HTPP systems are designed based on principles of automation and standardization. However, several critical aspects must be considered and discussed in order to effectively design and implement these systems. These include the required scale of the experiments, the selection of sensors that will be installed, the necessary software infrastructure, and the expected return on investment in terms of financial and human resources. These aspects can only be fruitfully discussed if guided by the question what purpose the system should serve and what goal(s) one would like to achieve through this platform. This is even more important in the case where multiple research groups with diverse interests are involved, as there is a risk that the intended platform becomes a compromise that in the end does not satisfy any.

HTPP systems involve a complex combination of logistics and technology, particularly when plants need to be moved to sensors. Such machines require the use of conveyor belts, gantries, or mobile robots to transport plants through growth chambers or glasshouses to an imaging station. At the imaging station, plants are generally subjected to controlled conditions, such as a consistent light spectrum and intensity, allowing for a consistent acquisition of images or other sensor-based data. Additionally, plants in pots or carriers can be automatically weighed and watered. Measurement equipment must be linked to large-scale data storage facilities and databases, in order to store and retrieve the collected measurement information and other metadata. Clearly, expert knowledge of automation, logistics, error control, fine mechanics, non-destructive sensing, image analysis, and database management is required to build and operate these systems properly. Companies specializing in these different fields offer to construct customized machines on-site. The costs of these systems can be substantial. Prices typically range from approximately €60,000 to €120,000 for small systems with a lower degree of automation and throughput, €350,000 – €500,000 for fully automated sensor-to-plant or plant-to-sensor systems, and up to €3,000,000 for high-end systems with extensive optimization and automation of the workflow. The costs vary depending on factors such as the types and number of sensors used, system size, quality of the product and service provided by the supplier. In addition, these custom-built facilities require expert maintenance, including sensor calibration, replacement of worn-out parts and preventive check-ups. Yearly servicing contracts often amount to 5-10% of the acquisition price and are necessary for the duration of system usage.

Given the above considerations, certain laboratories opt to construct HTPP systems themselves, either utilizing off-the-shelf products or with the support of robotics companies (Bagley et al., 2020). Building customized platforms can allow for increased flexibility (as discussed in section 3.4) and better fit the specific requirements of the involved labs. However, it is important to acknowledge that the path to a fully functional machine is challenging and time-consuming. While the direct hardware costs may be lower than those of commercial HTPP systems, the substantial human resources required for the development and maintenance of such automated systems should not be underestimated (Reynolds et al., 2019).

There are two additional points to consider regarding personnel investment. Firstly, it is advisable to involve expert users, project managers, and engineers throughout the planning and building phase. This ensures that any necessary compromises that almost inevitably have to be made during the whole process do not strongly constrain the desired goals of the end-users. Secondly, operating these systems can be complex, particularly when troubleshooting errors, as there are numerous variables to consider within the platform software, platform mechanics and network communications. Therefore, it is recommended to have an operational team led by an expert who is responsible for the platform. This lead expert can receive training from the supplier or the constructors, and subsequently train all other individuals who will be utilizing this platform. Clearly, such a person requires both technological as well as plant biological expertise to effectively manage and operate the HTPP system.

An HTPP platform allows researchers to significantly scale up their experiments, often by an order of magnitude. This implies that also an order of magnitude more plants are processed, more containers have to be filled, more consumables are used, more electricity and irrigation are required, and additional cleaning has to be done at the end of the experiment. While these aspects are not part of the phenotyping platform per se, they contribute to a significant workload. As a result, there is often a need for subsequent adjustments to the workflow’s ergonomics, such as automating tasks like pot filling, seed sowing (Jahnke et al., 2016) or plant transplantation. Together with the sampling for further physiological characterization and disinfection or disposal of used components, they bring about additional planning and investments on top of those for the platform itself.

Another aspect to consider is the data infrastructure. Given the massive amount of data generated by HTPP systems, effective storage and accessibility of data for both short-term and long-term use are essential. Such utilization of data requires a good documentation of data and metadata, as described in the ‘Minimal Information About Plant Phenotyping Experiments’ (MIAPPE) guidelines (Krajewski et al., 2015; Papoutsoglou et al., 2020). Also, data and metadata should be stored in accordance with the FAIR principle (findable, accessible, interoperable, and reusable; Wilkinson et al., 2016) as in GnpIS (Pommier et al., 2019) and as currently established in research data management infrastructures like DataPLANT (www.nfdi4plants.de).

