EN中文

Journal of Resources and Ecology >

2019 , Vol. 10 >Issue 5: 481 - 493

DOI: https://doi.org/10.5814/j.issn.1674-764X.2019.05.003

Plant Ecosystem

Greenness Index from Phenocams Performs Well in Linking Climatic Factors and Monitoring Grass Phenology in a Temperate Prairie Ecosystem

  • ZHOU Yuke , *
Expand
  • Key Laboratory of Ecosystem Network Observation and Modeling, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China
* ZHOU Yuke, E-mail:

Received date: 2019-03-19

  Accepted date: 2019-05-13

  Online published: 2019-10-11

Supported by

National Natural Science Foundation of China(41601478)

National Key Research and Development Program of China(2018YFB0505301)

National Key Research and Development Program of China(2016YFC0500103)

Copyright

Copyright reserved © 2019
Fold

Abstract

Near-surface remote sensing (e.g., digital cameras) has played an important role in capturing plant phenological metrics at either a focal or landscape scale. Exploring the relationship of the digital image-based greenness index (e.g., Gcc, green chromatic coordinate) with that derived from satellites is critical for land surface process research. Moreover, our understanding of how well Gcc time series associate with environmental variables at field stations in North American prairies remains limited. This paper investigated the response of grass Gcc to daily environmental factors in 2018, such as soil moisture (temperature), air temperature, and solar radiation. Thereafter, using a derivative-based phenology extraction method, we evaluated the correspondence between key phenological events (mainly including start, end and length of growing season, and date with maximum greenness value) derived from Gcc, MODIS and VIIRS NDVI (EVI) for the period 2015-2018. The results showed that daily Gcc was in good agreement with ground-level environmental variables. Additionally, multivariate regression analysis identified that the grass growth in the study area was mainly affected by soil temperature and solar radiation, but not by air temperature. High frequency Gcc time series can respond immediately to precipitation events. In the same year, the phenological metrics retrieved from digital cameras and multiple satellites are similar, with spring phenology having a larger relative difference. There are distinct divergences between changing rates in the greenup and senescence stages. Gcc also shows a close relationship with growing degree days (GDD) derived from air temperature. This study evaluated the performance of a digital camera for monitoring vegetation phenological metrics and related climatic factors. This research will enable multiscale modeling of plant phenology and grassland resource management of temperate prairie ecosystems.

Key words： vegetation phenology; green chromatic coordinate; vegetation indices; PhenoCam; near-surface remote sensing

Cite this article

ZHOU Yuke . Greenness Index from Phenocams Performs Well in Linking Climatic Factors and Monitoring Grass Phenology in a Temperate Prairie Ecosystem[J]. Journal of Resources and Ecology, 2019 , 10(5) : 481 -493 . DOI: 10.5814/j.issn.1674-764X.2019.05.003

