Experiments were conducted in the field at the ICRISAT research station in Sadoré (13°N, 2°E). Four experiments were conducted: two in the dry seasons of 2019 and 2020 (Ds19, Ds20) and two in the rainy seasons of 2019 and 2020 (Rs19, Rs20). These periods were chosen to expose the germplasm groups to high VPD in the dry season (high temperature, low relative humidity) and low VPD in the rainy season (low temperature and high relative humidity). Trials were planted in 2019 on end March and end July and were repeated at approximately the same dates in 2020. The experimental layout was a completely randomized block design with four replications for each genotype. Plot was 1 m × 3 m with a plant density of 15 plants/plot.

Experimentation during the dry seasons of 2019 (Ds19) and 2020 (Ds20) was conducted from March to June in both years. All plants were regularly irrigated to avoid any water stress until harvest. Rainy season trials in 2019 (Rs19) and 2020 (Rs20) were conducted under the rainfall regime from July to October. When the rain was interrupted, an irrigation was applied if necessary to avoid soil water deficit for the crop. Temperature and relative humidity (RH) were recorded with a data logger (Tiny tag) installed in the field. Recorded temperature and RH data were used to calculate the VPD with the following formula:

The photoperiods and radiation data, which have great impact on pearl millet phenology, were obtained from NASA.

To score the VPD effect on growth of the three germplasm groups, agronomic characteristics were assessed during the experiment and at maturity. The stem height was measured and the number of tillers was counted until the head emergence. On the date of female flower appearance, three representative plants were harvested in each plot to measure the leaf area using a leaf area meter (LI-3100C area meter, LI-COR BIOSCIENCES, USA). The stems and leaves of harvested plants were oven dried for 72 h at 70°C to determine the dry biomass. The grain yield was estimated at maturity on three plants representative of each genotype chosen in the center of the plot.

The different seasons received different amounts of radiation, in part because of cloud cover difference but also because of season duration differences. This could have an impact on agronomic performance, and normalizing by the cumulated radiation is a standard practice for comparing agronomic trials (Sadras et al., 2015). Therefore, a first normalization of the data was done by dividing, for each trial, each replication-wise value (for a given genotype and variable) by the cumulated radiation received during the season. The data were then expressed in their usual unit per MJoule (MJ⁻¹). The software R (R: A language and environment for statistical computing, R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/) was used to perform the statistical analysis.

The seasonal measurement of temperature ( Figure 1A ) and RH ( Figure 1B ) showed that the VPD mean values across the four experiments varied between 3.62 kPa (highest-VPD season) and 0.61 kPa (lowest-VPD season, Table 1 ). Dry season trials that were characterized by high temperature and low relative humidity (RH) had high VPD ( Figure 1C ). Ds19 was the hottest and driest season, whose average VPD was 3.62 kPa, whereas Ds20 was a bit cooler/wetter (2.92 kPa) due to an early onset of the rainy season. The rainy seasons had much lower VPD because of rainfall and lower temperature. The mean VPD of Rs19 was 1.14 kPa whereas that of Rs20 was slightly lower (0.61 kPa). Solar radiation was also lower in the two rainy seasons than in the two dry seasons ( Figure 1D ). Precipitations ( Figure 1E ) and wind ( Figure 1F ) varied largely among seasons.

After data normalization with radiation, it appeared that the dry season performance was lower than the rainy season one (when comparison for each year was made separately) for key traits such as leaf area and grain yield. As such, the radiation-normalized leaf area was lower in the dry season than in the rainy season ( Figure 2A ) and this in both years. Similarly, the radiation-normalized seed weight was lower in the dry season (Ds19, Ds20) than in the rainy season in both years ( Figure 2D ), whereas stem height and tiller number changed in opposite directions ( Figures 2B, C ).

However, the high performance of the 2020 trials over the 2019 trials suggested a possible field effect on millet performance as trials were conducted in different fields, i.e., one field for the Ds19 and Rs19 trials, and another field for the Ds20 and Rs20 trials. This made the comparison of season effect on trait performance difficult because of that possible field effect. Due to a lack of data on soil chemical composition that could have explained the differences in field performance, we performed a two-way ANOVA with year (2019 and 2020) and season (Ds and Rs) as factors to check for an eventual field effect on data collected across seasons. The year effect represented the field effect, whereas the season effect could be explained by the differences in VPD. The ANOVA showed a highly significant field effect on growth and yield data and also an interaction between the field and year for seed weight ( Table 2 ). These interactions indicated that, at least for seed yield, certain germplasm perform better in specific year-season combinations.