Remote support by the different suppliers requires a safe, secure, and reliable access policy. Moreover, in the analysis of data from such platforms, there is a growing emphasis on integrating genotypic and phenotypic data, as well as incorporating detailed environmental characterization of the experiments. Bringing all these data together often requires a considerable a-posteriori effort from trained personnel, which also needs to be budgeted in advance to make full use of the potential of these HTPP systems (Reynolds et al., 2019).

Automation can replace a considerable amount of tedious repetitive manual labor, especially when routine operations are customized to the research question and species of interest. However, this increased automation often comes at the cost of reduced flexibility. For instance, a system that is perfectly fit for small rosette plants like Arabidopsis thaliana, may not be as suitable for a large species with different architecture, such as Zea mays. Increased flexibility of the system can be achieved by planning for modular components that can be interchanged depending on the prevailing research question. However, increased flexibility to accommodate the needs of a variety of researchers also implies that the system becomes increasingly complicated, with a higher likelihood of failures and more efforts to solve these problems. A ‘jack-of-all-trades’ system will hardly provide the same level of detail and raw data resolution across the entire range of plant sizes and architectures as a dedicated high-resolution system designed specifically for either small or big plants, due to limitations imposed by the optical properties of the sensor and system layout. In those cases, a potential solution could be to have two smaller and targeted systems, rather than relying on a one-size-fits-all solution.

Another aspect of flexibility is the expected lifespan of system components. Technological advancements occur rapidly, and new sensors, for example, will generally be more powerful and informative than their predecessors. However, owners of commercially-acquired phenotyping systems often face problems in that the software to run the whole system is proprietary to the company, and therefore not accessible for further development. In such cases, it is complicated or impossible to integrate new sensors or apply other modifications to existing HTPP systems. New buyers are advised to discuss with their suppliers what support they can get in that respect. Self-builders are suggested to set up their software interfaces as flexible as possible, to easier adjust their system when new sensors or updates become available.

Plants are strongly influenced by their environment and different genotypes or species may show varying degrees of genotype x environment interactions for many of their phenotypic traits. This becomes particularly critical, because in most (semi-)controlled environments we impose abiotic conditions that significantly deviate from the natural conditions plants experience outdoors (Poorter et al., 2016; Chiang et al., 2020). Considerations about the location of the HTPP system (growth chamber, glasshouse, or field) and the range of environmental conditions provided to the plants are therefore an integral part of the design process.

Compared to traditional experiments, the use of high-throughput phenotyping systems often introduces additional constraints, that can impact the growth environment of the plants and, consequently, the outcomes of experiments. For instance, in a typical plant-to-sensor system, the combined weight of plants plus pots is limited by the strength of the conveyor belt used, as well as the scales used for gravimetric measurements. This limitation results in experiments being confined to plants in relatively small pots with substrates of low specific mass, which clearly affects plant growth and experimental outcomes (Passioura, 2002; Poorter et al., 2012). Transportation may also have other consequences. Tall plants, such as Zea mays, may topple over if the conveyor belts move too quickly. During transport, leaves of sensitive species (e.g. Brassica rapa, Hordeum vulgare) can get damaged, causing them to droop downwards along the pot. To mitigate damage, researchers may choose to place suitable support structures next to or around each plant, except in cases where wilting would be a phenotypic trait of interest. Apart from direct damage, one would also expect thigmo-morphogenetic responses to occur as a reaction to the mechanical perturbations during transport of plants, such as thicker and shorter stems (Anten et al., 2005). However, plant height was not negatively affected in the experiment of Brien et al. (2013), and neither was shoot biomass or leaf area in various experiments ( Table 1 ).

Fixed distances between cameras/sensors and plants may restrict the range of plant sizes that can be investigated, thereby limiting the developmental stages that can be studied in such HTPP systems to young vegetative plants. A last example pertains to the watering of the plants. Well-watered containers, especially with peat substrate, often show algal growth on the soil surface, which hampers the non-destructive derivation of plant size through image analysis ( Figure 2A ). To avoid this complication, researchers may opt to cover the top of the pots with a plastic sheet of contrasting color (Junker et al., 2015). Alternatively, for small rosette species like Arabidopsis, reducing the volume and/or frequency of watering can keep the topsoil dry for a longer duration. Although this suppresses algal growth, it also has unknown consequences for the growth rate and phenotype of the plants. It is recommended to consider a-priori whether the additional constraints imposed by an HTPP system are acceptable within the scope of the research question.

