High-Throughput, In-Field Screening of Photosynthetic Efficiency in Crop Plants Using an Autonomous Robot

Photosynthesis provides plants not only with the energy basis for the production of biomass and the maintenance of their own metabolism, but also for nitrogen (N) fixation in legumes¹ and other symbiotic processes². The yield potential of any crop depends on the proportion of photosynthetic photon flux rate (PPFR) intercepted by the canopy (ε_i), the proportion of this radiation that is actually used in photosynthesis and transformed into biomass (ε_c), and the proportion of the biomass energy that is partitioned into the harvested product (ε_p)³. The ε_p can be set equal to the harvest index (HI) if the latter is defined as the biomass of the harvested product divided by the total shoot and root biomass³. Past breeding efforts have mainly focused on increasing ε_i and ε_c and were highly successful in doing so³: For example, modern cereal and grain legume genotypes can reach an ε_i in the order of 90% and a HI in the order of 60%^3,4, leaving little prospect to further increase ε_i and ε_p^3,5. By contrast, values of ε_c observed in C₃ and C₄ crops still rarely exceed 1/3 of the theoretical maximum of 9.4% and 12.3%, respectively. This indicates a large potential to further increase crop productivity if ε_c can be increased through selection based on natural genetic variation and/or targeted optimization of elements of the photosynthesis apparatus, such as the enzyme ribulose-1,5-bisphosphate carboxylase/oxygenase (RuBisCO), through genetic engineering^3,5,6,7. However, it has been shown that i) photosynthetic efficiency is highly dynamic under field conditions as plants constantly adapt to changing light and other environmental covariates^8,9 ii) photosynthetic traits measured under steady-state conditions (e.g., in indoor growth chambers with artificial lighting¹⁰) show different patterns of heritability than the non-steady-state photosynthesis observed in the field¹¹. This limits the usefulness of indoor photosynthesis phenotyping for the selection of higher-yielding genotypes of staple food crops that in practice are generally grown outdoors in the field. As a result, it is difficult to conduct a targeted selection for photosynthetic traits unless high spatial and temporal resolution data from a large number of crop genotypes are available^6,12.

Past attempts to conduct non-destructive measurements of photosynthesis in the field have largely relied on handheld gas exchange^13,14,15 or pulse-amplitude modulation (PAM)-type chlorophyll fluorescence (ChlF)^16,17measurement devices. These approaches have the shared disadvantage that leaves on which measurements should be performed need to be manually clipped or positioned into a measurement chamber^18,19. Moreover, gas exchange measurements in the field take several minutes to ensure an equilibrium in the measuring chamber. This makes the measurements laborious and slow, resulting in great difficulties to achieve a sufficient throughput to study photosynthetic regulation under fluctuating conditions in general, and particularly for a meaningful screening of hundreds or even thousands of genotypes in breeding programs¹⁹. Like the aforementioned PAM ChlF measurement devices¹⁷, the light-induced fluorescence transient (LIFT) method²⁰ makes use of the facts that i) light energy that reaches the chlorophyll in photosystem II (PS II) can either be used for photosynthesis, dissipated as heat in a process called non-photochemical quenching (NPQ) or by ChlF and ii) blocking the photosynthesis pathway by saturation from a strong light pulse resulting in reduction of electron acceptors downstream of PS II will result in a corresponding increase in ChlF. Based on this induced variable ChlF, we can calculate the photosynthetic quantum efficiency of PS II (F_q'/F_m') under ambient light conditions or the maximum quantum efficiency of PS II (F_v/F_m) if leaves were kept in the dark before the measurements.

The parameter F_v/F_m was established in the early 1980s by Kitajima and Butler²¹, Butler²², and Björkman and Demmig²³, who reported an optimal value of ≈ 0.83 in non-stressed leaves across diverse species. The development of F_q'/F_m' (Φ_{PS II}) as a diagnostic of photosynthetic performance began with Genty et al.²⁴, who showed its near-linear relationship with CO₂ assimilation under non-stressed conditions. Maxwell and Johnson²⁵ and Baker²⁶ later provided a practical framework for applying F_q'/F_m', F_v/F_m, and NPQ, emphasizing their sensitivity to environmental stress and their value in identifying damage symptoms such as photoinhibition. Murchie and Lawson²⁷ highlighted the potential (and limitations) of F_q'/F_m' for field phenotyping and crop improvement. More recently, Long et al.⁸ stressed that photosynthesis in crops occurs under fluctuating light, where dynamic changes in PS II efficiency and NPQ regulation strongly affect carbon gain, underscoring the need to interpret fluorescence parameters in relation to temporal environmental variability.

