Predicting key water stress indicators of Eucalyptus viminalis and Callitris rhomboidea using high-resolution visible to short-wave infrared spectroscopy

2.1 Species studied and sampling locations

This study focused on two species, Callitris rhomboidea R.Br.Ex Rich. & A.Rich and Eucalyptus viminalis labill. C. rhomboidea is a needle-leafed coniferous tree from the family Cupressaceae that grows predominantly in dry eucalypt forests in the south-eastern states of Australia (Harris & Kirkpatrick, 1991). E. viminalis is from the family Myrtaceae and grows throughout the south-eastern states of Australia, predominately in dry Eucalypt forests with variants also found in wet eucalypt forests (Barson, 1978; Williams & Potts, 1996). E. viminalis is a flowering plant with vastly different leaf structure and life cycle to C. rhomboidea. C. rhomboidea was chosen for this study as it has been widely studied for its drought tolerance and ability to survive in dry environments (Brodribb et al., 2010; Brodribb & Cochard, 2008; Johnson et al., 2022). Whilst E. viminalis was chosen as it offers a contrasting ability to tolerate drought, having suffered significant die-back in past years (Ross & Brack, 2015; Tonet et al., 2023). For this study, 82 separate leaf samples with accompanying spectral measurements were taken from three E. viminalis trees. Additionally, 84 separate leaf and spectral samples were taken from three C. rhomboidea trees. The E. viminalis samples were taken from mature, fully grown trees located in natural conditions on the University of Tasmania campus, Sandy Bay. Similarly, all three of the C. rhomboidea trees were mature and fully grown, with two of the three sampled trees being located on the University campus. Due to no other individuals being located on campus, the third tree was sampled at a study site in Pelverata, approximately 30 km to the south of Sandy Bay. The C. rhomboidea tree located at the Pelverata site was deemed a suitable addition due to the site holding similar environmental conditions to the University campus.

2.2 Dehydration experiment

From each of the sampled trees, three branches (approximately 1 m in length) were taken from varying positions on the tree and were chosen to include predominately sunlit, mature leaves. The branches were sampled between the months of May and July 2021. Specifically, the branches were sampled before sunrise and during wet weather to ensure the plants were close to full turgor. For the E. viminalis samples, this was between −0.12 and −0.72 MPa Ψ_leaf, whilst the full turgor measurements for the C. rhomboidea samples were between −0.35 and −0.69 MPa Ψ_leaf. The branches were then placed in a dark air-sealed bag and transported to the laboratory. On the same day as sampling, the branches were dried using the dry bench method aided by ceiling-mounted fluorescent lighting in a laboratory environment at 22°C over 1−3 days. They were dried out gradually and were subsampled as the branch dried to around −5.0 MPa. Periodically, as the branches were drying out, subsamples of leaves were taken to measure their Ψ_leaf using a Scholander pressure chamber. It was assumed that all leaves on the branch were at similar Ψ_leaf, due to low transpiration rates in low light. Once Ψ_leaf was determined, that sample was discarded and a further subsample from neighbouring leaves was taken for spectral analysis. Using different subsamples for Ψ_leaf and spectral analysis was performed to prevent rehydration errors after pressure chamber analysis and to allow timely weight and area measurements before spectral analysis and drying of the leaf sample. For E. viminalis, the subsample consisted of 1−2 leaves and for C. rhomboidea, the subsample consisted of 1−2 small branchlets (Figure 1b). The subsample was then weighed for fresh weight, scanned using a flatbed scanner and placed in a makeshift darkroom for spectral measurement within 1−2 min to ensure the leaf did not dry significantly between detachment and measurement. Following this procedure, each sample was placed in a laboratory oven at 70°C for 72 h to obtain their dry weights. Each subsample for spectral analysis came from new leaves for each measurement. This increased the robustness of the sampling approach, as each new leaf represented an individual sample with different structural and spectral characteristics. Overall, 27 individual samples were taken from two of the three branches for E. viminalis with the third having 28 samples taken. For the C. rhomboidea samples, 39 individual samples were taken from the first branch, 27 from the second and 18 from the third branch. In total, there was a combined sample number of 82 for E. viminalis and 84 for C. rhomboidea.