Once an operational HTPP system is in place, attention can shift to the performance of experiments. Given their typical large scale, careful consideration of experimental design is essential. For a comprehensive discussion on this topic, the reader is referred to Thompson et al. (2022). One notable advantage of high-throughput phenotyping systems is that individual plants (or microplots) can be measured frequently and non-destructively. This allows for the repeated measurements of the same plants, enabling the tracking of their growth and development over time. Such analyses can yield valuable insights, with a good example discussed in Box 1 . However, in cases where the research question focuses on identifying the best-performing genotype at the end of the experiment, repeated measurements may be unnecessary. In those cases, researchers could also opt for one final destructive harvest, which might be more simple, cheaper, and more informative, as also illustrated in Box 1 . A lower measurement frequency may also be advantageous if the measurements have the potential to interfere with plant growth. This is particularly relevant when plants are taken out of their growth environment for longer-duration measurements, such as magnetic resonance imaging (MRI) or computer tomography (CT) scans. A higher measurement frequency then provides better insight into plant development, but may also have a stronger negative effect on plant growth.

A second point to consider is the time of day when plants will be measured. When using a drone to fly over an experimental field, it can capture images of many microplots simultaneously. Acquiring images or other data for these plots using sensors mounted on vehicles will take longer. Depending on the type of measurements taken, it may require minutes or more per microplot. With plant-to-sensor systems within a glasshouse, where it takes minutes to move a plant through the imaging station and measure its characteristics, phenotyping all plants in one experiment could take a full day. During that period, large changes in environmental variables, such as light intensity and temperature, may occur, likely resulting in large variation in physiological variables such as stomatal conductance and photosynthesis as well. One morphological trait that can exhibit considerable diurnal variation is leaf angle (Rosa and Forseth, 1996), which has obvious implications for the projected leaf area (PLA; Dobrescu et al., 2017) as used in many HTPP systems. For example, during the diurnal part of the diel cycle, Arabidopsis plants may increase their leaf area and biomass by 20% (Wiese et al., 2007). However, in a similar experiment, PLA values were found to decrease by 18-35% during the light period ( Figures 4A, B , 5 ), due to upward movements of leaves and petioles. Consequently, when conducting consecutive 2D measurements of plants day after day, it is important to measure them at the same time of day, to avoid bias caused by the diurnal rhythm of leaf movements. Moreover, different genotypes or treatments should be blocked into the same time window, to ensure that no confounding effects occur. One fast alternative approach is to perform a 3D laser scan (Dornbusch et al., 2012). Other options include using multiple imaging stations, or employing a gantry system where sensors are brought to the plants, enabling parallel measurements of many plants in short time.

In some cases, the research question requires a high frequency of measurements on the same plants. Examples are physiological responses of plants following the application of a compound, exposure to a pathogen, or exposure to abiotic stress (Jansen et al., 2009; Mahlein et al., 2019). In some species, capturing images at higher frequencies and analyzing them in almost real-time can be used to detect the onset of potentially undesired drought stress during the experiment ( Figures 4C, D ; Eberius and Lima-Guerra, 2009; Janni et al., 2019).

