Linking rainforest ecophysiology and microclimate through fusion of airborne LiDAR and hyperspectral imagery

We develop and validate a high-resolution three-dimensional model of light and air temperature for a tropical forest interior in Hawaii along an elevation gradient varying greatly in structure but maintaining a consistent species composition. Our microclimate models integrate highresolution airborne waveform light detection and ranging data (LiDAR) and hyperspectral imagery with detailed microclimate measurements. We then use modeled microclimate and forest structural and compositional variables to explain variation in spatially explicit measurements of leaf traits, including gas exchange and structure. Our results highlight the importance of: (1) species differences in leaf traits, with species explaining up to 65% of the variation in some leaf traits; (2) differences between exotic and native species, with exotic species having greater maximum rates of assimilation and foliar dN values; (3) structural factors, with foliar %N and light saturation of photosynthesis decreasing in mid-canopy locations; (4) microclimate factors, with foliar %N and light saturation increasing with growth environment illumination; and (5) decreases in mean annual temperature with elevation resulting in closure of the nitrogen cycle, as indicated through decreases in foliar dN values. The dominant overstory species (Metrosideros polymorpha) did not show plasticity in photosynthetic capacity, whereas the dominant understory species (Cibotium glaucum) had higher maximum rates of assimilation in more illuminated growth environments. The approach developed in this study highlights the potential of new airborne sensors to quantify forest productivity at spatial and temporal scales not previously possible. Our results provide insight into the function of a Hawaiian forest dominated by native species undergoing simultaneous biological invasion and climatic change.


INTRODUCTION
Tropical forests cover 11.7% of the global land surface area (Potter et al. 1993), contain 57% of above-and 27% of belowground carbon (Dixon et al. 1994), and are important contributors to the terrestrial global carbon cycle (Field et al. 1998, Cramer et al. 2004).Carbon fluxes and overall productivity within these forests are highly dependent upon light and temperature regimes (Boisvenue and Running 2006), which are predicted to undergo changes in the future (Hulme and Viner 1998, Mercado et al. 2009, Hansen et al. 2010).Light within a forest understory is considered to be the most important (Mercado et al. 2009), and limiting (Stadt et al. 2005), environmental factor influencing photosynthesis and carbon gain (Ellsworth and Reich 1992, Kull 2002, Araujo et al. 2008).Overall, light penetration to a tropical forest understory is among the lowest of terrestrial ecosystems (Chazdon and Pearcy 1991).The light which does exist is temporally variable, due to time of day, season or climatic conditions, and is highly dependent on the structure and density of both the forest under-and over-story (Montgomery 2004).Despite light limitation, the understory can make substantial contributions to overall forest productivity.For example, Sampson et al. (2006) calculated understory plants contributed up to 28% of a deciduous forest's gross primary productivity due to understory penetration by diffuse radiation.
Variation in microclimate often results in predictable changes in leaf traits (Poorter et al. 2006).The established paradigm is that photosynthesis is N limited (Evans 1989), and numerous studies have shown the significant positive relationship between leaf N content and maximum photosynthesis capacity (A max ) (Chazdon and Field 1987, Evans 1989, Evans and Poorter 2001), independent of species differences (Walters and Field 1987).Given this, forest canopies should optimize both the distribution of their leaves for high light capture efficiency and leaf photosynthetic rates according to their irradiance growth environment (Field 1983, Meir et al. 2002, Laisk et al. 2005).Canopy optimization of N distribution and photosynthetic capacity to light availability has been shown within a variety of crop and forest stands (Hirose et al. 1989, Hollinger 1989, Ellsworth and Reich 1993, Dang et al. 1997).
It remains unclear, however, how foliar acclimation and development adjusts to differing types and variability of irradiance (Meir et al. 2002, Bai et al. 2008).This is especially relevant when considering communities of diverse species, although the importance of such differences has been demonstrated (Chazdon and Field 1987).Understanding drivers of leaf trait variation in different species or functional groups (Poorter et al. 2006) is especially relevant in taxonomically and architecturally diverse, but light limited, tropical forest understory environments.A number of factors diminish the strength of relationships between light availability and investment in photosynthetic capacity, including light saturation, partitioning of nitrogen for nonphotosynthesis activities, leaf aging, and position (Field 1983), or variation in temperature, wind speed, precipitation and nutrient availability, as well as species differences (Dang et al. 1997).
Ecosystem processes within tropical forests, such as overall productivity (Baldocchi and Harley 1995), occur within a complex threedimensional architecture (Koetz et al. 2007).Interactions between architecture and microclimate require further study (Gastellu-Etchegorry and Trichon 1998).Failure to include spatial data on forest architecture, for example, can result in large errors from simple big-leaf models (Baldocchi andHarley 1995, Knohl andBaldocchi 2008).Attempts to estimate daily light regimes using traditional methods, such as hemispherical photographs, have resulted in inaccurate values, up to 107% greater than those shown from understory photosynthetic active radiation (PAR) sensors (Johnson and Smith 2006).Light regime modeling approaches explicitly integrating manually collected leaf area distributions showed greatly improved results (Aubin et al. 2000, Gersonde et al. 2004).
High-resolution light detection and ranging (LiDAR) sensors allow incorporation of spatially explicit information into microclimate models at scales infeasible through field data collection.Airborne LiDAR has recently been used to accurately estimate forest height (Hudak et al. 2002, Sexton et al. 2009, Dubayah et al. 2010), biomass (Asner et al. 2008a, Boudreau et al. 2008, Asner et al. 2012, Meyer et al. 2013), and architecture (Omasa et al. 2007), including gap dynamics (Koukoulas andBlackburn 2004, Kellner andAsner 2009).The capacity to quantify forest structure over large areas at high resolutions has led to insights into ecosystem function (Asner et al. 2008a, b), including patterns of canopy height heterogeneity not visible at smaller scales (Vitousek et al. 2009).Discrete LiDAR, in which a small number of individual laser pulses are used (Lim et al. 2003), has more recently, been combined with hyperspectral imagery and used to generate maps of leaf chlorophyll (Thomas et al. 2006), providing insights into flux tower measurements of gross ecosystem productivity (Thomas et al. 2009).Waveform LiDAR (i.e., wLiDAR) differs from discrete LiDAR sensors as it records a higher point cloud per area, approximating the complete waveform of the backscattered echo signal (Mallet and Bretar 2009), allowing more accurate estimations of forest understory architecture (Asner et al. 2007).This could lead to a better understanding of forest productivity if data approximates the fine scales at which canopy microclimate and ecophysiology are determined, but comes at the expense of greatly increased storage capacity and subsequent post-processing requirements (Mallet and Bretar 2009).Parker et al. (2001) used wLiDAR, one of the first attempts integrating this technology, to estimate nadir light transmittance statistics for two forest stands; however, a horizontal resolution of 10 m made detailed forest interior studies infeasible.Koetz et al. (2006) used physically-based radiative transfer models to invert large footprint waveform LiDAR (wLiDAR) accurately estimating forest biophysical parameters, including leaf area index (LAI), tree height, and general interior forest architecture.
In this research, we develop a new approach to map forest leaf area (2D) and leaf density (3D) at very high spatial scales.Using these maps, we develop and validate a three-dimensional model of direct and diffuse light transmittance and air temperature throughout a tropical rainforest in Hawaii.We then couple the microclimate models with detailed spatially explicit measurements of plant ecophysiological characteristics across a community of native and invasive species to understand structural, taxonomic, and climatic determinants of ecophysiological properties.The selected study forest, a model ecosystem having a near mono-dominant canopy species and both invasive and native species coupled with an extraordinary elevation gradient along the slope of Mauna Kea volcano, enables addressing questions related to relationships among forest structure, climate and ecophysiology not feasible in other systems.The specific research objectives of this study are to: (1) develop and validate a high resolution three-dimensional model of forest microclimate using a coupled airborne LiDAR-hyperspectral sensor; and then to (2) integrate remote sensing information and modeled microclimate data to better understand the taxonomic, structural and microclimatic determinants of foliar ecophysiology in our study area.

Study design
Fig. 1 presents the overall study design, and we describe in detail each component of the flowchart below.We combined airborne remote sensing data with spatially explicit measurements of forest microclimate and ecophysiology.We then developed detailed spatio-temporal models of microclimate and used these models to understand variation in foliar ecophysiology.We parameterized and validated remote sensing and modeling components using extensive field data.