As already stated in²⁸, F_q'/F_m' measured with a stationary LIFT sensor correlated well with PAM (R² = 0.89) and CO₂ assimilation rate measurements (R² = 0.89)^29,30,31. In line with this, modelling F_q'/F_m' responses during field seasons has been shown to allow good predictions of ε_c, crop productivity^12,32 and stress tolerance³³. However, the PAM-type ChlF measurement devices use a single flash of light (saturating pulse) to inhibit the photosynthesis pathway and weaker flashes to measure minimum and maximum ChlF (measuring pulse), which takes about 1 s in total and is thus not suitable for very high throughput. In contrast, the LIFT sensor produces a fast sequence of 300 high-intensity flashlets (with 40,000 µmol photons m^-2 s^-1) to achieve a gradual saturation of the photosynthesis pathway within 750 µs (= 0.00075 s) while the ChlF yield is discretely measured²⁰. This allows for a larger distance between the measuring device and the target leaf¹⁹ and enables rapid, high-throughput measurements using an autonomous carrier vehicle on which the sensor is rigidly mounted³². Previous studies (e.g., ³²) were often limited in throughput under field conditions or restricted to glasshouse experiments. Although these approaches were successful in identifying candidate genes for photosynthetic regulation¹², they likely missed important regulatory processes occurring under field conditions. Therefore, we hypothesize that a LIFT sensor mounted on an autonomous robot constitutes a suitable tool to measure F_q'/F_m' of large numbers of genotypes (several measurements per second on thousands of plants or plots per day) in the field.

1. Setting up the PPFR sensor and data logger

1. Connect the PPFR sensor and the powerbank to the data logger. Connect the data logger to the laptop computer and start the data logger using the data logger desktop client.
2. Install the camera tripod next to the experimental field with the PPFR sensor on top and the data logger and powerbank on the ground below.
   NOTE: Make sure that the PPFR sensor is in an upright position and not shaded or otherwise affected by any nearby object. The PPFR sensor would need a gimbal if mounted to the robot directly to measure accurate PPFR values.

2. Setting up the robot and LIFT sensor assembly

1. Load the waypoint coordinates (.geojson format) for the measurement onto the robot. They can be the same as those used for GNSS-guided sowing of the trial.
2. Mount the LIFT sensor assembly on the front of the robot at a height of about 60 cm above the crop canopy (see Figure 1A).
3. Place the laptop computer, car battery, and power inverter on top of the robot and connect the power inverter to the car battery and to the LIFT sensor assembly.
   CAUTION: The excitation beam of the LIFT sensor is dangerous to the eyes. Looking directly into the excitation beam has to be strictly avoided.
4. Position the GNSS antenna kit on top of the robot, connect it to the laptop computer, and start the GNSS logger desktop client.
5. Connect the LIFT sensor to the laptop computer and start the LIFT desktop client with the following settings:
   LIFT sensor: excitation power = 40,000 µmol photons m>^-2 s^-1; flashlet length = 1.6 µs; number of excitation flashlets per single measurement = 300; number of relaxation flashlets = 80; time between excitation flashlets = 2.5 µs; time between relaxation flashlets: j_i = 10^{1.28 + 0.0215 × i} μs, where j_i  is the interval length of the i^th flashlet; sensor gain = 10 or 25; measuring interval = 0 s.
   Spectrometer: spectral integration time = 100 ms, spectral range from 400 to 800 nm.
   RGB cameras: exposure = automatic; shutter = automatic; gain = automatic; brightness = 0; frame rate = 20 s^-1; triggering mode = allowed (triggers the camera at each LIFT measurement).
   NOTE: The quantity directly measured by the LIFT sensor is the ChlF yield (increase in total ChlF due to the excitation beam for each flashlet²⁰. Then, calculate F_q'/F_m' as
   F_q'/F_m' = (F_m' - F') / F_m' ²⁶
   where F' is defined as the ChlF yield of the 1^st flashlet and F_m' as the average of the ChlF yields of the 301^st and the 302^nd flashlets²⁸. The total duration of one ChlF measurement (including the relaxation phase) is around 21 ms, but the duration of the crucial 300 flashlet excitation phase is only 750 µs.

