Hyperspectral and Photodiode Retrievals of Nighttime LED‐Induced Chlorophyll Fluorescence (LEDIF) for Tracking Photosynthetic Phenology in a Vineyard

The magnitude of chlorophyll fluorescence emission represents both chlorophyll content and energy quenching processes enabling its application to serve as a proxy of photosynthetic activity. Thus, there is interest in advancing methods for canopy‐scale monitoring of chlorophyll fluorescence. Remotely sensed solar‐induced fluorescence (SIF) retrievals offer daytime monitoring of chlorophyll fluorescence, which can serve as an indicator of photosynthesis. However, it represents an instantaneous measurement during the day, which is strongly influenced by incoming radiation, solar angle, and sun/shade fraction—making it difficult to tease out baseline information on plant health and potential photosynthetic capacity—which could be tracked by changes in fluorescence yield (independent of sunlight). Recent advances have demonstrated the potential for inducing nighttime chlorophyll fluorescence via LED light sources at the canopy‐scale, which can be retrieved as LED‐induced chlorophyll fluorescence (LEDIF), potentially serving as a baseline indicator of plant health and photochemical capacity, independent of daytime conditions. In this study, we explored two methods of LEDIF retrievals: (a) hyperspectral sensor (1.33 nm full‐width half max) and (b) low‐cost Red‐Far‐Red photodiode sensor. LEDIF retrieved by the hyperspectral sensor demonstrated strong correlations with daytime SIF and gross primary productivity during mid to end of season phenology (R2 > 0.70). In contrast, phenological dynamics of LEDIF retrieved by the photodiode sensor was more subtle, likely due to weaker signal‐to‐noise ratio, but still demonstrated some potential. Overall, LEDIF offers a technique to monitor nighttime chlorophyll fluorescence emissions (and changes in its spectral shape with a hyperspectral sensor) to assess canopy‐scale phenology of photosynthetic potential.