Study site
This study was undertaken in the 5,016 ha State of Hawaii Hilo Forest Reserve and Laupahoehoe Natural Area Reserve, designated as a Hawaii Experimental Tropical Forest (HETF) of the US Forest Service (USFS), located on the North Hilo coast of the island of Hawai'i, Hawai'i.This reserve is also the location of a newly established Hawaii Permanent Plot Network (HIPPNET) and Center for Tropical Forest Science (CTFS) research plot (www.ctfs.si.edu).The reserve encompasses an elevation gradient from 600 to 1800 m elevation, with overall gradients in temperature and precipitation of 13-188C and2000-3500 mm, respectively (Giambelluca et al. 2011).A 2.5 km long by 800 m wide study transect was established in the northern central portion of the reserve extending from 1005 to 1343 m elevation (Fig. 2), corresponding to a mean annual temperature of 16.2-17.58C,respectively, comparable to the projected increases in global temperature over the next century (Nozawa et al. 2001).While three distinct substrate ages exist within the reserve: 4-14, 14-25 and 25-65 ty (ty ¼ 1000 yrs.), resulting from previous lava flows, the study transect was located entirely on youngest flow (4-14 ty).The transect consisted of two soil types, the lower half resting on the Akaka soil (rAK) and the upper half on Honokaa silty clay loam (HTD), both considered well drained with moderate available water capacity (websoilsurvey.nrcs.usda.gov,accessed 06/02/2011).The study transect was situated to keep the native Hawaiian tree Metrosideros polymorpha v. glaberrima (Myrtaceae) constant as the dominant canopy species, to the near exclusion of all others.Aboveground biomass (AGB) across the study transect ranged from approximately 500 MgÁha À1 at 1000 m to 250 MgÁha À1 at 1300 m (Asner et al. 2008a), simultaneous to a reduction in average canopy height from 24 to 14 m (Fig. 3).
The lower portion of the transect begins above Coprosma rhynchocarpa (Rubiaceae).The primary animal source of disturbance-constant throughout the study transect-consists of non-native feral pigs (Sus scrofa) which root the forest floor and propagate invasive species (Stone et al. 1992).

Study plots
Study plots were established at low (;1000 m) and high (;1300 m) elevations and positioned to encompass the range of forest structure found in the study transect.Six plots were located between 1000 and 1050 m and five between 1250 and 1300 m elevations.We established a 2 m by 30 m transect within each study plot (N ¼ 8).Data were collected for each stem greater than 0.5 m in height and included elevation (1000 or 1300 m), species, native vs. non-native status, height (m), and diameter (cm; at breast height v www.esajournals.orgwhen applicable, i.e., DBH).Volume (cm 3 ) was calculated as basal area (cm 2 ) multiplied by height (cm).Density (D ¼ no. of individuals/ 1000 m 2 ), dominance (Do ¼ P volume of all individuals/1000 m 2 ) and frequency (F ¼ number of transects containing the species) were calculated for each species.An importance value (IV) was calculated for each species using relative percent values (R) as compared to the median of all species, calculated as: Busby et al. (2010) and Curtis and McIntosh (1951).Georeferenced marker stakes were established within each study plot using a differentially corrected geographic positioning system (GPS) unit (GS-50þ, Leica Geosystems, St. Gallen, Switzerland) incorporating multiplebounce filtering.Following 6 to 8 hours of 5second interval GPS data collection per marker (N ¼ 10) final post-differential correction horizontal (XY) and vertical (Z) uncertainty were 19 6 13 and 33 6 24 cm (mean 6 SD), respectively.

Georeferencing
All interior forest measurements were georeferenced vertically and horizontally for integration with remote sensing data.The georeferencing procedure consisted of mounting a laser rangefinder with integrated inclinometer and 3D compass (Trupulse 360B, Laser Technology, Centennial, Colorado, USA) using filter mode and reflectors to avoid erroneous pulse returns on a tripod a known height (mh; cm) directly above a study plot marker stake.Position data returned from the Trupulse included the straight-line distance (sd; in meters), inclination (inc; in degrees) and azimuth (az; in degrees) from magnetic north.Prior to offset calculations, azimuth was adjusted to degrees from true north by adding a declination of 9.758 (www.ngdc.noaa.gov/geomagmodels/struts/calcIGRFWMM).Locations of offset locations were then calculated from the marker stake as: x offset ¼ hd 3 sinðazÞ ð 3Þ where hd ¼ horizontal distance (in meters), and x, y and z offsets are in meters from the marker stake.An accuracy assessment of geolocation offsets showed single offsets were accurate to ,30 cm in the vertical and horizontal dimensions over a wide range of distances (11-22 m), whereas double offsets, required in only a few instances when the marker stake had an obstructed view of the measurement location, were accurate to ,64 cm vertically and horizontally.

Climate measurements
Both top-of-canopy (TOC) climate and interior forest microclimate measurements were collected.TOC measurements were acquired continuously by stations at 1052 (i.e., low), 1180 (i.e., mid), and 1353 (i.e., high) m elevation, evenly spaced along the transect.TOC sensors at low and high-elevations consisted of a total quantum sensor (SQ-110, Apogee Instruments, Logan, Utah, USA), temperature and relative humidity, a sonic anemometer and precipitation (WXT-510, Vaisala, Helsinki, Finland) downloaded to a datalogger (CR-200, Campbell Scientific, Logan, Utah, USA).The mid-elevation sensor array consisted of a direct/diffuse quantum sensor (BF3, Delta-T Devices, Cambridge, UK), a total quantum sensor (LI-190, LI-COR, Lincoln, Nebraska, USA), and temperature and relative humidity sensors (HMP45C-L20, Vaisala) downloaded to a datalogger Campbell Scientific).Climate data collected every 15 seconds was averaged to a one-minute interval, with the exception of rainfall data that was the sum total each minute.Four mobile interior forest micro-climate stations were constructed, each consisting of a quantum sensor (SQ-110, Apogee Instruments), a temperature and relative humidity sensor (HOBO U23-002, Onset Computer, Bourne, Massachusetts, USA) and a cup anemometer (200-WS-01, Novalynx, Auburn, California, USA).PAR measurements, from the quantum sensors, and wind speed data were downloaded to a datalogger (CR-10x, Campbell Scientific) while temperature and relative humidity data were internally logged.Microclimate data were logged every 15 seconds and averaged to one-minute intervals.In addition, PAR data were logged every three seconds for the initial five minutes of each hour.
TOC PAR sensors were intercalibrated using known clear sky days to the mid-elevation quantum sensor, which was recalibrated annually, and calibration drift was removed using a clear sky PAR model coded in the R language (R Development Core Team 2013) and modified from equations provided by Apogee Inc. and as developed by the American Society of Civil Engineers (2005) (see Supplement for code).This model uses day of year, time of day, latitude, longitude, elevation, air temperature, and relative humidity as input variables and has been validated to estimate clear sky PAR within 3% at solar noon.Interior forest quantum sensors were intercalibrated weekly in an open field for two hours with data logged every 15 seconds averaged to one-minute intervals and returned for recalibration several times a year.