1 Introduction

Vegetation phenology is the timing and duration of recurrent biological events during plant growth and development, such as leaf unfolding, flowering and leaf senescence. Phenology is strongly linked to seasonal changes in ecosystem nutrients and carbon cycles and is considered as an essential biological indicator of climate change. With vegetation phenology as a first-order control of ecosystem productivity, it is critical to accurately model vegetation phenology in a changing climate (Hufkens et al., 2018). In addition, the relationships between plant phenology and climatic variables across various spatial and temporal scales remain uncertain. Field-based human observation of plant phenology has been used for a long time, but it is limited to specific species and focal scales. Currently, satellite remote sensing enables the measurement of plant phenology by using the spectral vegetation greenness index (e.g., NDVI) or productivity (e.g., GPP) at a landscape scale. To date, the ground-based digital camera has emerged as an effective tool for providing human- and satellite-based phenological measurements. Thus, a cross-scalar comparison of different phenological observations and climatic responses is still necessary to reveal the limitations of observations and to validate the performance of models.
Satellite remote sensing platforms have been a popular method to track vegetation growth metrics since the 1980s (Myneni et al., 1997; Nemani et al., 2003). A variety of greenness indices (e.g., NDVI and EVI) and plant canopy characteristics (e.g., LAI) derived from satellite sensors have become the prevailing data sources for depicting vegetation seasonal dynamics. Based on the long-term time series of greenness indices, key phenological events, such as the start and end of the growing season, can be extracted. It should be noted that these phenological events provide the seasonal changing signals of vegetation for the entire land surface at the landscape level, which is known as the mixed pixel effect and it is caused by the coarser resolution of satellite-based imagery. In one pixel, satellite-derived phenology reflects the growth dynamics of multiple canopies and vegetation types. The bare soil also affects the accuracy of landscape-level phenology determination. Although limited by the coarse spatiotemporal resolution, satellite-based phenological products usually span the wide temporal interval that is essential for analyzing the long-term trend and interannual variability of land surface phenology. For example, researchers have applied the third version of the GIMMS NDVI dataset to derive long-term vegetation phenology and its interaction with the biosphere over the past three decades (Atzberger et al., 2013; Piao et al., 2015; Zhao et al., 2015; Garonna et al., 2016; Marshall et al., 2016). In addition, other satellite greenness index products with high spatial resolution, such as MODIS, VIIRS and Sentinel2A data, are commonly used for determining vegetation phenological metrics and their responses to climate change (Zhang et al., 2018). A relatively large number of studies have focused on spring phenology and its driving factors. An earlier spring onset has been identified in the high northern latitudes, and this earlier onset has been mostly attributed to rising temperatures (Myneni et al., 1997; Piao et al., 2011). In the Canadian prairie grasslands, GIMMS NDVI3g-derived phenology has found that drought is a significant factor impacting the phenological timing of both spring and autumn (Cui et al., 2017). Compared with spring phenology, which has been well studied, autumn phenology is still poorly understood. Garonna et al. reported a strong contribution of autumn phenology to changes in satellite-derived growing season lengths across Europe during 1982-2011 (Garonna et al., 2014). Liu Q. et al. assessed the effects of natural forces (e.g., temperature, precipitation, and insolation) on autumn vegetation phenology in temperate regions, and highlighted the importance of temperature on phenological processes in autumn (Liu et al., 2016). In addition to climatic factors, Fu et al. (2018) used spring phenology to model autumn vegetation phenology and found that an earlier spring led to a delayed autumn in northern China. In North America, a delayed autumn was found to play an important role in the extension of the growing season (Zhu et al., 2012). The above findings show that satellite remote sensing metrics have become widely used indicators of vegetation growing seasons at regional and continental scales.
Currently, vegetation metrics derived from high-frequency digital photography provide a linkage between satellite and field observations and have been widely applied for refining phenophase models for various biomes. Compared with remote sensing data, digital repeat photography can provide imagery that is nearly continuous in time, rarely obscured by clouds, and robust to variation in illumination conditions (Sonnentag et al., 2012). The RGB digital images can be used to calculate a suite of greenness indices based on the red, green and blue color channel information stored in each image, which can also be used to track plant phenological stages (Sonnentag et al., 2012; Petach et al., 2014). The most widely used greenness index derived from digital images is the green chromatic coordinate (Gcc), which is calculated from the green channel divided by the sum of the R-G-B channels. Digital repeat photography is also able to obtain a consistent visual assessment of phenology over broad geographic ranges, especially after the construction of the PhenoCam network, which is a continental-scale phenological observatory spanning a wide range of ecozones (Klosterman et al., 2014; Brown et al., 2016). It has been well documented that phenocams can be employed for exploring species- and landscape-level phenological trends. Previous studies using multiple site datasets have demonstrated that digital repeat photography is able to quantify both the duration of the photosynthesis period and the total GPP in deciduous broadleaf forest, grass and crops (Toomey et al., 2015). Other studies have also documented that digital camera derived Gcc data are highly correlated with carbon flux data (Richardson et al., 2010; Richardson et al., 2013). For deciduous forest, Gcc time-series estimates for SOS are found to closely match those derived from visual assessments of leaf-out, as well as satellite-derived SOS (Klosterman et al., 2014). Digital camera greenness index (e.g., Gcc) derived phenological dates were found to be more relevant to field observations than those derived from satellite NDVI across rangeland and forest ecosystems (Browning et al., 2017; St Peter et al., 2018).
Vegetation in grasslands is sensitive to variability in weather and climate. In particular, understanding its distinctive response to local weather conditions is meaningful for improving grassland management. Due to their high frequency and quality, ground-level camera greenness indices enable us to examine the responses of plant greenness to changes in local environmental variables. Multiple natural factor observations have been used in ecological studies and can provide more information about the biosphere. The camera-based greenness index exhibits a significant relationship with the aboveground green biomass and productivity for various types of grasslands (Inoue et al., 2015; Hufkens et al., 2016). At the station scale, near-surface observations based on a digital camera represent a suitable method to bridge the gap between satellite observations and ground-level observations. Automated comprehensive observations for plant status and environmental elements should be linked together. Currently, the majority of in situ ground observation stations have been mounted with automated observation equipment for monitoring climatic factors and carbon-water cycling factors. From the vertical perspective, this equipment covers a large space, from near-surface towers to detectors of deep soil characteristics. In combination with observations from space-borne or aircraft sensors, a full-space ecological observation system was constructed (Fig. 1). Thus, multiple natural factors can be combined to explore the changing mechanisms driving plant phenology at a fine scale. The relationships among satellite vegetation indices, climatic factors, and phenology have been determined across various biomes at broad scales (Zhu et al., 2015; Vicente-Serrano et al., 2013; Buermann et al., 2013; Ganjurjav et al., 2016). However, the extent to which Gcc can be related to multiple climatic variables remains poorly understood.

Fig. 1 Concept of full-space plant monitoring for multiple environmental factors, from site scale to landscape scale.

Although a considerable number of studies have focused on digital photography derived phenology and its relationship with various climatic factors, few of these studies pay close attention to a full-space observational view at the station scale. In other words, a gap exists in the comprehensive assessment of phenocam Gcc and its relationships with various natural forces. This study aims to assess 1) how daily greenness indices derived from Phenocams (ground-based or tower-based digital cameras) are linked to various ground-observed environmental variables; and 2) the performance of this approach in capturing plant phenological metrics compared to satellite imagery time series over a four-year period (2015-2018) in a North American prairie ecosystem.

2 Materials and methods

2.1 Study area

This study focuses on the region of Oakville Prairie Biological Field Station (hereafter, Oakville Station), which is a member station of the PhenoCam network located in North Dakota, USA (Fig. 2). Oakville Station comprises 3.88 km2 of native prairie grassland and has a high water table and alkaline soil (Table 1). It belongs to the North America ecoregion of temperate prairie and has a humid continental climate (Köppen-Geiger climate classification code, Dfb). In addition to the near surface PhenoCam digital camera, Oakville Station has been equipped with soil moisture and temperature sensors, a camera calibration panel, and a weather station. This station was selected because of its observation of multiple natural environmental factors, from deep-soil temperature and water content to satellite-based vegetation indices. Although it is a single station, the manner in which observations are made construct a full space observation system, consisting of underground, near-surface and outer space measurements of rainfall, solar radiation, temperature, and vegetation conditions (Fig. 1). Full-space observation refers not only to multiple sensors monitoring multiple elements of the natural environment, but also to studies which span multiple spatial scales, such as from a local scale to a landscape scale.

Fig. 2 Overview of the Oakville Prairie Biological Field Station (a) and its surrounding land cover (b)

Table 1 Meta data for Oakville Prairie field station
Location Elevation Camera
description		 Camera
orientation		 Vegetation
type		 North America
ecoregion		 Dominant
species
47.8993°N,
97.3161°W		 268 m StarDot
NetCam SC		 North Grassland No. 9
Temperate prairies		 Andropogon gerardii, Distichlis spicata, Muhlenbergia richardsonis