such that an increase in one parameter often results in a decrease in the others (Porcar-Castell et al., 2014).This enables the utilization of chlorophyll fluorescence measurements to infer the dynamics of photochemistry and NPQ processes (Schreiber et al., 1995).
Pulse amplitude modulated (PAM) fluorescence systems have been instrumental in advancing the applications and understandings of leaf photosynthetic activity (Schreiber, 2004).The advantage of PAM fluorescence is the ability to measure both dark and light acclimated plants.Dark acclimated plant measurements enable a baseline status of potential photosynthetic capacity, which often serves as an indicator of the presence of plant stress.Light acclimated measurements probes the instantaneous status of a plant under ambient conditions (e.g., light intensity, temperature, water availability, etc.).To take full advantage of PAM fluorescence, both dark and light acclimated measurements can be combined to fully assess the fate of photons across both baseline and instantaneous processes (Hendrickson et al., 2004;Porcar-Castell, Juurola, Ensminger, et al., 2008;Porcar-Castell, Juurola, Nikinmaa, et al., 2008).Thus, PAM fluorescence has helped advance our understanding of the link between chlorophyll fluorescence and photosynthetic activity (Baker & Rosenqvist, 2004;Zavafer et al., 2020).However, PAM fluorescence is generally restricted to the leaf scale, but there is growing interest in technical advancements for scaling chlorophyll fluorescence techniques to larger spatial scales (Porcar-Castell et al., 2014, 2021;Mohammed et al., 2019;Sun, Gu, et al., 2023;Sun, Wen, et al., 2023).
Remote sensing methods with ultrafine spectral resolutions in the far-red spectral region have demonstrated their capability to retrieve a solar-induced fluorescence (SIF) signal from satellites (Frankenberg et al., 2011;Guanter et al., 2012;Joiner et al., 2011).SIF is linked to instantaneous chlorophyll fluorescence and has been shown to be a good indicator of photosynthesis and GPP across environmental conditions (Porcar-Castell et al., 2021).For example, at seasonal timescales in evergreen conifers, photosynthetic performance undergoes baseline changes in photochemistry, NPQ, and chlorophyll fluorescence, where in the winter NPQ is upregulated and photochemistry and chlorophyll fluorescence are downregulated, which reverses by the summer (Ensminger et al., 2004;Porcar-Castell, 2011).For short-term stresses such as drought, instantaneous photochemistry and chlorophyll fluorescence decrease while NPQ increases, enabling SIF to track drought-induced response of photosynthetic activity (Helm et al., 2020;Sun et al., 2015;Zhang et al., 2023).However under severe stresses such as heatwaves or extreme drought, SIF may decouple with photosynthetic activity due to the saturation of NPQ resulting in an increase of energy allocation towards chlorophyll fluorescence for energy dissipation (Magney et al., 2020;Marrs et al., 2020;Martini et al., 2022;Wohlfahrt et al., 2018).Thus, careful consideration of what the SIF signal represents mechanistically must be taken into account as SIF is currently only measured during the day and therefore offers information on the plant status encompassing both dynamic and sustained processes.
Recently, a new method has emerged to measure nighttime chlorophyll fluorescence spectra of canopies (Atherton et al., 2019;Romero et al., 2018Romero et al., , 2021)).Here, light emitting diode (LED) light sources can be utilized to induce chlorophyll fluorescence (LEDIF).A pure chlorophyll fluorescence signal can be detected by using blue LEDs, which induces chlorophyll fluorescence emissions in the red spectral region, thus the LEDIF signal is unconfounded by spectrally overlapping photosynthetically active radiation region (PAR; 400-700 nm) (Atherton et al., 2019;Romero et al., 2021).A key strength of LEDIF is that it enables dark acclimated measurements of canopy chlorophyll fluorescence, providing inference into vegetation baselines from sustained environmental conditions to assess presence of plant stress impacting fluorescence and potentially, leaf photosynthetic capacity (Rajewicz et al., 2023;Van Wittenberghe et al., 2019).In addition, using hyperspectral sensors to retrieve chlorophyll fluorescence spectra can enable evaluation of red to far-red fluorescence ratios (R:FR), which can be an indicator of plant stress or changes in chlorophyll concentration (Ač et al., 2015;Buschmann, 2007;Magney, Bowling, et al., 2019;Ortiz-Bustos et al., 2016).In addition to hyperspectral sensors, LEDIF opens the potential for relatively low-cost Red-Far-Red photodiode sensors for chlorophyll fluorescence retrievals.Brissette et al. (2023) explore the use of a Red-Far-Red photodiode sensor for tracking drought response.They find strong agreements between LEDIF retrieved from a hyperspectral sensor and a Red-Far-Red photodiode sensor in addition to strong correlations between LEDIF with PAM fluorescence metrics of chlorophyll fluorescence, photochemistry, and porometer measurements of stomatal conductance (Brissette et al., 2023).This highlights the potential for low-cost nighttime measurements of canopy chlorophyll fluorescence for tracking the dynamics of photosynthetic activity.Therefore, we seek to expand the applications of LEDIF for tracking photosynthetic phenology and physiological stress.
In this study, we used blue LEDs to induce canopy chlorophyll fluorescence emissions and retrieved LEDIF using a hyperspectral sensor and low-cost Red-Far-Red photodiode paired with Quantum sensors located 0.9 m above a grapevine canopy.In combination with eddy covariance estimates of GPP and a tower-based remote sensing

Site Description
This study was conducted at a commercial vineyard in California's Central Valley.The vineyard contains Vitis vinifera L. cv.Merlot vines planted in 2010 and trained on a split trellis.Plants were positioned in an east-west row orientation with 3.35 m of space between rows and approximately 1.5 m of space between plants with a continuous canopy within the row.The vineyard was divided into four subplots approximately four ha each as part of the Grape Remote Sensing Atmospheric Profile and Evapotranspiration eXperiment (GRAPEX) project (Kustas et al., 2018(Kustas et al., , 2022)).The canopy height ranged from 1.5 to 2.2 m.Soil type is classified as loam/sandy loam.More details of the vineyard site and irrigation treatments can be found in Knipper et al. (2019).