The amount of information collected from a single experiment can be substantial, especially when imaging of any kind is involved. The question is how well we, as experimenters, can handle this vast quantity of data (Eberius and Lima-Guerra, 2009; Tardieu et al., 2017). Based on our own experience, we have observed several issues that can arise during an experiment. Sensors, especially environmental sensors that are deployed throughout the year, may have not been calibrated for a long time or show failures of various kinds during the experimental period. With larger amounts of plants, problems might more easily go undetected. For example, some plants could topple over or do not receive adequate watering. In image analysis systems, leaves from neighboring plants might appear in pictures taken, photos might not cover the full plant, or masking might not function properly ( Figures 2B, C ). These errors are easily noticed if they occur frequently, but in the midst of hundreds of plants, thousands of pictures, and tens of thousands of other collected data points, such errors can easily go unnoticed. Thorough data inspection and double-checking for mistakes are therefore crucial, but can be cumbersome without the assistance of digital tools. Dedicated software programs for data visualization and targeted image retrieval, such as Azure, iRods, PHIS, Fairdom, Zegami, or similar solutions, enable fast selection of images of specific plants over time, aiding in the identification of potential outlier data. Automated quality control procedures should routinely flag instances where parts of leaves are outside the picture, or leaves of neighboring plants are distorting the results. Graphical analysis of data distribution, time courses, or dose-response curves can provide insights into potential issues with specific plants or entire groups of plants (Xu et al., 2015). This is particularly useful in ongoing experiments, when possible problems can be detected and solved by data analysis at an early stage. Real-time reporting, along with easy and user-friendly (remote) access to visualizations and resulting analyses, facilitates early detection and problem-solving.

For decades, scientists have relied on spectrophotometric measurements to asses enzyme activity, wet digestion or pyrolysis for leaf nitrogen determination, infra-red gas analyzers to determine photosynthesis, and manual harvesting to measure plant biomass. However, these conventional measurements all require significant manual effort, and are therefore not suitable for high-throughput phenotyping. Efforts have been made to automate such measurements (e.g. Gibon et al., 2004; Gomez et al., 2010), but challenges remain in automating processes such as grinding and weighing, particularly under low-temperature conditions to prevent chemical degradation (Hall et al., 2022). In search of alternatives, scientists have explored measurements that are easier to perform, yet still provide valuable information. For example, leaf nitrogen content can be estimated non-destructively using multispectral analysis (Ye et al., 2020), photosynthesis can be assessed through fluorescence, and biomass by counting green pixels in plant images.

Are these proxies informative enough? The chlorophyll fluorescence parameter (F_v/F_m), for instance, is often measured, but in many cases variation in F_v/F_m does not reflect variation in the actual rate of photosynthesis (Poorter et al., 2019). The electron transport rate offers a better approximation, but still does not capture the true rate of C-fixation (Kalaji et al., 2014). Similarly, estimating digital biomass based on the number of green pixels in an image can provide an indication of plant size, but does not give the actual biomass, or information on biomass allocation to leaves, stems, and roots. This lack of information hampers comparisons across experiments, platforms, and the published literature. However, advancements in machine learning techniques now enable the segmentation of 2D or 3D images into leaves and stems (Golbach et al., 2016; Jin et al., 2019; Shi et al., 2019), so far only for smaller and/or specific species. Moreover, combined shoot and root phenotyping is feasible in rhizotron or agar-based platforms (Nagel et al., 2012; Nagel et al., 2020). These developments hold the potential to extract more comprehensive information from these photographs or raw sensor data.

In all cases, users must maintain a critical approach to their data, particularly when changes in plant morphology occur. A clear example is observed in drought-stressed plants where a loss in turgor can result in leaf rolling or wilting, leading to a noticeable decrease in projected leaf area (PLA; Figures 4C, D ), while actual dry biomass is little affected. Even when the variable of interest is directly provided by sensors, it is wise to verify the definitions used. For one researcher, plant height may be the distance from the shoot basis to the highest plant part, for another it could be the distance from the basis to the apical meristem. In some non-destructive systems, a more statistical approach is taken, where this variable is defined as the average value of the pixels (or voxels) from the 80^th till the 90^th percentile with respect to height (Kjaer and Ottosen, 2015). Ideally, these definitions are included in the meta-data provided by the HTPP system.

The phenotype of plants is multifaceted, composed of hundreds of variables related to anatomy, morphology, chemical composition, carbon and water economy, growth, as well as reproduction (Lambers and Oliveira, 2019; Poorter et al., 2022; Zafaver et al., 2023). While some of these variables can be estimated non-destructively, the majority of plant traits require destructive sampling or harvesting. Consequently, HTPP systems, which are primarily non-destructive by nature, can only cover a subset of the phenotypic traits that researchers would ideally like to measure. However, by bringing sensors to the plant, additional measurements may become feasible. For example, continuous monitoring of transpiration over a plant’s lifespan in real time can be achieved by placing plants on a balance (Tardieu et al., 2017; Dalal et al., 2020). Nevertheless, many other traits can only be measured through destructive sampling or harvesting, which necessitates additional planning and manpower.

