysis to derive temporal trends in LU onset from SHMI and MODIS data. The input was the time series of the median values of the LU onset calculated for indi- vidual years of the period 2000–2021. We assessed the signic fi ance of the trends by testing the signic fi ance of the sample correlation coefficients. Under hypothesis H0, Fig. 4. Comparison of the equality of the cumulative altitu- the correlation coefficient in the population is equal to 0. dinal distribution of ground-based phenological plots of the It is proven that testing the signic fi ance of the correlation SHMI stations and MODIS pixels. coefficient is identical to testing the significance of the regression coefc fi ient (Šmelko 1991). The test character - istic t was used. For t > t , the hypothesis H0 is rejected at α,f 3.2. Validation of LU onset from MODIS α % level of significance, and with P = 1 – α% confidence using ground-based observations it is considered that both the correlation and the regres- sion sample coefficients are different from 0. Input data on the LU onset from ground-based SHMI We verie fi d whether the trends derived from MODIS observations (LU_SHMI) and MODIS (LU_MODIS) and SHMI were significantly different by testing the are presented in Table 4. The course of both onsets is hypothesis about the equality of the two regression coef- graphically compared in Fig. 5a–c, where the medians, ficients. 5th and 95th percentiles express the day of LU onset for individual years of the period 2000–2021. 195 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Table 4. Median, 5th and 95th percentiles of leaf unfolding from MODIS (413 pixels) and from phenological observations in the SHMI network (22 stations). Year LU_MODIS median 5th percentile 95th percentile LU_SHMI median 5th percentile 95th percentile 2000 111 109 113 114 105 119 2001 119 116 122 117 101 125 2002 119 111 121 114 104 125 2003 117 115 121 118 111 125 2004 116 112 122 119 110 128 2005 115 110 119 122 110 134 2006 115 112 121 120 110 129 2007 105 101 111 109 99.2 120 2008 115 110 122 115 103 126 2009 104 102 111 108 103 119 2010 115 111 119 118 105 125 2011 108 105 113 111 105 120 2012 115 110 119 116 102 125 2013 117 115 121 117 110 123 2014 96 91 102 103 89 119 2015 111 108 117 113 90.3 123 2016 105 102 109 108 90.5 121 2017 106 97 119 107 93.1 123 2018 107 105 108 108 96.3 117 2019 106 102 114 110 93.1 119 2020 109 106 116 110 97.1 121 2021 124 119 129 123 105 135 Average 111.59 107.68 116.77 113.64 101.48 123.68 Table 5 shows the differences (∆) in days between the with the original parameters. Therefore we consider the LU onset from ground observations and MODIS using LU_MODIS onset DOY as confirmed. MBE and MAE statistics derived according to relations The regression dependence between the onset of [4] and [5]. We found the smallest differences in the LU LU_SHMI and LU_MODIS is expressed in Figures 5d–f. onsets between the medians and the largest for the 95th The highest correlation coefc fi ient and the concentration percentile. The difference between the 5th percentiles of the points around the line y = x was observed for the was also substantial. The median MBE of −2.05 days median LU values (Fig. 5d). indicates an earlier determination of LU onset from The calculated differences between the 95th and 5th MODIS compared to ground observations. Similarly, percentiles, which determine the duration of the LU MBE of −6.91 days for the 95th percentile represents an onset phase, are shown in Fig. 6a for the entire monitored earlier determination of LU onset from MODIS. Con- period 2000–2021. The graph shows that the duration versely, the MBE of 6.20 days for the 5th percentile rep- of LU onset from MODIS was systematically under- resents the later onset from MODIS versus SHMI. estimated compared to ground observations of SHMI. The significant difference results from the two already described opposite tendencies of LU determination at the Table 5. Mean bias error (MBE) and Mean absolute error (MAE) in days between SHMI and MODIS leaf unfolding on- 5th and 95th percentiles. set for median, 5th percentile and 95th