MIDNIGHTS-Monitoring Instrument Detecting NIGHT Spectra: Hyperspectral and Photodiode Sensors for LEDIF Retrieval
The canopy-based LEDIF system consisted of a blue LED light source, and two LEDIF retrieval setups: (a) Quantum (SQ-100X, Apogee Instruments, Utah, USA), and Red-Far-Red (S2-131, Apogee Instruments, Utah, USA), and (b) hyperspectral spectrometer (Flame-S-VIS-NIR, Ocean Insight, Florida, USA) (Figure 1).The light source consisted of five blue LED lights (two 15 W and three 5 W) which were controlled by an electric outlet timer.The hyperspectral sensor was operated via a Raspberry Pi 4 (Raspberry Pi Foundation, Cambridge, England, UK), which were both enclosed in a weatherproof case (Figure 1b).A 2 m fiber optic cable with a 25.4° field of view extended out of the weatherproof case.The LEDs were positioned next to each other facing down toward the canopy (Figure 1).The photodiode sensors and the fiber optic cable were positioned between the LEDs facing downwards toward the canopy.The sensors were positioned slightly below the LEDs to prevent direct light contamination from the irradiance signal.This MIDNIGHTS system was mounted in a fixed position at the base of the TSWIFT tower (see Section 2.3 TSWIFT) from July to October 2021 measuring one of the four subplots.Here, the sensor and LED setup was extended 1.8 m away from the tower with an initial height of 0.9 m above the top of the canopy (Figures 1c-1e).The electric outlet timer was programmed to turn the LEDs on for 30 min between 0 and 0030 hr local time.The photodiode diode sensors were programmed to measure reflected and emitted radiation every 5 min continuously throughout the full day and night.The Raspberry Pi was programmed to take hyperspectral measurements every 1 min during the night (23-02 hr) and day (10-15 hr) with a fixed integration times of 0.6 and 0.1 s, respectively.This integration time was chosen to optimize signal-to-noise ratio and prevent detector saturation.In this study, only the nighttime data was used.
The MIDNIGHTS LEDIF system illuminates the canopy when the blue LEDs are on, during which a clear optical signal can be observed using the hyperspectral sensor (Figures 1e, 2, and 3).A strong blue signal between 400 and 500 nm represents reflected blue light radiance from the canopy since the canopy was illuminated with blue LEDs (Figure 2a).A relatively smaller red-far-red chlorophyll fluorescence signal can also be observed (Figure 2b).A dark correction was applied to each individual measurement by subtracting the mean radiance between 600 and 650 nm, where there is no reflected or emitted signal from the hyperspectral data.The dark correction was applied to each measurement to correct for a small signal drift over the course of the night (Figures 3c and 3e).To quantify PAR and the red and far-red signals with the hyperspectral sensor, we calculated mean radiance between 400-650, 675-695, and 725-750 nm for LED-induced PAR (LED PAR ), red (LEDIF Red ), and far-red (LEDIF FR ) signals, respectively (Figure 2).We chose a modified PAR range for LED PAR (compared to the traditional 400-700 nm range) to avoid including any chlorophyll fluorescence signal in our PAR estimation (we note that optionally expanding the range to 700 nm likely has negligible impacts given the difference in radiance signal between red vs. blue regions [Figures 2a and 2b]).LED PAR serves as an indicator of both physiological and canopy structural changes since it represents reflected blue light.In contrast, LEDIF Red and LEDIF FR , represent photon re-emissions.Canopy structural variation due to growth can still impact LEDIF signal via variation in leaf area within the sensor field of view.Therefore, to account for canopy structural variation in the LEDIF signal, LEDI-F Red and LEDIF FR can be normalized by LED PAR as LEDIF Red/PAR and LEDIF FR/PAR , respectively, to better represent physiological variation in the LEDIF signal unconfounded by structural variation.To evaluate changes in the fluorescence spectra, we calculated LED-induced red:far-red ratios (LEDIF R:FR ) using LEDIF Red :LEDIF FR ratios.
For the photodiode sensors, the Quantum sensor has a relative response signal from 300 to 700 nm, and the red-far-red sensor captures red signal from 645 to 665 nm and far-red signal from 720 to 740 nm (Figures 2c and 2d).We note that the Quantum sensor captures a low amount of red-far-red emission, but assume this to be negligible due to low sensitivity and signal compared to the rest of the PAR region.The red spectral range is also mismatched from the red emission signal shown from the hyperspectral sensor (Figure 2b) indicating that the red signal output will not capture the red emissions, thus was not used in this study.Therefore, only the far-red signal from the red-far-red sensor was used to retrieve photodiode-based LEDIF FR along with LED PAR from the Quantum sensor.To dark correct the Quantum and red-far-red sensors, we calculated the mean signal from 0030 to 0130 hr, representing an hour after the LEDs turned off, to subtract from the PAR and far-red signal (Figures 3b  and 3d).