A highly promising advancement is the development of robots capable of approaching a plant and taking a leaf punch from a specific leaf blade (Alenyà et al., 2013; Foix et al., 2018). By promptly storing these samples in liquid nitrogen, a broad array of relevant biochemical analyses can be conducted, including the assessment of key metabolites and RNA expression levels (Hall et al., 2022).

In certain cases, well-calibrated phenotyping equipment can directly provide data on physiologically relevant variables of interest. For instance, measurements such as leaf temperature or F_v/F_m yield output that is readily biologically interpretable and can be easily related to published work in the literature. In other cases, however, additional calibration is required to transform acquired data (e.g., number of green pixels) into biological meaningful variables (e.g., shoot dry mass). In such cases, a common procedure involves periodic measurements of a subset of plants, initially through non-destructive imaging, and subsequently destructively by determining leaf area, shoot dry mass or other traits of interest. An example is shown in Figure 6 , where the projected leaf area of Plantago major plants grown under different [CO₂] levels was assessed non-destructively through imaging, followed by destructive measurement of total leaf area and shoot biomass (see Material & Methods).

The first step in establishing a calibration curve involves plotting the variable of interest against the measured variable. We illustrate this process with a graph that depicts the relationship between total leaf area (TLA) and projected leaf area (PLA). The graph demonstrates that for small plants (< 30 cm² in this case, for plants up to 4 weeks old), TLA and PLA exhibit largely similar values. However, in larger plants, TLA increases at a faster rate than PLA, as newly-grown leaves will inevitably overlap partially or even fully with older leaves. In the experiment presented here, the TLA at the final harvest was approximately 2.3 times larger than the corresponding PLA, with no clear difference between plants of the two treatments.

The second step in constructing the calibration curve involves computing a regression line. Taken over both treatments, a linear regression yielded highly significant results (P < 0.001). Based on the calculated r², we found that variation in PLA accounted for 92% of the variation in TLA. While this initial outcome may appear very satisfactory, further examination showed that the regression line underestimated TLA at very small and high TLA values, while overestimating TLA at intermediate PLA values. Given the gradual increase in leaf overlap with plant size, a curved relationship appears to be a more appropriate model. Subsequent analysis with a second-order polynomial confirmed the high significance (P<0.001) of the quadratic term, resulting in a slightly improved r² ( Table 2 ).

Although many users are satisfied with the aforementioned correlative approach and the high r² values (e.g. Nagel et al., 2012; Vadez et al., 2015; Banerjee et al., 2020), certain aspects warrant further inspection. For instance, the growth of young plants often follows an exponential pattern, characterized by smaller absolute size increases in small plants, and larger increases as plants grow bigger. As a consequence, in the experiment we are discussing with weekly harvests, the first half of the calibration curve is determined by 82% of the observations, while the remaining 18% contribute to the second half. To achieve a more balanced distribution, we could log-transform both PLA and TLA. In our experiment, this log-transformation resulted in the first half of the curve containing 32% of the data, with the remaining 68% in the second half ( Figure 6B ). Although the distribution is still not perfectly equal, it improved considerably compared to the non-transformed dataset. Performing a linear regression on the log-transformed data yielded a highly significant fit, with an r² of 0.986. However, it is important to acknowledge the biological phenomenon of overlapping leaves. Incorporating a quadratic term into the equation further improved the fit, resulting in an r² value of 0.989. Clearly, it pays to analyze which function is most appropriate, and whether log-transformation of the data and/or non-linear fits can provide a more robust basis for the calibration curve than a standard linear regression on non-transformed data.