2.2 Data sources

The main data sources used in the study were satellite-based remote sensing vegetation indices, near-surface camera- derived Gcc (green chromatic coordinate), water and temperature records for soil and air. The meta-information for the data is described in Table 2, primarily including the beginning and end dates for the observation records, as well as the spatiotemporal resolution.
Table 2 Description of the datasets
Data name Start date End date Frequency Integrated frequency Spatial scope
Gcc 2015-4-1 2018-9-11 30 min Daily Landscape scale
Modis NDVI(EVI) 2015-1-1 2018-9-11 16 day Semimonthly 2.5 km×2.5 km
VIIRS NDVI(EVI) 2015-1-1
Air Temperature (Oakville) 2018-3-18 2018-9-11 15 min Daily Point
Air Temperature (Grand Forks) 2015-4-1 2018-9-11 Daily Daily Point
Precipitation (Oakville) 2018-3-18 2018-9-11 15 min Daily Point
Precipitation (Grand Forks) 2015-4-1 2018-9-11 Daily Daily Point
GDD (Grand Forks) 2015-4-1 2018-9-11 Daily Daily Point
Soil Water Content 2018-3-18 2018-9-11 15 min Daily Point
Soil Temperature 2" 2018-3-18 2018-9-11 15 min Daily Point
Soil Temperature 4" 2018-3-18 2018-9-11 15 min Daily Point
Soil Temperature 8" 2018-3-18 2018-9-11 15 min Daily Point
Soil Temperature 20" 2018-3-18 2018-9-11 15 min Daily Point
Soil Temperature 40" 2018-3-18 2018-9-11 15 min Daily Point
2.2.1 Near-surface digital camera imagery
Digital camera images were acquired between May 2015 and August 2018 from the PhenoCam data product for Oakville Station with a time frequency of 30 minutes. The digital camera was deployed on a holder at a height of 2 m above the ground and was configured to a fixed white balance and automated exposure using an embedded Linux system. Every 30 minutes during the local time of 4:00 AM-10:00 PM it recorded an image. Then, the images were uploaded to the PhenoCam server through a network connection using the ftp protocol. Samples of the digital images taken from Oakville Station are shown in Fig. 3. The repeated images are compressed in a JPEG format with RGB layers. The Gcc index was retrieved from these RGB pictures and the detailed Gcc computation procedure is described in the Methods section.

Fig. 3 Sample photography and grass growing dynamic at Oakville station

Note: Red polygon represents the ROI (Region of Interesting) used to calculate greenness index.

2.2.2 Satellite vegetation indices
Grass growth activity expressed through the satellite-based greenness indices were used to evaluate and validate the performance of Gcc in this study. Two types of satellite-based vegetation indices (VIs), NDVI (Normalized Difference Vegetation Index), and EVI (Enhanced Vegetation Index), were applied as references. The NDVI, as a proxy of vegetation canopy greenness, is widely used in climate change, vegetation, and process studies. However, it has a saturation effect in monitoring plant conditions in high biomass regions, due to the impacts of the canopy background and atmospheric conditions (Karkauskaite et al., 2017). Therefore, the EVI, which improves the sensitivity to variations in plant density, was also used in this paper. A comparison of different phenology extraction methods also confirmed that the EVI-derived phenological data has a smaller uncertainty than the NDVI-derived data (Klosterman et al., 2014). To reduce the uncertainty associated with using a vegetation index from one platform, NDVI and EVI data obtained from MODIS, VIIRS and Sentinel2 sensors were acquired for Oakville Station from Jan 2015 to Sep 2018, covering a spatial extent of 2.5 km×2.5 km around the center of Oakville Station. The VIIRS data products are designed after the MODIS Terra and Aqua vegetation indices product suite to promote the continuity of the Earth Observation System (EOS) mission. Modis MOD13Q1 and VIIRS VNP13A1 vegetation products were pulled from the MODIS subset web service. The original images of MOD13Q1 and VNP13A1 have a spatiotemporal resolution of 250m-16days and 500m-16days, respectively. The Sentinel-2 mission acquires spectral data globally at 10 to 60 m resolution every five (Sentinel-2B) or ten (Sentinel-2A) days. Therefore, Sentinel2 NDVI outperforms the equivalent NDVI products from Landsat, MODIS and VIIRS. In this study, Sentinel-2A NDVI with 10 m spatial resolution and 10-day temporal frequency around the station was obtained. It is worth investigating how Sentinel2 NDVI performs in delineating plant phenophase. For a better match with the spatial footprint of the site, the averaged VI value for the pixels that passed the quality control in the 2.5 km×2.5 km area was calculated as the VI value for the study field.
2.2.3 Field station climatic data
The weather records, mainly including air temperature, solar radiation, and precipitation, were applied to assess their impacts on grass growth at Oakville Station. These data were observed by the automatic weather station (model: HOBO RX3000 Station - Ethernet) at a frequency of 15 minutes. In addition, the soil temperature at five different depths and soil water contents (soil moisture, 10 cm underground) were acquired to investigate the relationships between underground soil physical properties and grass growth processes (Table 2). The average values of high-frequency environmental factors from one day were calculated to represent daily values, which could match the temporal frequency of the Gcc index. Due to the data sharing policy of Oakville Station, only the weather data that spanned the time interval between Mar 2018 and Sep 2018 (approximately 180 days) were included in the study of the daily relationship test with the vegetation indices.
To match the longer time series data of Gcc and satellite-based vegetation indices, we always obtained daily weather data from the nearest national meteorological station (Grand Forks) as a proxy for the Oakville Station long-term weather records. The Grand Forks national meteorological station is only 1 km from the Oakville station, and they have the same land cover type (i.e., grassland). The Grand Forks weather data source used in the study covered the time span of Jan 2015 - Sep 2018, and included air temperature and precipitation. Based on daily temperature, the growing degree days (GDD) were calculated as a measure of heat accumulation that would influence vegetation growth.