percentile. ∆ LU ∆ LU ∆ LU median 5th_percentile 95th_percentile MBE −2.05 6.20 −6.91 3.3. Temporal trend of leaf unfolding onset MAE 2.77 6.29 6.91 Median values of the day of LU onset, intra-annual vari- A close relationship between the median LU onset ability expressed by the 5th and 95th percentiles for the values from MODIS and SHMU is evident from the individual years of the period 2000–2021 and the calcu- course of DOY during the entire observed period (Fig. lated linear trends of the LU onset are shown in Fig. 7a 5a). It confirms the suitability of the GSD metric for and 7b for SHMI and MODIS selection of oak stands, determining LU onset from MODIS. At the same time, respectively. it was possible to identify the years with greater differ- Fig. 7a and 7b show considerable intra- and inter- ences between satellite and ground-based DOYs that annual variability of DOY of leaf unfolding onset. In may indicate possible errors in determined values that the Carpathian conditions, the intra-annual variability need to be checked. These are mainly the years 2002, is mainly conditioned by the altitude and the related air 2005, 2006 and 2014. For these years, we re-derived temperature. The inter-annual variability is determined the LU_MODIS onset day (GSD metric) by changing by the spring weather. Both SHMI and MODIS obser- the modelling parameters. We expanded the parameter vations show a tendency towards an earlier LU onset ranges of function [2] and doubled the subspaces within over the period 2000 to 2021. The regression coefficient which the iteration process took place. The calculation determines the magnitude of the change per year. For extended to more than three days, but in no case did the ground-based SHMI observations, this change was −1 difference exceeds 0.2 days compared to the estimate −0.32 days.year (p = 0.08) and for MODIS it was 196 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Fig. 5. Left: Comparison of the leaf unfolding onset from SHMI and MODIS observations for the period 2000–2021: a) median, b) 5th percentile, c) 95th percentile. X–axis: year; Y–axis: Day of the year (DOY). Right: Dependence between the leaf unfolding onset from SHMI and MODIS observations for d) median, e) 5th percentile, f) 95th percentile. X– and Y–axes: Day of the year (DOY). The dashed line corresponds to the line y = x. Fig. 6. Course (a) and dependence (b) between the duration of the LU onset on 22 SHMI phenological stations and 413 MODIS pixels. Duration is expressed in days and is calculated as the difference between the 95th and 5th percentiles for each year of the period 2000–2021. 197 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Fig. 7. Leaf unfolding onset day (median), duration (5th and 95th percentile) and onset trends of oak stands for the period 2000–2021 according to a) SHMI observations and b) MODIS. X–axis: year; Y–axis: Day of the year (DOY). −1 −0.33 days.year (p = 0.13). The test of the hypoth- 111.6 DOY in MODIS and 113.6 DOY in SHMI obser- esis about the equality of regression coefficients proved vations. that the slope of both regression lines was identical In both data sets, we recorded the earliest onset of (p < 0.001). LU in 2014, when the median DOY for Slovakia was 96 (MODIS), or 103 (SHMI). The year 2014 was historically the first year in Slovakia when the cumulative value of the average daily air temperature exceeded all previous 4. Discussion observations since 1951 (Bochníček 2014 – available 4.1. Notes on phenological modelling online). and validation of results In 2021, we observed the atypical late onset of LU DOY equal to 124 DOY (MODIS) and 123 DOY (SHMI). The day of the leaf unfolding onset was derived from According to the SHMI meteorological report, the tem- NDVI annual trajectories fitted with the double logis - perature in March 2021 was normal, but the tempera- tic function according to formula [2]. Only the spring ture in April was below average in the entire territory of period from about 70 to 200 DOY was analysed. The Slovakia and the weather was unstable and cold. This parameters m , n of equation [2] are decisive