3. Conducting LIFT measurements

1. Manually drive the robot to the beginning of the first row of plots in the experimental field using the remote controller.
2. Activate the measuring script for the LIFT sensor and the spectrometer in continuous mode.
3. Start the autonomous robot navigation at 0.5 m s^-1 velocity using the robot control website.
4. During measurement, periodically check the plot representing the ChlF yield over time in the LIFT sensor assembly desktop client to confirm that it has the expected shape (see Figure 1B). Adjust the sensor gain if signals are too weak.
5. Periodically hold the white reference panel under the excitation beam at the height of the crop canopy while the robot is turning at the field border.

4. Data integration and preprocessing

NOTE: The R code is available on GitHub (https://github.com/beat2keller/lift_data_processing)

1. Read in the LIFT transient and spectral data from all *_data.csv and *_spectral.csv using R data.table package³⁴.
2. Read in the GNSS and weather data together with the .geojson plot maps and the experimental design.
3. Combine the datasets and perform statistical analysis, e.g., as described in¹², or below.
   1. Extract F_q'/F_m' from each recorded transient.
   2. Join each GNSS point to its corresponding plot based on .geojson polygons using spatial containment. Do not forget to account for the distance between the LIFT sensor assembly and the GNSS antenna on the robot.
   3. Determine the heading of the robot from consecutive positions. Merge the experimental design with the GNSS data by the assigned plot identifiers.
      ​NOTE: This step links each spatial position to the respective treatment (e.g., genotype) and replication (e.g., block) information.
   4. Combine the high-resolution data from the PPFR sensor with low-resolution PPFR data from the nearest weather station to produce a fused incident PPFR time series (optional).
      NOTE: Low-resolution meteorological data for Switzerland can be derived from the agrometeo.ch (https://agrometeo.ch/de) network as described in​³⁵.
   5. Filter the white reference measurements from the spectral reflectance dataset and merge them with the PPFR data using time stamps to create a lookup table for the spectral reflectance under different incident light intensities as described in²⁸.
   6. Correct the raw spectral reflectance data based on the lookup table.
   7. Calculate the MERIS terrestrial chlorophyll index (MTCI) and the normalized difference vegetation index (NDVI) from the corrected spectral reflectance data using the formulas
      
      ​where R_λ denotes the mean reflection at wavelength λ nm.
   8. Merge the ChlF, GNSS position/experimental design and PPFR/spectral reflectance data by performing a nearest neighbor join with a defined tolerance window (e.g., 1 ms) based on time stamps.
   9. Identify and remove rows with outliers in F_q'/F_m' from the dataset.
   10. Extract genotype-specific photosynthetic response trends (G: PPFR) from the outlier-corrected ChlF data by fitting the model
       F_q'/F_m' = β₀ + β₁ Date +  β₂ (Heading x Hour) + β₃ (Genotype × PPFR) +  β₄ MTCI + β₅ PPFR + ε
       to the outlier-filtered dataset.
   11. Estimate the slope (β₃) of the (Genotype × PPFR) predictor term using estimated marginal trends (emtrends in R) to obtain a quantitative measure of the genotype-specific response to irradiance. Including the MTCI in the model accounts for differences in chlorophyll content and canopy status that influence F_q'/F_m' as described previously¹².
       NOTE: Heading denotes the driving direction of the robot (derived from GNSS bearings in step 4.3.3) to account for potential directional effects during measurements.

In total, 91,205 ChlF transient measurements were acquired of 36 soybean (Glycine max (L.) Merr.) breeding lines in 7 measuring days from June 12 to 27, 2025. 74,927 data points could be georeferenced to plots, and 58,916 were filtered for data quality and successfully matched with spectral and weather data. An overview of the measurement setup and representative data is shown in Figure 1, including an image of the instrument in operation (Figure 1A), ChlF induction kinetics across 12 soybean genotypes (Figure 1B), and corresponding spectral reflectance curves (Figure 1C).

The georeferenced data revealed pronounced spatial variation in both Fq'/Fm' and NDVI across the two experimental fields (Figure 2). These spatial patterns were partly associated with the driving direction of the robot (Figure 3), which was therefore accounted for in subsequent modeling to reduce the potential effects of shading on the target leaves.

Linear mixed-effects modeling of Fq'/Fm' responses to incident PPFR (Figure 4A) explained a substantial proportion of the variance (R = 0.45, R = 0.60). The extracted genotype-specific slopes (Response G: PPFR) showed clear differences among breeding lines (Figure 4B), with several lines exhibiting steeper or flatter response curves compared to the panel average (Figure 4C).