2.3 Data processing

2.3.1 The calculation of Green Chromatic Coordinate (Gcc)
The Gcc is a widely used measure of canopy greenness that is retrieved from the visible spectrum RGB band. Previous studies have reported that Gcc is a robust metric used to characterize seasonal changes in the state of the canopy (Richardson et al., 2018a). The computing process of Gcc, based on digital photography, consisted of several steps. First, after constructing a time series of digital photographs, an appropriate “region of interest” (ROI) is defined on a high-quality template image (Fig. 2a). Second, all images are masked by the ROI to calculate the RGB color channel information for each pixel. The mean Gcc value is computed according to Equation (1), where Gdn, Rdn and Bdn are the mean digital numbers of the R-G-B spectral band for the pixels in the ROI, respectively. In this study, Gcc was calculated for ROI within each image during the day time 10:00 AM - 4:00 PM, in which the camera could have better light conditions for exposure.
${{G}_{cc}}=~\frac{{{G}_{dn}}}{{{R}_{dn}}+{{G}_{dn}}+{{B}_{dn}}}$
Third, to reduce noise and detect seasonal cycles from high-frequency Gcc data, the Gcc time series was smoothed using a three-day moving window method, named the 90th percentile method (Richardson et al., 2018b). The half-hourly Gcc results were aggregated to obtain a daily value. A previous study reported that this approach is generally effective for minimizing day-to-day variation due to weather related scene illumination changes or digital image format (with a 34% lower RMSE between observed Gcc and fitted Loess curves) (Sonnentag et al., 2012).
2.3.2 Time series data smoothing and correlation analysis
The raw materials, or pre-processed climatic or vegetation indices data, were further processed for their time series. The daily time series of weather data, including air temperature, soil temperature, soil water content, GDD, and solar radiation (at both Oakville and Grand Forks stations), were filtered by employing a moving average method with a window width of 9 days. To determine the relationships of the time series between the weather and Gcc data, the daily Gcc values were also smoothed using the same moving average approach. The 16-day Gcc, satellite EVI and NDVI were each smoothed using a moving average method with a window width of three 16-day values to assess their relationships. It should be noted that a specific smoothing and interpolating method was applied to the time series vegetation indices for key phenological event extraction (see details in section 2.3.3).
The correlations between the Gcc and climatic data are characterized in Fig. 5 by drawing the growth curves and examining their Pearson correlation coefficients. For Oakville Station, the influences of natural environment factors on grass growing conditions (represented by Gcc), obtained from above- and below-ground observations, were examined using backward stepwise regression. The final regression factors were selected using the minimum AIC (Akaike Information Criterion) standard.
2.3.3 Vegetation phenology measurement
Measuring plant phenology from ground to space is meaningful for reducing the uncertainty of vegetation phenology analyses. A suite of phenological metrics for Oakville prairie was estimated based on the smoothed and interpolated time series of ground-level and satellite data. The semi- monthly satellite NDVI and EVI time series were smoothed and interpolated to daily data using the double logistic method to fit the grass seasonal trajectory (Beck et al., 2006). The daily Gcc data were also smoothed using the double logistic method. Then, the Beck method handled the outliers effectively and estimated the parameters that are related to phenological events, such as the timing of spring and autumn. The Beck method first estimates the VI during winter (WVI), then models the VI as a double logistic function of time (t) to the growth curve (Equation (2)). Winter VI is calculated as the minimum value of the positive VI within the winter.
$\begin{matrix} & VI\left( t \right)=\left( {{m}_{VI}}-{{W}_{VI}} \right)\left( \frac{1}{\left[ 1+\exp \left( -mS\times \left( t-S \right) \right) \right]}+ \right. \\ & \ \ \ \ \left. \frac{1}{\left[ 1+\exp \left( -mA\times \left( t-A \right) \right) \right]} \right)-1+{{W}_{VI}} \\ \end{matrix}$
where mVI represents the maximum VI during the growing year; S and A are the inflection points representing the start and end of season (SOS, EOS), respectively; and mS and mA are the rates at points S and A, indicating the greenup rate in spring (RSP) and the senescence rate in autumn (RAU), respectively. The key phenological parameters and their descriptions are shown in Fig. 4.

Fig. 4 A sample vegetation growth curve and the key phenological events, derived from NDVI

2.3.4 Relative differences between phenological metrics derived from Gcc and VI
In addition to the correlation analysis between the remote sensing greenness curves and near surface Gcc curves, the relative differences between the grass phenological metrics derived from the VI and Gcc curves were calculated (Equation (3)). P_VIi and P_Gcci denote the ith phenological event derived from the time series of satellite remote sensing VI and Gcc, respectively. Dev refers to the deviation of phenological metrics calculated from the Gcc and VI. Because a few remote sensing VIs were employed in this study, P_VIi was set to the average value of the ith phenological event dates obtained from Modis NDVI (EVI) and VIIRS NDVI (EVI).
$De{{v}_{i}}=~\left| \frac{P\_V{{I}_{i}}-P\_Gc{{c}_{i}}}{P\_V{{I}_{i}}} \right|$
2.3.5 Principal component analysis and stepwise regression
Multiple environmental variables may influence the grass growth condition that is reflected by the near-surface vegetation index. But the extent to which these factors would impact plant growth needs to be accurately estimated. The daily climatic variables of Oakville were integrated into a data matrix, then analyzed using Principal Component Analysis (PCA). PCA is a multivariate statistical technique that uses orthogonal transformation to convert a set of correlated variables into a set of orthogonal, uncorrelated axes called principal components (Legendre et al., 1998; Jolliffe et al., 2011). PCA is the most commonly used technique to identify linear combinations of variables in a high-dimensional space, and would best represent the variance in our Oakville climatic data. This is implemented by considering each variable to be an axis in a high-dimensional space.
To quantitatively determine the relationships of Gcc with various environmental variables, a backward stepwise linear regression analysis was applied. The final explanatory variables were determined by the step with the minimal AIC (Akaike’s Information Criterion) value. Then, the relative importance of the environmental factors with respect to Gcc was calculated using the R package ‘relaimpo’ (Grömping et al., 2006).

