Monitoring vegetation phenology using an infrared-enabled security camera

1 Sensor-based monitoring of vegetation phenology is being widely used to quantify 2 phenological responses to climate variability and change. Digital repeat photography, in 3 particular, can characterize the seasonality of canopy greenness. However, these data cannot be 4 directly compared to satellite vegetation indices (e.g., NDVI, the normalized difference 5 vegetation index) that require information about vegetation properties at near-infrared (NIR) 6 wavelengths. Here, we develop a new method, using an inexpensive, NIR-enabled camera 7 originally designed for security monitoring, to calculate a “camera NDVI” from sequential 8 visible and visible+NIR photographs. We use a lab experiment for proof-of-concept, and then 9 test the method using a year of data from an ongoing field campaign in a mixed temperate forest. 10 Our analysis shows that the seasonal cycle of camera NDVI is almost identical to that of NDVI 11 measured using narrow-band radiometric instruments, or as observed from space by the MODIS 12 platform. This camera NDVI thus provides different information about the state of the canopy 13 than can be obtained using only visible-wavelength imagery. In addition to phenological 14 monitoring, our method should be useful for a variety of applications, including continuous 15 monitoring of plant stress and quantifying vegetation responses to manipulative treatments in 16 large field experiments. 17

phenology is important because it mediates many of the feedbacks between terrestrial vegetation 4 and the climate system (Richardson et al., 2013a).From an ecological perspective, phenology 5 plays an important role in both competitive interactions and trophic dynamics, as well as in 6 reproductive biology, primary production, and nutrient cycling (Morisette et al., 2009).7 Satellite remote sensing can provide global coverage of vegetation phenology, but suffers 8 from tradeoffs between spatial and temporal resolution (Zhang et al., 2003;White et al., 2009).9 Thus, over the last decade, there has been great enthusiasm for increased on-the-ground 10 monitoring of phenology (Betancourt et al., 2005;Morisette et al., 2009;Polgar and Primack, 11 2011).The general objective of these efforts is to better understand spatial and temporal 12 variation in phenology, and how this variability is driven by environmental factors such as 13 temperature, precipitation, and photoperiod (or insolation).Citizen science networks, such as the 14 USA National Phenology Network (http://www.usanpn.org)and Project Budburst 15 (http://budburst.org),are playing an important role in this monitoring, by engaging large 16 numbers of motivated volunteers and establishing standardized protocols.17 Instrument-based approaches (Richardson et al., 2013b) provide a compelling alternative 18 to observer-based phenology, because of the potential for high frequency, automated data 19 collection in a manner that is scalable for regional or continental monitoring.In this context, 20 digital repeat photography (e.g.Richardson et al., 2007Richardson et al., , 2009; Sonnentag et al., 2012) is an 21 attractive option because images can be analyzed either qualitatively or quantitatively, and 22 analysis can focus on individual organisms or integrate across the field of view to obtain a 23 community-or canopy-level perspective.Compared to data collected by a human observer, 24 which tend to focus on discrete phenophases, such as flowering or budburst, the entire seasonal 25 trajectory of canopy greenness can be characterized from digital camera imagery.Additionally, 26 the archived images provide a permanent visual record that can be reanalyzed as new tools and 27 questions are developed.Camera-based monitoring (e.g. the PhenoCam network, To date, most camera-based monitoring of vegetation phenology has been conducted 1 using standard, consumer-grade digital cameras (e.g.Sonnentag et al., 2012).These typically 2 record a three-layer image (red, green and blue: RGB), which is sufficient for the representation 3 of colors in the visible spectrum (VIS, λ = 400-700 nm) as perceived by the human eye.For 4 quantitative analysis, the average value of each color layer for all pixels within a user-defined 5 region of interest (ROI) is extracted from each image to yield a digital number