for autumn 2 2 type of weather had a contrary character compared to the phenophases and do not affect the calculation of spring above average air temperatures in April in the previous parameters in eq. [2]. Therefore, the NDVI images from decade. It was related to atypical circulation conditions DOY above 200 were not included in the modelling. in the northern hemisphere resulting from the global cli- The LU onset determined from 413 MODIS pixels mate change (SHMÚ, 2021). and 22 SHMI phenological stations and expressed by the median had a comparable course during the monitored period 2000–2021 (Fig. 5a). We checked the four more substantial differences in 2002, 2005 and 2006 and 2014 4.3. Intra-annual variability of leaf unfolding by recalculating the DOY of LU onset from MODIS. The onset refined DOY derivation did not differ substantially from The intra-annual variability of the LU onset is mainly the original one. We did not n fi d an exact explanation for conditioned by climatic conditions resulting from the the identified differences, as significant differences did altitude and the current development of the weather not occur in other years. It is also not possible to rule out in a given year. The air temperature plays a decisive the error of the ground-based phenological observations role (Škvareninová 2008). Gafenco et al. (2022) found due to the subjectivity of the observers. a longitudinal shift in the budburst of Quercus petraea in Romania. Besides, the variation in leaf flush timing involves some ecological trade-offs in temperate regions. 4.2. Inter-annual variability of the leaf For example, the phenological form of pedunculate oak unfolding onset (Quercus robur var. praecox) in Eastern Poland flushes early in the season to avoid summer droughts but con- The days of LU onset expressed by the median during sequently suffers from later spring frost and insect her- 2000–2021 naturally varied depending on the spring bivory, whereas the other oak form (Quercus robur var. weather. Annual median values ranged from 96 to 124 tardiflora) flushes up to five weeks later in the season, DOY for MODIS and from 103 to 123 DOY for SHMI (Fig. 5a). The mean value for the period 2000–2021 was avoiding insect herbivory but suffering from summer 198 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 droughts (Wesołowski & Rowiński 2008; Puchałka et longitude has not yet been quantified in the oak forests. al. 2017). Considering the mask with the oak share of at least We expressed the range of oak stands LU onset in 70%, it is not possible to exclude the influence of other individual years using the 5th and 95th percentiles (Fig. tree species occurring in a given pixel on the LU onset. 5b and 5c). For both percentiles, we found differences in The tree species with the highest representation in Slo- determining the LU onset between MODIS and SHMI. vakia could have the most significant impact: beech at Moreover, the nature of the difference was contradic- the 95th percentile, hornbeam and Turkey oak, especially tory. The 5th percentile of the onset determined from at the 5th percentile due to their altitudinal occurrence. MODIS occurred later than from ground observations. Schieber et al. (2009) showed an earlier beech and horn- For the 95th percentile, we observed the opposite. When beam LU onset of up to 6 days (average for 1995–2007) determining the duration of the leaf unfolding onset in compared to Sessile oak in central Slovakia. On the con- Slovakia from MODIS, the phenophase was the shortest trary, a later LU onset is reported in the case of Turkey in 2000 and 2018, lasting only 5 or 4 days. The longest oak. When analysing LU onset with an 85% oak mask phenophase duration was found in 2017, equal to 23 (136 pixels), we found no shift for the 5th percentile com- days. In other years, it varied from 7 to 13 days. Accord- pared to the 70% mask. For the 95th percentile, there was ing to ground-based phenological observations of SHMI, a 1-day shift to a later LU onset and thus a prolongation the shortest phenophase duration was in 2013, 2000 of the phenophase. and 