TSWIFT
A tower-based hyperspectral and solar-induced fluorescence system, TSWIFT (Tower Spectrometer on Wheels for Investigating Frequent Timeseries) (Wong et al., 2023), was set up in the center of the four subplots.The system consisted of a 2D scanning telescope unit and RGB camera mounted at the top of a tower (10 m).This enables a 360° pan and −45°-90° tilt to enable spot targeting of vegetation (0.7° field of view) and sky references.During sky reference measurements, an opal diffuser was used to increase the field of view to 180°.The telescope was connected to a fiber optic cable that connected to two spectrometers located in a temperature-controlled unit at the base of the tower.A Flame VIS-NIR spectrometer captured spectral data from 350 to 1000 nm with 1.33 nm full width half maximum (Ocean Optics, Orlando, USA).A QE Pro spectrometer captured ultrafine Gray region shows the period when the blue LED light sources are on.For the hyperspectral sensor, dark corrections were applied per observation by using a spectral region from 600 to 650 nm, unconfounded by the blue LEDs and red-far-red re-emission.For the photodiode sensors, dark correction was obtained using a period 1 hour after the LED lights shut off to subtract from the LEDIF signal.Note that there are minor differences in the sensor logging times and the electric outlet timer to true time.
spectral data from 730 to 780 nm with 0.3 nm full width half maximum (Ocean Optics, Orlando, USA).See Wong et al. (2023) for technical details of the TSWIFT system, but note this installment was not on wheels, but on a fixed tower with line power.
TSWIFT measured six vegetation target locations per subplot.Sky measurements occurred after every 6 vegetation target scans (approximately 30 s).To optimize signal-to-noise ratio and prevent spectrometer saturation, integration time was automatically adjusted per scan to achieve 80% to signal saturation.For the Flame spectrometer, reflectance was calculated by dividing the target radiation by nearest in time sky irradiance.With reflectance, the normalized difference vegetation index (NDVI) was calculated as NDVI = (R NIR − R Red )/(R NIR + R Red ), where R NIR is the average reflectance from 830 to 860 nm and R Red is the average reflectance from 620 to 670 nm.For the QE Pro spectrometer, solar-induced fluorescence (SIF) was retrieved using Differential Optical Absorption Spectroscopy (DOAS) (Grossmann et al., 2018;Platt & Stutz, 2008).Here, far-red SIF retrieval exploits the optical depth of Fraunhofer lines where they remain unchanged as sunlight passes through the atmosphere.Thus, any changes observed in the Fraunhofer line optical depth between sky irradiance and vegetation target radiance is caused by the addition of a plant's fluorescence signal (Frankenberg et al., 2011;Grossmann et al., 2018;Joiner et al., 2011).

Micrometeorology and Eddy Covariance System
Eddy covariance towers were set up in the southeast corner of each subplot (four in total).All eddy covariance towers were equipped with identical instrumentation to measure micrometeorological and carbon flux data.For carbon fluxes, an integrated open path infrared gas analyzer and sonic anemometer, IRGASON (Campbell Scientific., Logan, USA) was mounted near the top of the tower, 4.5 m above local ground level facing northwest.Micrometeorological instruments consisted of NR01 net radiometer (Hukseflux), EE08 temperature and relative humidity probe (E + E Elektronik) in an aspirated shield (Apogee Instruments, Logan, USA), and soil moisture and temperature sensor (Stevens HydraProbe, Oregon, USA) installed at 5 cm depth.Data were collected 20 Hz and surface fluxes were estimated over 30 min time periods.Anomalous records in the high frequency data were removed using a de-spiking moving window algorithm.Fluxes were corrected using a two-dimensional coordinate rotation of the three wind components as well as for sensor displacement and frequency response attenuation.Sonic temperatures were corrected based on Schotanus et al. (1983), and the fluxes were adjusted by the Webb, Pearman and Leuning (WPL) density corrections (Webb et al., 1980) More details of the post processing of the eddy covariance data can be found in Bambach et al. (2022).