triplet (R DN , G DN , 6 B DN ).Then seasonal variation in the state of the canopy is characterized by the use of several 7 color indices, such as the green chromatic coordinate (g CC, , Eq. 1a) and excess green (G ex , Eq. 8 1b) (Sonnentag et al., 2012;Richardson et al., 2013b): Eq. 1a 10 Conversely, satellite remote sensing of vegetation has traditionally used both visible and 12 near-infrared (NIR, λ = 700-1400 nm) wavelengths.The reason for this is that healthy vegetation 13 can be distinguished from other land cover types by its unique spectral signature, which 14 combines low reflectance in the VIS with high reflectance in the NIR.Thus, the camera indices 15 presented above, which are based only VIS wavelengths, are not directly comparable to standard 16 satellite vegetation indices such as NDVI (Normalized Difference Vegetation Index, Eq. 1c), 17 calculated from red band and NIR band reflectances (ρ R and ρ NIR , respectively) (Tucker, 1979).18 Intriguingly, the CCD (charge-coupled device) or CMOS (complementary metal-oxide-20 semiconductor) imaging sensors used in most digital cameras are sensitive to wavelengths in the 21 NIR portion of the spectrum.An infrared cut filter is typically used to block these wavelengths 22 from reaching the imaging sensor, as they are beyond the spectral range to which the human eye 23 is sensitive and are thus not necessary for conventional color photography.Customized cameras 24 have been used in the past to leverage this NIR sensitivity (Shibayama et al., 2009(Shibayama et al., , 2011;;25 Sakamoto et al., 201025 Sakamoto et al., , 2012;;Nijland et al., 2013).For example, using a two-camera system 26 Sakamoto et al. ( 2012) calculated an NDVI-style index that was more akin to the conventional 27 NDVI than either g CC or G EX .The two-camera approach allows for simultaneous recording of 28 information about the VIS and NIR properties of vegetation, but creates challenges related to 29 camera alignment, cross-calibration, and synchronization of image capture.Very recently, 30 relatively low-cost NDVI cameras have become available (e.g.MaxMax, Event-38, and Regent brands), but these have not been produced with long-term monitoring in mind, and such cameras 1 are unable to also produce conventional RGB imagery-that is, infrared wavelengths are 2 recorded at the expense of information in one of the RGB channels.3 Here, we show that a commercially-available, network-enabled camera ("webcam") with 4 a software-controlled infrared cut filter overcomes the above limitations.With the cut filter in 5 place, standard 3-layer RGB imagery is recorded; with the filter removed, a monochrome 6 RGB+NIR image is obtained.We develop a method to compute an NDVI-style vegetation index, 7 which we call "camera NDVI", from this imagery.A lab experiment, conducted under controlled 8 conditions, is used as a proof-of-concept.We then apply the method to a one-year archive of 9 images from the Harvard Forest to demonstrate the feasibility of employ this method for field 10 monitoring of vegetation phenology, where day-to-day variation in weather and lighting cause 11 additional challenges.As a final test, we compare the seasonality of camera NDVI from the 12 Harvard Forest data with that obtained using co-located narrow-band radiometric instruments 13 and from satellite sensors.Data from our camera system will be of value for quality assessment 14 of phenology products derived from satellite imagery (e.g.White et al., 2009).15 16

Camera 18
We used a NetCam SC IR (StarDot Technologies, Buena Park, CA) camera, featuring a 19 Micron ½" CMOS active-pixel digital imaging sensor and configured for 1.3 megapixel (1296 x 20 976) output.The camera was set at manual (fixed) white balance and, unless otherwise noted, 21 automatic exposure.With a built-in uClinux operating system, the camera operates as a with a wide range of spectral signatures (Figure 1).Each sample was illuminated from above 1 with a 50 W Halogen lamp designed for indoor diffuse reflectance measurements (ASD 2 ProLamp, Analytical Spectral Devices Inc., Boulder, CO).The StarDot camera was mounted on 3 a tripod to the side of the sample and inclined downward at an angle of about 45°.Each sample 4 filled approximately one-quarter of the camera's field of view.For quality assurance, we 5 included a multi-color reference panel in each