Finally, canopy-level 3D reconstructions derived from RGB imaging and the Matching And Stereo 3D Reconstruction (MASt3R) algorithm36 (Figure 5) demonstrated the potential of integrating LIFT-based physiological measurements with structural phenotyping to capture canopy architecture in three dimensions.

Figure 1: Overview of light-induced fluorescence transient (LIFT) and spectral reflectance measurements for soybean genotypes. (A) Example image of the LIFT device in operation. The inlet shows the blue excitation beam to induce chlorophyll fluorescence (ChlF). (B) Transient ChlF induction curves measured on 27 June 2025 for 12 soybean genotypes (mean ± SD, n between 88 and 553 per genotype, ntotal = 2,035). (C) Corresponding leaf reflectance spectra measured on 25 June 2025 for the same genotypes (mean ± SE, ntotal = 2,035). Colors indicate individual genotypes; error bars represent variation among replicate measurements. Please click here to view a larger version of this figure.

Figure 2: Spatial distribution of quantum efficiency of photosystem II and normalized difference vegetation index (NDVI) values. Spatial distribution of (top) quantum efficiency of photosystem II (Fq'/Fm') and (bottom) normalized difference vegetation index (NDVI) values was measured across two experimental soybean fields comprising 120 plots and 36 breeding lines (n = 58,916). Points represent measurement locations, with colors indicating the observed value and shapes indicating the heading direction of the robot: northwest (NW), northeast (NE), southwest (SW), and southeast (SE)-ward. Black lines delineate field plot boundaries. Black dots represent the global navigation satellite system (GNSS) coordinates of the field robot. Please click here to view a larger version of this figure.

Figure 3: Quantum efficiency of photosystem II. Boxplots show the quantum efficiency of photosystem II (Fq'/Fm') and the logarithm of incident irradiance at 680 nm across measurement hours (9 - 15 h) for both north-east (NE)-ward and south-west (SW)-ward rover heading direction (nNE-ward = 14,339, nSE-ward = 25,669, ntotal = 40,008). The robot was shading the measurement spot while heading SW-ward in the morning and NE-ward in the afternoon. Each box represents the interquartile range (IQR) with the horizontal line and whiskers indicating the median and 1.5 × IQR limits, respectively. Please click here to view a larger version of this figure.

Figure 4: Photosynthetic response slopes. (A) Spatial variation in estimated photosynthetic response slopes (interaction between genotype and photosynthetic photon flux rate (G: PPFR)) across two soybean experimental fields, derived from linear mixed-effects modeling. (B) Adjusted means and model-estimated photosynthetic response slopes for individual breeding lines, with highlighted genotypes of interest shown in color. (C) Relationship between incident PPFR and the quantum efficiency of photosystem II (Fq'/Fm') for the highlighted genotypes, with fitted square-root regression curves. Please click here to view a larger version of this figure.

Figure 5: Canopy architecture measurement by LIFT. The canopy architecture of the soybean variety Gallec is shown. (A) The red, green, and blue (RGB) cameras of the light-induced fluorescence transient (LIFT) device acquired images while screening for photosynthesis. (B, C) The matching and stereo 3D reconstruction (MASt3R) algorithm36 was used to reconstruct 3D canopy architecture. The pyramids indicate estimated camera positions. Please click here to view a larger version of this figure.

Throughput of the method

The LIFT method allows for a much greater measuring distance between the sensor and the target leaves than previous gas exchange and PAM-type ChlF measurement devices^18,19, which in turn enables automated high-throughput measurements being conducted both in the field and in indoor environments^12,32,37. This eliminates the need for physical interaction with the canopy. The throughput of the method increases with robot driving speed. To ensure accurate measurement of ChlF yield over time, the distance travelled during one F_q'/F_m' measurement must remain negligible compared to the illuminated area diameter (20 mm for the LIFT sensor). When the robot moves during excitation, the illuminated spot also shifts, so that ChlF yield per flashlet decreases, ultimately causing erroneous estimates of F_m' and F_q'/F_m' at excessive speeds. At a robot speed of 0.5 m s^-1, only about 2.4% of the area illuminated by the last excitation flashlet was not already illuminated by the first excitation flashlet, meaning that the systematic measurement error on F_m' due to the motion of the robot will be numerically negligible (full calculation on GitHub, https://github.com/beat2keller/lift_data_processing). At a velocity of 0.5 m s^-1, we measured 300 plots (1.5 × 2 m) per hour on a 40 × 36 m field, corresponding to several thousand plots per day. Moreover, the relatively low total weight (200 kg) and compact dimensions of our measuring setup allow transport by light vehicles, facilitating multi-site trials to assess photosynthetic efficiency across environments³⁸.