Data Analysis
The LEDIF data was averaged daily from 0015 to 0030 hr to ensure that the LEDIF signal was stable, acclimated to the LEDs, and avoided time differences between LED light timer and sensors at the beginning of illumination.The TSWIFT NDVI and SIF data was averaged daily with a four hour window surrounding solar noon from 11 to 15 hr to ensure stable solar conditions.Each target from all four subplots were averaged together as well.For the EC data, daily midday means of GPP, air temperature (T air ), solar radiation, vapor pressure deficit (VPD) and soil moisture at 5 cm depth were computed by averaging half-hourly values from 11 to 15 hr PST.Daily minimum and maximum of T air (T air min and T air max, respectively) were also determined.LEDIF data was lagged by 1 day so that daily TSWIFT and EC data would be compared with LEDIF data for the same night rather than prior night.All analysis was performed using R (R Development Core Team, 2022).

Results
The LED-induced chlorophyll fluorescence (LEDIF) spectra demonstrated a seasonal decline in the red and far-red regions (Figure 4b).In addition, reflected radiance in the blue region also showed a decline (Figure 4a).For both LEDIF Red and LEDIF FR the relative percent difference between the start and end of the experiment was about −200% (Figure 4c).When comparing the shape of the spectra, the red and far-red LEDIF signals were relatively similar in magnitude from July to August.However, starting in September, the far-red LEDIF signal decreased more, relative to the red LEDIF signal (Figures 4c and 4d).When evaluating the overall time series, all hyperspectral-based LEDIF metrics showed seasonal declines (Figure 5).For the photodiode-based sensors, the seasonal variation of LED PAR varied less compared to the hyperspectral LED PAR (Figure 5b).We note that there was a sudden decrease in late August, most noticeably in both hyperspectral-and photodiode-based LED PAR , which was due to manual sensor repositioning by increasing the MIDNIGHTS system's height relative to the canopy due to canopy growth (Figures 5a and 5b).Normalizing LEDIF by LED PAR , accounts for structural variation from sensor reposition but also for canopy structural change from growth enabling a more physiologically driven LEDIF signal.Here, LEDIF Red/PAR and LEDIF FR/PAR show a seasonal decline and also remove the sharp late August decline from sensor repositioning (Figures 5e and 5h).We note that the photodiode LEDIF FR and LEDIF FR/PAR show similar patterns even with the LED PAR normalization, this is attributed from the low LEDIF FR signal relative to LED PAR resulting in a value shift but retaining similar seasonal patterns (Figures 5b-5d, and 5f).The seasonal declines of the LEDIF signals tracked similar patterns of the vegetation (GPP, SIF, NDVI) and meteorology (irradiation, temperature, water availability) (Figure 6).During this study, solar radiation and soil moisture showed a large seasonal decrease, while T air was more subtle (Figures 6a-6c).The vegetation dynamics of the vineyard also reflected this trend shown by GPP, SIF, and a more subtle change for NDVI (Figures 6d-6f).
Comparing nighttime LEDIF signal from the hyperspectral sensor to daytime SIF from TSWIFT, we observed strong correlations (Figure 7; R 2 > 0.70, p-value < 0.001).The highest R 2 with SIF was with hyperspectral-based LEDIF Red and LEDIF FR (Figures 7b and 7d; 0.89 and 0.91, respectively).A PAR correction for the LEDIF Red .We note that in late August, the MIDNIGHTS system was repositioned higher to account for canopy growth, resulting in a sudden signal shift, most noticeably from LED PAR from both hyperspectral and photodiode sensors.Shaded region represents the daily standard error of the mean from 0015 to 0030 hr when LEDs were powered.Line colors, purple and orange, represent the hyperspectral and photodiode sensors, respectively.and LEDIF FR signal leads to a slightly lower R 2 (0.77 and 0.87, respectively).LEDIF R:FR to SIF relationship was also nonlinear (Figure 7f).For the photodiode sensors, only LED PAR tracked SIF, performing similarly to the hyperspectral LED PAR (Figure 7a).LEDIF FR from the photodiode sensor had a low R 2 when compared to SIF of 0.12 (Figure 7d).
For tracking the photosynthetic phenology of grapevines, we used a correlation coefficient matrix between fluorescence parameters (SIF and LEDIF) and reflected PAR radiance (LED PAR ) to GPP, NDVI, and environmental conditions to evaluate respective correlations (Figure 8).SIF and most LEDIF retrievals demonstrate high correlation coefficients with GPP and NDVI (r > 0.77).Only the photodiode-based LEDIF demonstrated noticeably weaker performance.In comparison with environmental parameters, soil moisture followed by solar radiation had the highest correlations with SIF and LEDIF.Temperature and VPD were relatively weaker.

