Trait data collection was conducted at four coffee plantations in Java Island, Indonesia (Figure 1; Table 1). The first site (‘AE-W’, for Arabica hybrid–Eucalyptus–West Java) is located at Riung Gunung, Pangalengan, Bandung Regency, West Java (7° 9’58.25”S, 107°31’8.47”E). The coffee farm belongs to Java Frinsa Estate and is situated at the foothill of Mount Tilu. The coffee variety grown at this site is Timor hybrid (‘hibrido de timor’), which is a natural cross of arabica and robusta originating from East Timor in the 1920s, known to be resistant to leaf rust. Timor hybrid’s leaf length is typically around 21 cm, with greenish-brown young leaves and dark green mature leaves (Budiasih et al., 2025). The coffee plants are grown at 1 × 3 m spacing together with timber species shade trees with dense canopy (83.4% cover), such as Eucalyptus sp. (78 trees/ha; 33 m height) and Agathis dammara saplings (78 trees/ha; 4 m height). This site has the highest elevation (1621 m) among all sites, making it the coolest and wettest site. This site is also the most intensely managed, where it receives around 750 kg/ha of nitrogen-phosphorus-potassium (N-P-K) fertilization and 500 kg/ha of dolomite per year, and is regularly pruned to ca. 1.75 m height and weeded after the harvesting season.

The second site (‘AM-W’, for Arabica–Mixed trees–West Java) is located at Pasir Sereh, Cisurupan District, Garut Regency, West Java (7°17’28.44”S, 107°45’56.25”E), which is situated on the foothill of Mount Papandayan. This site grows arabica var. Typica, which was historically originated from in Malabar, Java in the 18^th century. This variety has low yield and is less resistant to leaf rust, but is adapted to high elevation (World Coffee Research, 2025). Arabica var. Typica generally has a leaf length of around 16 cm and dark green-colored mature leaves, and could naturally grow up to 3–5 m (Budiasih et al., 2025). Coffee plants at this site are grown together with mixed fruit tree species as shade trees, such as citrus (133 trees/ha; 3.44 m height), jackfruit (48 trees/ha; 5.22 m height), and mango (26 trees/ha; 16.55 m height). It is characterised by a reduced shade (31% canopy cover) and minimum management measures, where coffee plants are not pruned but given organic fertilizer once or twice a year during the flowering period. Further details on shade tree characteristics are provided in Supplementary Materials, Supplementary Figure 3.

The third site (‘AP-E’, for Arabica hybrid–Pine–East Java) is located at Brawijaya University education forest (UB Forest), Sumbersari, Karangploso, Malang Regency, East Java (7°49’28.10”S, 112°34’29.06”E) at the foothill of Mount Arjuna. This site is characterized by the dense canopy (94% cover) of pine trees (ca. 800 trees/ha) with ‘Catimor’ coffee growing underneath. ‘Catimor’ variety is a cross between Caturra (C. arabica) and Timor (C. arabica x C. canephora), and is resistant to coffee leaf rust and berry diseases (World Coffee Research, 2025). This variety generally has a leaf dimension of 18.6 × 7.1 cm, with plant height around 140 cm (Budiasih et al., 2025). Coffee plants at this site were recently rejuvenated, but with low management intensity.