Leaf versus canopy photosynthesis

The LIFT measurements in this method are spatially restricted to the uppermost, sunlit layer of the canopy, which is responsible for about 50% to 70% of total photosynthesis³⁹. Furthermore, the amount of light each leaf can absorb critically depends on its angle relative to the sun, with more vertical leaves generally increasing leaf area index, enhancing light penetration and canopy photosynthesis^40,41. With the LIFT sensor assembly used in this method, effects of 3D canopy architecture on single-leaf or whole-canopy photosynthesis can be modelled using the reflectance measurements from the spectrometer²⁸ and/or an explicit reconstruction of 3D canopy geometry⁴² based on the images of the stereo RGB camera system. New algorithms, such as MASt3R³⁶, learn dense, geometry-aware feature correspondences using deep transformers, enabling more robust and accurate 3D reconstruction across wide or repetitive regions. However, even without such additional corrections, the usefulness of automated, high-throughput methods, such as the LIFT, to identify more productive and resilient crop varieties has been demonstrated^11,43.

Sink limitation

Sink limitation, i.e., active downregulation of photosynthesis due to an inability of the plant to use the amount of photosynthates it could produce^44,45, can mask differences in ε_c among genotypes. Sink-limited downregulation of photosynthesis likely accentuates with increasing CO₂ concentration in the atmosphere⁴⁵. Our method can detect sink limitation only when it leads to a decrease also in F_q'/F_m'; there are indications that this can in fact occur⁴⁶. In any case, the relevance of sink limitation appears to strongly depend on the crop and developmental stage, with wheat during grain filling being much more affected than grain legumes⁴⁵ because the symbiotic N fixation of the latter constitutes a strong additional sink^1,44,46. If sink limitation is expected to be of high relevance in a specific crop during the time period in which LIFT measurements are performed, further studies are needed to compare the resulting F_q'/F_m' values to simultaneously collected gas exchange data and check the validity of the assumption that F_q'/F_m' constitutes a good proxy for ε_c.

Measurement distance and variability in canopy height

The LIFT sensor assembly used in this method is designed to conduct measurements from a distance of approximately 60 cm to the crop canopy^32,47. The LIFT sensor used in the method is fairly robust against minor differences in measuring distance, but a larger measuring distance generally appears to result in somewhat smaller values of F_q'/F_m'⁴⁷. To prevent such bias, we recommend using the response of F_q'/F_m' to PPFR (which is more robust to changes in measurement distance) instead of the absolute values of F_q'/F_m' as in¹² and / or to explicitly account for differences in canopy height during data analysis.

Unintended, artificial shading of leaves

Increased light intensity generally decreases F_q'/F_m'^3,8,12. Thus, unnecessary shading by operators or the measuring setup should be avoided^3,8,12. To minimize shading from the robot, the driving direction can either be set to prevent it entirely³² (at the cost of reduced throughput due to unproductive return trips), or measurements can be taken in alternating directions and then statistically corrected for shading effects. In this study, the term Heading × Hour within the statistical model of F_q'/F_m' (see step 4.3.10) ensures that shading of the target leaves by the robot is accounted for.

Conclusion

In summary, a LIFT sensor mounted on an autonomous robot enables fast and automated measurements of photosynthetic efficiency under field conditions, overcoming throughput limitations of previous approaches. Whereas factors such as canopy structure and sink limitation require careful consideration, our results demonstrate that reliable and scalable photosynthesis screenings in agronomic field trials and plant breeding nurseries are feasible. The ensemble of a high-resolution stereo RGB camera system and the LIFT as a point sensor will allow for a further increase in precision: AI-driven 3D localization of leaves will enable sampling leaves with similar orientation towards the sun and taking genotype-specific leaf inclination into account. From a research perspective, it may be particularly interesting to compare the photosynthesis of genotypes with varying canopy architecture and conduct high-throughput LIFT measurements under free-air CO₂ enrichment (FACE) to better understand how F_q'/F_m' is affected by processes downstream of PS II under field conditions and potential sink limitations. In practical application, screening for photosynthetic performance and stress tolerance in breeding nurseries up to the farmer's field for precision farming applications has great potential.