2.3 Determinations of leaf water

2.4 Spectroscopy measurements

The spectral measurements of the leaf samples were measured using the RS-8800 spectroradiometer from Spectral Evolution (RS-8800; SpectralEvolution). This spectroradiometer measures spectral reflectance of all wavelengths between 350 and 2500 nm at a spectral resolution (full width at half maximum, FWHM) of 2.8 nm at wavelength 700 nm, 8 nm at 1500 nm and 6 nm at 2100 nm. The spectroradiometer was set up in a custom darkroom with the sensaprobe grip facing downwards, 15 cm above the sample (Figure 1a). Attached to the fibre-optic cable on the sensaprobe grip was an eight-degree field of view fore-optic, resulting in a circular footprint on the sample with a diameter of approximately 2.1 cm. Within the darkroom, at the same height as the sensaprobe grip (15 cm above the sample), two 50 W halogen lights (ILM 550-Tungsten Halogen Light Source; SpectralEvolution) were pointed towards the sample at a 45° angle, producing a spatially homogenous light field across the entire leaf tray. At every measurement, the spectroradiometer was set up to compute the average of 10 spectral observations. Leaf samples were placed as shown in Figure 1, and four spectral measurements were acquired for each leaf sample, one observation for every 90° clockwise rotation. This increased the robustness of the spectral measurements and compensated for any directional scattering effect caused by the structural properties of the leaf. The final reflectance of each sample was then computed as the average of the four rotated observations. A reference measurement of a Spectralon® panel with known reflectance was made between every set of samples for calibration. Example spectrums from both species are shown in Figure 1c.

2.5 Spectral analysis

2.5.2 Continuum removal

Continuum removal is a mathematical technique used in high-resolution spectroscopy to isolate the absorption features of interest by dividing out the underlying continuum (Clark & Roush, 1984). In their study focussing primarily on chlorophyll content, Malenovský et al. (2015) found the continuum removed absorption feature between 650 and 720 nm to be a valuable indicator of Antarctic moss water content but discussed the need to cover the absorption features located in the NIR and SWIR region of the spectrum. As found in previous literature, the absorption features at 1200 and 1450 nm have been linked to changes in plant water status (Pasqualotto et al., 2018; Pu et al., 2003; Wang et al., 2008). For this study, the area (f1200_area, f1450_area) and depth (f1200_depth, f1450_depth) of these features were calculated and analysed using the HSDAR package (version 1.0.4) in R (Lehnart, 2021). Band depth is the depth of the feature at its maximum point and area is the area under the curve of that feature, calculated using Equation 4. An additional area normalised band depth (ANMB) metric was calculated for the features at 1200 and 1450 nm (Equation 5). These were calculated following the success of ANMB in estimating chlorophyll content (Cimoli et al., 2020; Kováč et al., 2013; Malenovský et al., 2006).

$$<math altimg="urn:x-wiley:01407791:media:pce15083:pce15083-math-0004" display="block" wiley:location="equation/pce15083-math-0004.png" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mrow><msub><mi>Area</mi><mi mathvariant="normal">f</mi></msub><mo>\unicode{x0003D}</mo><munderover><mo>\unicode{x02211}</mo><mrow><mi>Min</mi><mrow><mo>(</mo><mi>\unicode{x003BB}</mi><mo>)</mo></mrow></mrow><mrow><mi>Max</mi><mrow><mo>(</mo><mi>\unicode{x003BB}</mi><mo>)</mo></mrow></mrow></munderover><mi>BD</mi><mi>\unicode{x003BB}</mi></mrow></mrow></math>$$(4)

Where Area_f is the area of the feature, min($<math altimg="urn:x-wiley:01407791:media:pce15083:pce15083-math-0005" wiley:location="equation/pce15083-math-0005.png" display="inline" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mrow><mi>\unicode{x003BB}</mi></mrow></mrow></math>$) is the minimum wavelength of the spectrum, max($<math altimg="urn:x-wiley:01407791:media:pce15083:pce15083-math-0006" wiley:location="equation/pce15083-math-0006.png" display="inline" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mrow><mi>\unicode{x003BB}</mi><mo>)</mo></mrow></mrow></math>$ is the maximum wavelength of the spectrum and BD is the band depth (Lehnart, 2021).