image, made by painting red, green, blue, white 6 and grey strips on a flat piece of plastic.We recorded four images of each sample: one image at 7 fixed exposure (1/300 s) for both color RGB and monochrome RGB+NIR images, and one 8 image at automatic exposure for both color RGB and monochrome RGB+NIR.Automatic 9 exposure values were determined by the camera.The mean automatic exposure for the color 10 RGB images was 1/30 s (minimum 1/120 s), compared with 1/200 s (minimum 1/350 s) for the 11 monochrome RGB+NIR images.Thus the fixed exposure images were almost always under-12 exposed compared to the automatic exposure images.13 We measured the reflectance spectrum (λ = 350-2500 nm) of each sample using a 14 spectroradiometer (ASD FieldSpec 3, Analytical Spectral Devices Inc.) connected to a 5 cm (2 15 inch) three-port integrating sphere (SphereOptics, Concord, NH) and a 10 W hemispheric 16 collimated light source with a 6 V regulated power supply.White Spectralon discs were used as 17 reference standards.The sphere featured an 8° near-normal incidence port, such that reflectance 18 measurements include both diffuse and specular components.We processed the raw data to 19 reflectances (1 nm increment) using ViewSpecPro software (Analytical Spectral Devices Inc.).20 Spectroradiometer NDVI was calculated using red and NIR band reflectances; specific 21 wavelength ranges are reported below.22 Most of the samples (42) for this experiment were leaves picked from a selection of trees 23 and shrubs native to New England, representing a wide range of leaf health and corresponding 24 colors, from fresh dark green to senesced red to fallen brown leaves (Figure 1).We included 25 other materials, including asphalt, cardboard, wood, and several paint color chips, for a total of 26 51 samples.27 28

Field data 29
We used field data from an ongoing measurement campaign at the 40 m "Barn Tower" Boston.Mixed forest stands surrounding the tower are dominated by the deciduous species red 1 oak (Quercus rubra L., ~40% of basal area) and red maple (Acer rubrum L., ~20% of basal 2 area), with evergreen white pine (Pinus strobus L.) the dominant conifer.The MODIS land cover 3 classification for the tower, and the land immediately surrounding the tower, is deciduous 4 broadleaf forest.5 We used imagery (April 1, 2012 through March 30, 2013) from a StarDot camera that is 6 mounted at the top of the tower.The camera points north and is inclined ~30° below horizontal.7 As in the lab experiments, command scripts on the camera trigger the infrared cut filter and 8 obtain successive (about 30 s apart) color RGB and monochrome RGB+NIR imagery.Automatic 9 exposure is used for each image.Images are uploaded by FTP to a remote server every 15 10 minutes between 4 a.m. and 10 p.m. 11 Also mounted atop the same tower is a pair of 4-channel (blue: 470 ± 20 nm, green: 557 12 ± 25 nm, red: 605 ± 35 nm, NIR: 750 ± 42 nm) narrowband radiometric sensors (Model 1850, 13 Skye Instruments, Llandrindod Wells, UK).One unit with a cosine diffuser is pointed upwards to 14 measure incident solar radiation, while the second unit, with a 25° field of view, is pointed in the 15 same direction as the StarDot camera to measure radiation reflected by the canopy.We log 16 measurements every 15 seconds and record 30 minute averages on a datalogger (CR1000, 17 Campbell Scientific, Logan, UT).From these data, we calculated canopy reflectance (ρ i ) as in 18 Eq. 2, where Q i ↓ and Q i ↑ are measurements of the incident and reflected quantum flux, 19 respectively, for each band i, and the calibration constant (k i ) determined under natural (sunlit) 20 conditions using a Spectralon panel.21

Image processing 28
Briefly, image analysis for both lab and field data included manual delineation of an 29 appropriate ROI and extraction of color channel information for that ROI in each image.We also used an optical character recognition algorithm to read the image exposure from the text overlay 1 at the top of each image. 