Leaf traits
Leaf trait measurements included light, CO 2 and induction gas exchange response curves, foliar mass per area, elemental C and N percentage, and d 13 C and d 15 N stable isotopes.Foliar gas exchange measurements were acquired using a LI-6400 portable infrared gas analyzer (LI-COR) on the dominant species identified by the species importance values.Single and double rope tree climbing techniques were used to collect in situ foliar gas exchange above 2.5 m in height, while tripods were used below that height.Additional data collected at each measurement location were: (1) species, (2) time and date, (3) DBH, (4) height of measurement, and (5) total height of plant.Photographs were collected for identification by botanists at the University of Hawaii at Hilo in cases where the species was not identified in the field.Gas exchange measurements were acquired at ambient leaf temperature, between 238 and 278C, on mature leaves with relative humidity maintained between 65% and 75% and following a minimum 30-minute LI-COR 6400 stabilization period.Most measurements were conducted with the cuvette leaf area at capacity (6 cm 2 ); however, when leaves smaller than 6 cm 2 were used, leaf area was measured in the field and gas exchange measurements were adjusted accordingly.Each of the three response curves were collected on separate leaves located immediately adjacent to each other and having similar characteristics.Curves were measured between the hours of 09:00 and 16:00 at a flow rate of 400 molÁairÁs À1 .
Light response curves were collected at a constant reference chamber CO 2 concentration (lmolÁCO 2 Ámol À1 air) of 400 and by increasing the photosynthetic photon flux density (PPFD or Q; lmolÁm À2 Ás À1 ), i.e., encompassing the 400-700 nm wavebands, stepwise from zero through saturating PPFD using the following increments: 0, 20, 40, 60, 80, 100, 130, 160, 200, 250, 300, 400, 800, and 1600.Measurements at each PPFD were logged when gas exchange was stable as indicatv www.esajournals.orged by: (1) visually stable intracellular CO 2 concentration (C i ; lmolÁCO 2 ÁmolÁair À1 ) and net CO 2 assimilation rates (A; lmolÁCO 2 Ám À2 Ás À1 ) values, (2) a total coefficient of variation (CV) percentage (calculated as the sum of CO 2 and H 2 O CV percent) of less than 0.1% and (3) following a min-max wait time of 3-10 minutes, respectively.CO 2 response curves, CO 2 assimilation rates versus the intracellular CO 2 concentration (AC i ), were collected at saturating PPFD þ 200 (lmolÁm À2 Ás À1 ) identified by the light response curve.Following a five-minute stabilization period at a reference chamber CO 2 concentration (lmolÁCO 2 Ámol À1 Áair) of 100, CO 2 concentration was increased stepwise through the following increments: 100, 300, 600, 900, 1200, and 1500.Measurements were logged at each increment using the same criteria as for light response curves, but with min-max time adjusted to 3-5 minutes, respectively.Induction response curves were collected following a five-minute stabilization period at a PPFD of 20.During the last 30 seconds of stabilization, measurements were logged every two seconds, following which PPFD was increased directly to 1300, and logging continued every two seconds for 3-5 minutes.Prior to analysis measurements from the light and CO 2 response curves were normalized for differences in leaf temperature to A at 258C through a custom version of the SiB2 photosynthesis model (Sellers et al. 1996) where A (i.e., A area ) is the net CO 2 assimilation (lmolÁCO 2 Ám À2 Ás À1 ) per area, A max is the maximum rate of A (the asymptote), [ is the apparent quantum efficiency (i.e., the initial slope of the fit hyperbola), p is the curve convexity parameter, A 0 is the dark respiration rate (lmolÁCO 2 Ám À2 Ás À1 ) and Q is the incident PPFD (lmolÁm À2 Ás À1 ).In addition, the light compensation point and light saturation estimate (lmolÁm À2 Ás À1 ) were calculated as Q value at which A ¼ zero and the linear intersection of [ and A 0 with A max , respectively.AC i curves were fit to a biochemical model of photosynthesis developed by Farquhar et al. (1980) and updated to account for triose-phosphate limitation (TPU) as described in Long and Bernacchi (2003), where net CO 2 assimilation (A) per area, dependent solely on mesophyll processes, is determined by the minimum of three potential limiters: Rubisco activity (V cmax ; W c ), RuBP regeneration (J max ; W j ) or the regeneration and utilization of inorganic triose-phosphate (V TPU ; W p ). Limitation typically shifts from W c to W j to W p with increasing C i , calculated by: where V cmax is the maximum rate of carboxylation by Rubisco (lmolÁCO 2 Ám À2 Ás À1 ), K c and c are the Michaelis-Menten constants of Rubisco for CO 2 and O 2 , respectively, and O is the stroma O 2 concentration (Pa).
where C ¼ 0.5 3 O/s; s is the specificity factor for Rubisco, and J, the whole chain electron transport rate, is: with h PSII ¼ curvature factor, Q 2 ¼ incident quanta available to PSII, and: where } 1 ¼ leaf absorptance, [ PSII,max ¼ max quantum yield of PSII, and b ¼ fraction absorbed light accessible by PSII.
where V o is the rate of oxygenation of Rubisco and TPU is rate of triose phosphate utilization (lmolÁCO 2 Ám À2 Ás À1 ).The determination of A at C i (x) is: where R day represents the CO 2 released through non-photorespiration processes (lmolÁCO 2 Á m À2 Ás À1 ).Induction response data were analyzed through non-linear parameterization of a two parameter modified rectangular hyperbola model in JMP software (Version 7; SAS Institute, Cary, North Carolina, 1989-2007) developed for induction response analysis (Hunt et al. 1991, Poorter andOberbauer 1993).The model is defined as: where A t is assimilation at time t (seconds), k i is the induction curve convexity, t is seconds post-PPFD increase and A max is net CO 2 assimilation per unit area.Using the output A max value, the input data were then converted to induction state, redefining A as percentage of A max at time t (seconds) post-induction, and the equation was then reparameterized using an A max equal to 100% to obtain k is , a value comparable across leaves varying in A max .Output results for statistical analysis were the estimated A max , k i , k is , and time (seconds) to 50% induction state (IS50%) calculated as: where t is time post induction in seconds, IS equals the induction state (%), and A max and k i are as defined above.
All leaves on which gas exchange measurements were conducted were collected and scanned at 600 dpi (x9575, Lexmark International, Lexington, Kentucky, USA) for leaf area calculation in Photoshop (CS, Adobe Systems, San Jose, California, USA) within 12 hours.Leaves were then oven dried at 558C for 48-56 hours, weighed (0.01 mg; Mettler Toledo AG245), and ground to a fine powder using a Wiley Mill (Thomas Scientific, Swedesboro, New Jersey, USA) fitted with a 40-mesh screen.Samples were analyzed for C and N concentration, and d 13 C and d 15 N, using a Vario Microcube elemental analyzer (Elementar Analysensysteme, Hanau, Germany) coupled to an isotope ratio mass spectrometer (Isoprime, Manchester, UK) operating in continuous flow mode at the Stable Isotope Laboratory at Tulane University.Samples were normalized to international isotope scales by bracketing with USGS-40 and USGS-41 glutamic acid standards (calibrated to the international VPDB (d 13 C) and AIR (d 15 N) scales) and repeated analysis of sorghum flour was used to assess instrumental drift during runs as well as differences between runs.Stable isotope data are expressed using ''delta'' notation (Ometto et al. 2006).
Additional variables were defined as follows: specific leaf area (SLA) is the projected leaf area per unit leaf dry mass (cm 2 /g) (Evans and Poorter 2001, Martin and Asner 2009, Liu et al. 2010), leaf mass per area (LMA; g/m 2 ) (Cordell et al. 1998), N area is nitrogen content per area (g/m 2 ) (Ellsworth and Reich 1993, Dang et al. 1997, Cordell et al. 1998), photosynthetic nitrogen use efficiency (PNUE ) is the ratio of A max to N area (lmolÁCO 2 Ás À1 Ámol À1 N) (Cordell et al. 1998, Funk andVitousek 2007), water use efficiency (WUE) is the ratio of A area to transpiration rate (lmolÁCO 2 per mmol H 2 O) under saturating PAR (Funk and Vitousek 2007), and A mass is the ratio of A max to unit leaf dry mass (nmolÁCO 2 Ág À1 Ás À1 ) (Ellsworth and Reich 1992).We also calculated the ratio of IS50%/A max , to understand optimization of induction response time, which we included in the Akaike Information Criteria (AIC) and best subsets regression analyses described in Data integration, below.