The fourth site (RM-E, for Robusta–Mixed trees–East Java) is located in Jabon village, Ngantang District, Malang Regency, East Java (7°50’39.99”S, 112°22’48.06”E). Coffee plants grown here are robusta clone of BP (Besoekisch Proefstation), which is one of the national superior clones suitable for areas with elevation along 400–900 m. BP clones have green-colored elliptical leaves with wavy surface, with a length distinctively larger than arabica leaves of ca. 24 cm (Ramadiana et al., 2018). This site is characterized by robusta coffee grown under moderate canopy (71% cover) of mixed fruit and mixed timber shade trees, such as durian (81 trees/ha, tree height 5.81 m), avocado (74 trees/ha, tree height 6.57 m), cacao (85 trees/ha, 4.12 m). This site is situated at a lower elevation (725 m) and has a higher mean temperature compared to other sites. The management intensity is medium, where coffee plants are pruned and given organic fertilizer once or twice a year. At each site, we set up three plots with a size of approximately 30 × 30 m². Within each plot, we randomly chose five healthy coffee plants (i.e., no visible damage and minimum discoloration of leaves) to measure their functional traits. To capture the seasonality, we conducted a round of measurement in the dry season (September – October 2024) and another round in the wet season (February – March 2025), which coincides with the flowering and the fruit-filling period, respectively. The measurement was conducted over a period of one week per site and season. For the AE-W site, four individuals were cut for thinning in the middle of the study period, so we replaced them with the neighbouring individual for the wet season measurement.

At each plot, we installed a Temperature-Moisture-Sensor (TMS) datalogger (TOMST, Prague), which measures air temperature at 15 cm above the soil surface (T_a,15cm), surface temperature, and soil temperature and moisture at 8 cm depth from September to March 2025. Unfortunately, all three TMS dataloggers at the AE-W site and one at AM-W went missing, and their data could not be retrieved. To address these gaps, we estimated missing environmental variables using ERA5-Land hourly data (Muñoz-Sabater, 2019) for the period 1 April 2024 to 31 March 2025. ERA5-Land is a global gridded land-surface dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) at 0.1° spatial resolution. We specifically estimated:

To measure leaf physiological traits, we selected one healthy, sunlit, fully expanded leaf from the upper 30% of the plant canopy to minimize variation associated with leaf position and uneven sun exposure throughout the day. Physiological trait measurements were conducted on site on one attached leaf from five coffee plants per plot. We measured light response curves (LRCs) and CO₂ response curves (A/C_i) using a portable gas exchange system Li-6800 (LiCor, Lincoln, NE, USA). LRC was recorded under a reference CO₂ concentration of 400 μmol mol^-1, at a leaf temperature of 25 °C, and relative humidity (RH) between 50–60%. To build LRCs, we measured CO₂ assimilation rate (A) under 14 decreasing steps of photosynthetic photon flux density (PPFD), at 1200, 1000, 800, 700, 600, 500, 400, 350, 300, 250, 200, 150, 100, 50, 0 µmol m⁻² s⁻¹ following (Xiao et al., 2024). From the LRCs, we inferred dark respiration (R_d), light-saturated photosynthetic rate per unit area (A_max), light-saturated photosynthetic rate per unit dry mass (A_mass), transpiration (E), and stomatal conductance (g_sw) at saturating PPFD and calculated intrinsic water use efficiency (iWUE) as A_max/g_sw, and instantaneous water use efficiency (WUE) as A_max/E. We also derived photosynthetic quantum yield (α) by fitting an exponential model to the LRCs (Bassman and Zwier, 1991; He et al., 2024).

When measuring A/C_i curves, we set the PPFD to 800 µmol m^-2 s^-1, where photosynthesis of most coffee plants appears to saturate without signs of photoinhibition. We then measured A at 15 steps of reference CO₂ of 400, 300, 200, 100, 150, 250, 350, 400, 450, 550, 650, 750, 850, 950, 1100 μmol mol^-1. In total, we obtained 60 coffee leaf LRCs and 60 A/C_i curves for each season (note: 55 A/C_i curves for the wet season due to equipment malfunction). Using the ‘Plantecophys’ package in R (Duursma, 2015), we estimated photosynthetic parameters, maximum carboxylation rate (V_cmax)and maximum electron transport rate (J_max), by fitting the Farquhar-von Caemmerer-Berry (FvCB) model of leaf photosynthesis (Farquhar et al., 1980) to the A/C_icurves.