2 For the lab experiment, the ROI was adjusted for each sample to include as much of the 3 sample as possible.For the field data, we defined separate deciduous (predominantly red oak and 4 red maple) and conifer (white pine) ROIs, which were each roughly 440 x 440 pixels in size.The 5 camera field of view did not change over time, and thus these ROIs were fixed over the period of 6 analysis.7 We processed the camera imagery as follows.For each sample, we defined the ROI and 8 determined the average pixel value (digital number) across the ROI for three channels in the 9 color RGB images (R DN , G DN , B DN ) and one channel in the monochromatic RGB+IR images 10 (Z DN ).If both images were taken at the same exposure, then the monochrome RGB+IR images 11 could be partitioned to a visible component (Y DN ) plus a NIR component (X DN ) according to Eq. 12 4a, with the visible component calculated from the color RGB images as in Eq. 4b (Daniel 13 Lawton, StarDot Technologies, personal communication).Then the NIR component, X DN , was 14 Technologies, personal communication).However, our results were essentially the same 26 regardless of whether we used 2.0 or 1.7.27 Eq. 5a 28 We calculated camera NDVI as in Eq. 6, in terms of exposure-adjusted digital numbers. 2 For fixed-exposure images, the same calculation was used, ignoring the primes (´). 3 For both radiometer and camera data, there was substantial variability in the derived 5 indices at the 30 minute time step, which may be associated with factors such as overall light 6 levels, cloudiness, and illumination geometry (see further analysis in Results).After comparing 7 various averaging, quantile, and filtering methods, we derived daily estimates by calculating the 8 arithmetic mean across all observations where the incident photosynthetic photon flux density 9 (PPFD; measured at the top of the tower using a PQS-1 quantum sensor, Kipp & Zonen, Delft, 10 the Netherlands) was greater than 200 µmol m -2 s -1 .This method reduced day-to-day variability 11 in the resulting time series better than the 90 th percentile approach used by Sonnentag et al. 12 (2012), although an obvious advantage of the latter approach is that it does not require solar 13 radiation data.

Lab experiment 26
Spectral reflectance signatures differ markedly among the 51 samples scanned by the 27 spectroradiometer (Figure 1).There are pronounced differences in reflectance spectra between 28 green, yellowing, and red leaves, but the non-foliar samples added greatly to the variability across samples is smallest for violet and blue (400-500 nm) wavelengths, and largest for red 1 (620-700 nm) and near infrared (700-1000 nm) wavelengths.2 For the fixed exposure imagery (results not illustrated), digital numbers extracted from 3 camera imagery for the red channel (R DN ) are well correlated with mean reflectance over red 4 wavelengths (620-700 nm) measured with the spectroradiometer (r = 0.91).Similarly, digital 5 numbers for the near infrared (X DN ) component of the RGB+NIR imagery are well correlated 6 with mean reflectance over NIR wavelengths (700-1000 nm) measured with the 7 spectroradiometer (r = 0.87).Camera NDVI is well correlated with spectroradiometer NDVI 8 using these broad bands (r = 0.91), or using the spectral range of MODIS bands (band 1 = 620-9 670 nm, band 2 = 841-867 nm) (r = 0.93).10 We used an iterative procedure to identify the wavelengths across which reflectance 11 measured by the spectroradiometer is most highly correlated with R DN and X DN from the fixed 12 exposure imagery.For R DN , we obtained a correlation of r = 0.96 across the range from 570-660 13 nm (Figure 2A), whereas for X DN , we find a correlation of r = 0.91 across the range from 805-14 815 nm (Figure 2B).For the fixed exposure imagery, there is an excellent correlation (r = 0.99) 15 between camera NDVI and spectroradiometer NDVI calculated using these particular bands 16 (Figure 2C).17 We conducted a similar analysis for the automatic exposure imagery.Exposure-adjusted 18 digital numbers (Eq.5b, 5d) for each channel are best correlated with mean reflectance, 19 measured by the spectroradiometer, across an appropriate range of wavelengths (Figure 3).For 20 example,  !" ′ is most strongly correlated with mean reflectance across violet and blue 21 wavelengths (430-515 nm, r = 0.92),  !" ′ with green wavelengths (510-570 nm, r = 0.94),  !" ′ 22 with yellow and red wavelength (575-710 nm, r = 0.96; Figure 4A), and  !" ′ with near infrared 23 wavelengths (800-815 nm, r = 0.88; Figure 4B).The contour plots in Figure 3 show how the 24 strength of these correlations tended to fall off rapidly outside the optimal range.For example, 25  !" ′ is not well correlated with