$$<math altimg="urn:x-wiley:01407791:media:pce15083:pce15083-math-0007" display="block" wiley:location="equation/pce15083-math-0007.png" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mrow><msub><mi>ANMB</mi><mi mathvariant="normal">f</mi></msub><mo>\unicode{x0003D}</mo><mfrac><msub><mi>Area</mi><mi mathvariant="normal">f</mi></msub><mrow><mi>Band</mi><mspace width="0.25em"/><msub><mi>Depth</mi><mi mathvariant="normal">f</mi></msub></mrow></mfrac></mrow></mrow></math>$$(5)

For calculating the area and depth of both absorption features, a continuum removal procedure was performed by first creating a continuum line connecting the local maxima of the reflectance spectra (Clark & Roush, 1984; Malenovský et al., 2013). As each spectral measurement had slightly differing local maxima, all spectra were subsetted at specific wavelengths to ensure a systematic application of the continuum removal algorithm. The wavelength ranges for the spectral subsets were manually defined for both individual species. For both absorption features, the end point of the subset was determined as the wavelength for the maximum reflectance point on the spectrum after the absorption feature (i.e., the highest point of the spectral peak after the absorption feature and before the next feature). The start point was manually selected, due to issues arising from the continuum line intersecting with the spectral curve using automated methods. For the E. viminalis samples, the spectrum was subset for between wavelengths 1120 and 1268 nm to isolate the feature centred at wavelength 1200 nm, and 1305 and 1648 nm to isolate the feature at 1450 nm. For the C. rhomboidea samples, the spectrum was subset for between wavelengths 1120 and 1278 nm, as well as 1305 and 1654 nm to isolate the features at 1200 and 1450 nm, respectively. Once the continuum lines were established and the spectra was subset at the appropriate wavelength ranges, the continuum removal method was applied. As an example, the subset points of a C. rhomboidea sample can be seen in Figure 2a with Figure 2b showing the continuum removed spectrum for the subsetted sample.

2.6 Segmented regression and TLP measurement

The primary aim of this study was to examine the relationships between spectral variables and the measures of plant water stress. Through observation, we found distinct slope changes in the linear relationships between the spectral variables and Ψ_leaf, suggesting a breakpoint. Therefore, an additional objective of this study was to explore this through the use of segmented regression and examine whether the associated breakpoints could be indicative of TLP. The segmented regression was done using the ‘segmented’ package in R (version 1.6-0), which implements a linearisation technique to find the breakpoint of a regression data set (Muggeo, 2008). The best performing VIs, as well as, all absorption feature properties were analysed for both species. The breakpoint in the relationship between the spectral variables and Ψ_leaf is defined as the position in the scatterplot where the relationship is distinctively different before and after the breakpoint (Muggeo, 2008). Due to the segmented package picking up breakpoints that may be deemed outside what is possible, in relation to TLP, breakpoints that were greater than −1.0 or less than −4.0 were deemed unviable. This is in line with research by Bartlett, Scoffoni, and Sack (2012), showing the TLP of plants from many biomes not exceeding this range.

3.1 Reflectance changes induced by water stress

The changes in reflectance across all wavelengths measured from all samples from both species are shown in Figure 3, sorted and coloured by the measured Ψ_leaf value. In the visible range (between 350 and 700 nm) the reflectance of leaves from both species decreases with Ψ_leaf. This decrease in Ψ_leaf is more significant in the SWIR range (between 1300 and 2500 nm) with a gradual decrease of reflectance across all wavelengths proportional to the decline of Ψ_leaf. The absorption feature at 1450 nm clearly displays this gradual decrease in both species. The reflectance of the wavelengths in the NIR region (between 700 and 1300 nm) as well as the absorption features at 970 and 1200 nm did not follow this trend, showing high variance. When compared to the E. viminalis spectra in Figure 3a, the VIS and SWIR regions of the C. rhomboidea spectra in Figure 3b have a lower variance of reflectance values, whilst the NIR region has higher variance. Additionally, the overall reflectance is shown to be lower in the C. rhomboidea spectra.

