1. Introduction
Plant phenology is the study of recurring life cycle events, such as growth reactivation and dormancy, leaf emergence and senescence, and flowering. Phenology is considered as a sensitive bio-indicator of climate and its changes [
1,
2,
3]. Over the past three decades, phenology studies have been conducted using field-based observations, bioclimatic models, near-surface remote sensing and satellite remote sensing based techniques [
4,
5,
6,
7]. Researchers have applied different approaches depending on the question addressed and the spatial scale of investigation. Satellite data provide wide coverage with varying temporal, spectral and spatial resolutions. Remote sensing is therefore a flexible, reliable and widely recognized tool to study phenology in a number of ecosystems, from forests to grasslands, which may be too difficult using other data collection methods [
8,
9,
10]. Current sensors on board satellite platforms record spectral signals that can be used to monitor seasonal and interannual variations in vegetation cover and determine the timings of phenological events and the growing season across the landscape [
11,
12].
Imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS) on board the Terra and Aqua satellites is widely used to study vegetation phenology [
4,
13]. Among the available tools, the Normalized Difference Vegetation Index (NDVI) [
14,
15] is a reliable spectral index for the reconstruction of phenological transitions of different vegetation types, including croplands, forests and grasslands, because it is related to the amount of green-leaf biomass [
13,
16,
17]. NDVI is considered as a potential remote sensing proxy to investigate the effect of climate on vegetation phenology at regional to continental scales due to its close relationship with plant activity [
18,
19].
In the last two decades, substantial advances have been made in predicting length and start/end dates of the growing season from both field observations and satellite remote sensing data [
20,
21,
22]. However, the scarcity of ground observations in some locations is a challenge for studies of growth patterns over large or remote regions of the world [
23,
24]. Another constraint exists in stands of evergreen species, which show only small seasonal variations in canopy greenness compared to deciduous species [
21,
25,
26]. Phenology and growing season in North America have been studied extensively using the NDVI at medium (500 m) and coarse (1 and 8 km) spatial resolutions [
22,
27,
28]. However, direct observations in the field confirming the results from remote sensing in remote areas remain scarce.
To address these issues, we propose an innovative statistical approach to calibrate MODIS NDVI time series and demonstrate how it can be used to describe annual changes in black spruce [Picea mariana (Mill.) bud phenology across a large portion of the coniferous forest in a study area in Quebec, Canada. This approach provides a robust alternative to field-based monitoring, especially in areas where field observations are difficult due to the remoteness or extreme weather conditions.
4. Discussion
Time series of satellite-derived vegetation indices are a reliable tool to describe the patterns of bud break and bud set dates across large and remote regions. In this study, we used field observations of all black spruce phenology phases to calibrate time series of MODIS NDVI, which we used to spatialize bud break and bud set dates across the black spruce stands in Quebec, Canada. The first phase of bud break (B1), occurring around mid-May, corresponded to a standardized NDVI of 80.5%. NDVI culminated at the end of July when the process of bud set was occurring. Winter bud (i.e., the last phase of bud set) was reached when standardized NDVI reached 92.2% of its maximum amplitude at the beginning of September. Our study provided a new statistical approach based on ordinal logit modelling to statistically connect the sequential events of phenological phases with fortnightly NDVI data.
Remotely sensed phenology of the boreal forest has been performed across North America or in combination with the entire northern hemisphere [
9,
19,
22,
39]. Wide area coverage provided by satellite remote sensing offers a unique point of view at synoptic scale but lacks in fine spatial details. To our knowledge, species–specific calibrations of the major parameters of boreal forest bud phenology based on satellite imagery are still missing. Fu et al. [
40] calibrated MODIS vegetation phenology at global scales (i.e., Global Vegetation Phenology) using only remote sensing phenological observations from 2001 to 2010, observing the summer culmination in NDVI and late beginning of the growing season (in mid-April to mid-May) in the boreal and cool regions of North American evergreen coniferous forests. White et al. [
27] calibrated NDVI at 8-km resolution using ground phenology records from 1982 to 2006 for North American forests but they explicitly removed evergreen species from their analysis. Our results improve the phenological description by calibrating MODIS NDVI time series on field observations of bud phenology at higher spatial resolution (250 m), enhancing the reliability of the estimations at the stand level for two important phenological phases, the beginning and end of bud activity.