wavelengths < 575 nm or >710 nm, and  !" ′ is not well 26 correlated with wavelengths < 700 nm.For the automatic exposure imagery, camera NDVI is 27 very well correlated with spectroradiometer NDVI using either the most-highly-correlated bands Results from the lab experiment demonstrate the potential of the StarDot camera imagery 1 for characterizing the spectral properties of diverse materials, particularly in red and near 2 infrared bands.By processing sequential VIS and VIS+NIR images, we are able to back-3 calculate the NIR component.Furthermore, this experiment shows that camera NDVI is strongly 4 correlated with spectroradiometer NDVI when using either fixed exposure or automatic exposure 5 imagery.Since field images are recorded with automatic exposures to optimize dynamic range 6 under varying illumination conditions, our ability to correct for variations in exposure is critical.7 These lab results prove that camera NDVI is sensitive to the variation in reflectances of a wide 8 range of materials and surfaces.This indicates the potential for monitoring canopy phenology in 9 the field using a similar approach to characterize the seasonal variation in canopy optical 10 properties, as described in the following section.11 12

Field measurements 13
The green chromatic coordinate calculated from the narrowband radiometric sensors, 14 radiometer g CC (Figure 5A), exhibits a seasonal pattern that is typical of deciduous forests (e.g., 15 declining rapidly (day 250) in autumn with leaf coloration and abscission (day 300).For the 19 deciduous ROI, camera g CC (Figure 5B) follows essentially the same seasonal pattern, with the 20 notable exception of a more pronounced dip in greenness around day 290, corresponding to the 21 peak of autumn colors and a marked increase in canopy redness that is clearly visible in the RGB 22 camera imagery.The coniferous ROI shows a seasonally varying signal in camera g CC (Figure 23 5C), but with substantially smaller amplitude than that for the deciduous ROI, reflecting the 24 year-round presence of foliage but nevertheless indicating seasonal variation in foliar chlorophyll 25 concentrations (Richardson et al., 2009).The start of spring green-up also begins about 30 days 26 earlier, and the end of the autumn decline ends about 60 days later, for the coniferous ROI 27 compared to the deciduous ROI.28 By contrast, the seasonal cycle of radiometer NDVI differs from that of radiometer (or 29 camera) g CC (Figure 5D).The primary difference is the absence of the spike seen in g CC around 30 day 140.A secondary difference is the presence of a broad plateau in radiometer NDVI from about day 130 to day 240.As a result, radiometer NDVI gives a better representation of the 1 seasonal dynamics of canopy leaf area index (LAI) than does radiometer g CC (e.g.compare with 2 Fig. 1 in Richardson et al., 2012).Additionally, the seasonal cycle of radiometer NDVI roughly 3 parallels (and is comparable in magnitude to) that of MODIS NDVI (Figure 5D).In particular, 4 the timing of the spring increase and autumn decrease in NDVI is similar in both time series.5 However, during the winter months the MODIS data are somewhat noisier than the radiometer 6 data.7 There is a strong linear relationship (R 2 = 0.89) between radiometer NDVI and camera 8 NDVI for the deciduous ROI.The best-fit linear scaling coefficients (Eq.7) are a = 0.53 ± 0.02 9 SE and b = 0.84 ± 0.01, with σ(ε) = 0.041.Henceforth, we focus on the rescaled time series, 10 camera NDVI R , for which the seasonal cycle is much more similar in shape to that of radiometer 11 NDVI than camera g CC .For example, the relative rate of increase in both radiometer NDVI and 12 camera NDVI R for the deciduous region of interest in spring is more gradual than the 13 corresponding rate of increase in radiometer or camera g CC , and conspicuously absent in the 14 camera NDVI R data for the deciduous region of interest is the g CC spike that occurs around day 15 140.For the coniferous region of interest, camera g CC shows a seasonal pattern that is less 16 pronounced than that for the deciduous region of interest.However, there is no clear seasonal 17 cycle in the camera NDVI R data for the coniferous region of interest.Together, these results 18 suggest that camera NDVI is capturing different aspects of seasonal canopy dynamics than 19 indices, such as g CC , that are based only on visible wavelengths (cf.Nijland et al. 2013).The use 20 of both g CC and camera NDVI together give more information than can be obtained from either 21 index on its own.The advantage of g CC is that it is sensitive to leaf color, which is related to 22 pigmentation (Keenan et al. 2014), while camera NDVI is a better proxy for LAI.23 24