3.2 Linear prediction of EWT, FMC and Ψ_leaf

The strongest linear regression against Ψ_leaf for both species came from the newly created NRIs (Table 2). Figure 4 visualises the R² value from the regression between the computed NRIs of the specified band combinations against Ψ_leaf for both species. For the E. viminalis measurements, the band combinations of 406 and 1888 nm (NRI_{406, 1888}) showed the highest regression against Ψ_leaf with an R² of 0.851. As seen in Figure 4a, there are a considerable number of band combinations with a high R², particularly in the SWIR region of the spectrum (1300–2500 nm). The results for the C. rhomboidea measurements are considerably different to those of E. viminalis and are shown in Figure 4b. Unlike the E. viminalis measurements, only few regions of combinations showed high R² results for C. rhomboidea, primarily around the 1850 nm wavelength region. The band combinations of 1817 and 1845 nm (NRI_{1817, 1845}) had the highest regression against Ψ_leaf with an R² of 0.850. Additionally, NRI_{1817, 1845} was the best performing variable when predicting FMC for both C. rhomboidea (R² = 0.802) and E. viminalis (R² = 0.910).

For the E. viminalis samples, the best performing established VI (those shown in Table 1) against EWT and Ψ_leaf was Datt8 (R² = 0.875 and 0.673, respectively), and against FMC was PWI (R² = 0.586). For the C. rhomboidea samples, the best performing established VI against EWT, FMC and Ψ_leaf was PWI (R² = 0.664, 0.443 and 0.381, respectively). Across both species, the best-performing absorption feature was the feature centred at 1450 nm. The depth, area or ANMB of the feature at 1450 nm had the best regression result against all three plant water variables for both species (Table 2). For the E. viminalis measurements, the best performing absorption feature property against EWT was ANMB₁₄₅₀ (R² = 0.902), against FMC was F1450_Depth (R² = 0.733) and against Ψ_leaf was F1450_Area (R² = 0.752). For the C. rhomboidea measurements, the best performing property against EWT, FMC and Ψ_leaf was F1450_Area (R² = 0.671, 0.472 and 0.412, respectively). Most relationships assessed were deemed statistically significant with the majority having a p-value of <0.05. Furthermore, we evaluated the relationships between EWT and FMC in relation to Ψ_leaf for both species (Figure S2). The E. viminalis Ψ_leaf measurements exhibited a stronger correlation with EWT (R² = 0.70), while the C. rhomboidea Ψ_leaf measurements showed a higher association with FMC (R² = 0.73).

3.3 Convolved and masked spectral prediction

The convolved and masked spectra simulate satellite observations by the EnMAP hyperspectral satellite sensor (assuming pure, non-mixed spectral observations) with a reduced number of spectral bands and removal of the atmospheric water absorption regions. For the convolved and masked E. viminalis measurements, the band combinations of 1567 nm (band 123) and 1579 nm (band 124) (NRI_{1567, 1579}) showed the highest R² (0.792) against Ψ_leaf. The band combinations of 1251 nm (band 113) and 1299 nm (band 117) (NRI_{1251, 1299}) were the best performing for the convolved and masked C. rhomboidea measurements with an R² of 0.783. These results show a similar trend to the previous NRI results in which there is a larger number of band combinations with a high R² against Ψ_leaf in the E. viminalis measurements when compared against the C. rhomboidea combinations. The convolved and masked spectra along with the NRI results is shown in Figure S3.

3.4 TLP and breakpoint analysis

The LDMC, π₀ and TLP of all three trees from both measured species is shown in Table 3. Both species exhibited similar ranges of measured TLP. Including the standard error, the range of TLP found for C. rhomboidea is −1.65 to −2.37 MPa and for E. viminalis is −1.64 to −2.47 MPa.

The strong linear regressions against Ψ_leaf identified in the previous section additionally exhibited a breakpoint in some of the relationships. The breakpoints for all assessed variables at the individual and combined tree level for both species are shown in Table S1. The highest performing linear regressions against Ψ_leaf, from the newly created NRIs for both species, are visualised at the individual and combined tree level in Figure 5. As seen in Figure 5, the segmented linear fit for NRI_{406, 1888} amongst the combined E. viminalis samples had an R² of 0.897 and for NRI_{1817, 1845} in the combined C. rhomboidea samples had an R² of 0.875. For the E. viminalis samples, NRI_{406, 1888} had a breakpoint for tree 1, 2 and 3 of −1.92, −2.07 and −1.12 MPa, respectively (Table S1). The breakpoints for tree 1, 2 and 3 of NRI_{1817, 1845} for the C. rhomboidea samples was −1.52, −2.77 and −1.83 MPa, respectively.