After collecting the leaf physiological traits data, we sampled the measured leaf along with four additional leaves from the same plant to measure their structural traits. We measured the leaf thickness, leaf wet mass without the petiole, and leaf area. Leaf thickness was measured using a 0–25 mm digital micrometer (Mitutoyo, IL, USA) averaged across three locations of the leaf. To obtain leaf area, we scanned the leaf using a digital scanner and estimated the area using ImageJ software (Schneider et al., 2012). We then dried the leaves in the oven at 70–80 °C for 84 hours, recorded their dry mass, and calculated their LMA as leaf dry mass divided by leaf area, and leaf dry matter content (LDMC) as dry mass divided by wet mass.

The dry leaves were then ground for further chemical analysis. Leaf carbon (C) content, leaf nitrogen (N) content, C isotope (δ¹³C), and N isotope (δ¹⁵N) were analysed using the Elemental Analyzer-Isotope Ratio Mass Spectrometry (EA-IRMS) technique (EA: EuroVector, model EA3000; IRMS: Nu Instruments, model Perspective). In total, we measured the total C and N content of 300 coffee leaves for each season. We also analysed leaf phosphorus (P) content, but only for the 60 leaves that were measured for their physiological traits. For P analysis, we conducted acid digestion using HNO₃ on the dry ground leaves. P contents were then analysed using Inductively Coupled Plasma Optical Emission Spectroscopy (ICP-OES) (Agilent).

In addition to leaf traits, we measured plant height, stem diameter at 15 cm above ground, canopy cover above the plant, and the number of fruits per plant. The canopy cover above the coffee plant was measured using a densiometer (0 – 100%, with higher percentages having more shade). Fruit counts were estimated following the protocol described by Cilas and Descroix (2004) using a multiplicative method based on plant architecture:

Fruit count per plant=(number of stems per plant)×(fruit bearing branches per stem)×(glomerules per branch)×(fruits per glomerule)

For each plant, the number of stems was counted directly. One representative stem was selected per plant to assess reproductive architecture. On this stem, we counted the number of fruit-bearing branches, then randomly selected four of these branches to estimate the number of glomerules per branch. On each of the four branches, two glomerules were randomly chosen (eight glomerules per plant in total) to count the number of fruits per glomerule.

For each plant, leaf traits were averaged across sampled leaves to obtain individual-level trait estimates. For site-level comparisons, we further averaged these individual values across the two sampling seasons (wet and dry) to derive a single site-level trait mean. Descriptive statistics (mean and (plant-level) standard error) were then calculated for each trait at each site. Trait variability was quantified using the coefficient of variation (CV) calculated at two scales: (1) the overall CV across all plants in the dataset, and (2) the site-level CV based on plants within each site. To test whether site-level trait variation differed significantly from overall trait variation, we used a null model approach based on a bootstrapping procedure. Specifically, we generated 500 randomized datasets following the method described in Buchanan et al. (2019) and compared observed site-level CVs against the distribution of bootstrapped values. To assess seasonal effects on traits, we compared within-site trait means between wet and dry seasons using a one-way analysis of variance (ANOVA), followed by a post-hoc pairwise T-test. All statistical analysis was carried out using R 4.3.2 (R Core Team, 2023).

To examine the coordination of leaf traits, we conducted a principal component analysis (PCA) using ‘prcomp’ function within the ‘stats’ package using the plant-level data aggregated across sites and seasons. We excluded traits that are highly correlated with each other (Supplementary Material, Supplementary Figure 2). Thus, the traits used in the PCA were LMA, LDMC, V_cmax, A_mass, g_sw, leaf N, leaf C, leaf K, and leaf P. We then evaluated the significance of the PCA by comparing the observed correlation structure to that expected by chance (Björklund, 2019) using the ‘PCAtest’ package (Camargo, 2025). Using the resulting two principal components (PCs), we analysed the relationship between trait strategies, canopy cover, and reproductive output. We fitted the scores of the two PCs to canopy cover and fruit count using a linear model. We tested the linear model with and without season as an interaction term using analysis of variance (ANOVA).