Other studies have used satellite-based retrievals of solar-induced fluorescence (SIF) to study phenological changes in evergreen canopies [
41,
42]. However, apart from the spatial resolution of the current satellites used for SIF (from ~3 km
2 (OCO-2) to >1500 km
2 (GOME-2)), the main difference between the approach presented here and SIF-based phenology resides is their respective definitions of phenological events. For example, SIF, which is mostly driven by plant fluorescence yield and by the amount of incoming photosynthetically active radiation (PAR), is an indicator of plant carbon uptake (gross primary production, photosynthetic light-use efficiency). Consequently, SIF- and other phenological indicators based on plant physiology [
43,
44] define, for example, the start of the growing season (SOS) as the beginning of photosynthetic activity (e.g., date of spring recovery in evergreen trees). In contrast, with our approach, SOS is defined as the bud break date which, from a spectral perspective, will result in visible changes in canopy color, namely greenness. Seasonally, photosynthetic activity and changes in canopy greenness and structure are not well synchronized, either in the spring [
45,
46,
47] or in the fall and as a result, this may lead to differences in estimates of the growing season length between physiology-based and structure-based indicators of plant phenology [
42].
Compared with autumn, we observed higher slopes of the double-logistic function in spring. The estimated winter NDVI (coefficient
min of the function) was able to reduce the influence of snowmelt during winter, thus representing an effective greening up of evergreen trees during spring [
20,
48]. In addition, bud break in black spruce starts at the end of May when snowmelt is completed [
30]. These results are in accordance with a previous study that used NDVI time series from satellite remote sensing to investigate xylem growth and timing of plant phenology across the boreal forests of Quebec [
49]. In this study, we demonstrated the existence of a relationship between black spruce bud phenology phases at a wide geographical scale and a remote sensing-derived vegetation index (NDVI). Our results show that establishing a link between bud phenology observations and NDVI time series provides a comprehensive view of boreal forest activity patterns and trends in remote areas where direct observations or recurrent samplings are unachievable.
Methods to detect the beginning and end of the growing season from NDVI are already available in the literature [
13,
50]. In optical remote sensing, the variability in spectral signatures is considered as a noise disturbance. In general, these noises are related to varying atmospheric conditions, the presence of snow on coniferous canopies and the spatial and spectral variability of understorey vegetation [
51,
52,
53]. Phenological events may therefore be harder to estimate for boreal evergreen coniferous forests, which also exhibit lower seasonal changes in foliage biomass and thus, in vegetation indices such as the NDVI [
4,
20].
In this study, the double-logistic model was able to suitably describe the seasonal pattern of NDVI, despite the wide spectral variability within and between forest stands. It is well known that such a function can effectively reduce the noise in remote sensing time series, thus adequately describing the timings and duration of the growing season [
53,
54]. In previous studies, the inflection points of double logistic models fitted to NDVI data more accurately estimated spring than autumnal events in boreal species [
20,
22]. This was attributed to the longer and slower decrease in canopy greenness occurring in autumn [
4,
55,
56]. In this study, we were able to adequately describe and detect the inflection point in autumn using the double-logistic function that represents the ending phase of bud set. Over the last two decades, NDVI has been the most commonly used vegetation index for remote sensing-based modelling of phenology [
13,
17,
27,
56]. Our method has great potential to be applied to a wider range of species worldwide, after species–specific calibration. By providing a novel and reliable statistical approach, this study could allow the improvement of the performance of phenological models for evergreen forests to the same levels as has been achieved for grasslands and deciduous forests [
57,
58].