The coefficient of determination (r²) is a convenient parameter to describe the goodness of fit of a statistical relationship between variables, and because it (generally) scales between 0 and 1, it allows easy comparisons across various experiments (Chicco et al., 2021). The calibration curves discussed in the previous section all exhibit relatively high r² values ( Table 2 ), but does this r² truly represent the desired accuracy? Different expressions of plant size, such as leaf number, leaf area, and total biomass, are generally well correlated. How well the calibration curve works depends partly on the appropriateness of the chosen proxy trait. For instance: a top-view picture might provide more information for a rosette plant, while side views or views from different angles could be more informative for species with a single stem. Furthermore, it is important to note that r² provides information on the total variation in the y-variable that can be explained by the total variation in the x-variable. Hence, in monotonically increasing relationships like the one depicted in Figure 6 , the larger the span in size in both x and y, the higher the r² will be. If we restrict the calculation to plants with a PLA >100 cm² instead of considering all data, the r² value decreases from 0.92 to 0.76. Consequently, a high r² for a calibration curve like the one shown in Figure 6 , indicates that we are able to effectively distinguish between small and large plants. However, if a researcher’s prime interest lies in understanding the variation in final size across genotypes, relying solely on the r² of the full calibration curve may provide a somewhat misleading sense of accuracy.

An alternative measure to assess the goodness of fit is the root mean square error (RMSE), which quantifies the average distance in the y-direction between the observed data points and the fitted line. It gives more weight to points that are further away from the line, compared to those that are closer (Hodson, 2022). The advantage of RMSE is that, all else being equal, it is not influenced by the total variation range in x and y, as is the case with r². Additionally, RMSE provides an absolute error in the units of the Y-axis. If the residuals follow a normal distribution, it informs the researcher that there is a 68% probability that the estimated total leaf area (TLA) deviates by less than the RMSE from the true TLA. However, RMSE may not be suitable for calibration curves that cover a wider range of plant sizes, as the error is not equal for plants of all sizes. For instance, an RMSE of 5 cm² may represent a minor deviation for a plant of 1000 cm², but a huge variation for a plant with a leaf area of 1 cm².

To assess the accuracy of the estimates, we calculated for each plant the absolute percentage error. This involved determining the absolute difference between the actual TLA, and the TLA value estimated from the calibration curve, normalized to the actual TLA measured. These values ranged from nearly 0%, indicating a highly accurate estimate, to over 1000% in the specific case of very young plants where TLA was fitted with a straight line across all plant sizes ( Figure 7A ). In the last case, the actual TLA value was 1 cm², whereas the estimated TLA value was calculated to be -10 cm². This illustrates that even an r² value exceeding 0.90 does not necessarily guarantee accurate estimates for every individual plant. The median absolute percentage error (MdAPE) serves as a useful summary descriptor for non-normally distributed data. For the linear calibration curve, MdAPE was approximately 37% ( Table 2 ; Figure 7A ), indicating that the accuracy fell short of our expectations. However, when utilizing a quadratic fit with log-log transformed values, the MdAPE decreased to 11%, signifying a substantial improvement in accuracy. Additionally, in the quadratic fit, MdAPE values were lower for smaller compared to larger ones ( Figure 7B ), which is logical given the increased leaf overlap in larger plants.

The mean absolute percentage error (MAPE) is increasingly utilized in the field of high-throughput plant phenotyping (e.g. Paproki et al., 2012; Paulus, 2019; Rossi et al., 2022). However, given the log-normal distribution of these data, we advocate for the use of the median absolute percentage error (MdAPE) as a more informative measure of the general accuracy. By employing the MdAPE, we aim to capture a representative summary of the actual accuracy, rather than relying solely on the r² of a calibration curve.

So far, we have considered a common calibration curve for both low and high CO₂ plants. For the relationship between projected leaf area (PLA) and total leaf area (TLA), this approach may seem reasonable as long as the treatment does not influence leaf angle or any other aspect of leaf display. However, what would happen if we aim to use PLA to estimate shoot biomass ( Figure 8A )? The relationship between leaf dry mass and leaf area is known to shift, as plants exposed to elevated CO₂ almost invariably exhibit higher leaf mass per area (LMA; Poorter et al., 2022). Using quadratic polynomials on log-transformed PLA and shoot dry mass, we indeed found different curves ( Figure 8B ). They indicated that for a given PLA, elevated CO₂ plants were 20-30% heavier, although not for the smallest or largest plants. These findings align with the LMA data, which also showed LMA averages to be 20-30% heavier, except for the first and last harvest (cf. Figure 8B in Poorter et al., 1988). However, statistically, this did not show up as a significant effect, neither for CO₂ as a main factor nor for the interaction of CO₂ with the linear and quadratic components of PLA.