Chlorophyll Fluorescence Spectra
The spectral signature of chlorophyll fluorescence can provide information about changes in the photosynthetic apparatus (PSII and PSI) and chlorophyll content of plants, which can be used to assess changes in plant pigment pools, their stress tolerance and growth stage/phenology (Buschmann, 2007;Gitelson et al., 1999;Pedrós et al., 2008).Both LEDIF Red and LEDIF FR showed similar seasonal declines (Figures 4 and 5).This is likely indicative of both decreasing photosynthetic capacity and chlorophyll content (Atherton et al., 2019;Rajewicz et al., 2023;Romero et al., 2021;Van Wittenberghe et al., 2021).Normalizing LEDIF with LED PAR , which represents changes in reflected radiation, can account for changes in canopy structure due to absorbed photosynthetically active radiation (APAR).Thus, LEDIF Red/PAR and LEDIF FR/PAR represent only the physiological processes of light energy balance and chlorophyll content, which ultimately demonstrates a seasonal pattern and good performance for tracking photosynthetic phenology (Figures 6-8).The spectral shape of chlorophyll fluorescence and the red:far-red ratio (R:FR), can also be used for chlorophyll content determination (Buschmann, 2007).The red fluorescence maxima is strongly reabsorbed in a canopy whereas reabsorption is minimal in the far-red fluorescence maxima (Buschmann, 2007).Thus, as chlorophyll content decreases, the far-red maxima decreases more relative to the red maxima leading to an increase in R:FR (Gitelson et al., 1998).Our LEDIF R:FR shows an expected increase near the end of the growing season in September (Figure 4d), where chlorophyll content is decreasing due to senescence (Magney, Frankenberg, et al., 2019;Yang & van der Tol, 2018).

LEDIF Retrieval
The MIDNIGHTS LEDIF system offers a low-cost build designed for automated and continuous long-term deployment for retrieving canopy chlorophyll fluorescence.The MIDNIGHTS LEDIF system builds upon  2019) using multi-color LEDs (blue, green, and red) to explore nighttime canopy chlorophyll fluorescence spectra signals.However, the spectrometers from these studies are more costly (tens of thousands of dollars) and not designed for automated and long-term deployments.Thus, the MIDNIGHTS LEDIF system tests a more low-cost hyperspectral sensor (thousands of dollars) as well as the most low-cost photodiode Quantum and Red-Far-Red photodiode sensors (hundreds of dollars) that are suitable for automated and long-term deployments.The hyperspectral-based LEDIF retrieval performed best, exhibiting a clear seasonal decline that closely tracked SIF (Figures 5 and 7).When comparing the raw LEDI-F Red and LEDIF FR signal to the normalized LEDIF Red/PAR and LEDIF FR/PAR signal, we saw a slight decrease in correlation with SIF (Figures 7b-7d, and 7f).This is because LEDIF Red/PAR and LEDIF FR/PAR represent a physiological signal, whereas LEDIF Red and LEDIF FR are sensitive to both physiology and canopy structure.SIF itself is sensitive to both physiology and canopy structure (Porcar-Castell et al., 2021;Sun, Gu, et al., 2023;Sun, Wen, et al., 2023), which likely drives the stronger phenological correlations with LEDIF Red and LEDIF FR .
Compared to the hyperspectral-based LEDIF signal, the photodiode-based LEDIF signal demonstrated minimal seasonality and did not correlate well with SIF (Figures 5 and 7).This is in contrast to Brissette et al. (2023), who found close correspondence between the photodiode and hyperspectral derived LEDIF values.This is likely due to lower signal-to-noise ratio for the photodiode sensors (low intensity ∼0.01 μmol m −2 s −1 signal and ∼2.5 signal-to-noise ratio; Figure 2), which they were placed further away from the canopy in a less controlled environment as was done in the experiment by Brissette et al. (2023) (∼0.05 μmol m −2 s −1 signal).For the hyperspectral sensor, we manually set a 0.6 s integration time to maximize signal-to-noise ratio while avoiding saturation (low intensity ∼2.5 signal-to-noise ratio Figure 2).We were unable to adjust the integration time away from default instantaneous sampling for the photodiode sensors, so the radiance signal was noticeably weaker.If possible, one could increase the integration time for the photodiode sensors and/or increase LED intensity with stronger blue lamps to induce a stronger chlorophyll fluorescence emission.However, we urge caution for using too strong of LED lights as it may result in light acclimation inducing photosynthesis and thus a chlorophyll fluorescence response.Thus, a fine balance of light intensity is required to excite chlorophyll to induce a fluorescence signal, but not sufficient enough to induce light acclimation and electron transport/photosynthesis (Murchie & Lawson, 2013).