To eliminate this field effect on plant performance, and to be able to robustly assess the season effect on plant performance, we then proceeded to a second normalization for each trial, using the data that came out from the first normalization done above against season’s cumulative radiation. This second normalization was based on the grand mean of each of the traits that were measured. These grand means of plant performance for the different traits measured (for example leaf area or yield) represented the overall trait performance in a given field for the whole set of genotypes. Therefore, the second normalization consisted, in a given trial, in dividing replication-wise values (for all genotypes and for a given trait) by the grand mean of the same trait for the considered trial. This second normalization was a way to remove the field effect on that performance, and then to keep only the relative genotypic differences. Therefore, these double-normalized data allowed to compare genotypic performances within trials, and also to test season effects across trials. It was a necessary step to correctly compare the groups of genotypes with regard to the response of growth and agronomic parameters to season effects. The data that were analyzed for the remaining part of the paper were then unit-less data that represented the variation against a grand mean for a given variable.

Figure 3 represents the performance of the three germplasm groups across the four trials, which is without considering any season in particular. Results showed large differences among groups of germplasm. For instance, leaf area was significantly higher in G3 than G2 and G1 (1.29, 0.92, and 0.79, respectively, p value < 0.0001) ( Figure 3A ). Similar results were observed for basal stem height and seed weight (p value <0.0001 and 0.0001, respectively) with G3 showing the highest values ( Figures 3C, D ). By contrast, the high tillering ability was lower in G3 and G2 as compared with G1 (p value < 0.0001) ( Figure 3B ).

The ANOVA shown in Table 2 revealed significant season-by-field interactions, indicating that certain germplasm performed better in specific seasons. Hence, Figure 4 compares the performances among germplasm groups within each of the seasons. Results indicate that across all seasons (with high and low VPD), leaf area was significantly higher in G3 than in G1 and G2 ( Figure 4A ). By contrast, the tiller number was significantly higher in G1 in the two low VPD seasons (1.14 kPa and 0.61 kPa), and then in G2 in the one high VPD season (3.62 kPa), as compared with G3 ( Figure 4B ). Stem height showed no clear group difference across the different seasons, except for the higher stem height of the G3 group than G1 and G2 in the two low-VPD seasons (1.14 kPa and 0.61 kPa) and one high-VPD season (2.92 kPa) ( Figure 4C ). Notably, while there was no significant yield difference among the three germplasm groups in the two rainy seasons, G3 had a higher grain yield than G1 in both dry seasons, and then G2 in one of the two dry seasons ( Figure 4D ).

Figure 5 represents a comparison within the germplasm group of their performance across the four seasons. Figure 5A shows that across the four environments (Ds19, Rs19, Ds20, Rs20), only G3 showed a variability for leaf area with a significant decrease in Rs20, which was one of the two low-VPD environments. By contrast, even though the leaf areas for G1 and G2 seem to respond more favorably to a VPD decrease, this difference was not statistically significant ( Figure 5A ). Results revealed that when the VPD is above (2.9 kPa), the tillering ability of G1 decreased whereas the environmental change seemed to have no effect on G3 tiller number ( Figure 5B ). Unexpectedly, a high tiller number was observed for G2 under high VPD in the dry season (Ds19) ( Figure 5B ). Results revealed also that under high-VPD conditions, the seed weight was reduced in G1 whereas it tended to increase in G3, although the statistical test showed that this increase was not significant ( Figure 5B ).

Figure 6 is a principal component analysis for each of the four trials. The first two vectors explained 75%–80% of the variation. Clearly, the first vector highlighted high yields and came with high positive loading of leaf area and to some extent stem height. By contrast, tiller number had no weight on this vector. Germplasm having high loading on that vector were mostly group 3 germplasm ( Table 3 ), especially in the trials taking place in dry years ( Figures 6A, C ), for instance IP13840, IP5441, IP6460, IP10543, or IP18168.