We also assessed pairwise correlations among leaf traits and compared these relationships with those reported in other published datasets. Specifically, we queried the TRY plant trait database (Kattge et al., 2019) for records of Coffea arabica and Coffea canephora that included at least one combination of LMA, leaf N, and A_mass. The search resulted in three datasets (Martin et al., 2017; Buchanan et al., 2019; Gagliardi et al., 2023). For each trait pair, we then fitted a linear model across all datasets, including ours, and included dataset identity as an interaction term to test whether trait relationships differed among studies.

The four sites showed clear variation in environmental conditions (Figure 2). AE-W was the coolest and wettest among all sites, with a mean T_a,15cm of 17.88 °C on average (range: 13.77 – 25.64 °C), and total precipitation of 4052 mm/yr during the study period. AM-W experienced similar annual precipitation (3673.7 mm/yr), but a higher average T_a,15cm of 20.52 °C (range: 13.56 – 27.44 °C) than AE-W. In contrast, the two East Java sites (AP-E and RM-E) were at lower elevations and received less annual precipitation—3174mm/yr and 3382 mm/yr, respectively. Among them, AP-E had a cooler T_a,15cm (20.36 °C, range: 15.18 – 24.76 °C). RM-E was the warmest site overall, with a mean T_a,15cm of 23.18 °C (range: 18.18 – 29.18 °C).

Seasonal variation in temperature was relatively weak at all sites, while precipitation and soil moisture showed stronger seasonality. AP-E and RM-E experienced a pronounced dry period from May to October, in which the minimum monthly precipitation reached 13 mm/month for AP-E and 14 mm/month for RM-E, compared to the wetter minimum monthly precipitation in West Java (AE-W: 44.74 mm, AP-W: 21.72 mm). However, despite the lower rainfall, AP-E retained soil moisture more effectively during the dry season. Overall, the two less-shaded sites (AM-W and RM-E) exhibited larger seasonal variations in soil moisture, while the two shaded sites had weaker seasonality, reflecting the buffering effect of shade trees, regardless of the amount of precipitation they receive.

We examined site-level means (Table 2; Supplementary Material, Supplementary Figure 3), as well as the extent of variation in coffee leaf traits among coffee plants of all sites (overall CV) and within each site (site-level CV) (Figure 3). We found that structural traits showed low within-site variation but clear site differences in trait means, while chemical and physiological traits showed larger within-site variation with limited between-site differences in trait means.

Structural traits (e.g., LDMC and LMA) exhibited small within-site variation (site CV<20%), but showed clear differences between sites. In particular, leaf area and dry mass are the largest at the robusta site (RM-E), followed by the hybrids and arabica sites. In contrast, leaf thickness and LMA are the highest at AM-W. Within-site variations in structural traits for all sites are significantly smaller than expected by chance.

In contrast, leaf chemical traits have a large range of within-site variation, in which leaf N and C:N is less varied (site CV< 20%), while leaf P, Mg, and K content are more varied (site CV ~30%). Leaf chemical (N, C:N, and P) shows similarity between three sites (AE-W, AM-W, and RM-E), while AP-E shows greater N and P and lower C:N. Within-site variations in leaf chemical traits are significantly smaller than expected by chance for AE-W.

Leaf physiological traits (V_cmax, J_max, and A_mass) have the largest variation (CV>30%) among all traits. The East Java sites (AP-E and RM-E) have significantly higher A_mass, A_max, J_max, V_cmax, E, and g_sw, and lower R_d than the West Java sites (AE-W and AM-W). At the same time, the East Java sites also exhibited significantly smaller within-site variation than expected in most physiological traits, while α and WUE did not differ across sites.