Non-phenological sources of variability in camera NDVI 25
Even with averaging of 30-minute data to a daily product, the signal-to-noise ratio of 26 camera NDVI R is somewhat higher than that of camera g CC (day-to-day variability being equal to 27 about 8% of the seasonal amplitude for camera NDVI R , compared with 5% for camera g CC ; 28 compare Figures 5B and 5E), and indeed also higher than that of radiometer NDVI (compare which can change over the course of the day or between one day and the next.And, some of this 1 variability is undoubtedly the result of the brief (30 s) lag between the VIS and VIS+NIR 2 imagery, which is unavoidable as the camera needs some time to adjust after the cut filter has 3 been triggered on or off. 4 We further investigate some of the factors associated with variability in camera NDVI R 5 using the 30-minute data, focusing on the period from day 160 to day 200, when camera NDVI R 6 for the deciduous region of interest is essentially stable at its maximum summertime value.7 During this period, camera NDVI R at the 30-minute time step (1 SD = 0.13) was 10 times more 8 variable than at the daily time step (1 SD = 0.01).Much of the variability in the 30-minute data 9 results from imagery captured under low-to-intermediate light levels (1 SD = 0.17 for PPFD < 10 500 µmol m -2 s -1 ; 1 SD = 0.06 for PPFD ≥ 500 µmol m -2 s -1 ).At both dawn and dusk, camera 11 NDVI R is considerably lower, and markedly more variable, than at mid-day, indicating both a 12 systematic bias and a lower signal-to-noise ratio under low-light conditions.However, excluding 13 periods with PPFD < 200 µmol m -2 s -1 , the diurnal pattern (data from 6 a.m. to 6 p.m.) in camera 14 NDVI R is negligible: a polynomial function of time of day accounts for no more than 2% of the 15 total variance in camera NDVI R .This suggests that illumination geometry has only a minimal 16 effect on camera NDVI R .Similarly, although camera NDVI R is highly variable when the ratio of 17 diffuse/total PPFD is ≥ 0.90, the noisy observations are all associated with PPFD < 500 µmol m -2 18 s -1 .19 A final-and important, because it can not be entirely eliminated by filtering for lower-20 light conditions-source of variation in camera NDVI R is the ratio of the exposure times for the 21 RGB and RGB+NIR imagery, i.e.E ratio = E Y /E Z .A second-order polynomial of E ratio explains 22 73% of the variation in 30-minute camera NDVI R , with both low (< 2) and high (>5) values of 23 E ratio contributing to this pattern.Filtering for periods with PPFD ≥ 500 µmol m -2 s -1 eliminates 24 most instances of low E ratio .High values of E ratio are most often associated with anomalously long 25 RGB exposures (high E Y ).The RGB images associated with these long exposures are 26 characterized by an anomalously bluish cast.We are unable to identify the specific lighting 27 conditions (total PPFD, PPFD variance, direct PPFD, diffuse PPFD, or the direct/total PPFD analysis to high-light, mid-day conditions, it should be possible to eliminate this source of 1 variability in camera NDVI R . 2 3 Discussion 4 We have proposed a method by which an off-the-shelf networked digital camera, 5 originally marketed for security monitoring applications, can be repurposed and used to obtain 6 information about the spectral properties of vegetation in both visible and NIR wavelengths, and 7 thus to calculate NDVI-style indices (cf.Nijland et al. 2013, who found that imagery from a 8 filtered, infrared-sensitive camera was of no more value than conventional RGB imagery for 9 tracking plant phenology and health).The lab experiment described here is used as a proof-of-10 concept (Figs. 2, 4), and shows that not only can we back out the IR component from the 11 RGB+NIR imagery, but also that our method for exposure-correction of auto-exposed imagery is 12 effective.The field data, on the other hand, show how this method can be applied for long term 13 monitoring of vegetation phenology in real-world conditions