2003, namely 14 and 15 days. The most extended Both effects (a shift due to the earlier LU onset of duration of the phenophase was found in 2015, namely beech in analysed pixels and possible latitudinal shift) 33 days. In other years, it varied from 16 to 31 days. could extend the average 10.1 days-long duration of the Due to the opposite nature of determining the 5th and LU onset, found in our work, by 2–2.5 days. Despite this, 95th percentiles of LU onset from MODIS compared to most of the differences remain unexplained. SHMI, the resulting range of the phenophase duration was even more different than for the individual percen- tiles. This was manifested by a substantially shorter range 4.4. Statistics to determine the leaf unfolding of the LU onset in MODIS than in SHMI observations. onset The average duration of the phenophase from MODIS for From the analysis above of the intra- and inter-annual var- the entire monitored period was 10.1 days, while accord- iability of the LU onset and from the regression depend- ing to SHMI it was up to 23.2 days. ence between LU_MODIS and LU_SHMI expressed in A partial explanation for the significant difference Fig. 5d–f, it follows that for determining the onset, it is is the uneven geographical distribution of SHMI sta- most appropriate to use the median value. tions and pixels from MODIS. From Fig. 1 we can see The coefficient of determination between satellite that some SHMI stations are located further south (sta- and ground observations was R = 0.82 (Fig. 5d), and tions no. 2, 4, 6–9, 21) but also further north (stations the Mean bias error (MBE) was −2.05 days, indicating no. 5, 13, 16) than most MODIS pixels. Considering the the earlier detection of the LU onset from MODIS (Table south-north and partly the west-east dependence of the 5). The 5th and 95th percentiles are inappropriate met- onset of spring phenophases (with increasing latitude rics if the presented methodical approach is used. In the and continentality, the phenophase is delayed) and the case of the 5th percentile (R = 0.53, MBE = 6.20 days), smaller number of SHMI stations, it cannot be ruled out we revealed a substantial shift to a later onset derived that the mentioned factors affected the comparability of from MODIS compared to SHMI. On the contrary, for SHMI and MODIS data despite the agreement of their the 95th percentile (R = 0.63, MBE = −6.91 days), there altitudinal distribution. For example, Stakčín and Zborov was a substantially earlier determination of the LU onset stations, which are the most northern and at the same from MODIS compared to SHMI. time the most eastern, but also Bytča-Starovec, which The magnitude of the deviation in the median values is located in the northwest of the country, regularly had of LU onset between satellite and ground assessments the late LU onset. This affectes the 95th percentile of LU (MBE −2.05 days and MAE 2.77 days) confirms our onset. The southerly located SHMI stations, Hvezdáreň assumption that the GSD satellite metric represents the in the west, Hajnáčka in the middle and Veľké Kapušany phenophase LU onset and is a suitable metric for satellite- in eastern Slovakia, usually had the earliest onsets of the based observations of the mentioned phenophase. The phenophase. This was manifested in the 5th percentile determination of the LU onset at marginal conditions towards the earlier LU onset compared to MODIS. Barka of oak stand occurrence, expressed by the 5th and 95th et al. (2019) quantified a delay in LU onset towards the percentiles, and the duration of the LU onset were not North by 1.05 days per 100 km from MODIS data for sufficiently validated. The differences between ground beech forests in the Pannonian-Carpathian macro- and satellite observations should be the subject of further region. The longitudinal delay was observed as well but research. was not significant. The LU onset shift due to latitude/ 199 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 to an earlier onset of LU in 22 years. The differences 4.5. Temporal trend of the leaf unfolding onset between ground and satellite observations