We also found that traits vary between dry and wet seasons, although the patterns are not consistent across sites (Figure 4). The arabica (AM-W) and hybrid (AE-W and AP-E) sites had similar seasonality in leaf structural and chemical traits, while the robusta coffee site (RM-E) experienced a more pronounced seasonality in physiological traits. In the wet season, AE-W experienced a significant decline in leaf area, leaf thickness, and leaf dry weight (p<0.05). Since the decline in leaf dry mass is larger than that of leaf area at AE-W, it resulted in a significant decrease in LMA. This is accompanied by a lower leaf N and a higher leaf C:N during the wet season. Similarly, AM-W and AP-E also followed this pattern, although AM-W had a significant decline in LMA (p<0.01) with no significant increase in C:N, while AP-E had the opposite pattern. For RM-E, both leaf area and leaf dry mass declined in the wet season, resulting in no significant change in LMA. Instead, RM-E experienced around a 40–50% increase in A_max and V_cmax(p<0.05). While this increase in photosynthetic capacity is not accompanied by an increase in leaf N, we observed an increase in leaf P and leaf C:N at RM-E in the wet season.

We examined the trait coordination at our sites using PCA of all traits (Figure 5), in which the first two principal components (PC) together accounted for 61.72% of the trait variation. PC1 explained 42.89% variation in leaf traits and reflects traits associated with leaf-level carbon gain, where it is positively correlated with A_mass, V_cmax, and g_sw; and is negatively correlated with LMA and LDMC. PC2 explained 18.83% of the variation in leaf traits, and is positively correlated with leaf chemical traits (leaf K, N, P); and is negatively correlated with g_sw, V_cmax, and leaf C.

The two sites in West Java (AE-W and AM-W) overlapped with each other in the two-dimensional trait space, and are characterized by a negative PC1 (high LMA and LDMC; low photosynthetic capacity), although the variation along the PC1 axis is larger for AE-W. In contrast, the arabica site in East Java (AP-E) is characterized by a positive PC1 (high photosynthetic capacity and low LMA). Meanwhile, the robusta site (RM-E) is characterized by a negative PC2 (high g_sw and leaf C), although the variation along this axis is high.

To understand the drivers for the differences in trait strategies between sites, we evaluated the relationship between canopy cover of individual coffee plants and their PC1 scores (Figure 6). Canopy cover correlates positively with PC1 scores when data are aggregated across all sites and seasons (R²=0.35, p<0.05; Figure 6). This indicates that coffee trees grown under higher canopy cover tend to have higher A_mass and V_cmax, but lower LMA and LDMC. PC2 scores, however, have no significant relationship with canopy cover.

We also evaluated the relationship between PC scores and the estimated number of fruits of individual coffee plants. The number of fruits is negatively correlated to both seasons’ PC1 score, although the relationship is stronger between fruit and the dry season PC1 score (R²=0.21, p<0.05). This indicates that coffee plants with higher A_mass and V_cmax but lower LMA and LDMC tend to have fewer fruits. Similarly, the number of fruits also correlates negatively with the dry and wet season PC2 scores (R²=0.098–0.2; p<0.05). This indicates that plants with higher g_sw but lower leaf chemical traits are more likely to have less fruit.

Leaf structural, physiological, and chemical traits in our study showed substantial differences across sites. In particular, we observed clear site-level differences in LMA, leaf area, A_max, V_cmax, and to a lesser extent N and P. The observed values broadly fall within the reported range for arabica and robusta coffee from other studies in South and Central America, Africa, and Thailand. For example, our average A_max ranges between 4–7 µmol CO₂ m² s⁻¹, within the reported range for arabica (3–9 µmol CO₂ m² s⁻¹) (Bote and Jan, 2016; Rodrigues et al., 2016) and in the lower end for robusta (5–15 µmol CO₂ m² s⁻¹) (Rodrigues et al., 2016; Chiarawipa et al., 2025). Our leaf N (2.55–3.18%) is at the lower end of the reported range in arabica (2.1–4.4%) (Martin and Isaac, 2021; Gagliardi et al., 2023), possibly due to lower fertilizer inputs and nutrient availability. Arabica leaves showed the highest LMA (~113.27 g m^-2), followed by the hybrids and robusta (81.52–94.7 g m⁻²), all within the reported range of both species (arabica: 36–110 g m⁻², robusta: 78–93 g m⁻²) (Bote and Struik, 2011; Rodríguez-López et al., 2013; Buchanan et al., 2019; Gagliardi et al., 2023). V_cmax and J_max at our robusta site are within the range of reported values of 8.7–109 µmol CO₂ m² s⁻¹ for V_cmax and 30–102 µmol m² s⁻¹ for J_max (Rodríguez-López et al., 2013; Chiarawipa et al., 2025). These comparisons suggest that while our systems fall within global trait spectra, they may be constrained at the lower end by nutrient and management limitations.