For the combined E. viminalis samples, the mean breakpoint across all evaluated spectral variables was −2.08 MPa and the median −1.89 MPa with a range of between −1.72 and −3.89 MPa (Table S1). For the combined C. rhomboidea samples, the range of breakpoints across the evaluated spectral variables was more varied, with six breakpoints of >−1.0 MPa and two showing <−4.0 MPa. Excluding these breakpoints, the range was between −1.30 and −2.50 MPa with a mean and median of −2.28 and −2.41 MPa, respectively.

4.1 Spectral predictive capacity

Amongst the purely linear regressions, the best performing spectral variables against Ψ_leaf came from the NRIs created from the species-specific regression of combined wavelengths (see Table 2). This method evaluated features across the entire 350–2500 nm spectral range and was able to individuate the wavelength combinations that had the highest predictive power against Ψ_leaf. This technique is advantageous in new datasets when compared to predefined VIs calculated from wavelength-specific ratios, as it makes use of the entire available spectrum. These newly created NRIs proved more accurate than the broader VIs, suggesting that creating species-specific indices enhances the robustness of spectrally predicting Ψ_leaf. The high predictive capacity of the NRIs can facilitate noninvasive prediction of Ψ_leaf for these two species in both laboratory and in situ field settings. As demonstrated, our method can be adjusted for sensors with different spectral specifications and accounting for the strong influence of atmospheric absorption. Although initial model development may take time, spectroscopic measurements can be near instantaneous and non-destructive. As similar sensors become more cost-effective and common across field-based applications and remote sensing platforms (Cavaco et al., 2022), our method could facilitate scaling from the lab to field, airborne and spaceborne approaches. This enables the monitoring of Ψ_leaf at broader spatial scales and over longer continuous temporal series than would be achievable with standard methods.

From the NRIs calculated under laboratory and simulated atmospheric conditions, six of the eight bands used in the indices were located in the SWIR region of the EM spectrum (Figure 5 and Figure S3). The success of these bands could be explained by the absorption of incoming radiation by water in the SWIR (Hale & Querry, 1973), with absorption being particularly strong at the spectral ranges centred around 1450 and 1940 nm (Datt, 1999; Tucker, 1980; Woolley, 1971). Studies have established significant correlations between specific wavelength bands or absorption feature properties and leaf water content estimates derived from features centred at 1450 and 1940 nm (Datt, 1999; Wang et al., 2008), as well as those centred at 970 and 1200 nm (Pasqualotto et al., 2018; Peñuelas et al., 1993). In this study, the properties derived from the absorption feature centred at 1450 nm exhibited superior performance compared to all other assessed variables in predicting EWT. Moreover, they outperformed the established VIs in predicting FMC and Ψ_leaf. In contrast, the feature centred at 1200 nm demonstrated comparatively weaker performance, suggesting that the 1450 nm feature serves as a stronger indicator of plant water stress for these specific species. The advantage of taking spectroscopy measurements at the leaf level is that it removes any atmospheric absorption effects seen in wavelengths in the SWIR region of the spectrum, which are problematic in typical airborne and spaceborne remote sensing applications (Gao et al., 2009). Though problematic, through our simulation of atmospheric conditions, we found that high predictive relationships against Ψ_leaf can still be found when removing these absorption features and convolving the spectra to the same resolution as a satellite-based sensor (see Figure S3). This suggests Ψ_leaf can be predicted at broader spatial scales using the techniques put forward by this study and further work should be focussed on validating these methods with in situ measurements. New and emerging spectroscopic technologies on-board low flying platforms such as Unoccupied Aerial Systems (UASs) may give the opportunity to limit the impact of the atmospheric water column yet capture the spatial variability of spectral water traits over relatively large areas (e.g., Turner et al., 2023). The findings in this study will inform further work on low-altitude, high-resolution spectroscopy acquisitions and future research from UAS platforms may determine the impact of atmospheric water on spectral observations. Identifying species-specific indices, as found in this study, may be important for more robust predictions of Ψ_leaf, especially at larger scales. Future research should build upon these relationship databases and develop techniques to apply them at the individual species or functional group level. This will enhance the robustness of monitoring Ψ_leaf over broader spatial scales and continuous time series. These studies and any predictions made will need to be validated by robust in situ water potential measurements taken from leaf samples during spectral acquisition.