at the 5th The length of the monitored period (2000–2021) is suf- and 95th percentiles, i.e. at the beginning and the end of ficient for a statistically relevant derivation of temporal the duration of the LU onset in Slovakia, were not suf- trends in the LU onset. The statistical evaluation results ficiently explained. at the national level showed a tendency towards an ear- Based on the achieved results, it will be possible to lier LU onset of oak stands. The rate of change derived expand the scope of modelling the spring phenophase −1 from ground observations was −0.32 days.year , and leaf unfolding onset using continuous satellite data −1 −0.33 days.year in MODIS observations (see regres- from MODIS. However, the derivation of more accu- sion coefficients in Fig. 7a and 7b). Over the entire rate results on the reaction of trees to the manifestations monitored period of 22 years, this represented a shift of of climate change along the altitudinal gradients requires ~7 days to an earlier start of leaf unfolding. This result a clarification of the differences in the LU onset between is in line with Škvareninová’s (2014) finding based on ground-based and satellite data at the peripheral occur- ground observations in Slovakia from 1988–2013, which rence of oak stands. Only subsequently it will be possi- showed ane earlier LU onset of Quercus robur by 7 days ble to apply the acquired knowledge, for example, in the −1 (−0.27 days.year ). The values of the regression coeffi- forestry decision-making about changes in the transfer cients (LU_SHMI and LU_MODIS) are almost identical, of forest reproductive material between forest seed the difference between them was statistically insignifi - regions or about changes in the regeneration and target cant. This conr fi ms our assumption that MODIS satellite tree species composition in management models appli- data and the GSD metric derived from them are suitable cable in the adaptive forestry. for temporal analyses and derivation of trends in the LU onset of oak stands. This finding is important because the shift to an earlier onset of the phenophase is a consid- Acknowledgement ered response of woody plants to climate change (Berra This publication is the result of the project implementation Cen- & Gaulton 2021). Thus, the use of satellite data makes it tre of Excellence of Forest-based Industry, ITMS: 313011S735, possible to reliably analyse the reaction of oak stands even supported by the Operational Programme Integrated Infrastruc- at regional or local levels due to the significantly larger ture funded by the ERDF. Study was partially supported by the MODIS dataset compared to ground surveys. Slovak Research and Development Agency under the contract No. APVV–20–0365. 5. Conclusion In the presented study, we validated the accuracy of the references satellite-based modelling of the leaf unfolding onset Barka, I., Bucha, T., Molnár, T., Móricz, T., Somogyi, Z., spring phenophase. LU onset of a large homogeneous Koreň, M., 2019: Suitability of MODIS-based NDVI set of oak stands (413 pixels) from the MODIS was com- index for forest monitoring and its seasonal applica- pared with data from 22 ground-based phenological sta- tions in Central Europe. Central European Forestry tions in the Slovak part of the Carpathians territory. In Journal, 65:206–217. the case of satellite data, the LU onset corresponds to the Beck, P. S. A, Atzberger, T. C., Hogda, K. A., Johansen, extreme of the first derivative of the double logistic func- B., Skidmore, A. K., 2006: Improved monitoring of tion [2] in its spring phase, represented by the GSD met- vegetation dynamics at very high latitudes: a new ric. From the verie fi d statistics, median, 5-percentile and method using MODIS NDVI. Remote Sensing of 95-percentile, we found the closest relationship between Environment, 100:321–334. ground-based and satellite determination of LU onset at Berra, E., Gaulton, R., 2021: Remote sensing of temper- the median calculated from onset days in individual pixels ate and boreal forest phenology: A review of progress, and SHMI stations. The coefc fi ient of