The phenotypic variation observed in crop yield includes genetic and environmental and the interaction between them (Hamidou et al., 2013), although in many cases, the largest sources of variation come from the environment (Sadras et al., 2015). The 2019 and 2020 trials were conducted on different fields inside the research station. While no information on the chemical and mineral composition of these soils was available—it was assumed that the two fields had similar fertility levels. However, the two-way ANOVA with year (2019 and 2020) and season (Ds and Rs) as factors showed that there was a field effect on the measured variates ( Table 2 ). In similar trials which are conducted on different fields and in several seasons, classifying varieties according to yield difference (or any measured variate) is likely to be biased (Sadras et al., 2015). This is why it was important to normalize the data in order to separate the part of variation related to fields and that due to VPD. In addition, another normalization was done by dividing the first normalization data by the grand mean for each of the variables. The normalization by the cumulated radiation then by the grand mean of each trial allowed to compare crop performance as affected by other factors such as VPD, independently from the field and the radiation effects.

In pearl millet, both temperature and photoperiod are known to influence growth (Begg and Burton, 1971; Sanon et al., 2014; Vadez et al., 2012b). In our experiments, the average temperatures (34.29°C, 33.29°C, 28.9°C, and 26.76°C, respectively, for Ds19, Ds20, Rs19, and Rs20) were within the limit of good growth of millet (Ong and Monteith, 1985) as well as radiation (Singh et al., 1998). Rain was not deficient during the two rainy seasons (Marteau et al., 2011), and the trials were conducted without any water deficit in the soil. Table 1 shows that the two dry seasons differed very little in terms of average temperature (34.29°C and 33.29°C) and radiation (23.27 MJ m⁻² day⁻¹ and 22.96 MJ m⁻² day⁻¹), so did the two rainy seasons 28.9°C 26.76°C and 20.51 MJ m⁻² day⁻¹ and 20.68 MJ m⁻² day⁻¹ for mean temperature and radiation, respectively. The photoperiod was almost the same for all seasons (13 h). Hence, most differences came from VPD and genotypes.

After first normalization, growth and yield were higher in the season with the lowest VPD, indicating a positive effect of the low-VPD condition on both growth and yield in pearl millet. Several reports show a positive effect of low evaporative demand on growth (Darlington et al., 1997; Devi et al., 2015; Leuschner, 2002; Yuan et al., 2019). Leaf expansion of maize leaf number 6 is sensitive to high VPD, and there is genetic variation for that trait (Reymond et al., 2003; Welcker et al., 2006). Here, our measurements were coarser since they were done in the field in a destructive manner. Yet, they were able to pinpoint a VPD effect on the leaf area development after normalizing field and radiation differences.

High evaporative demand has an impact on both growth and yield process depending on the germplasm group. The G1 group showed a decrease in leaf area and seed weight under high VPD. This could be explained by the leaf expansion sensitivity to high VPD which decreased leaf area and by the same occasion radiation interception then yield. In addition to less radiation capture, the yield decline in G1 group could also be temperature-related, since the reproduction period is sensitive to high temperature, and high VPD and high temperature often occur together (Hamidou et al., 2013). Figures 1A, C indeed show that high temperature occurred during a period when most entries came to flowering (between 40 and 50 DAS, data not showed). So, the drastic yield decrease during high-VPD season may be due to reproductive sterility and/or flower abortion. These results are similar to previous works, which reported that high VPD had a negative effect on seed set at higher temperature in crops including pearl millet (Hamidou et al., 2013; Gupta et al., 2015). G2 did not show any clear response to VPD but was slightly less tolerant to high VPD as compared with G3.

In this work, tolerance to high-VPD conditions was akin to sustaining leaf area and stem height under high VPD conditions, as this was related to higher yield ( Figure 6 ), which also validated our second hypothesis. From this standpoint, the G3 group would be the most tolerant one as it showed the highest yield (biomass and grain) under the highest-VPD seasons ( Figure 4D ) as compared with G1 which decreased the same traits during the same seasons. These results are similar to previous works on water stress (Blum and Sullivan, 1986). These authors reported that drought resistance, measured as the degree of growth inhibition under stress, was higher in races from dry regions than in races from humid regions in a landrace comparison of sorghum and pearl millet from dry and humid regions. The decrease of leaf area in G1 and G2 is a morphological adaptation that enables them to withstand high-VPD periods. Our results differed from Zhang et al. (2005) who compared a population of Populus divana to different soil water deficit intensities. They showed that water stress affected dry matter accumulation and allocation more in the dry climate ecotype than in the wet climate ecotype. Although the two types of droughts, atmospheric drought in our experimentation and soil drought in Zhang et al.’s study, are different, the consequences on plant were reported to be similar (Sulman et al., 2016).