Having different calibration curves for different treatments, or maybe for different genotypes, can be quite inconvenient, as it requires more manual harvesting, partially nullifying the intended savings in human effort. What could we do about that? A well-cited method paper by Golzarian et al. (2011) discussed the phenotyping aspects of calibration curves, using the example of an experiment where plants were exposed to salinity or to control conditions. The study found that the two calibration curves were significantly different, with salt-stressed plants exhibiting higher biomass estimates for a given number of pixels compared to control plants. That would fit with the general understanding that salt stress increases leaf mass per area (LMA; Poorter et al., 2009; Kamanga et al., 2023). Golzarian et al. (2011), however, reasoned that salt-stressed plants were smaller, and considered them to be of ‘younger’ age. By adding the factor age to the equation, a single calibration function could be achieved for plants of both treatments. However, this approach mixes plant ontogeny with the direct effects of salinity, and is likely not broadly valid, especially when multiple treatments with varying salinity concentrations would be involved. Consequently, if the treatment of interest affects LMA, leaf angle, or other relevant morphological parameters, different calibration curves for different treatments may indeed be unavoidable.

Another important question to consider is the validity of a calibration curve that has been developed for a particular species, and whether it can be applied to other experiments involving the same species. In growth chambers, where light and temperature stay fixed to the same level, the transferability of a calibration curve seems more likely compared to glasshouses or field settings, where seasonal variation in light and temperature may strongly impact both LMA and stem thickness (and consequently stem mass per projected area). To use HTPP systems more effectively, calibration curves deserve more attention than they got so far. As is custom in many other laboratory methods, it might be good to regularly validate the measuring pipeline with a couple of reference samples that are measured using destructive methods.

Plants grown singly in pots may show quite some variability in plant mass or other traits, which can negatively impact the statistical power to detect differences between genotypes or treatments (Poorter and Garnier, 1996). When planning the size of an HTPP system, it is important to consider not only the number of species or genotypes to be tested, but also the number of replicates per genotype that will be required. Genome-wide association studies (GWAS) or Quantitative Trait Loci (QTL) experiments often benefit more from including additional genotypes rather than increasing the number of replicates per genotype (Zou et al., 2006). If treatments are compared across many genotypes, the large number of plants grown in HTPP systems will provide sufficient statistical power for general conclusions. However, when researchers are also interested in testing specific differences between individual genotypes, the number of replicates becomes more critical. This is particularly true when a calibration curve is used to estimate the values of the trait of interest. Calibration curves with low r² and high MdAPE introduce additional variability on top of the inherent variation that will already be present among plants. In the case of the CO₂ experiment, a t-test conducted at the final harvest revealed that the actual shoot dry mass for the two treatments was only marginally significant (P = 0.10). However, the difference was far further from significance when the shoot dry mass estimates based on PLA values were used (P = 0.24). In situations where HTPP system users are interested in specific contrasts, the utilization of calibration curves implies that they may need to include more replicates than in traditional experiments to achieve the same level of statistical power.

Despite the robotized and computerized systems, HTPP experiments often push limits and compare large amounts of genotypes in a standardized manner. In as far as these experiments have a background in (eco)physiological approaches, with single plants growing in pots under controlled conditions, data from a single experiment is often considered sufficient for publication. However, in agriculture it is generally regarded as the gold standard to repeat an experiment in multiple years or locations, before any importance is attached to the results.

Possibly, HTPP in controlled conditions can be seen as the initial, important step in a two-phase approach. During this first step, a wide range of genotypes or species can be tested, either in their own right or in combination with specific treatment factors. Without being overly concerned about genotypic effects on calibration curves (as mentioned in point 12), this step can be used to identify the worst-performing and best-performing genotypes, simply based on green pixel counts or similar proxy traits (‘forward phenomics’ sensu Mir et al., 2019; cf. Merlot et al., 2002). The most extreme and interesting genotypes, for example those carrying contrasting alleles for important QTLs can then be further investigated in a targeted experiment. This subsequent phase would involve non-destructive phenotyping complemented by more labor-intensive physiological analyses such as gas exchange and chemical characterization on the one hand, and destructive harvesting of both shoot and root biomass on the other hand (Poudyal et al., 2018). Such a two-step approach would also be helpful in screening a wide range of germplasm for contrasting genotypes.