Airborne remote sensing
The study transect was imaged by the Carnegie Airborne Observatory (CAO) Alpha sensor system in January 2008 at a height of 500 6 50 m.The CAO-Alpha integrated a high-fidelity hyperspectral imager (HiFIS) having 72 bands distributed from 368 to 1040 nm, a waveform LiDAR (wLIDAR) scanner operating at 1064 nm and 100 kHz and a Global Positioning System-Inertial Measurement Unit (GPS-IMU).An automated processing stream incorporated ortho-georectification and atmospheric correction for a final spatial accuracy of ,15 cm in the vertical or horizontal dimension (Asner et al. 2007).HiFIS data had a final spatial resolution of 1.25 m while wLiDAR data were collected at 0.56 m. wLiDAR pre-processing included noise reduction, deconvolution, waveform registration and angular rectification (Wu et al. 2011).wLiDAR point clouds were processed to proportional data by summing points within 0.56 3 0.56 3 0.15 m (XYZ ) voxels, and dividing each voxel's value by the summed total points in each 0.56 3 0.56 m vertical profile throughout the study transect.Ground and tree crown topography maps were generated through analysis of point cloud data.Solar azimuth at time of data collection was calculated in IDL using solar geometry and tree crown topography.
LAI (m 2 leaf area/m 2 ground area) was calculated using paired LAI-2000 (LI-COR) units in remote mode for 49 locations randomly distributed throughout the study transect at a height of 155 cm.We calculated a map of the topof-canopy cosine angle to the sun using the tree (primarily M. polymorpha) canopy topography maps and per-pixel acquisition time.HiFIS data were resampled to 0.56 3 0.56 m from 1.25 3 1.25 m spatial resolution to be directly comparable with the wLiDAR data using the nearest neighbor method.The modified red edge normalized difference vegetation index (mNDVI; see Gitelson andMerzlyak 1996 andGamon andSurfus 1999): was applied to the image and we then removed the effect of shade through a linear regression between mNDVI extracted for a 2 m radius surrounding each LAI field location and the cosine value (P , 0.0001; R 2 ¼ 0.5367), then calculated LAI as the linear relationship between field calculated LAI values and the difference between the cosine predicted mNDVI value and that calculated from the image.The final relationship was highly significant (P , 0.01; R 2 ¼ 0.3401; N ¼ 49) and was applied to the entire HiFIS image to derive a detailed LAI map of the study area.
Leaf area density (LAD), defined as m 2 of leaf area per m 3 of volume, was calculated throughout the study area using the proportional data derived from the corrected wLiDAR points.Three-dimensional maps of LAD were calculated by converting two-dimensional LAI values from m 2 to pixel scale (0.56 3 0.56 m ¼ 0.3136 m 2 ) and distributing the leaf area across the vertical profile according to the proportion values obtained from the wLiDAR proportion maps at a vertical resolution of 0.15 m, as below: where LAD ¼ leaf area (m 2 ) within the vertical profile (0.56 3 0.56 m) at height x to x þ 0.15 m, LAI ¼ leaf area index (m 2 leaf/m 2 ground area), and pwf ¼ proportion waveform LiDAR points occurring at x.The wLiDAR correction process was validated for this study site using field leaf area density profiles (N ¼ 13) ranging from 10 to 24 m in height with horizontal and vertical resolution of a 0.2463 m 2 and 0.5 m, respectively.Field LAD profiles were collected across diverse forest structure types by establishing and rappelling off horizontal Tyrolean rope traverses between tree canopies, collecting all leaves and measuring their collection height using an ultrasonic range finder (SONIN, Charlotte, North Carolina, USA) with a sonic target to reduce erroneous returns.Leaves were stored in zip lock bags with moist paper towels until leaf area (cm 2 ) was calculated within 24 hours using a LI-3100 (LI-COR).A significant linear relationship was shown between the cumulative percentage of leaf area identified rebinned to comparable vertical (0.5 m) and horizontal (pixel area of 0.3136 m 2 ) resolution (P , 0.0001; R 2 ¼ 0.4960; N ¼ 404), and no significant difference in LAD values was identified in a matched pairs analysis (P .jtj ¼ 0.1253; N ¼ 404).Virtual forests were then generated through integration of the surface elevation and 3D LAD (cm 2 ) data, which were used in the subsequent modeling analyses.

Interior forest microclimate modeling
Interior forest climate data were compared to interpolated TOC climate data at each study plot center point.TOC values were derived from the weighted average-based on elevation-of the most proximate pair of TOC sensors.For example, for a position located between the mid and high TOC towers the interpolated TOC values would be: v www.esajournals.orgTOC interp ¼ðTOC high 3 pÞ þ TOC mid 3ð1 À pÞ : Interpolated total PAR values were proportioned according to diffuse and direct PAR measured at the mid elevation tower which was the only location with a BF3 direct vs. diffuse sensor installed.Separate direct and diffuse PAR models were developed.Direct PAR utilized solar azimuth and elevation calculated through interpolation of National Aeronautics and Space Administration (NASA) Jet Propulsion Laboratory (JPL) planetary ephemeris (DE405; http:// ssd.jpl.nasa.gov)using IDL code modified from that provided by Craig Markwardt (http://cow.physics.wisc.edu/;craigm).
The direct PAR model calculated the distribution of leaf area density at 0.25 m increments from zero to 100 m distance from the sensor directly towards the sun position, which was calculated at one-minute intervals.LAD was then adjusted according to its distance from the sensor using the following equation: where DirSF is the direct structure factor, k is the extinction parameter set to 0.025, sd is the straight-line Euclidean distance from the sensor (x), and LAD is the leaf area density (cm 2 ) encountered at distance (x).DirSF calculation was limited to the daytime, defined as solar elevations !258.The diffuse PAR model used the same approach for each location to calculate the diffuse structure factor (DifSF) but averaged DirSF values from 36 combinations of azimuth and elevation (.22.58) equally distributed across a hemisphere above the sensor.Total interior forest PAR was modeled at onehour intervals using the averaged PAR and structure data.Interior PAR was calculated: The constants a-d equal: 2.3638, 0.3633, 0.0304 and 3.792E-04, respectively.tPAR toc and tDirSF represent power transformed (t; i.e., (UÙk) À 1/k) versions of the raw variables conducted to normalize residual distributions using lambda values of 0.28 and 0.44, respectively.Parameterization was conducted in JMP on a randomly selected 50% of the available interior forest data during daylight hours.The remaining data were used to validate the model.Both the parameterization model (P , 0.0001, R 2 ¼ 0.6566, N ¼ 355) and validation (P , 0.0001, R 2 ¼ 0.6622, N ¼ 371) were highly significant.
Average air temperature was predicted at each sensor location using the environmental lapse rate calculated at 30-minute increments as follows: where Temp pred is the predicted air temperature (8C) at time t and INT elev is the elevation at the interior forest sensor.MID elev and MID temp is the elevation and air temperature at the mid elevation climate tower, and MLR is the mean environmental lapse rate calculated among the high, mid and low elevation climate towers.This relationship was highly significant (P , 0.0001, The final forest interior air temperature model compared the measured versus the predicated air temperatures as influenced by DirSF and DifSF as: The constants a-c equal: 1.014, 6.1687E-04, and 4.984E-06, respectively.The sample size was reduced as DirSF was calculated for daytime hours only.The model was parameterized and validated as described for PAR above and was highly significant (P , 0.0001, R 2 ¼ 0.8914, N ¼ 1529).

Data integration
An overview of the data integration approach is illustrated in Fig. 1.For each location where ecophysiological data were collected, 4400 minutes of daytime climate data were randomly selected between December 17, 2010 and June 1, 2011, averaged to half hour intervals, and the following variables were calculated using the field data collection, climate models and remote sensing data: (1) ground elevation (m), (2) canopy, plant and leaf height above forest floor (m), (3) DBH (cm), (4) mean and standard deviation of modeled available PPFD, and (5) modeled average air temperature (8C).Highly correlated variable groups were identified using Pearson correlation analysis (.0.7), resulting in the final selection of the leaf trait predictor variables: (1) leaf height, (2) PPFD mean and (3) SD, and (4) mean air temperature.Ground elevation, although highly correlated with mean air temperature, was included as well to account for potential unquantified variation in climate and forest dynamics along the elevation gradient.
Predictor variables were transformed, when appropriate, to create a normal distribution, and the importance of all variables on each ecophysiological variable was assessed using: (1) community scale; best subsets multiple regressions with the most significant combinations of predictor variables identified using the adjusted R 2 value, and (2) for the dominant canopy and understory species-M.polymorpha (Ohia) and C. glaucum, respectively-and the entire species community, linear regressions between selected ecophysiological variables and leaf height, PPFD mean, and air temperature mean.Separate general linear models were fit to understand differences due to the following classifications: (1) species, (2) life-form, (3) exotic vs. native, (4) height strata, (5) canopy position, and (6) M. polymorpha or other.Life forms were defined as herb, fern, liana, shrub, tree fern, understory tree or canopy tree.Height strata were defined as ground, mid or upper.Canopy positions were defined as understory or canopy.Models were compared using AIC values (Mazerolle 2006, Mutua 1994) and weight (Anderson 2008), which adjusts for differences in parameter size, using identical data sets.AIC provides a method to compare relative model goodness of fit for a specific foliar variable, with models D i .10 above the minimum AIC having little support (Burnham and Anderson 2004).For our analysis, we kept the two best models and discarded those having AIC values D i .20.
We sought to identify groups having similar ecophysiological characteristics using two approaches.We used principal components analysis (PCA; Reich et al. 1999) to assess if general trends in ecophysiological variables (N ¼ 22) existed.We then identified significant correlations between PCA axes and box-cox transformed, for increase normality, PPFD and air temperature.We clustered the foliar dataset into three groups through K-means analysis using the first three PCA axes and compared ecophysiological, structural, and climatic variables among these groups using one-way ANOVAs and Pearson tests.