The measurement of Ψ_leaf is the primary technique in determining water stress in plant physiology research as it is a direct measure of xylem pressure, strongly indicating leaf water deficit (Choat et al., 2018). Remote sensing and spectroscopy research primarily uses leaf water content measures, with EWT being one of the most popular variables for plant water status determination (Govender et al., 2009; Ma et al., 2019). This likely explains the results of the established VIs assessed in this study, showing better predictive ability against EWT than the other plant water determinations. The use of water content determinations in remote sensing and spectroscopy is popular due to their direct representation of the physical amount of water within the leaf. A spectroscopic sensor can only measure the reflectance of the inherent physical and chemical properties of the sensed object, thus, direct measures of water content within the leaf are likely better correlated to spectrally derived variables (Ripple, 1986). The findings from this study show that spectroscopy can accurately predict Ψ_leaf of mature trees from two distinctively different species within the measured range. This finding could be attributed to an underlying relationship between water content determinations and Ψ_leaf (see Figure S2), but the high predictive capacity against Ψ_leaf provides encouraging evidence for future research to derive physiological measures of plant water stress through predictive models based on spectral data. Additionally, while this study focused on a large number of samples from a small number of trees, allowing for a wide range of physiologically diverse samples, future research can further enhance the representativeness of the population by incorporating broader sampling strategies. Expanding the number of replicate trees will increase the robustness of the species-level spectral predictions of Ψ_leaf.

Due to the spectroscopy sensor measuring the reflectance of the leaf's physical and chemical properties, variations in leaf reflectance are influenced by a multitude of factors, extending beyond biochemical properties like water content and pigment concentration. The structural properties of a leaf, including its shape, size, cell arrangement, surface features (such as hairs or wax coatings) and internal structures (such as stomata or specialised tissues), also play a crucial role in determining its reflectance at specific wavelength ranges in the EM spectrum and therefore influence predictive capabilities (Eitel et al., 2006; Woolley, 1971). These structural differences could be further amplified through leaf stacking, a sampling approach which has been shown to affect spectral reflectance, particularly in the NIR region, with less impact on the SWIR and VIS regions of the EM spectrum (Neuwirthová et al., 2017). Although leaf stacking was not used in this study, the slight overlap of samples may have affected their reflectance and in the case of C. rhomboidea branchlets, the structural properties of the leaves and their overlap likely contributed to these variations. The disparity in performance observed between C. rhomboidea and E. viminalis could partly be attributed to the complex leaf structure of C. rhomboidea. The cylindrical and needle-like nature of C. rhomboidea's branchlets likely resulted in increased light scattering, leading to higher spectral variation (see Forsström et al., 2021; Hovi et al., 2022; Markiet et al., 2017; Rautiainen et al., 2018) and in combination with gaps between the branchlets lead to a decrease in overall reflectance detected by the sensor. The disparity in performance could additionally be attributed to the inherent stress response of C. rhomboidea. Due to C. rhomboidea being more tolerant to drought conditions (Brodribb et al., 2010; Brodribb & Cochard, 2008; Johnson et al., 2022), the species is likely to experience less changes in leaf water content in comparison to E. viminalis within the measured Ψ_leaf range as covered in this study. As noted previously, spectroscopy measures the direct optical properties of the leaf and thus, the Ψ_leaf predictions are likely influenced by the underlying relationship with water content. With less changes in water content due to the ability of C. rhomboidea to cope better with drying conditions, the relationship between certain spectral variables and Ψ_leaf is notably less linear. In future research, exploring the spectral prediction of Ψ_leaf in C. rhomboidea could involve incorporating measurements obtained at more negative water potentials, broadening the scope for a comprehensive understanding of the plant's behaviour under increased drying conditions.

4.2 Linear breakpoints and TLP

This study showed that Ψ_leaf can be linearly predicted with very high confidence using variables calculated from spectral data. The presence of a breakpoint in the relationships between the majority of spectral variables and Ψ_leaf identifies that this relationship is not purely linear, adding complexity to linearly predicting Ψ_leaf. Through accounting for this breakpoint and applying a segmented linear fit to the relationship, the predictive performance of the best performing indices increased from an R² of 0.85 to an R² of 0.875 for C. rhomboidea and 0.897 for E. viminalis (Figure 5). Though the initial linear regression may still be important to identify new indices, such as the NRIs in this study, future studies should be aware of this breakpoint and potentially account for it in their predictive models.