We found that structural traits showed low overall variation among individuals, while chemical and physiological traits showed greater variability. This pattern reflects how traits respond to environmental heterogeneity: structural traits are less sensitive to variation in environmental conditions because they are largely determined during leaf development and thus more costly (Grime and Mackey, 2002; Stotz et al., 2021); chemical traits respond to moderate variability in soil condition; and physiological traits are sensitive to the spatial variability of microclimate (Martin et al., 2017; Odé et al., 2025). At the site level, trait variation was generally low, consistent with other studies on arabica coffee, where homogenous management constrains divergence among individuals compared to natural systems (Martin et al., 2017; Buchanan et al., 2019). However, notable exceptions exist. For example, our A_mass variability was the highest for site AM-W, which has the least shade, in contrast to Buchanan et al. (2019), who found the highest A_mass variability under permanent shade. This is likely related to the more heterogeneous light availability at our site due to the varying canopy cover (31.18 ± 6.13%), despite the low mean shade.

Studies on evergreen species show mixed evidence of trait seasonality, with certain traits exhibiting larger variation between seasons than others. At the species level, Gotsch et al. (2010) reported an average decrease in LMA and an inconsistent change in leaf N in the wet season, whereas at the community level Bloomfield et al. (2018) found that seasonality explained a substantial proportion of variation in V_cmax but not in LMA and leaf N. Such responses are often species-specific, and in coffee, previous studies have demonstrated an apparent seasonal shift in photosynthetic traits of both arabica and robusta (Silva et al., 2004; Baliza et al., 2012; Rodrigues et al., 2016).

In our study, seasonal patterns of trait variations differed between coffee types. At the robusta site (RM-E), A_max and V_cmax increased in the wet season, accompanied by an increase in leaf P, and no detected changes in leaf N and LMA. The positive correlation between A_max and leaf P (but not leaf N) suggests that photosynthetic capacity may be P-limited. This limitation may be exacerbated by reduced P uptake under RM-E’s particularly low dry-season soil moisture (Radersma et al., 2005). Additionally, the influence of P on photosynthetic traits has been highlighted by cross-species studies, which reported a larger sensitivity of A_max and V_cmax to leaf N under higher P (Reich et al., 2009; Walker et al., 2014).

In contrast, our arabica and hybrid sites showed no changes in photosynthetic traits between seasons. Instead, we observed a decrease in structural traits (LMA, leaf area, and leaf thickness) accompanied by a decrease in leaf N during the wet season. Higher LMA in drier conditions is commonly linked to thicker leaf cell walls to reduce transpiration (Matos et al., 2009; Buchanan et al., 2019). Accordingly, lower LMA in the wet season is accompanied by lower leaf N, since N makes up a substantial proportion of the cell wall (Onoda et al., 2017). The decrease in leaf N could also reflect translocation of photosynthates to fruits in the wet season, which coincides with the fruit-filling stage.

The absence of A_max seasonality at our arabica sites contrasts with findings from Southeastern Brazil, where both arabica (Silva et al., 2004; Baliza et al., 2012; Rodrigues et al., 2016) and robusta (Rodrigues et al., 2016) show higher A_max during the wet season. These contrasting results likely reflect differences in seasonal climate drivers: in Southeastern Brazil, the wet season coincides with higher radiation and warmer temperatures, whereas in Java the wet season brings cooler, cloudier conditions that may suppress photosynthesis despite abundant water.