Plant diversity and structure
Top-of-canopy (TOC) height declined from 20.4 6 8.1 m at 1000-1049 m elevation to 13.2 6 5.7 m at 1300-1349 m elevation (Fig. 3).LAI remained constant across the study area (3.9 6 1.4 m 2 /m 2 ).A total of 24 species were identified in the plant diversity transects (Table 1).Importance values ranged widely, with M. polymorpha, C. glaucum and C. trigynum identified as the three most important species (Table 2).The canopy was comprised almost exclusively of M. polymorpha, although its seedlings and saplings also existed in more open understory environments, with the next strata comprised mostly of C. trigynum, C. rhynocarpa, and I. anomala.A final strata occurring at 3-5 m height consisted almost entirely of C. glaucum.Species growing below 2 m included an abundance of the exotic species H. gardnerianumi and occasional P. cattleianum individuals.Open wet areas at low elevations were dominated by the exotic species P. punctata (Smartweed) below 1 m height, with shrubs including young M. polymorpha and the exotic species C. hirta.

Microclimate
Table 3 summarizes mean TOC climatic conditions recorded at the low, mid and high elevation towers between December 17, 2010 and May 16, 2011.Seasonal dynamics of PPFD and air temperature are provided in Fig. 4 and diurnal dynamics are provided in Fig. 5. Measured mean air temperature dropped from 15.938 to 14.138C from the low to high elevation tower respectively.The low elevation tower received 58% more rainfall than the high elevation tower.Total daytime PAR was nearly equal among the low, mid and high elevation locations with the mean  diffuse percentage equaling 66% of daily PAR.
The best subsets multiple regression analysis revealed strong climatic and structural determinants of foliar ecophysiology (Tables 6 and 7).Most leaf traits were correlated with leaf height (20 of 24), followed by modeled mean PPFD (14 of 24), elevation (7 of 23), and modeled mean air temperature (7 of 24).Of all leaf traits, leaf height was most significantly correlated with foliar C (positive; adjusted R 2 ¼ 0.40), followed by SLA (negative; adjusted R 2 ¼ 0.35) and IS50% (negative; adjusted R 2 ¼ 0.36).A max , and associated %N, were strongly positively correlated with modeled mean PPFD, and foliar C:N was negatively correlated with modeled mean PPFD.While d 15 N was significantly positively correlated with modeled mean air temperature (adjusted R 2 ¼ 0.28), d 13 C was not correlated with any structural or climatic variables.Rates of respiration increased with modeled air temperature but decreased with increasing leaf height.
Linear regressions between leaf traits and the predictor variables defined in Table 8 were calculated for M. polymorpha and C. glaucum (Table 8A, B), the overstory and understory species having the greatest importance values (Table 2), as well as across the entire community (Table 8C).Significant positive relationships existed between modeled mean PPFD and A max across the community (adjusted R 2 ¼ 0.26), and at the species scale for M. polymorpha (adjusted R 2 ¼ 0.13) and C. glaucum (adjusted R 2 ¼ 0.40).Light saturation was strongly correlated with modeled air temperature for C. glaucum.The dynamic response time (IS50%) of C. glaucum increased in higher modeled light environments, while that of M. polymorpha did not.Although increases in foliar C:N were shown at the community level (adjusted R 2 ¼ 0.18), no such relationships existed within the individual species.
K-means analysis of these axes revealed three distinct groups within the foliar dataset (Table 10 and Fig. 6), which following analysis were found to be sorted by growth light environment (Table 11).The low light cluster had lower maximum rates of photosynthesis (i.e., A max and A mass ), including light saturation, but was able to reach A max quickly as compared to the medium and high light clusters.Leaves in this cluster had higher leaf mass per area, foliar %C, and water use efficiency (WUE) and lower foliar %N, and photosynthetic nitrogen use efficiency (PNUE).The high light cluster had very high rates of A max (and A mass ), PNUE, light saturation, triose phosphate utilization (TPU), V cmax , and J max .While many values were similar to other clusters, the medium light cluster was distinguished by intermediate values of A max , A mass , PNUE, induction response time (IS50%), and lower values of V cmax , J max , light compensation and saturation, and WUE.The medium light cluster had the lowest leaf mass per area.
We identified taxonomic, climatic and structural differences among the clusters (Table 11).The low light cluster was comprised of all tall plants with large DBH values whose sampled leaves occurred in low and less variable (i.e., low modeled standard deviation of PPFD) light environments.This cluster was composed entirely of native species and dominated by the dominant canopy species M. polymorpha and the understory tree fern C. glaucum.The medium light cluster was dominated by native species growing in light environments intermediate between the low-and high-modeled light clusters and often situated within the mid height strata.The high light cluster, similar in many respects to the medium light cluster, was composed of low height plants-the lowest strata In general, leaves from any individual species were grouped in the same cluster (78 6 18%; Table 2).A significant positive relationship between A max (log-transformed) and mean modeled PPFD was found across the community the entire dataset (adjusted R 2 ¼ 0.25, P 0.0001, N ¼ 152).However, individual clusters had different A max to mean modeled PPFD relationships.No significant relationship existed for cluster one (P ¼ 0.2766, N ¼ 152), but a significant positive relationship existed for cluster two (adjusted R 2 ¼ 0.16, P 0.0072, N ¼ 44) and cluster three (adjusted R 2 ¼ 0.26, P 0.0040, N ¼ 30; Fig. 7).

DISCUSSION
The main components of this project were performed as follows: (1) development of a twodimensional map of LAI using airborne hyperspectral imagery; (2) derivation of a threedimensional LAD map through integration of the two-dimensional LAI map with vertical profiles provided through airborne wLiDAR; (3) coding and validation of PAR and air temperature microclimate models integrating TOC climate measurements with forest structure from the LAD map; and (4) integrating modeled microclimate information with remote sensing and detailed field data to predict leaf traits and gas exchange dynamics for a suite of species occurring within a range of forest structural types (i.e., closed, open) along an elevation gradient.