with varying solar illumination and 14 weather conditions.Specifically, we show good agreement between data obtained using the 15 camera NDVI method and the seasonal trajectory of NDVI measured using radiometric sensors 16 (Fig. 5).This approach is highly economical: the camera used here retails for about US$1200, 17 which is a fraction of the cost of a pair (upward-and downward-looking) of multi-channel 18 radiometric sensors.While we did not rely on a reference panel for standardization, inclusion of 19 a grey Spectralon (or other diffuse reflector) panel within the field of view of the camera would 20 potentially be of value for normalization under changing illumination conditions (e.g.cloudy vs. 21 sunny days); however, the results shown here suggest that even without this kind of calibration Networked cameras are well suited to field monitoring applications because with Internet 1 connectivity (using cell phone modems, this is now possible even at remote field sites) images 2 can be archived to an off-site server, and camera functionality can be monitored remotely.3 Furthermore, since this eliminates the need for manually swapping out memory cards, the 4 potential for shifts in camera alignment are minimized, making it is easier to maintain a constant 5 field of view.This facilitates image processing and improves data quality.6 We acknowledge that cameras with red and NIR sensitivity, primarily designed for 7 precision agriculture applications, have been developed and are commercially available (e.g. the 8 Agricultural Digital Camera by Tetracam, Inc., Chatsworth, CA, which retails for US$4800).9 These have been used for ecological studies (e.g.Steltzer and Welker, 2006;Higgins et al., 10 2011), but we are not aware of this type of camera being installed in the field for continuous, 11 long-term monitoring applications.Conventional digital cameras have also been customized and 12 used for similar work, but these have typically made use of two-camera systems, with one 13 camera filtered for visible wavelengths and the other for near infrared wavelengths (e.g.unique about the present approach is that by controlling the camera's infrared cut filter we 16 instead obtained sequential images from a single sensor, rather than simultaneous images from 17 two sensors.Not only does the single-camera approach reduce costs, it also eliminates issues 18 related to parallax, sensor calibration, and image alignment.Our method could, in principle, be 19 used with other camera systems although the linear scaling coefficients (a, b) reported here are 20 probably specific to the StarDot NetCam SC IR.21 cameras, e.g.Surface Optics SOC-710 or TetraCam MCA.However, for budget-limited 1 observational and experimental studies, the system proposed here may represent an acceptable 2 compromise, given its substantially lower cost and proven performance.3 Our camera NDVI approach is conceptually similar to that used to obtain broadband 4 NDVI from readily available radiometric measurements of incident and canopy-reflected visible 5 (PPFD) and total shortwave radiation (Huemmrich et al., 1999;Wang et al., 2004;Jenkins et al., 6 2007; see review in Richardson et al., 2013b).The similarity of camera NDVI to radiometer 7 NDVI, and the dissimilarity between camera NDVI and camera g CC , highlights the potential for 8 camera NDVI to provide different information about the state of the canopy than can be obtained 9 using only visible-wavelength (RGB) imagery.Furthermore, indices such as EVI (the enhanced 10 vegetation index, which also uses blue channel information) can be calculated from the camera 11 imagery in a similar manner.The resulting data should therefore be of great value for "apples-to-12 apples" evaluation of landscape phenology products derived from satellite remote sensing, as 13 suitable data for this kind of analysis are currently lacking (cf.Hufkens et al., 2012; note that 14 landscape heterogeneity and the mismatch between the camera field of view and the satellite 15 pixel to which it is being compared remain outstanding challenges).More generally, camera 16 NDVI could be used for continuous monitoring of plant stress in greenhouse or nursery 17 applications, or even quantifying responses to experimental manipulations in large field 18 experiments (e.g.nutrient additions, elevated CO 2 , rainfall exclusion, etc.).