3 Results

3.1 Relationships among Gcc and environmental factors

The relationships between Gcc and environmental factors for the year 2018 were examined using the Pearson correlation test (Fig. 5). The growth trajectories of Gcc are similar to that of air temperature, which was indicated by a pronounced Pearson correlation coefficient of 0.86 (P < 0.001). While the Gcc demonstrates a similar response to soil temperature at various depths, the Gcc increased with the increasing soil temperature. However, the soil temperatures in different depths have different influences on Oakville grass according to Pearson’s r values. Thus, the soil temperatures at greater depths have weaker correlations with Gcc (Fig. 5b). Although various soil temperatures demonstrate similar development trajectories, at the same time point, the upper-level soil has a higher temperature than the lower-level soil. In contrast, soil water has a negative feedback to Gcc (i.e., negative Pearson R), as the grass canopy greenness shows a declining trend with soil water (Fig. 5c), especially during the greenness rising transition period. This result may be attributed to the pressure that grass growth puts on water demand, but the soil water cannot obtain effective supplementation from precipitation due to the continental climate in Oakville. The lower correlation between the Gcc and soil water may be caused by the variation in soil water influenced by the abruptness of rainfall. By comparing rainfall (red line in Fig. 5c, 5e) and soil water, we found that soil moisture had a positive response to rainfall. For example, the abrupt increase in soil moisture during June-August occurred simultaneously with the rainfall. The solar radiation time series also exhibited a coherence pattern with the Gcc (Pearson R = 0.6). We found that the Gcc of prairie grass had a slight relationship with air wetness (Fig. 5f, Pearson R = 0.119). Because there were so few rainy days at Oakville Station, we did not conduct a correlation test between the Gcc and precipitation. However, there is little apparent change in the Gcc on rainy days, as reflected by the slight increase at the end of June, and in the middle of August. These phenomena demonstrate the ability of digital camera-based Gcc to instantly capture the grass growth dynamics.

Fig. 5 Relationships between Gcc and climatic factors in 2018

Over a relatively longer temporal period (2015-2018), the Gcc also had a close correlation with air temperature (Pearson R = 0.815, Fig. 6a). During the winter time, it is obvious that the Gcc was maintained at a relatively constant level. The GDD (Growing Degree Days), which is an indicator of the plant growth period, was highly consistent with the Gcc (Pearson R = 0.919). For the period of GDD with zero values, the prairie grass was also in a dormant state, which was reflected by the stagnant canopy greenness. This finding also demonstrates the good performance of the Gcc in characterizing the growth trajectory of prairie grass. The Gcc time series for Oakville Station showed a strong correlation with the MODIS and VIIRS NDVI time series (Fig. 6c-6d). Additionally, the new and high resolution Sentinel2 satellite NDVI for Oakville Station was extracted from the Google Earth Engine cloud computing platform. A grass growth curve of Sentinel2 NDVI for 2016-2018 was constructed and compared with the Gcc index (Fig. 6e). The results show that the Gcc had a high correlation with the Sentinel2 NDVI (Pearson R = 0.81).

Fig. 6 Comparison of Gcc with air temperature (2015-2018) (a), and GDD (b); and Gcc index compared with MODIS VI (c), VIIRS VI (d) and Sentinel2 VI (e).

3.2 Regression analysis of Gcc and natural factors

The PCA analysis was performed to examine the correlations among climatic factors at Oakville Station. The first two standardized PCs and environmental variables are illustrated in Fig. 7. A projection of the variables on the factor plane revealed that the 1st and the 2nd axes of the PCs explained 74.5% and 13.0% of the total variance, respectively. Significant variance could be found among these climatic factors. There were five major directions (groups) in the plots, composed of soil water, solar radiation, wetness, soil temperature, and air temperature. This distribution reflected the fact that these variables have distinguishing representations of the samples. In other words, this plot could help to highlight groups of homogeneous individuals. The plot also reflected that, at different time points during the growing season, different climatic factors had key influences on the grass. For example, in the spring, water resources might be the dominant element affecting on the grass growth. However, solar radiation might be more important for grass activity in the summer. The loads of the climatic factors on the first principle component indicated that soil water and solar radiation played the most important roles among these natural elements that might affect grass growth. Air and soil temperature were not highly significant in the first principle component.

Fig. 7 Projection of the nine environmental variables on the 1st and 2nd principal component axes. Black points refer to the observed records for the variables.

In the backward stepwise regression, air temperature, soil water content, soil temperature, and solar radiation were selected as the independent variables and regressed step-by-step with the Gcc (Equation (4)). The final regression result was obtained when the stepwise regression achieved the minimum AIC value (AIC = -1418.41). The key factors related to the Gcc consisted of solar radiation, soil temperature (4, 8, 20, and 40 inches underground) and soil water content, which could explain 95% of the total variation in the Gcc at the significance level of 0.001. The air temperature, air wetness, and soil temperature at a 2-inch depth were excluded from the multivariate linear model.
$\begin{matrix} & Gcc\text{ }=(25.27\times Sol8.647\times S{{T}_{4}}+8.264\times S{{T}_{8}}+ \\ & \ \ \ \ 0.376\times S{{T}_{20}}-0.368S{{T}_{40}}+61.0\times SW)\times 0.001+0.313 \\ \end{matrix}$
Soil temperature was the most important variable, accounting for 82% (ST4 21%, ST8 20%, ST20 20%, ST40 21%) of the Gcc variation. Solar radiation and soil water contributed 10% and 3%, respectively, to the Gcc variation. The sum of these relative importance values equals a total R2 of 95%. This result indicated that the growth activity of grass at Oakville Station was mainly controlled by soil temperature.

3.3 Grass phenology derived from digital cameras and satellites

Phenology is a valuable diagnostic of ecosystem health, and has applications in environmental monitoring and management. Here, the dates of key phenological events derived from time series of the Gcc and satellite vegetation indices were compared. First, the key phenological metrics during 2015-2018, including not only the key phenological dates, but also the greening ratio in spring and dormant ratio in autumn, were derived from the daily Gcc index using the double logistic method (Table 3). SOS, EOS and LOS presented significant variations in these years. In 2018, the EOS and LOS had more significant differences than the values obtained for other years, which may be caused by the data missing after DOY 250 in 2018. In addition, the quality of digital images that may have been affected by snow or bad weather conditions in spring, could cause SOS differences. The same POP dates were found in 2016 vs. 2018 (167) and in 2015 vs. 2017 (172). The mean growing season Gcc had a multi-year averaged value of 0.416. RSP and RAU in the four years had little variation, indicating that the prairie grass growing rate was not critically affected by the daily data quality. The absolute Gcc values in the maximum point, spring and autumn also showed low levels of variation.
Table 3 Statistical results of stepwise regression coefficients
Coefficients Estimate Std. error t value Pr (>|t|)
Intercept 3.132e-01 7.881e-03 39.798 < 2e-16 ***
Solar 2.527e-04 5.277e-05 4.788 4.28e-06 ***
SoilTemp4 -8.647e-03 3.546e-03 -2.439 0.0160 *
SoilTemp8 8.264e-03 4.011e-03 2.061 0.0412 *
SoilTemp20 3.756e-02 2.804e-03 13.396 < 2e-16 ***
SoilTemp40 -3.677e-02 2.066e-03 -17.801 < 2e-16 ***
Soil Water 6.107e-02 1.167e-02 5.232 6.10e-07 ***