Remote sensing
While passive and active remote sensing techniques have proven useful for large-scale analyses (Asner et al. 2005), remote sensing studies have historically been limited in their ability for finer scale analyses of ecosystem function.Recent advances have increased the capacity of remote sensing to integrate with ecosystems at scales appropriate for detailed functional analysis (Chambers et al. 2007).Space-borne hyperspectral imaging resulted in development of techniques to link remote sensing more directly to plant physiological traits (Asner et al. 2004(Asner et al. , 2005)).Airborne hyperspectral analysis provided finer spatial resolution studies, allowing detection of species composition and foliar properties (Carlson et al. 2007).Studies using large footprint LiDAR (Koetz et al. 2007) showed the utility of three-dimensional structural information, and airborne platforms have now integrated hyperspectral sensors with LiDAR systems (Asner et al. 2007).With this fusion, Notes: See Appendix: Table A3 for an explanation of abbreviations.The curve from which variables are calculated is provided in parenthesis following the variable name.
A max values combined from light response and induction response curves.
v www.esajournals.orgsimultaneous analysis of ecosystem structure and foliar traits has become feasible (Asner and Martin et al. 2008).There is now an increasing application of sensor fusion to elucidate forest properties at scales ranging from individual trees to global analyses (Todd et al. 2003, Reitberger et al. 2009, Zolkos et al. 2013).
A primary objective of many remote sensing studies has been to understand forest function in three dimensions spatially (Omasa et al. 2007).Efforts to better understand spatial properties of ecosystem dynamics, including productivity and canopy chemistry (Asner and Martin 2008), however, has been limited due to the difficulty of acquiring maps of forest interior structure and leaf area distribution (Houldcroft et al. 2005).In tropical forests especially, this difficulty has stemmed from the rapid extinction of the LiDAR signal in the dense overstory, and by the difficulties associated with collecting field pa-rameterization and validation data within forest canopies (Laman 1995).Methods developed in this project have helped surmount some canopy access issues, in particular, intra-crown access using vertical transects off horizontal Tyrolean traverses.
Spatial resolution represents a serious obstacle to fully understanding forest dynamics.Interior forest light is known to be the primary limiter of photosynthesis in many tropical forests understories (Whitmore 1996).The distribution of foliage within a forest plays an integral role in light distribution (Chazdon et al. 1988, Chen et al. 1994, Montgomery 2004).Light that does reach the understory arrives as either direct or diffuse radiation, and global increases in diffuse radiation are predicted for the next 100 years (Mercado et al. 2009).Baldocchi and Wilson (2001) highlighted this importance through a modeling analysis which showed that differences in leaf distribution throughout the forest vertical profile can alter forest net primary productivity (NPP) by up to 50%, largely through alteration of available PAR (Chazdon and Pearcy 1986a, b).
Without an improved understanding of interactions between leaf area density distribution and light major errors will continue to exist in forest productivity models.However, no significant advances have been made in modeling 3D interior forest light dynamics at fine scales, although advances using medium-large footprint LiDAR are ongoing (Parker et al. 2001, Thomas et al. 2006).This has been due to several reasons including: (1) fine spatial scale of leaf area density determining direct light, and (2) the lack of a wLiDAR equipped remote sensing platform (Mallet and Bretar 2009) and decomposition algorithms (Wu et al. 2011) enabling detailed 3D analyses at a high spatial resolution.However, even field-based approaches have encountered difficulties and required inclusion of extensive stand structural information (Sonohat et al. 2004).
The development of the Carnegie Airborne Observatory integrating high pulse rate waveform LiDAR with a hyperspectral sensor has begun to overcome limitations to forest interior Notes: See Appendix: Table A3 for an explanation of abbreviations.The curve from which variables are calculated is provided in parenthesis following the variable name.
A max values combined from light response and induction response curves.A3 for an explanation of abbreviations.The curve from which variables are calculated is provided in parenthesis following the variable name.
v www.esajournals.orgmicroclimate modeling (Asner et al. 2007).The data collected for this study is among the highest spatial resolution available and allows us to test the feasibility of modeling microclimate dynamics at a temporal and spatial scale directly comparable to the scale of a leaf's growth environment.Future improvements to the PPFD and air temperature microclimate models we develop and parameterize in this study would include the use of interior forest PAR sensors distinguishing between direct and diffuse light and parameterization for other tropical forests with differing structure and gap dynamics (Kellner and Asner 2009).This would allow for an improved understanding of interactions between forest structure and the direct/diffuse light ratio versus the approach used in our study that combined direct and diffuse light as total PAR.
This study represents a development and validation step towards a rapid large-scale remote sensing based approach to model detailed forest productivity.Johnson and Smith (2006) have highlighted the need for such data.In addition, Alton et al. (2007) state that understanding how climate change will interact with plant photosynthesis is a key issue requiring further study.An approach built off an airborne system allows for rapid and economic collection of detailed forest structural measurements, and thereby models of microclimate and foliar ecophysiology, over a wide variety of forest types.Such efforts will enable a more unified understanding of climate change effects on the threedimensional dynamics of forest photosynthesis and physiology at larger scales.For example, such information would be appropriate for integration with flux towers, which are providing significant insight into forest productivity dynamics (Schwalm et al. 2010).

Sources of leaf trait variation
Many factors influence a leaf's physiological traits, including: (1) structural parameters, such as the plant's height (Kenzo et al. 2006); and (2) site specific differences, such as general climate (Reich et al. 1996) or soil fertility (Ordoñez et al. 2009).Inter-species differences (Hikosaka 2004), result from divergent competitive growth strategies (Poorter et al. 2006) or simply life form (Wright et al. 2005), while intra-species occur as individual plants optimize their nutrient alloca-Fig.6. Principal leaf trait clusters identified through K-means analysis.Low, medium and high light clusters are represented by the colors red, blue, and green, respectively.v www.esajournals.orgtion to maximize productivity (Field and Mooney 1986, Hirose and Werger 1987, Hollinger 1989).The importance of vertical distributions in leaf traits, representing broad changes in micro-climate (Kumagai et al. 2001) and hydraulic limitations (Taylor and Eamus 2008), has been identified in many studies (Domingues et al. 2005).Leaves positioned in the upper canopy generally increase net CO 2 uptake while those situated in the lower canopy have reduced (or negative) uptake as the proportion of maintenance respiration costs are increased relative to C gain (Ellsworth and Reich 1993).Surprisingly, our analysis showed no correlation between leaf height and mean daily PPFD (Table 12).This difference is likely due to the forest structure in our study area, having a low LAI open canopy with dispersed tall relatively small DBH trees, differing from those in many tropical areas that have very high LAI canopies with heavily shaded forest interiors.Given this, horizontal differences in topography, low to mid-story leaf area, and gap dynamics may be dominant controls over microclimate variation, rather than vertical gradients as found in other studies in tropical forests (Domingues et al. 2005).A broad suite of leaf traits have been shown to co-vary, including positive relationships between foliar %N and A max (Field andMooney 1986, Reich et al. 1997), as we likewise find in this study (Appendix: Tables A1 and A2).This relationship partly results from the availability of photosynthetic enzymes limiting photosynthetic capacity (Field 1983).In addition to leaf N concentration and A max , leaf photosynthetic induction rates, i.e., the activation and synthesis of photosynthesis related biochemical components and stomatal movements (Pearcy 1990) vary with both species and growth light environment (Bazzaz andCarlson 1982, Portes et al. 2008).

Species and structure
The single greatest source of leaf trait variation found in our study was inter-species differences (Table 4).Wright et al. (2004) used a global database of leaf traits, including 2,548 species, and found large variation among functional groups but strong co-variation among leaf traits, consistent with changes in species growth strategies along a continuous 'leaf economics spectrum' (Wright et al. 2004), constructing short lived, low LMA, high A max leaves to long lived, high LMA, low A max leaves.Likewise, Popma et al. (1992) found variation among species resulted from specialization to different growth environments, with gap-independent species producing nutrient poor leaves with low photosynthetic rates.In addition, they found that species adapted growing in a wide range of light environments show larger phenotypical plasticity in leaf traits.Markesteijn et al. (2007) also found leaf trait variation among 43 tropical forest tree species to be dominated by inter-species differences, with short-lived pioneer species having the greatest leaf trait plasticity.
Many studies in Hawaii have focused on differences between species of exotic and native origin (Hughes and Denslow 2005).This topic is of conservation importance as many native ecosystems are undergoing invasion in which native dominated species composition becomes altered or dominated by exotic species, resulting in changes in composition, structure, and other ecological properties (Vitousek et al. 1987).Among other reasons, exotic species may invade native ecosystems following disturbance or when native communities have low resource use resulting in resource availability, termed the Fig. 7. Relationships (log-log regressions) between A max (maximum CO 2 lmolÁm À2 Ás À1 ) and modeled mean daily PPFD (lmolÁm À2 Ás À1 ) for the entire community and modeled low, medium and high light leaf trait clusters.Regressions for the community and modeled medium and high light leaf trait clusters are significant (P , 0.05).
v www.esajournals.org'fluctuating resource hypothesis' (Funk and Vitousek 2007).Funk and Vitousek (2007) studied the resource use efficiency (RUE) of invasive and native species and found that invasive species used limiting resources more efficiently, as indicated by higher photosynthetic rates (A max ), higher photosynthetic nitrogen use efficiency (PNUE), although water use efficiency (WUE) was not different.Asner et al. (2006) used remote sensing coupled with the CASA carbon cycle model and found that growth rates of M. faya, an invasive species, were 16-44% higher than M. polymorpha.In our study, exotic species exhibited significantly greater rates of photosynthesis and nutrient use, including higher foliar N and A max .These findings are in line with other studies showing higher growth rates on nonresource limited sites, a reasonable assumption for our study area given the young age of the substrate.The lack of resource limitation is also exhibited in the non-significance of PNUE between our exotic and native species, a result found by Funk and Vitousek (2007).

