Hyperspectral reflectance integrates key traits for predicting leaf metabolism

There has been widespread interest in developing trait-based models to predict photosynthetic capacity from leaves to ecosystems (Walker et al., 2014; Xu & Trugman, 2021), but comparably less for nonphotorespiratory mitochondrial CO₂ release (dark respiration, R_dark). This is significant, given that about half of the CO₂ released from plants is via R_dark – which occurs day and night – and supports ATP production, redox balance, nitrogen assimilation and carbon skeleton synthesis (Atkin et al., 2015). Terrestrial biosphere models use simplified empirical relationships between the maximum rate of carboxylation (V_cmax) and R_dark – often derived from more easily measurable leaf traits such as leaf mass per area (LMA), leaf lifespan, nitrogen (N), and phosphorus (P), which have more extensive data availability (Reich et al., 1998; Tcherkez et al., 2024). Notably, these traits are measured across a unidimensional continuum, and there has yet to be solid evidence that the magnitude and direction of a leaf trait is highly predictive of a metabolic trait like R_dark. Leaf metabolic parameters change dramatically with their environment and encompass an integrated suite of traits – some which increase, some which decrease, and some that remain unchanged. This begs the question – is there an alternative approach, which integrates a large suite of the biochemical, structural and environmental traits, to predict R_dark on its own? A recent paper published in New Phytologist (Wu et al., 2024; doi:10.1111/nph.20267) addresses this question by comparing the utility of traditional trait-based approaches against hyperspectral reflectance data across three forest types.

‘By incorporating bidirectional variations across the visible to shortwave spectrum, hyperspectral reflectance effectively captures dynamic shifts in a broad array of leaf structural and biochemical traits…’

Wu et al. (2024) show that while trait-based models have provided valuable insights in some other studies, their predictive power of R_dark is underwhelming. The authors show that univariate trait–R_dark relationships are weak (r² ≤ 0.15), and even multivariate models explain only a fraction of the observed variability (r² = 0.30), leaving much of R_dark complexity unexplained. Beyond traditional leaf economic traits like LMA, N, and P, the authors investigate other elements such as magnesium (Mg), manganese (Mn), calcium (Ca), potassium (K), and sulfur (S), as they play crucial roles in respiratory metabolism but are rarely incorporated into predictive frameworks (Tcherkez et al., 2024). Despite the inclusion of more leaf traits for R_dark prediction, their poor performance highlights the need for alternative approaches that can more holistically capture the physiological complexity of R_dark.

By incorporating bidirectional variations across the visible to shortwave spectrum, hyperspectral reflectance effectively captures dynamic shifts in a broad array of leaf structural and biochemical traits, offering a rapid, scalable solution for characterizing physiological variability (Fig. 1; Ustin et al., 2009). Using data from Wu et al. (2024), there are subtle differences between the mean, lowest 10^th and highest 90^th percentile of R_dark samples (Fig. 1a). To understand the magnitude and direction of spectral changes between low R_dark (10^th percentile) and high-R_dark (90^th percentile) measurements, the percent difference from the mean spectra in the dataset is shown (Fig. 1b). In the visible spectrum, there is comparably less reflectance in the blue and red regions of the spectrum for high-R_dark measurements, associated with chlorophyll absorption features. Additionally, a change in the opposite direction occurs in the green region – centered c. 531 nm – which has been shown to be sensitive to photoprotective carotenoid pigments, that is the xanthophyll cycle (Gamon et al., 1992). Beyond this, we observe differences in the near infrared (NIR), indicating leaves with higher R_dark are likely thicker, or have higher LMA, but might also have higher leaf water content – as is highlighted by changes in water absorption features in the shortwave infrared (SWIR). While specific nutrients do not have an explicit spectral signature, it is likely that their concentration covaries with these leaf biochemical and structural attributes (Wong, 2023). Taken together, increases and decreases across the spectrum seem to match what we would theoretically assume for leaves with greater photosynthetic capacity, and potentially higher R_dark.

Fig. 1

Open in figure viewerPowerPoint

Subtle differences in hyperspectral reflectance curves from leaves representing a range of physiological conditions. (a) Reflectance data for the mean (50^th percentile, black dashed), high (90^th percentile, green), and low (10^th percentile, purple) R_dark samples from the Wu et al. (2024; doi: 10.1111/nph.20267) dataset, highlighting key traits associated with regions of interest in the visible (400–700 nm), near-infrared (c. 700–1400 nm), and shortwave infrared (c. 1400–2500 nm). (b) The percent difference between high- and low-R_dark samples from the mean, annotated with the direction of observed differences for key traits. The high-R_dark spectra show increased absorption in the chlorophyll (Chl) regions (c. 400–470 and c. 630–670 nm), while the low-R_dark spectra show decreased reflectance centered at 531 nm, a prominent xanthophyll absorption feature. Additionally, there is greater reflectance in the near-infrared region for high-R_dark spectra, suggesting higher leaf thickness and leaf mass per area (LMA), and higher absorption (lower reflectance) in the water absorption features in the shortwave infrared.