Note: Significance codes: *** means P<0.001, ** means P< 0.01 and * means P< 0.05

Table 4 Phenological metrics derived from daily Gcc. The unit of SOS, EOS, POP is day of year (DOY)
Year SOS EOS LOS POP MGS RSP RAU PEAK MSP MAU
2015 151 284 133 172 0.402 0.003 -0.0014 0.426 0.398 0.350
2016 141 283 142 167 0.412 0.0036 -0.0017 0.438 0.3981 0.352
2017 132 257 125 172 0.411 0.0032 -0.0013 0.427 0.376 0.374
2018 127 210 83 167 0.438 0.005 -0.0013 0.451 0.386 0.420
Mean 137 258 120 169 0.416 0.0037 -0.0014 0.4355 0.389 0.374
Fig. 8 depicts the fitted growing curve and phenological transition dates derived from the daily Gcc time series. From the four growth trajectories, we find that the seasonal metrics for Gcc varied annually, causing variation in the corresponding phenological metrics. For all these years, the range of POP date (DOY 167-172 in June) is narrower than the ranges of SOS and EOS dates.

Fig. 8 Daily Gcc curve fitting (blue line) and extraction for key phenological parameters (SOS, EOS, POP)

For validation, phenological metrics derived from the Gcc and various satellite remote sensing VIs at half-monthly frequencies were compared. Fig. 9 shows the sample results of phenological metrics derived from satellite VIs and Gcc in 2015. In Fig. 9a, although various satellite VIs had different growth amplitudes, POP dates were identified within a short time range. SOS and EOS were distributed in a broader time range. Specifically, the EOS dates from the four VIs had significant differences. This result might be generated by the sensitivity of SOS and EOS to the derivative method from which they were obtained. Because POP is related only to the position of the VI peak value that can be captured with relative precision, the data source and calculation method should not be the key impacting factors.

Fig. 9 Grass growth curve and key phenological parameters derived from satellite VI (a) and Gcc (b) with 16-days frequency, for the year 2015

From the relative difference (Dev) in the years 2015-2017, significantly larger values were found for Dev in RSP and RAU than in other phenological parameters, reflecting the uncertainty of the plant greenup or greendown rates derived from various data sources. In addition, Dev of the plant autumn senescence rate was more obvious, which might be attributed to the complexity of the plant growth process in autumn. Previous studies have reported the uncertainties and complexities of autumn phenological analysis (Wu et al., 2018). The MGS Dev in the three years had only small differences, because MGS is directly calculated from the vegetation index rather than from the curve-fitting derived value, which reduces the variation among these years.

4 Discussion

4.1 Comparision among Gcc, satellite VI and environmental factors

In this study, Gcc was strongly correlated with various natural variables. The Gcc time series showed that grass had a positive response to air temperature, precipitation, and solar radiation, which was in agreement with the plant growth conditions in the humid continental climate zone. The high temporal frequency of the Gcc demonstrated that plants had a rapid response to rainfall events. Here, air temperature was not found to be a dominant explanatory variable for Gcc in the multiple linear regression analysis. This finding may be in agreement with a previous study, which reported that air temperature might not be a direct factor determining GPP variations in the springtime (Piao et al., 2015). Warming air temperatures may trigger the earlier leaf onset of vegetation in spring, but the summer water deficiency induced by subsequently rising temperatures may suppress plant peak growth (Huang et al., 2018). Soil moisture available for plant growth makes up approximately 0.01% of the world's stored water. Soil water content also has a complex interaction with grass growth. Among our findings, soil moisture demonstrated a decreasing trend as the greenness index increased. Soil moisture also showed an apparent fluctuation corresponding to the rainfall events that could effectively supply soil water content. In contrast, soil water resources often stressed the growth of grass. For example, the free soil water would improve plant growth by impacting transpiration (Gardner et al., 1963). In addition, soil temperature played an important role in explaining grass greenness. A previous study revealed the positive influence of soil temperature on improving grass productivity (Xu et al., 2013). That finding was consistent with the results of the stepwise regression in this study, which indicate that soil water plays a key role in controlling plant greenness.
Although only a short timeframe of digital images was used in this study, the findings validate the performance of Gcc for monitoring phenological change at the station scale, as well as its relation to satellite-retrieved greenness indices. The seasonal changes and phenological parameters derived from Gcc were comparable with those derived from MODIS and VIIRS VIs, although there was a larger relative distinction in SOS dates among them for the year 2016‒2017 (Table 5). A recent study also reported that the average absolute difference between the VIIRS EVI2 and PhenoCam greenness index phenological dates was approximately 7-13 days in the greenup and senescence phases (Zhang et al., 2018). The uncertainty of satellite-derived SOS may be caused by several factors, such as the coarse temporal resolution, the snow cover in spring, and the sensitivity of the derivatives method to the rapid growing up process. A similar study of phenology monitoring at oak/grass ecological sites also confirmed the strong correlation between greenness indices from Phenocams and various satellites (Liu et al., 2017).
Table 5 Phenological metrics derived from various vegetation indices (with a temporal frequency of 16 days). The unit of SOS, EOS, POP is day of year (DOY), and the unit of LOS is Days.
Year VI SOS EOS LOS POP MGS RSP RAU Peak MSP MAU
2015 Gcc 149 271 122 191 0.420 0.003 -0.001 0.434 0.388 0.378
MODNDVI 143 243 100 194 0.661 0.007 -0.004 0.732 0.499 0.584
MODEVI 151 261 110 190 0.444 0.010 -0.004 0.504 0.324 0.332
VIIRSNDVI 137 250 113 194 0.690 0.006 -0.004 0.761 0.541 0.599
VIIRSEVI 142 278 136 192 0.445 0.009 -0.005 0.495 0.323 0.317
Dev 0.042 0.05 0.07 0.005 0.250 0.625 0.760 0.30 0.08 0.170
2016 Gcc 152 340 188 200 0.406 0.003 -0.001 0.447 0.393 0.348
MODNDVI 108 338 230 221 0.562 0.005 -0.007 0.721 0.372 0.209
MODEVI 133 312 179 221 0.382 0.003 -0.004 0.462 0.274 0.22
VIIRSNDVI 102 329 227 214 0.628 0.005 -0.007 0.79 0.419 0.284
VIIRSEVI 146 265 119 205 0.451 0.004 -0.004 0.51 0.350 0.354
Dev 0.25 0.09 0.005 0.07 0.20 0.29 0.82 0.28 0.11 0.30
2017 Gcc 176 295 119 213 0.423 0.003 -0.001 0.438 0.393 0.391
MODNDVI 128 296 168 214 0.620 0.005 -0.007 0.745 0.444 0.368
MODEVI 142 279 137 210 0.408 0.004 -0.005 0.481 0.301 0.27
VIIRSNDVI 120 278 158 199 0.670 0.006 -0.005 0.777 0.456 0.525
VIIRSEVI 145 256 111 200 0.516 0.006 -0.005 0.591 0.379 0.405
Dev 0.32 0.06 0.17 0.04 0.24 0.43 0.82 0.32 0.05 0.003