The breakpoint identified in the relationships between the majority of the spectral variables and Ψ_leaf additionally indicates a potential link to the TLP of the measured trees. Notably, we found evidence showing these breakpoints to be within the measured TLP range, particularly for the E. viminalis samples. Here, we combined the linear breakpoints for all three trees from both species across all assessed spectral variables, creating a more robust sample set when compared to using only samples from individual trees. For the E. viminalis measurements, the majority of the assessed spectral variables held linear breakpoints within the measured TLP range. Additionally, both the mean and median breakpoints for all combined variables resided within the measured range. These identified breakpoints are further supported by TLP measurements of E. viminalis in other studies. Tonet et al. (2023) found the mean TLP of E. viminalis seedlings to be −2.01 ± 0.01 MPa. Valentini et al. (1990) studied 6-year-old E. viminalis saplings and found their TLP to reach a minimum in the winter months of −2.15 MPa and a maximum in the summer months of −1.8 MPa. Salvi et al. (2022) found a TLP of −1.7 ± 0.2 MPa in E. viminalis seedlings, and Li et al. (2018) found the TLP of mature E. viminalis trees located in a wet sclerophyll forest to be −1.51 MPa. The E. viminalis trees assessed in this study were mature, which may explain the slightly more negative predictions when compared to the findings from Salvi et al. (2022). Further, the trees in Li et al. (2018) were likely more adapted to wetter conditions, potentially explaining the higher TLP. Additionally, at the individual tree level, the linear breakpoints found between NRI_{406, 1888} and Ψ_leaf were within the measured TLP range for two of the three E. viminalis trees. The measured TLP range in this study was obtained from samples taken at different times of the year but under similarly hydrated conditions. These TLP measurements may have been influenced by inherent differences in soil and air temperature as well as the overall stress response of the trees at those times of year (Bartlett et al., 2014). As this study focused on preliminary relationship exploration, future studies aiming to directly predict or interpret TLP from spectral variables should ensure that TLP measurements of the sampled trees are taken at the same time of year. TLP is a complex trait influenced by a multitude of factors, requiring careful consideration when attempting to predict or interpret it. However, recent results by Castillo-Argaez et al. (2024) show that TLP can be estimated using spectroscopy and together with the measurements from the cited literature provides evidence that the breakpoint observed for the E. viminalis samples could be attributed to the TLP of the measured trees.

For the C. rhomboidea samples, the majority of the assessed variables held a linear breakpoint outside the prescribed viable range (between –4.0 and –1.0 MPa). Of the viable breakpoints, f1450_area was the only variable within the measured TLP range. Additionally, the mean breakpoint of the combined variables was shown to be within the measured range. The measured range of TLP for C. rhomboidea in this study are in agreement with the TLP found by Mercado-Reyes et al. (2023) where plants that were never before stressed were found to have a TLP of −1.922 ± 0.13 MPa and in plants that had experienced drought stress of −2.016 ± 0.14 MPa. At the individual tree level, the breakpoints from the newly created NRI_{1817, 1845} were only within the measured TLP range for one of the three C. rhomboidea trees (tree 3) (Figure 5). The amount of nonviable breakpoints and the decline in the ability of the spectral variables to pick up a breakpoint within the measured range for the C. rhomboidea samples could again be attributed to the structurally induced scattering or the inherent stress response of the species, as explained previously. This likely affected the relationships between Ψ_leaf and the spectral variables, ultimately leading to a decrease in the ability of these variables in picking up an accurate breakpoint. Furthermore, the estimations of breakpoints using the segmented regression technique can be highly sensitive to outliers, noise, and slope changes along a nonlinear relationship. As shown here with the nonviable breakpoints, false breakpoints are often given at the edges of noisy or nonlinear datasets. Future research incorporating larger datasets should be aware of this and potentially implement methods such as noise reduction or smoothing techniques to address them in their segmented regression models (Muggeo, 2017). Additionally, as breakpoint measurements derived from segmented regression is dependent on the amount of data taken before and after the breakpoint, future studies may focus on including additional samples, which may lead to more robust breakpoint measurements that are within the measured TLP ranges and in agreement with past literature.