determination was challenges and opportunities in the intercomparison R = 0.82 with 2.05 days earlier detection of the LU onset of in-situ and satellite phenological metrics. Forest by MODIS. We can therefore consider the LU onset from Ecology and Management, 480:118663. MODIS expressed by the median as validated and reliable Braslavská, O., Kamenský, L., 1996: Fenologické pozo- for further analyses. rovania lesných drevín. Metodický predpis Sloven- The temporal analysis of the trend over the period ského hydrometeorologického ústavu. SHMÚ, 22 s. 2000–2021 also confirmed the consistency of determin- Bucha, T., 1999: Classification of tree species compo - ing LU onset using the median values. The statistical test sition in Slovakia from satellite images as a part of confirmed that the regression coefficients determining monitoring forest ecosystems biodiversity. Zvolen, the rate of the temporal change derived from SHMI and LVÚ Zvolen, Acta Instituti Forestalis Zvolen, Tomus MODIS data were identical. We found a shift of ~7 days 9:65–84. 200 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Bucha, T., Priwitzer, T., Koreň, M., 2011: Modelovanie Norris, J. R., Walker, J. J., 2020: Solar and sensor geom- fenologického vývoja lesných porastov pomocou etry, not vegetation response, drive satellite NDVI vegetačného indexu NDVI odvodeného zo satelitných phenology in widespread ecosystems of the western snímok MODIS. Lesnícky časopis – Forestry Journal, United States. Remote Sensing of Environment, 249: 57:187–196. 12013. Pavlendová, H., Snopková, Z., Priwitzer, T., Bucha, T., Bucha, T., Koreň, M., 2014: Fenológia dubových a buko- 2014: Using of long-term phenological observations vých porastov v období 2000–2014. In: Bucha, T. et of SHMI and NFC for validation of regional phenol- al. (ed.): Satelity v službách lesa. Bratislava, SAP, ogy model for European beech. In: Roznovsky, J., p. 101–115. Litschmann, T. (eds.): Proceedings of the Inter- Bucha, T., Koreň, M., 2017: Phenology of the beech national Conference Mendel and Bioclimatology. forests in western Carpathians from MODIS for Masaryk University (Brno, Czech Republic) 3–5 Sept 2000–2015. iForest - Biogeosciences and Forestry, 2014, p. 294–311. Available at: http://www.better- 10:537–546. cities.eu/bulletin/wp-content/uploads/2014/09/ Fisher, J. I., Mustard J. F., 2007: Cross-scalar satellite mendel_ConferenceProceedingsMendel.pdf phenology from ground, Landsat and MODIS data. Piao, S., Fang, J., Zhou, L., Ciais, P., Zhu, B., 2006: Remote Sensing of Environment, 109:261–273. Variations in satelite-derived phenology in China’s Gafenco (Pleşca), I. M., Pleşca, B. I., Apostol, E. N., temperate vegetations. Global Change Biology, Şofletea, N., 2022: Spring and Autumn Phenology 12:672–685. in Sessile Oak (Quercus petraea) Near the Eastern Piao, S. L., Liu, Q., Chen, A. P., Janssens, I. A., Fu, Y. Limit of Its Distribution Range. Forests, 13:1125. S., Dai, J. H. et al., 2019: Plant phenology and global Gray, R. E. J., Ewers, R. M., 2021: Monitoring Forest climate change: Current progresses and challenges. Phenology in a ChangingWorld. Forests, 12:297. Global Change Biology, 25:1922–1940. Heumann, B.W., Seaquist, J.W., Eklundh, L., Jönsson, Preushler, T., 1999: Manual on methods and criteria for P., 2007: AVHRR derived phenological change in the harmonized sampling, assessment, monitoring and Sahel and Soudan, Africa, 1982–2005. Remote Sens- analysis of the effects of air pollution on forest. Part ing of Environment, 108:385–392. IX, Phenological observation. UN-ECE, 35 p. Hmimina, G., Dufrêne, E., Pontailler, J. Y., Delpierre, N., Puchałka, R., Koprowski, M., Gričar, J., Przybylak, R., Aubinet, M., Caquet, B. et al., 2013: Evaluation of the 2017: Does tree-ring formation follow leaf phenol- potential of MODIS satellite data to predict vegeta- ogy in Pedunculate oak (Quercus robur L.)