4.2 Limitations

The major limitation of this research is that it is a case study focusing on only one ecological observation station. Although the environmental factors were observed at different spatial scales (stand plot to landscape scale) and could build a long-term time series in the future, the small volume of data used provides only limited potential information for plant growth monitoring. Phenological studies require long-term observations of vegetation status across plant species and various spatiotemporal scales (Sonnentag et al., 2012). With the further development of the PhenoCam network, more stations from a variety of ecological observation networks will be equipped with phenology cameras. Thus, a greater diversity of environmental variables and a larger number of repeat photographs from global observation sites will remarkably improve the study of phenology changes and ecosystem carbon cycling, due to their high temporal resolutions and fine spatial scales. In future work, more remote sensing images (e.g., Phenocams, Modis, Landsat8, and Sentinel2) and longer-term observation records will be collected to assess the impacts of natural factors on the phenological transition dates across various vegetation types.

5 Conclusions

This study investigated the capabilities of a digital camera-based vegetation index for monitoring vegetation growth dynamics, and its relationship with above-ground and below-ground observations of natural elements at a field station scale. In addition, vegetation phenological metrics derived from digital images and satellite data were examined. The conclusions drawn are as follows:
(1) In the North American temperate prairie ecosystem, Gcc retrieved from repeated digital photography indicated a strong relationship with several high-frequency environmental variables for air and soil conditions. Soil temperature demonstrated a marked impact on the daily Gcc variability. Although rainfall was not chosen in the regression analysis, the Gcc growth profile had a rapid response to rainfall events. From the long-term perspective, Gcc also had a significant positive correlation with air temperature derived from a national weather station neighboring the ecological station.
(2) Compared with the satellite greenness index, digital image-based Gcc could closely capture the intra-annual variation in grassland growth. Multi-year Gcc also presented a prominent correlation with MODIS and VIIRS vegetation indices. Regarding their performance for extracting phenological metrics, these greenness indices could consistently derive the transition dates of POP, EOS and LOS, with a larger relative difference in the SOS dates. Another phenophase parameter, GDD, derived from long-term daily air temperature, was also highly consistent with the Gcc time series during 2015-2018.

Acknowledgments

The producers of the PhenoCam dataset and processing tools are kindly acknowledged. I thank researchers at Oakville Station for their work on observations and freely published data. Dr. Li Renqiang and Xu Wenxin are acknowledged for helping me to improve the content of this article.

References

Publishing order | Descend order by publishing year | Descend order by cited within
[1]
Atzberger C, Klisch A, Mattiuzzi M , et al. 2013. Phenological metrics derived over the European continent from NDVI3g data and MODIS time series. Remote Sensing, 6(1):257-284.

[2]
Beck P S, Atzberger C, Høgda K A , et al. 2006. Improved monitoring of vegetation dynamics at very high latitudes: A new method using MODIS NDVI. Remote Sensing of Environment, 100(3):321-334.

[3]
Brown T B, Hultine K R, Steltzer H , et al. 2016. Using phenocams to monitor our changing earth: Toward a global phenocam network. Frontiers in Ecology and the Environment, 14(2):84-93.

[4]
Browning D M, Karl J W, Morin D , et al. 2017. Phenocams bridge the gap between field and satellite observations in an arid grassland ecosystem. Remote Sensing, 9(10):1071.

[5]
Buermann W, Bikash P R, Jung M , et al. 2013. Earlier springs decrease peak summer productivity in north American boreal forests. Environmental Research Letter, 8(2):024-027.

[6]
Cui T, Martz L, Guo X . 2017. Grassland phenology response to drought in the Canadian prairies. Remote Sensing, 9(12):1258.

[7]
Fu Y, He H S, Zhao J , et al. 2018. Climate and spring phenology effects on autumn phenology in the Greater Khingan Mountains, northeastern China. Remote Sensing, 10(3):449.

[8]
Ganjurjav H, Gao Q, Schwartz M W , et al. 2016. Complex responses of spring vegetation growth to climate in a moisture-limited alpine meadow. Scientific Reports, 6:23356.

[9]
Gardner W, Ehlig C . 1963. The influence of soil water on transpiration by plants. Journal of Geophysical Research, 68(20):5719-5724.

[10]
Garonna I, De Jong R, De Wit A J , et al. 2014. Strong contribution of autumn phenology to changes in satellite-derived growing season length estimates across Europe (1982-2011). Global Change Biology, 20(11):3457-3470.

[11]
Garonna I, De Jong R, Schaepman M E . 2016. Variability and evolution of global land surface phenology over the past three decades (1982-2012). Global Change Biology, 22(4):1456-1468.

[12]
Grömping U . 2006. Relative importance for linear regression in R: The package relaimpo. Journal of Statistical Software, 17(1):1-27.

