Predicting tropical tree mortality with leaf spectroscopy

  • Chris Doughty (Contributor)
  • Alexander Cheesman (Contributor)
  • Terhi Ruitta (Contributor)
  • Andrew Nottingham (Contributor)



Do tropical trees close to death have a distinct change to their leaf spectral signature? Tree mortality rates have been increasing in tropical forests globally, reducing the global carbon sink. Upcoming hyperspectral satellites could be used to predict regions close to experiencing extensive tree mortality during periods of stress, such as drought. Here we show, for a tropical rainforest in Borneo, how imminent tropical tree mortality impacts leaf physiological traits and reflectance. We measured leaf reflectance (400-2500 nm), light saturated photosynthesis (Asat), leaf dark respiration (Rdark), leaf mass area (LMA) and % leaf water across five campaigns in a six-month period during which there were two causes of tree mortality: a major natural drought and a co-incident tree stem girdling treatment. We find that prior to mortality, there were significant (P<0.05) leaf spectral changes in the red (650-700 nm), the NIR (1000 -1400 nm) and SWIR bands (2000-2400 nm) and significant reductions in the potential carbon balance of the leaves (increased Rdark and reduced Asat). We show that the partial least squares regression technique can predict mortality in tropical trees across different species and functional groups with medium precision but low accuracy (r2 of 0.65 and RMSE/mean of 0.58). However, most tree death in our study was due to girdling, which is not a natural form of death. More research is needed to determine if this spectroscopy technique can be applied to tropical forests in general.,Study sites Our study plots are in Kalabakan Forest Reserve in Sabah, Malaysian Borneo (Tower SAF‐05 4.716°, 117.609°) within the Stability of Altered Forest Ecosystems (SAFE) Project study site (Ewers et al., 2011; Riutta et al., 2018). A schematic of the study site is shown in figure 1C. Mean annual temperature is approximately 26.7°C and mean annual precipitation is 2,600–2,700 mm with no distinct dry season but, on average, ~12% of months with precipitation below 100 mm month-1 (Walsh & Newbery, 1999). The plot has been selectively logged four times since the 1970s, which represents a high logging intensity for this region. The soils are orthic Acrisols or Ultisols on undulating clay soil. The tree basal area is 13.9 m2/ha. Total NPP and autotrophic respiration have been measured at this plot since 2011 and there is an eddy covariance tower nearby (Riutta et al., 2018). The plot is split in half by a small stream. All the trees on one side of the stream were girdled in late January 2016 by removing the phloem tissue in a 10 cm band, as described below (note: the plot was in the process of conversion to oil palm agriculture production). This part of the study site is hereafter referred to as the “girdled plot.” The trees on the other side of the stream were not girdled and represent the treatment control. This part of the study site is hereafter referred to as the “drought plot”. Although all trees experienced drought, the “drought” plot only experienced drought and not the effects of girdling. We collected data during five field campaigns that took place from January to June 2016. Campaigns began on the following days and generally took several days: Campaign 1=21 Jan-16, Campaign 2=10 Feb-16, Campaign 3=01 Mar-16, Campaign 4=29 Mar-16, Campaign 5 08 Jun-16. The first field campaign (C1) was conducted before girdling occurred to determine pre-girdling conditions and process rates. Girdling experiment – In late Jan 2016, after the first field campaign, we further explored the causes of tree mortality by conducting a girdling experiment. Girdling involved removing a 10 cm strip of the periderm and phloem in a ring around the tree stem at ~1.2 m height (with exceptions for trees with buttresses, which were girdled above the buttress) above the soil (Figure 1a) in a plot that was scheduled for conversion to a palm oil plantation. This technique prevents carbohydrate transport to the roots but maintains hydraulic connectivity because xylem tissues are not severed. Tree death was determined visually, based on the absence of visible canopy, with regular (average 18-day period) visits to the plots for both the drought and the girdled plots. We give the species measured in both plots in Table 1. Leaf sampling strategy –In each plot, 20-25 trees were chosen during each campaign, and tree climbers with extendable tree pruners removed one branch per tree that was growing in full sunlight (Asner & Martin, 2008). These branches were quickly recut underwater and taken to the laboratory for further measurements. On each of these branches, five fully expanded non-senescent leaves in randomly selected locations were chosen for measurements of: leaf-gas exchange leaf spectral properties (measured within 1 hour of being cut) and LMA. Leaf area was determined immediately after collection using a digital 476 scanner (Canon LiDE 110). Leaves were then oven dried at 72 °C until constant mass was reached. We subtracted wet weight from dry weight to calculate % leaf water and used dry weight and leaf area in order to calculate LMA. Leaf-level gas exchange – We used a portable