estimated as Z DN -Y DN .15 Z DN = Y DN + X DN Eq. 4a 16 Y DN = 0.30 * R DN + 0.59 * G DN + 0.11 * B DN Eq. 4b 17 However, if the images were taken at different exposures (where E Y denotes the exposure 18 of the color RGB image and E Z the exposure of the RGB+IR image), then these exposure 19 differences had to be accounted for.Exploratory analyses indicated that division through by the 20 square root of the exposure time offered a straightforward solution (Eq.5a-d) to exposure 21 adjustment.Note that camera systems with different gamma values, where γ is the exponent in 22 the power law relationship between input and output signals of digital imaging systems, may be 23 different (cf.Sakamoto et al., 2010, 2012).Taking the square root of E assumes a γ = 2.0, a 24 reasonable approximation of the actual StarDot value of γ = 1.7 (Daniel Lawton, StarDot 25

14 Because
!" ′ and  !" ′ are not direct measurements of reflectance, the magnitude of 15 camera NDVI depends on the spectral distribution of the incident light.Thus, camera NDVI 16 values from the lab experiments are not directly comparable to those from the field experiment, 17 and those from the field experiment are not comparable to either radiometer or satellite NDVI 18 values.To compensate for this, we re-scaled camera NDVI (yielding camera NDVI R ) by 19 estimating the coefficients of a linear regression between camera NDVI (for the deciduous ROI) 20 and radiometer NDVI, where a is a slope coefficient, b is the y-axis intercept and ε is the model 21 residual: 22 radiometer NDVI =  camera NDVI +  +  Eq Sonnentag et al., 2012; Richardson et al., 2013b).Radiometer g CC rises rapidly in spring with 16 budburst (day 115-120) and leaf development to a pronounced spiky peak (day 140; see Keenan 17 et al., 2014, for discussion), before decreasing gradually over the course of the summer, and then 18

19 20Figure 1 . 5 Figure 2 . 7 Figure 3 .
Figure 1.Reflectance spectra of the 51 samples (thin grey lines) used in the laboratory 1 experiment.The heavier black lines indicate representative spectra from a healthy green leaf 2 (bottom), a yellowing leaf (middle), and a red (top) leaf.3

Figure 6 .
Figure 6.Winter (February 1, top) and summer (July 6, bottom) false-color images (XRG) 1 obtained from an infrared-enabled security camera mounted on the Harvard Forest Barn Tower. 2 The near infrared component (X) is mapped to the red (R) channel, the red channel is mapped to 3 the green (G) channel, and the green channel is mapped to the blue channel.Both images were 4 taken from the top of the tower at 4 PM local standard time.The evergreen trees that are clearly 5 visible in the top image are white pine (Pinus strobus).6 7 The phenology of terrestrial vegetation is highly sensitive to climate variability and 2 change (Rosenzweig et al., 2007; Migliavacca et al., 2012).In the context of climate change, 3