? European tion phenology in different biomes: An investigation Journal of Forest Research, 136:259–268. using ground-based NDVI measurements. Remote Richardson, A. D., Keenan, T. F., Migliavacca, M., Sensing of Environment, 132:145–158. Ryu, Y., Sonnentag, O., Toomey, M., 2013: Climate Kočický, D., Ivanič, B., 2011: Geomorfologické členenie change, phenology, and phenological control of veg- Slovenska. Štátny geologický ústav Dionýza Štúra, etation feedbacks to the climate system. Agricultural Bratislava, Available at: https://www.geology.sk/ and Forest Meteorology, 169:156–173. new/sites/default/files/media/geois/PrehladneM - Schieber, B., Janík, R., Snopková, Z., 2009: Phenology of apy/GM_mapa.pdf four broad-leaved forest trees in a submountain beech Lukasová, V., Lang, M., Škvarenina, J., 2014: Seasonal forest. Journal of Forest Science, 55:15–22. changes in NDVI in relation to phenological phases, Soudami, K., Maire, G. M., Dufrene, E., Francois, Ch., LAI and PAI of beech forests. Baltic Forestry, 20:248– Delpierre, N., Ulrich, E. et al., 2008: Evaluation of the onset of green-up in temperate deciduos broadleaf Lukasová, V., Bucha, T., Škvareninová, J., Škvarenina, forests derived from Moderate Resolution Imaging J., 2019: Validation and Application of European Spectroradiometer (MODIS) data. Remote Sensing beech Phenological Metrics Derived from MODIS of Environment, 122:2643–2655. data along an Altitudinal Gradient. Forests, 10:60. Stöckli, R., Vidale, P. L., 2004: European plant phenol- doi:10.3390/f10010060 ogy and climate as seen in a 20-year AVHRR land- Miklós, L. 2002: Atlas krajiny Slovenskej republiky : surface parameter dataset. International Journal of Landscape atlas of the Slovak Republic (1. vyd.). Remote Sensing, 25:303–3330. Bratislava, Ministerstvo životného prostredia SR. Škvareninová, J., 2008: Vyhodnotenie nástupu jarných Moravčík, M. et al., 2021: Report on the forest sector of fenofáz duba letného (Quercus robur L.) v Zvolenskej the Slovak republic 2020 – Green report. Ministry of kotline vo vzťahu k teplotným sumám. Meteorolo- Agriculture and Rural Development of the Slovak gický časopis, 11:15–20. Republic, Bratislava and National Forest Centre, Škvareninová, J., 2014: Phenological signs of pedun- Zvolen, 69 p. culate oak (Quercus robur L.) in Slovakia as a bio- indicator of forest ecosystems, weather extremes and climate change. Zprávy lesnického výzkumu, 59:250–255. 201 T. Bucha et al. / Cent. Eur. For. J. 68 (2022) 191–202 Štefančík, I., 1995: Fenológia v lesníctve: 1. Začiatok Zlatník, A., 1976: Přehled skupin typů geobiocénů vegetačnej činnosti. Lesnícky časopis – Forestry původně lesních a křovinných ČSSR. Zpravodaj Journal, 41:131–139. Geograc fi kého ústavu Československé akademie věd, Šmelko, Š., 1991: Štatistické metódy v lesníctve. VŠLD Brno, 13:55–64. Zvolen, skriptá, 276 s. Zhang, X., Friedl, H. A., Schaaf, B. S., Strahler A. H., Tholt, U., 2021: Modelovanie lesníckej fe n o l ó g i e Hodges, J. C. F., Gao, F. et al., 2003: Monitoring veg- s použitím vegetačných indexov spektrorádiometra etation phenology using MODIS. Remote Sensing of MODIS. Diplomová práca. Zvolen, Lesnícka fakulta Environment, 84:471–475. TU vo Zvolene. Internet Sources Townshend, J. R. G., Huang, S. N., Kalluri, V., Defries, SHMÚ, 2021: Bulletin – Meteorológia a klimatológia. R. S., Liang S., 2000: Beware of the per-pixel char- Slovenský hydrometeorologický ústav, 27:66. Availa- acterization of land cover. International Journal of ble at: https://www.shmu.sk/File/ExtraFiles/KMIS/ Remote Sensing, 21:839–843. publikacie/BMaK_0421.pdf Wesołowski, T., Rowiński, P., 2008: Late leaf develop- NFC Zvolen: Available at: http:/ / gis. nlcsk. org/ lgis/ ment in pedunculate oak (Quercus robur): An anti- Bochníček, O., 2015:Available at: https://www.shmu. herbivore defence? Scandinavian Journal of Forest sk/sk/?page=2049&id=592 Research, 23:386–394. Younes, N., Joyce, K. E., Maier, S. W., 2021: All models of satellite-derived phenology are wrong, but some are useful: A case study from northern Australia. International Journal of Applied Earth Observa- tion and Geoinformation, 97:102285. Available at: https://doi.org/10.1016/j.jag.2020.102285

Journal

Forestry Journal – de Gruyter

Published: Dec 1, 2022

Keywords: MOD09; MYD09; NDVI; leaf unfolding; validation

There are no references for this article.