LEDIF for Tracking Photosynthetic Phenology
LEDIF from the hyperspectral sensor and tower-based SIF both tracked the end of season phenology of GPP and NDVI (Figure 8).During this period, declining GPP and chlorophyll fluorescence was driven by decreasing solar radiation, soil moisture, and air temperature, likely attributed to decreasing chlorophyll content (Gitelson et al., 1999).As LEDIF is captured at night, it serves to represent the chlorophyll content and potential photosynthetic capacity (indicated by baseline fluorescence emission) without the influence of PAR (Atherton et al., 2019;Romero et al., 2021).Interestingly, the LEDIF correlations with GPP improved when LEDIF was normalized by LED PAR (i.e., LEDIF Red/PAR and LEDIF FR/PAR ) (Figure 8).This correction results in a more physiologically driven fluorescence signal (e.g., light energy balance and chlorophyll content), much like how relative SIF can account for changing canopy structure (Magney, Bowling, et al., 2019;Pierrat et al., 2021).This reduces contributions of canopy structure to the LEDIF signal from sources such as canopy structural change via growth leading to changing distance between canopy height to sensor, and leaf movement and wilting.GPP itself represents photosynthetic activity independent of canopy structure, which likely contributes to the improved performance of LEDIF when normalized with LED PAR .
In contrast, midday SIF represents instantaneous chlorophyll fluorescence (i.e., F s ), which is influenced by physiology, structure, and ambient PAR (Porcar-Castell et al., 2021;Sun, Gu, et al., 2023;Zhang et al., 2018).Thus, while both LEDIF and SIF capture canopy chlorophyll fluorescence, they represent different processes due to the timing at which they are captured, much like PAM fluorescence with dark and light acclimated leaf measurements (Porcar-Castell et al., 2014).Yet when evaluating phenological dynamics, both LEDIF and SIF demonstrate strong correlations with GPP and NDVI, supporting the covarying dynamics of chlorophyll content, chlorophyll fluorescence, and photosynthetic activity (Atherton et al., 2019;Rajewicz et al., 2023;Romero et al., 2021).As this paper focused on the application of LEDIF for tracking photosynthetic phenology, we suggest further exploration of LEDIF chlorophyll fluorescence spectra in response to environment across different conditions and stress events such as heat and drought stress.Perhaps, this could enable further applications of nighttime LEDIF for elucidating baseline and plant health information on photosystem II and I processes.

Conclusions
This study demonstrates the potential for retrieving a LEDIF signal using a hyperspectral sensor and low-cost photodiode sensors for tracking the phenology of GPP.While the photodiode sensors were limited due to poor sensitivity of the small chlorophyll fluorescence signal (weaker signal-to-noise ratio), we suggest further optimization with longer sensor integration times, more powerful blue LED light source albeit with careful consideration of optimal light intensity, or closer sensor to canopy distance.The hyperspectral LEDIF retrieval, which optimized signal-to-noise ratio, demonstrated its ability to track decreasing chlorophyll fluorescence which correlated with midday SIF and GPP.Thus, LEDIF offers a technique to monitor changes in chlorophyll content, and likely potential photosynthetic capacity uninhibited by instantaneous conditions such as PAR that would impact SIF dynamics.This work complements a recent study demonstrating the LEDIF system for detecting short-term drought stress response tracking pigment, PAM fluorescence, and stomatal conductance dynamics (Brissette et al., 2023), as well as a tractor mounted system (Romero et al., 2021).Beyond phenology, chlorophyll determination via R:FR has also been used to assess changes driven by plant growth development, health, stress events and stress tolerance (Buschmann, 2007).To this end, LEDIF can enable a baseline assessment of vegetation health via chlorophyll content, and the potential for photodiode sensors for LEDIF retrieval offers a low-cost system that can be deployed for continuous and automated monitoring to aid in applications for baseline validations of SIF and for real-time stress detection.This project received support from the American Vineyard Foundation (AVF).CYSW and TSM received support from USDA-NIFA (Award #2020-67013-30931).TSM acknowledges support from the State of California Hatch Project (CA-D_PLS-2705-H).Funding for AJM was provided by USDA-ARS CRIS project 2032-21220-008-000-D.We would like to thank E&J Gallo Winery and the vineyard management staff for logistical support of the tower measurements used in this study.The use of trade, firm, or coporation names in this article is for the information and conenience of the reader.Such use does not constitute official endorsement or approval by the U.S. Department of Agriculture (USDA) or the Agricultural Research Service of any product or service to the exclusion of others that may be suitable.USDA is an equal opportunity provider and employer.