gas exchange system (LI 6400, Li-Cor Biosciences, Lincoln, NE, USA) to measure leaf-level gas exchange. After returning to the laboratory, leaf dark respiration (Rdark) was measured by covering branches with an opaque bag for at least 20 minutes prior to measurement at a cuvette temperature of 30° C (Rowland et al., 2017). After this, branches were exposed to sunlight and light-saturated leaf photosynthesis was measured (Asat; 1200 µmol m-2 s-1 PPFD, 400 ppm CO2, at 30° C). We chose a light level of 1200 µmol m-2 s-1 for Asat because we tested photosynthetic capacity and found it generally saturated below light levels of ~1200 µmol m-2 s-1 PPFD, similar to other tropical studies (Both et al., 2019; Gvozdevaite et al., 2018; Doughty & Goulden, 2009b). We waited for gas exchange values to stabilize before starting a measurement, recorded data every two seconds and averaged the results after eliminating the first 20 measurements. We excluded photosynthesis measurements less than 0 µmol m-2 s-1 as this was indicative of a failure to maintain hydraulic connectivity in the sampled branch resulting in stomatal closure. We also excluded dark respiration measurements more negative than -1.5 µmol m-2 s-1 as this was considered indicative of a failure to truly represent Rd, or in some cases operator error. Most physiological measurements were collected between 07:00 and 14:00 local time and branches were cut from trees between 06:00 and 13:00 local time. An online supplement includes our averaged ± sd data for each leaf measured for transpiration rate (mmol H2O m-2 s-1), vapor pressure deficit based on leaf temperature (kPa), intercellular CO2 concentration (µmol CO2 mol-1), conductance to H2O (mol H2O m-2 s-1), and photosynthetic rate (µmol CO2 m-2 s-1). Leaf spectroscopy – We randomly selected five leaves within an hour of each branch being cut, and measured hemispherical reflectance near the mid-point between the main vein and the leaf edge (Asner & Martin, 2008). We used an ASD Fieldspec 4 with a fibre optic cable, contact probe and a leaf clip (Analytical Spectral Devices, Boulder, Colorado, USA). The spectrometer records 2175 bands spanning the 325–2500 nm wavelength region. We corrected for small discontinuities between spectral bands (~950 and ~1750 nm), where the instrument transitions from one sensor to another. Measurements were collected with 136-ms integration time per spectrum (Asner & Martin, 2008; Doughty, Asner, et al., 2011). To ensure measurement quality, the spectrometer was calibrated for dark current and stray light, and white-referenced to a calibration panel (Spectralon, Labsphere, Durham, New Hampshire, USA) after each branch(Asner & Martin, 2008; Doughty, Field, et al., 2011). The spectrometer was optimized after every branch so the light levels did not saturate. For each measurement, 25 spectra were averaged together to increase the signal-to-noise ratio of the data. Data analysis - We used the Partial Least Squares Regression (PLSR) modelling approach to predict leaf traits with spectral information (Geladi & Kowalski, 1986). PLSR incorporates all the spectral information within each leaf reflectance measurement, eventually reducing all spectral data (400-2500 nm) down to a relatively few, uncorrelated latent factors. This approach has been used successfully to predict plant traits across a wide range of ecosystems, including tropical forests (Asner & Martin, 2008; Serbin et al., 2014). We used the PLSregress command in Matlab (Matlab, MathWorks Inc., Natick, MA, USA) to establish predictive models for LMA, Asat, wood density (estimated with tree species and a lookup table (Chave et al., 2009)) and tree mortality (Doughty, Asner, et al., 2011). To avoid over-fitting the number of latent factors we minimized the mean square error with K-fold cross validation (set as an upper bound as 30). To avoid issues of pseudo replication, we emphasize that the unit of analysis in these analyses is the leaf. To create a completely independent testing dataset, we used the above method on 70% of our data to calibrate our model and then the remaining 30% to test the accuracy of our model. We evaluated the accuracy of our modelled estimates using two main metrics: r2 and root mean square error (RMSE)/mean. We graded our results as high precision and accuracy (r2 > 0.70; %RMSE < 15%), medium precision and accuracy (r2 > 0.50%; % RMSE < 30%), low precision and accuracy (r2 > 0.50; % RMSE > 30%). We also calculated NDVI for our five study periods as NDVI = (NIR-red)/(NIR+red) where we use 1000 nm for NIR and 650 nm for red. Statistical tests – For our leaf spectral measurements, for each 1 nm bandwidth, we determined statistical significance (P<0.05) between trees within 50 days of mortality and prior to this with a paired t-test (Matlab, Mathworks). To understand significant differences between % water, LMA, Rdark, and Asat, we used a t-test. To understand the impact of the girdling between % water, LMA, Rdark, and Asat over time, we used a repeated measures ANOVA.,To produce all figures in the paper, run Doughty_Biotropica_2020.m in the same folder with the datasets girdle_data1.mat and girdle_data2.mat. Raw spectral data are in the folder spectral_data. Raw physiology data are in the folder licro_data.,
Date made availableJan 1 2020

Cite this