More than half of the trait variation in our coffee plants was captured in a two-dimensional plane, with the primary axis aligning closely with the Leaf Economic Spectrum (LES) framework (Wright et al., 2004). One end of the LES represents coffee plants with a quick return on investment or ‘resource-acquisitive’ plants, as noted by the high photosynthetic capacity, and low LMA and LDMC (positive PC1). The other end of the LES represents plants with slow return of investment or ‘resource-conservative’ plants, as noted by the high LMA and LDMC, and low photosynthetic capacity (negative PC1) (Figure 5). PC2 represented a nutrient-stomatal trade-off, contrasting plants with high g_sw and photosynthetic capacity but low nutrient, against those with lower g_sw and photosynthetic capacity but higher nutrient.

Pairwise trait correlations among leaf traits (leaf N – A_mass – LMA) of coffee in our study largely follow the global pattern of LES of natural plants (Wright et al., 2004) and that of other coffee studies (Martin et al., 2017; Buchanan et al., 2019; Gagliardi et al., 2023) (Figure 5). As predicted by LES, leaf N was positively correlated with A_mass, since much of leaf N is invested in Rubisco. However, the N–A_mass relationship (r=0.27, p< 0.05) was weaker than in the global LES (Figure 7a), consistent with previous reports for coffee (Wright et al., 2004; Martin et al., 2017; Buchanan et al., 2019). A substantial proportion of leaf N may reside outside photosynthetic enzymes or cell walls, for example, in alkaloids (Onoda et al., 2017) such as caffeine. In coffee, N residing in caffeine constitutes approximately 0.35 to 0.46% of the leaf dry weight (Gonthier et al., 2011; Vega et al., 2020), which corresponds to 11-20% of total leaf N. Additionally, caffeine content is often higher in shade-grown coffee plants (Vaast et al., 2006), which may further decouple leaf N from leaf A_mass. We also found a strong negative LMA–A_mass relationship (r=-0.59, p< 0.05, Figure 7b) and a negative LMA–leaf N relationship (r=-0.28, p< 0.05, Figure 7c), which aligns well with LES prediction. Higher LMA is associated with a higher proportion of cell walls relative to leaf dry mass, which could limit photosynthesis by limiting CO₂ diffusion to chloroplasts and/or reallocating N to cell walls instead of Rubisco (Onoda et al., 2017). Overall, our LMA–N–A_mass relationships are more consistent with Martin et al. (2017) than Buchanan et al. (2019). One possible explanation is that the fertilization treatments from Buchanan et al. (2019) resulted in higher and less variable leaf N compared to our sites that are less-fertilized, which may have weakened their trait relationship relative to the LES pattern of natural ecosystems.

There are two hypotheses generally used to explain plant strategies to tolerate shade, which are (1) increasing investment in storage and defence (resource-conservative), or (2) maximizing carbon gain (resource-acquisitive) (Valladares et al., 2016). Shade-gradient studies in coffee agroecosystems reported conflicting findings, with some supporting the first hypothesis (Martins et al., 2014; Buchanan et al., 2019; Getachew et al., 2022), and the rest supporting the latter (Freitas et al., 2003; Bote and Struik, 2011; Gagliardi et al., 2015). Along the LES axis, we found that higher canopy cover, which likely limits light availability, is associated with resource-acquisitive traits in coffee plants, aligning with the second hypothesis. Shading may influence traits indirectly via microclimate buffering (Piato et al., 2020). For instance, higher humidity in the understory (Oliosi et al., 2016) could increase g_sw and subsequently increase photosynthesis and/or decrease LMA of coffee plants (Bote et al., 2018). Nevertheless, the effect of canopy cover on leaf traits is complex and may not be explained by light availability alone, with some studies reporting a non-linear increase in photosynthesis with increasing shade (Baliza et al., 2012).