Light and air temperature
Photosynthetic active radiation is a dominant limiting factor to total forest photosynthesis (Whitmore 1996, Kull 2002, Graham et al. 2003) and its availability plays a major role in survival, growth, ecology and physiology of forest plants (Chazdon andPearcy 1986a, Myneni andGanapol 1992), especially in the understory (Capers and Chazdon 2004).Photosynthetic active radiation dynamics vary greatly between forests of different stand architecture, even when species composition remains the same (Sonohat et al. 2004).In addition to light, air temperature is a primary determinant of photosynthesis and respiration (Berry and Bjorkman 1980) and increases in mean annual temperature (MAT) have been shown to correspond to increased total net primary productivity (Raich et al. 2006).We see the importance of these two environmental gradients expressed in our PCA analysis, with axis one being significantly correlated with modeled light and axis two correlated with modeled temperature.The fact that the first two axes encompass only 51% of the variability in leaf traits across our community may be partly explained by the exclusion of species, identified as a significant source of leaf trait variation, from the PCA.Preliminary analysis on our dataset does reveal the possibility of models with greater significance (PCA 1 vs. MM-PPFD; adjusted R 2 ¼ 0.58, P , 0.0001, df ¼ 77) and merits further investigation.
Differences in light and temperature in the plant growth environment are directly correlated with changes in leaf photosynthetic capacity, as well as a broad suite of ecophysiological characteristics (Wright et al. 2004(Wright et al. , 2005)).Field (1983) showed that plants optimize the distribution of N within their leaves according to the distribution of daily photosynthetic active radiation.Evans and Poorter (2001) found that photosynthesis was three times greater in ten dicotyledonous C 3 species grown under 1000 versus 200 lmolÁm À2 Ás À1 PPFD, however photosynthesis per unit leaf dry mass was not significantly different due to increased specific leaf area (SLA) but constant nitrogen concentration.Hollinger (1989) showed that leaf N content and A max followed vertical gradients according to available PPFD in the forest canopy.However, relation- ships are not linear and photosynthesis of plants located in the understory often become light saturated at less than half of full sunlight intensity (Lambers et al. 1998).This may partly explain the low correlation we found between mean PPFD and light saturation at the community scale.
In our study, most leaf traits did follow vertical gradients, as shown by the significance of leaf height, but were also simultaneously correlated with light and air temperature; likely resulting from inconsistencies in the leaf height to PPFD relationship, which will be focus of future studies.Chazdon and Field (1987) found that in understory plants the light environment was only able to explain a maximum of 35.1% of the variation in A max , indicating that other determinants, including climatic variation, topography or resource competition, were playing important roles.They also found that understory plants (versus plants growing in open gap environments) were less able to adjust A max to variation in the light intensity of their growth environment, although compensation through increased light use efficiency was possible (Chazdon and Pearcy 1986b).Similar to their results, however, we found that modeled light variability did not significantly correlate to Amax, while simpler measures, such as canopy openness in their study, or in the case of our study modeled mean daily PPFD, did.This is similar to the percentage of variation in A max explained for the C. glaucum, the dominant understory species in our study area.While direct radiation accounts for between 10 and 80% of total understory irradiance, and in some cases the majority of carbon fixation (Pearcy and Calkin 1983, Chazdon 1986, Pearcy 1990), it arrives to the understory of a forest within an intact canopy in the form of sun flecks lasting from seconds to minutes (Chazdon and Fletcher 1984).Thus, it is to the advantage of understory plants to develop the foliar capacity to rapidly use short temporal bursts of light (i.e., sunflecks) (Chazdon and Pearcy 1986a).
We found a positive correlation between modeled PPFD and time to reach 50% maximum assimilation rate (IS50%), differing from Rijkers et al. (2000), who found no differences in time to reach 90% maximum assimilation in spite of large differences in A max .At first, this relationship appears to indicate that the dominant control over induction response time is the maximum photosynthetic rate of the leaf, although the Pearson correlation of 0.22 is low (Appendix: Table A1), and A max is related to species differences and the illumination of the growth environment.However, further analysis shows that while A max is related positively with growth environment, IS50% is more related to leaf height where it declines with height, although leaf height does have a significant negative correlation with modeled illumination environment.Given the low Pearson correlation, there is an opportunity to identify where leaves simultaneously maintain high maximum photosynthesis rates and rapid induction response times using the ratio of IS50%/A max , which we consider a measure of induction response efficiency (i.e., lower values ¼ more efficient).Our results show increasing induction response efficiency with increasing leaf height and in higher modeled light growth environments, but decreasing efficiency with increasing modeled air temperature.The relationship with increasing leaf height is logical as PPFD actually decreases at the base of the M. polymorpha canopy prior to increasing near the TOC.Decreasing efficiency in warmer environments also makes sense as increased availability of nutrients could results in reduced requirement for resource use optimization.As a plant likely shifts a range of leaf traits simultaneously, we expected to find a pattern of increasing WUE simultaneous to induction response efficiency.A linear regression analysis showed this significant negative relationship between them (adjusted R 2 ¼ 0.17, P 0.0001, N ¼ 90), indicating simultaneous increases in light and water use efficiency.Similar to the findings of Funk and Vitousek (2007), we did not find significant differences in either water use efficiency or induction response efficiency between exotic and native species.

Stable isotopes
Foliar d 13 C is typically thought to represent changes in foliar water use efficiency (WUE) resulting from differences in water availability (Seibt et al. 2008) or structural differences in the leaf (Bonal et al. 2007).Variation in foliar d 13 C due to differences in atmospheric isotopic composition (i.e., increased respired versus atmospheric CO 2 ) would be unlikely to have an effect higher than 3-5 m (Ometto et al. 2002) and leaf height was a non-significant predictor of variation in foliar d 13 C.Our d 13 C values were similar to those reported in the Amazon (Ometto et al. 2006).However, for M. polymorpha, the canopy dominant in our study area, different conclusions regarding determinants of d 13 C variation have been found.Vitousek et al. (1990) found that, for Ohia (M.polymorpha), internal CO 2 resistance to diffusion resulting from increased LMA was a primary factor determining variation in d 13 C, similar to the results found by Ko ¨rner and Diemer (1987) who identified a pattern of increasing elevation, LMA, and d 13 C.In addition, a positive relationship between increased carboxylation efficiency and leaf area based N content, associated with increasing elevation, was found to result in significantly less negative foliar d 13 C values (Cordell et al. 1999).Both studies agree that species differences is likely a primary factor determining d 13 C, but is often overlooked during analysis, and Seibt et al. (2008) argues the potential for species differences in mesophyll conductance may mask trends in WUE.This is highlighted in our study, as only species differences (or as lumped into life-form groupings) were able to explain variation in foliar d 13 C, with no correlations found between foliar d 13 C and any other leaf trait or predictor variable, including modeled mean PPFD or air temperature.As water is in abundant supply across the entire study area this result most likely represents maximized species variability when no potentially cofounding variation in water availability exists.We did find, however, that photosynthetic WUE increased significantly with increasing height in the canopy.We explain this as representing simultaneous increases in hydraulic limitations to water availability (Panek 1996) and shifts in species composition from fast growing less efficient species in the understory to, primarily, M. polymorpha, a slow growing canopy species with low photosynthetic capacity and high LMA (Cordell et al. 1999).
Foliar d 15 N represents an integrative measure of ecosystem dynamics over time (Adams and Grierson 2001), among other things.In our study, foliar d 15 N had no correlation with species but was significantly explained by air temperature.As decreasing foliar d 15 N is generally considered to represent a tightening of the N cycle (Austin and Vitousek 1998), the positive relationship between foliar d 15 N increasing with air temperature may indicate increased N availability at lower elevations resulting from faster nutrient cycling, including decomposition (Vitousek et al. 1989).This is further indicated by the lack of species significance and the significant multiple regression (adjusted R 2 ¼ 0.30, P 0.0001, df ¼ 97) for foliar d 15 N which showed that leaf height was non-significant (F ¼ 0.69, P ¼ 0.41) while air temperature was highly significant (F ¼ 42.04, P 0.0001; see Fig. 8).These results are similar to those reported by Craine et al. (2009) and Amundson et al. (2003), who both reported a positive relationship between MAT and foliar d 15 N. Foliar d 15 N values in our study were more depleted than those reported in parts of the Amazon (Ometto et al. 2006), similar to those reported by Martinelli et al. (1999) for M. polymorpha, but more negative than those reported by Cordell et al. (1999) which were collected at a different study site in Hawaii.This may be partly due to our study area having a comparatively young substrate, as compared to older substrates in the Amazon, and therefore a more conservative nitrogen cycle (Martinelli et al. 1999).Other sources of variation may include differences in rainfall (Austin and Vitousek 1998) or increased microbial activities at higher MAT (i.e., lower elevations) resulting in soil 15 N enrichment through the preferential loss of isotopically lighter N gases (Martinelli et al. 1999).