4.1. Relationship between NDVI and Field-Based Phenology
Our study assesses the explicit link between satellite imagery and phenological events recorded in the field [
13,
59]. We observed a quick increase in NDVI during the bud break process from mid-May to the end of June and a slow decrease in NDVI during the bud set phases from July to October. The punctual phases of bud break were identified along the continuous process of gradual variation in greenness. The variability in NDVI at bud set was lower than that at bud break because NDVI was close to its maximum when bud set was already occurring. By comparing bud phenological phases with NDVI curves, the peak standardized NDVI occurred at S2. It was observed four weeks and one after the completion of bud burst and the start of bud set, respectively. This increase in NDVI could be related to changes in needle biomass resulting from a gradual increase during bud burst phases and shoot elongation during spring and summer. Furthermore, leaf expansion, i.e., increment in leaf length and width during spring, is relatively rapid due to the transition from dormancy to active photosynthesis [
60,
61,
62].
The decrease in NDVI during late summer and autumn may be associated with needle aging and declining pigmentation [
63]. Moreover, in evergreen species, some of the old needles, commonly those with ages varying from five to seven years, change color at the end of summer and fall in autumn. Given that MODIS vegetation indices (NDVI, EVI) are primarily sensitive to changes in leaf chlorophyll content and structure [
18,
19,
64], NDVI may be more representative of phenological changes in evergreen species during the first part of the growing season (from mid-spring to summer) than in the autumn [
4,
22]. This could be attributed to the reduction in photosynthetic efficiency and the lower carbohydrate demand during dormancy [
45], which results in changes in light absorbance and reflectance. For evergreen trees, detection of these changes is difficult as they usually require specific and very narrow wavebands, still lacking in most current orbiting sensors [
48,
65]. However, in this study, the calibration of NDVI time series with phenological observations and the resulting threshold values alleviates some of the difficulties encountered for evergreen species when using other approaches. This calibration is specific to black spruce stands because different thresholds of NDVI could be estimated for other tree species of the boreal forest, mainly for deciduous species. We expect this model to be reliable for closed stands strongly dominated by black spruce. Overall, our results suggest that satellite NDVI may be used as a large-scale complement to field observations of bud phenology, and therefore, support the general applicability of satellite-based vegetation indices to estimate both the beginning and end of the growing season in boreal evergreen forests, as often defined by bud break and bud set dates.
4.2. Spatial and Temporal Changes of Bud Phenological Phases
The beginning of bud break (phase B1) appeared later in the northeastern regions, which are also closer to sea, resulting in a shorter duration of the overall bud break process compared with bud set. On average, bud break lasted 51 days across the study area, which covered approximately 5° in latitude. This trend confirms results from other studies; however, we extended our analysis to 5000 sites compared to a previous study that analyzed altitudinal and latitudinal gradients of xylem phenology at only six sites black spruce sites [
49]. The timings of bud break were observed from the beginning of May to the end of June and are related to a number of factors, including the fulfilment of the requirement in winter chilling, the lengthening of photoperiod and warming in spring temperature [
66,
67]. In contrast, bud set lasted longer, 87 days, and lacked a clear and constant spatial pattern. Bud set is an important phenological event and explains most of the variation in tree growth [
68,
69], while the bud set in black spruce is a typical photoperiod-dependent process [
70]. In conifers, the cessation of shoot elongation and development of terminal buds indicates vegetative maturity. It therefore seems to result from exposure to the shortening days of late summer. During this time, the temperature remains warm so it could be a response to some endogenous signals triggered by the shortened photoperiod [
66,
71]. Overall, satellite remote sensing offers the possibility of modeling phenology by recording spectral information at regular time intervals due to their wider coverage area and high temporal resolution [
22,
72]. Our model utilized the potential of time series satellite data to provide the spatio-temporal patterns of bud phenology by calibrating NDVI across the Quebec region of Canada from 2009 to 2018. We expect that the presented calibration approach could be tested on a wider basis for other sites and tree species.