Figure 1 .
Figure 1.Pictures detailing the MIDNIGHTS (Monitoring Instrument Detecting NIGHT Spectra) LEDIF system.(a) Shows the blue LED light and sensor system (note that this picture does not show the fiber optic, and an additional blue LED light was added prior to field deployment); (b) shows the weatherproof case enclosing the hyperspectral spectrometer, Raspberry Pi and extending fiber optic and power cable; (c) shows an overhead view; (d) shows a side view; and (e) shows a side view of the MIDNIGHTS system at night with the blue LEDs on.

Figure 2 .
Figure 2. Top panels show the LED-induced reflected radiance in the PAR region (a) and emitted photons in the red and far-red regions (b) detected by the hyperspectral sensor.The gray, red, and dark red regions represent the wavelengths used for calculating PAR, red and far-red, respectively.Bottom panels show the relative response of the Quantum and Red-Far-Red photodiode sensors to photons.

Figure 3 .
Figure 3. Examples of non-dark corrected PAR, far-red, and red emissions from the hyperspectral instrument (left, purple) and from the Quantum and Red-Far-Red sensors (right, orange) from one night (2021-08-03).Gray region shows the period when the blue LED light sources are on.For the hyperspectral sensor, dark corrections were applied per observation by using a spectral region from 600 to 650 nm, unconfounded by the blue LEDs and red-far-red re-emission.For the photodiode sensors, dark correction was obtained using a period 1 hour after the LED lights shut off to subtract from the LEDIF signal.Note that there are minor differences in the sensor logging times and the electric outlet timer to true time.

Figure 4 .
Figure 4.The hyperspectral LEDIF signal across the visible-NIR (a) and Red-Far-Red (b) spectral regions.(c) Shows the relative percent difference of the Red-Far-Red emission relative to the first observation date (2021-07-25).(d) Shows the time series of the Red-Far-Red ratio (R:FR) and the standard error of the mean (shaded region).

Figure 5 .
Figure 5.Time series of the reflected PAR radiance and LEDIF signal from the hyperspectral (left, purple) and Quantum/Red-Far-Red sensors (right, orange).We note that in late August, the MIDNIGHTS system was repositioned higher to account for canopy growth, resulting in a sudden signal shift, most noticeably from LED PAR from both hyperspectral and photodiode sensors.Shaded region represents the daily standard error of the mean from 0015 to 0030 hr when LEDs were powered.Line colors, purple and orange, represent the hyperspectral and photodiode sensors, respectively.

Figure 6 .
Figure 6.Time series of midday (averaged between 11 and 15 hr PST) solar radiation, midday air temperature (T air ), midday soil moisture at 5 cm depth, midday GPP, midday SIF and midday NDVI and nighttime LEDIF FR/PAR, Red/PAR and R:FR.Colors represent the instrument (EC-black, TSWIFT-red, hyperspectral sensorpurple, Red-Far-Red sensor-orange).Shaded region represents the standard error of the mean between 11 to 15 hr and 0015-0030 hr, for the corresponding afternoon and nighttime periods, respectively.

Figure 8 .
Figure 8. Pearson's correlation matrix of fluorescence (SIF and LEDIF) and reflected PAR radiance (LED PAR ) versus midday GPP, midday NDVI and midday or daily minimum (min) and maximum (max) meteorological data from July 3 to October 7. Text colors represent the instrument (EC tower-black, TSWIFT-red, hyperspectral sensor-purple, photodiode Quantum and Red-Far-Red sensors-orange).