The use of reflectance spectra to infer physiological activity has often relied on the principle that subtle changes in pigments are directly tied to photosynthetic processes (Gamon et al., 1992) due to the strong coupling between the light-dependent and light-independent reactions of photosynthesis (Magney et al., 2020). However, traditional optical remote sensing has primarily focused on the development of vegetation indices (VIs) – or combinations of typically two bands – which ascribe a unidimensional change with a specific plant function, similar to traditional leaf traits. Building on foundational work with VIs, researchers have expanded to using the complete hyperspectral reflectance spectrum (400–2500 nm) to improve the detection of plant physiological dynamics (Serbin et al., 2012; Barnes et al., 2017). The ability of hyperspectral reflectance to leverage multiple signals – ranging from pigments in the visible region (400–700 nm) to structural, nutritional and water-related traits in NIR and SWIR – has enabled researchers to more accurately estimate photosynthetic capacity (Wu et al., 2019; Yan et al., 2021). This leap from traditional VIs to hyperspectral reflectance represents a paradigm shift in plant ecophysiology, unlocking the potential to track physiological dynamics across scales with greater precision and broader applicability.

Machine learning algorithms, such as Partial Least Squares Regression (PLSR), have been instrumental in identifying key spectral regions sensitive to photosynthetic parameters (Burnett et al., 2021). PLSR reduces the dimensionality of hyperspectral data by identifying latent components that summarize the relationships between spectral predictors and physiological traits. These components are linear combinations of the original spectral bands, designed to capture the maximum covariance between the predictors and the response variables (in this case, R_dark). Unlike traditional regression methods, which use raw variables directly, PLSR transforms the data into a smaller, uncorrelated set of variables, minimizing noise and redundancy. Latent components in PLSR are derived from the entire spectrum, providing a holistic approach to capture subtle spectral signals linked to leaf traits, even those without direct absorption features such as Mg, Ca, and Mn. While these components lack direct physiological interpretations, they effectively summarize complex spectral patterns, making PLSR a powerful tool for predicting physiological traits.

Notably, PLSR remains an empirical approach, reliant on the relationships in training data. Models must be validated across diverse datasets and conditions to ensure their generalizability. Despite this limitation, the identification of key spectral regions through PLSR provides a foundation for understanding wavelengths of interest, offering a scalable alternative for ecosystem monitoring. This is done by plotting variable importance projections (VIPs) of spectral features (as in fig. 8, Wu et al., 2024). Here, many spectral regions identified by VIPs exhibit sensitivity to multiple traits due to their shared physiological roles. For instance, regions associated with Mg and N overlap with those involved in photosynthesis, structural integrity, and respiration, reflecting the interconnected nature of these processes. This shared variability allows spectroscopy to capture broad functional relationships among traits, enabling simultaneous monitoring of multiple leaf characteristics. However, it also introduces complexity in interpreting VIP scores, as the spectral bands may reflect indirect predictions based on covarying traits rather than direct mechanistic links (Wong, 2023).

One of the most compelling aspects of hyperspectral reflectance is its potential for generalization across scales and ecosystems (Serbin & Townsend, 2020). Metabolic parameters such as R_dark are plastic traits influenced by dynamic environmental drivers, meaning datasets must span diverse plant functional types, biomes, and seasonal gradients to improve model robustness. Wu et al. (2024) emphasize that addressing these gaps requires comprehensive datasets that account for vertical canopy profiles and full-season dynamics (Niinemets et al., 2015; Lamour et al., 2023). These efforts will help bridge the trade-off between site-specific precision and cross-site applicability, a critical balance for scaling plant processes optically. While most hyperspectral predictions of metabolic processes have been done at the leaf scale, its implications extend to larger spatial scales – from towers (Pierrat et al., 2024), to aircraft (Wang et al., 2020) to satellites (Cawse-Nicholson et al., 2023). However, scaling introduces new challenges such as canopy heterogeneity, mixed pixels, solar/viewing angular effects, and background noise (Serbin & Townsend, 2020). Addressing these complexities requires hybrid modeling approaches that combine hyperspectral data with machine learning and radiative transfer models, as well as multi-scale measurements (Pierrat et al., 2024).

Going forward, standardizing hyperspectral datasets across species and ecosystems is critical for using these methods at scale. For example, the development of open-access databases such as the Global Spectra Trait Initiative (https://github.com/plantphys/gsti/tree/main), are essential for building a more robust framework. Ultimately, our growing need to rapidly detect change and track ecosystem function will benefit from tools that enable rapid, nondestructive, and scalable monitoring of plant metabolic traits. Hyperspectral reflectance bridges the gap between mechanistic understanding and large-scale ecological monitoring, offering new insights into the drivers of carbon cycling and ecosystem dynamics. The recent paper by Wu et al. (2024) demonstrates the power of hyperspectral reflectance to capture a wider suite of leaf traits and move beyond the limitations of traditional unidimensional approaches, which fail to capture the complexity of plant physiological dynamics.