[13]
Huang K, Xia J, Wang Y , et al. 2018. Enhanced peak growth of global vegetation and its key mechanisms. Nature Ecology & Evolution, 2(12):1897.

[14]
Hufkens K, Basler D, Milliman T , et al. 2018. An integrated phenology modelling framework in R. Methods in Ecology and Evolution, 9(5):1276-1285.

[15]
Hufkens K, Keenan T F, Flanagan L B , et al. 2016. Productivity of north american grasslands is increased under future climate scenarios despite rising aridity. Nature Climate Change, 6(7):710.

[16]
Inoue T, Nagai S, Kobayashi H , et al. 2015. Utilization of ground-based digital photography for the evaluation of seasonal changes in the aboveground green biomass and foliage phenology in a grassland ecosystem. Ecological Informatics, 25:1-9.

[17]
Jolliffe I. . 2017. Principal component analysis. In: International Encyclopedia of Statistical Science, Springer:1094-1096.

[18]
Karkauskaite P, Tagesson T, Fensholt R . 2017. Evaluation of the plant phenology index (ppi), NDVI and EVI for start-of-season trend analysis of the northern hemisphere boreal zone. Remote Sensing, 9(5):485.

[19]
Klosterman S, Hufkens K, Gray J , et al. 2014. Evaluating remote sensing of deciduous forest phenology at multiple spatial scales using phenocam imagery. Biogeosciences, 11(2):4305-4320.

[20]
Legendre P, Legendre L . 1998. Numerical Ecology. Second English edition. Elsevier, Amsterdam .

[21]
Liu Q, Fu Y H, Zeng Z , et al. 2016. Temperature, precipitation, and insolation effects on autumn vegetation phenology in temperate China. Global Change Biology, 22(2):644-655.

[22]
Liu Y, Hill M J, Zhang X , et al. 2017. Using data from landsat, modis, viirs and phenocams to monitor the phenology of California oak/grass savanna and open grassland across spatial scales. Agricultural and Forest Meteorology, 237:311-325.

[23]
Marshall M, Okuto E, Kang Y , et al. 2016. Global assessment of vegetation index and phenology lab (vip) and global inventory modeling and mapping studies (gimms) version 3 products. Biogeosciences, 13(3):625-639.

[24]
Myneni R B, Keeling C, Tucker C J , et al. 1997. Increased plant growth in the northern high latitudes from 1981 to 1991. Nature, 386(6626):698.

[25]
Nemani R R, Keeling C D, Hashimoto H , et al. 2003. Climate-driven increases in global terrestrial net primary production from 1982 to 1999. Science, 300(5625):1560-1563.

[26]
Petach A R, Toomey M, Aubrecht D M , et al. 2014. Monitoring vegetation phenology using an infrared-enabled security camera. Agricultural and forest meteorology, 195:143-151.

[27]
Piao S, Cui M, Chen A , et al. 2011. Altitude and temperature dependence of change in the spring vegetation green-up date from 1982 to 2006 in the Qinghai-Xizang Plateau. Agricultural and Forest Meteorology, 151(12):1599-1608.

[28]
Piao S, Tan J, Chen A , et al. 2015. Leaf onset in the northern hemisphere triggered by daytime temperature. Nature Communications, 6:6911.

[29]
Richardson A D, Black T A, Ciais P , et al. 2010. Influence of spring and autumn phenological transitions on forest ecosystem productivity. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 365(1555):3227-3246.

[30]
Richardson A D, Hufkens K, Milliman T , et al. 2018a. Tracking vegetation phenology across diverse north American biomes using phenocam imagery. Scientific Data, 5:180028.

[31]
Richardson A D, Hufkens K, Milliman T , et al. 2018b. Intercomparison of phenological transition dates derived from the phenocam dataset v1. 0 and MODIS satellite remote sensing. Scientific Reports, 8(1):5679.

[32]
Richardson A D, Keenan T F, Migliavacca M , et al. 2013. Climate change, phenology, and phenological control of vegetation feedbacks to the climate system. Agricultural and Forest Meteorology, 169:156-173.

[33]
Sonnentag O, Hufkens K, Teshera-Sterne C , et al. 2012. Digital repeat photography for phenological research in forest ecosystems. Agricultural and Forest Meteorology, 152:159-177.

[34]
St Peter J, Hogland J, Hebblewhite M , et al. 2018. Linking phenological indices from digital cameras in Idaho and Montana to MODIS NDVI. Remote Sensing, 10(10):1612.

[35]
Toomey M, Friedl M A, Frolking S , et al. 2015. Greenness indices from digital cameras predict the timing and seasonal dynamics of canopy-scale photosynthesis. Ecological Applications, 25(1):99-115.

[36]
Vicente-Serrano S M, Gouveia C, Camarero J J , et al. 2013. Response of vegetation to drought time-scales across global land biomes. Proceedings of the National Academy of Sciences , 110(1):52-57.

[37]
Wu C, Wang X, Wang H , et al. 2018. Contrasting responses of autumn-leaf senescence to daytime and night-time warming. Nature Climate Change, 8(12):1092.

[38]
Xu M . 2013. A research on summer vegetation characteristics & short-time responses to experimental warming of alpine meadow in the Qinghai-tibetan plateau. Acta Ecologica Sinica, 33:2071-2083.(in Chinese)

[39]
Zhang X, Jayavelu S, Liu L , et al. 2018. Evaluation of land surface phenology from viirs data using time series of phenocam imagery. Agricultural and Forest Meteorology, 256:137-149.

[40]
Zhao J, Zhang H, Zhang Z , et al. 2015. Spatial and temporal changes in vegetation phenology at middle and high latitudes of the northern hemi- sphere over the past three decades. Remote Sensing, 7(8):10973-10995.

[41]
Zhu L, Meng J . 2015. Determining the relative importance of climatic drivers on spring phenology in grassland ecosystems of semi-arid areas. International Journal of Biometeorology, 59(2):237-248.

[42]
Zhu W, Tian H, Xu X , et al. 2012. Extension of the growing season due to delayed autumn over mid and high latitudes in North America during 1982-2006. Global Ecology and Biogeography, 21(2):260-271.

Options
Outlines

/

〈 〉