In terms of reproductive output, we found that resource-acquisitive plants tend to have fewer fruits. This pattern aligns with previous studies that reported lower fruit count in coffee plants with higher leaf N and A_mass and lower LMA (Gagliardi et al., 2015; Bote and Jan, 2016). This inverse relationship between leaf carbon gain and fruits may be a result of the trade-off between vegetative and reproductive growth, as previously noted (Rodrigues et al., 2016; Piato et al., 2020). Meanwhile, along the nutrient-stomatal axis, we found that plants with higher g_sw and C content, coinciding with lower nutrients, tend to bear more fruits. Phenological studies explain the nutrient-fruit relationship through nutrient translocation, where photosynthetic products from the leaves are translocated to developing fruits, thus lowering leaf N, P, and K during the fruit-filling phase (León-Burgos et al., 2024b; Oliosi et al., 2020; Santos et al., 2021; Silva et al., 2022). While previous studies reported an increase in photosynthesis at a later fruiting stage due to rising carbon demand for fruit development (León-Burgos et al., 2024b), DaMatta et al. (2008) argued that this response is mainly due to changes in g_s, although the mechanism is not yet well understood. Other more direct physiological factors on flowering and fruit development are hormonal signalling (Sadka et al., 2023) and resource allocation/mass partitioning between vegetative and reproductive organs (León-Burgos et al., 2024a).

However, we did not observe a direct relationship between canopy cover and fruit count at our sites. Instead, the relationship is likely to be indirect, where: greater shade → more resource-acquisitive plants → fewer fruits. The effect of shade on fruit is often inconsistent, where some studies observed an increased productivity with increasing shade (Vaast et al., 2006), whereas others observed no effect (Oliosi et al., 2016). Overall, the effect of shade on growth and fruit appears to be complex, depending on factors such as plant age, cultivar, and other site conditions (Piato et al., 2020). It is also important to note that our estimates of fruit count do not capture total yield, bean size, bean quality, and multiple fruiting periods. These aspects can substantially influence yield outcomes, for example, due to planting density and the trade-offs between fruit loads and bean size (Vaast et al., 2006).

Additionally, while we assessed shading levels through canopy cover, we did not specifically look into the effects of shade tree species composition. Shade species can differ in their canopy architecture (compact vs. diffuse canopy, leaf angle, and leaf area), thus having different influence in microclimate buffering (e.g., light interception, wind buffering) and pathogen transmission (Gagliardi et al., 2020, 2023). Shade tree species also vary in their leaf traits such as leaf N, C, lignin, and phenols, leading to differences in litter decomposition, and consequently soil nutrient and water cycling (Isaac et al., 2024; Misgana et al., 2024). These shade species-specific effects are not captured by canopy cover measurements alone and thus could contribute to some of the site-specific trait variations we observed.

Trait coordination in agricultural plants is often weaker than natural plants, likely due to selection breeding for specific domestication traits (e.g., seed size) under an environment with abundant resources (Milla et al., 2014). Cultivation is hypothesized to lead to plants with a more resource-acquisitive trait strategy compared to their progenitor (Milla et al., 2015), but this has not been studied for coffee cultivars. Focusing breeding efforts for optimizing productivity and climate resilience will require a better understanding of coffee plant trait trade-offs and plasticity across genotypes and environments. While the effects of genotypes on trait trade-offs cannot be disentangled in our study, we provide a first step toward future comparative analysis.

Beyond site-level insights, the observed variations in traits provide valuable parameters for constraining process-based models of coffee productivity and carbon cycling (Van Noordwijk and Lusiana, 1999; van Oijen et al., 2010; Vezy et al., 2020). These models consider competition for light, water, and nutrients, thus can be used to assess the optimum management treatments (e.g., planting density, thinning, fertilization) that can maximize yield with minimum environmental impacts (van Oijen et al., 2010). Integrating traits into modelling frameworks can support sustainable coffee management and climate-smart agriculture by helping farmers design management strategies before implementing them in the field.