Clusters
Leaves may be stratified across light and temperature gradients due to changes in species composition based on each preferred growth environment or acclimation of individual leaves within a species (Reich et al. 1994).In our study area, simultaneous changes in species composition and leaf trait acclimation were evident.The community A max to light relationship occurred in part due to transitions from low to medium to high light clusters of species (Reich et al. 1994).The response of each cluster differed however, similar to findings by Reich et al. (1998a, b), with those species growing in the highest light environments having significant correlations between light and A max .Such differences are similar to those identified during forest succes-sion, with shifts occurring from pioneer to climax species having differing capacities to acclimate to their irradiance growth environments (Poorter et al. 2006). Ku ¨ppers et al. (1996) used both successional and light classes to show a reduction in the time to reach 50% induction state (IS50%) from pioneer through late successional species, but increasing time to reach IS50% with increasing light.Our results highlight that the acclimation capacity may differ regardless of the leaf trait cluster, based on mean values, to which a species is assigned.For example, M. polymorpha, a dominant overstory species, did not increase A max with increases in light, while C. glaucum did.Ecologically, such changes make sense as species adapted to low light growth environments will require greater acclimation ability to take advantage of forest disturbances (Reich et al. 1994), i.e., including tree fall gaps, which result in large, but potentially short, increases in incident radiation both within the gaps and within the surrounding forest (Denslow et al. 1990).

Conclusions
We developed and validated a high-resolution three-dimensional model of microclimate through airborne wLiDAR hyperspectral fusion in a native dominated Hawaiian rainforest.Using this model, we show that a broad suite of leaf traits, occurring across species clusters, as well as within individual species, can be predicted by the modeled light and modeled air temperature of the growth environment.At the community scale, we show that correlations between A max and light and air temperature of the growth environment changes with species clusters, as well as through acclimation of individuals leaves within a species to its unique growing environment.This relationship differs between species with different competitive growth strategies, with no acclimation occurring in the dominant overstory species M. polymorpha, but significant acclimation (i.e., plasticity) occurring in the dominant understory species C. glaucum.However, the greatest factor contributing to leaf trait variation was identified as inter-species variability.Analysis of stable isotopes shows that while foliar d 13 C is determined by inter-species differences (and in particular life form) in leaf physiology, foliar d 15 N is determined by ecosystem differences in nutrient cycling resulting from differences in MAT along the elevation gradient.While no significant differences in RUE were identified between exotic and native species, we did find that species in general increased their light and water RUE simultaneously related to v www.esajournals.orgleaf height in general, the illumination state of their growth environment, and more broadly, to ecosystem changes in nutrient availability related to decreasing modeled air temperature associated with increasing elevation.The results of this study serve to develop and validate new tools useful for investigating ecosystem function at high spatial and temporal resolutions as well as provide insights into an ecosystem undergoing rapid degradation through species invasion and climate change.Future work built off this study will include full forest productivity modeling under different disturbance and climate change scenarios.

Fig. 1 .
Fig. 1.Overview of remote sensing and field data integration and analysis.Hyperspectral and waveform light detection and ranging (LiDAR) data were collected simultaneously using the Carnegie Airborne Observatory (CAO) while discrete LiDAR was collected separately.Field data collected for parameterization and validation included: (1) LAI-2000 for leaf area index (LAI; two-dimensional), (2) vertical leaf area density (LAD; threedimensional) transects, (3) microclimatic data, and (4) leaf trait measurements throughout the study transect.Leaf traits included chemical and gas exchange analyses.Microclimate data included modeled daytime mean and standard deviation photosynthetic photon flux density (PPFD) and modeled mean daytime air temperature.Spatial data included location, elevation, and forest structural information.Taxonomic data included species, native vs. exotic status, and life form.Principal component analysis (PCA) axes were input into K-means analysis to identify ecophysiological similar clusters that were explained through differences in microclimate, taxonomy, and spatial location.

Fig. 2 .
Fig. 2. Study area (C) located within the Hawaii Experimental Tropical Forest (B) in Laupahoehoe, Hawai'i (A).Inset C provides tree height at 1.25 3 1.25 m resolution, with heights ranging from 0 m (black) to 40 m (white).

v
www.esajournals.orgSUPPLEMENT R scripts for conducting clear sky PAR simulations described in the main text (Ecological Archives C005-006-S1).

Table 1 .
Plant species occurring within diversity transects.

Table 2 .
Plant density (no.individuals/1000 m 2 ), volume (basal area 3 height cm 3 ), frequency (% transects occurring) of species found within diversity transects.Species are ranked by their importance value (IV).Sampled (cluster) ¼ whether foliar samples were collected and to which cluster the species was assigned.Species taxonomic information is provided in Table1.Height and DBH values are expressed as mean 6 SD.

Table 3 .
Daytime (sun elevation .258C; mean 6 SD) climatic conditions at top of canopy tower low, mid, and high locations between December 17, 2010 and June 19, 2011.Rainfall is the total over this period.
PCA axis as the response revealed PCA axis one to have a significant positive relationship with modeled mean PPFD but not with modeled mean air temperature [model adjusted R 2 ¼ 0.35, P , 0.0001; MM-PPFD and MM-air temperature F ¼ 22.2 (P , 0.0001) and 0.04 (P ¼ 0.85), respectively] while PCA axis two had a significant positive relationship with air temperature, but not PPFD [model adjusted R 2 ¼ 0.17, P ¼ 0.0003; MM-PPFD and MM-air temperature F ¼

Table 4 .
General linear model results for leaf traits versus taxonomic, ecological and structural categories (no.parameters).Data are the Akaike Information Criteria (AIC) (adjusted R 2 ) degrees of freedom and P. AIC allows intra-row comparisons and models .20 above the minimum AIC have been removed.Status ¼ invasive or native; Strata ¼ low, mid or overstory position; Position ¼ canopy or not; M. polymorpha ¼ yes or no.Smartweed (Persicaria punctata) is not included in these analyses.The curve from which variables are calculated is provided in parenthesis following the variable name.

Table 5 .
General linear model results for leaf traits versus taxonomic, ecological and structural categories (no. parameters).Data are the Akaike Information Criteria (AIC) (adjusted R 2 ) degrees of freedom and P. AIC allows intra-row comparisons and models .20 above the minimum AIC have been removed.Notes: See Appendix: TableA3for an explanation of abbreviations.* P, 0.05; ** P , 0.01; *** P , 0.001;**** P , 0.0001.Status ¼ invasive or native; Strata ¼ low, mid or overstory position; Position ¼ canopy or not; M. polymorpha ¼ yes or no.Smartweed (Persicaria punctata) is not included in these analyses.The curve from which variables are calculated is provided in parenthesis following the variable name.

Table 6 .
Best subsets models of leaf traits versus structural and modeled mean and standard deviation (SD) photosynthetic photon flux density (PPFD; lmolÁm À2 Ás À1 ) and air temperature (8C).Data represents the t-ratio (F-ratio) and P-value significance, with increasing *, **, ***, ****, and / representing P-values of 0.1, 0.05, 0.01, 0.001, and ,0.0001, respectively.PPFD SD was not found significant for any leaf traits and not included in the table.Notes: See Appendix: TableA3for an explanation of abbreviations.The curve from which variables are calculated is provided in parenthesis following the variable name.A max values combined from light response and induction response curves.

Table 7 .
Best subsets models of leaf traits versus structural and modeled mean and standard deviation (SD) photosynthetic photon flux density (PPFD; lmolÁm À2 Ás À1 ) and air temperature (8C).Data represents the t-ratio (F-ratio) and P-value significance,

Table 8 .
Predictors of Hapu'u (Cibotium glaucum) and Ohia (Metrosideros polymorpha) leaf traits.Data represent the slope (adjusted R 2 ) degrees of freedom and P-value of linear regressions.PPFD and air temp are modeled for each growth environment.

Table 10 .
Leaf trait values (mean 6 SD) for K-means low, medium and high modeled light clusters and results of among cluster one-way ANOVAs sorted by the adjusted R 2 value.
Notes: See Appendix: Table

Table 11 .
Forest elevation, structure and micro-climate values (mean 6 SD) for low, medium and high light environment K-means clusters and results of among cluster one-way ANOVAs.Categorical classifications are % true and differences among clusters are assessed using the Pearson test.

Table A2 .
Pearson correlations between foliar variables 11-20.Only correlations having P-values ,0.05 are shown.P-values ,0.01, ,0.001, and ,0.0001 are represented by *, **, and /, respectively.Column heading numbers correspond to numbered variables in the first column here and in TableA1.Note: See TableA3for an explanation of abbreviations.

Table A1 .
Pearson correlations between foliar variables 1-10.Only correlations having P-values , 0.05 are shown.P-values , 0.01, , 0.001, and , 0.0001 are represented by *, **, and /, respectively.Column heading numbers correspond to numbered variables in the first column.Notes: See TableA3for an explanation of abbreviations.The curve from which variables are calculated is provided in parenthesis following the variable name.

Table A3 .
Glossary of selected abbreviations as used in this study.