James R. Kellner, David B. Clark, and Michelle A. Hofton. 2009. Canopy height and ground elevation in a mixed-land-use lowland Neotropical rain forest landscape. Ecology 90:3274.

INTRODUCTION

Light detection and ranging (LiDAR) is powerful remote sensing technology for quantifying the structure and dynamics of forests at large spatial scales (Drake et al. 2003, Roth et al. 2007, Kellner et al. 2009). By recording the return time of emitted laser pulses, aircraft-mounted sensors acquire measurements of canopy and ground elevation at high spatial resolution, providing detailed measurements of vertical and horizontal vegetation structure (Fig. 1). In raw form, LiDAR measurements record the time interval required for an emitted laser pulse to travel from the sensor to reflecting surfaces (e.g., leaves, branches and the underlying ground surface) and return to the sensor. By combining the travel time estimate with three dimensional information on the orientation and position of the LiDAR system at the time of laser firing, the elevation and horizontal position of each reflecting surface can be calculated relative to a reference frame, such as WGS 1984 (e.g., Hofton et al. 2000). Elevation measurements can be processed further to generate images, such as digital elevation or canopy models (Clark et al. 2004), or analyzed directly (Lefsky et al. 1999, Asner et al. 2008, Kellner et al. 2009). The data presented here have been processed, screened, and validated using extensive field studies to generate easting and northing mapped coordinate locations (UTM zone 16 North), and canopy and ground elevation that are precise and accurate.

This data set contains 127,849,839 observations of canopy and ground elevation in 128 km² of mixed-land-use lowland Neotropical rain forest in the Atlantic lowlands of Costa Rica (Fig. 2). The study area includes all 16 km² of La Selva Biological Station, a site of long-term research in basic and applied tropical forest ecology, current and abandoned plantations outside the La Selva property, and a portion of Braulio Carrillo National Park along an elevation gradient on Barva Volcano. The study area is defined by a mixed history of historical land use that encompasses terre-firme old-growth forest, secondary forest, swamp forest, and a wide range of current and former human land uses that are representative of contemporary continental rain forests. Previous studies have used these data to quantify size-frequency distributions of canopy gaps, and to determine whether canopy height within old-growth forest at La Selva was close to steady-state equilibrium under the 1997–2006 disturbance regime (Kellner et al. 2009, Kellner and Asner 2009). LiDAR data can be applied to a wide range of questions in basic and applied science, including calculation of above-ground carbon stocks (Drake et al. 2002), photosynthetically-active radiation (Chasmer et al. 2008), rainfall interception and water budgets (Roth et al. 2007), canopy and gap dynamics (Vepakomma et al. 2008, Kellner et al. 2009), individual-based tree demography (Yu et al. 2006), understory light environments and vegetation structure (Asner et al. 2008, Chasmer et al. 2008) and other questions (Turner et al. 2003, Chambers et al. 2007). The data are also a valuable resource from a well-studied tropical rain forest for teaching and education, and provide a quantitative baseline against which future conditions can be assessed.

METADATA

CLASS I. DATA SET DESCRIPTORS

A. Data set identity: Canopy height and ground elevation in a mixed-land-use lowland Neotropical rain forest landscape.

B. Data set identification code: The data are in 511 comma-delimited ASCII text files named tilexxx.txt, where xxx is replaced with the respective tile number from 001–511. Download the entire data set here, or to download selected data tiles, rather than the entire data set, please click here to go to a page listing all 511 files with links. Each file represents a 500 × 500 m contiguous area. Tile locations are shown graphically in Fig. 3, and can be indexed in a GIS using the shapefile (see section V.C.2), or in text format in Table 1.

C. Data set description:

1. Principal investigators:

James R. Kellner, Department of Plant Biology, The University of Georgia, Athens, Georgia, 30602 USA (Department of Global Ecology, Carnegie Institution for Science, Stanford, California, USA). jkellner@stanford.edu

David B. Clark, Department of Biology, University of Missouri, St-Louis, St. Louis, Missouri, 63121 USA and La Selva Biological Station, Puerto Viejo de Sarapiquí, Costa Rica. dbclark@sloth.ots.ac.cr

Michelle A. Hofton, Department of Geography, University of Maryland, College Park, Maryland, 20742 USA. mhofton@umd.edu

2.  Abstract: We obtained spatially extensive canopy height measurements using airborne remote sensing to characterize the structure and dynamics of a tropical rain forest landscape. Light detection and ranging (LiDAR) is a remote-sensing technology that acquires measurements of canopy height and ground elevation. By recording the return time of laser pulses emitted by aircraft-mounted sensors, LiDAR systems quantify the structure and geometry of individual trees and canopy height and enable estimation of the vertical and horizontal distribution of biomass using millions of accurate height measurements. This data set contains 127,849,839 records from 128 square kilometers of tropical wet forest in the Atlantic lowlands of Costa Rica (mean sampling density is 1.99 observations/m²). The study area includes all 16 square kilometers of mixed-land-use forest at the La Selva Biological Station and the lower flanks of Braulio Carrillo National Park. It contains a mosaic of historical and contemporary land use that is representative of contemporary tropical forest landscapes. Field studies demonstrated that LiDAR measurements were precise and accurate throughout the topographic and structural conditions at the site. Each record includes: easting and northing mapped coordinates (UTM Zone 16 North), height above ground (m), interpolated ground elevation (m), and six remote-sensing descriptors (point classification, the return number, the number of returns for the given pulse, intensity, scan angle, and the time of emission of the laser pulse). The data can be applied to a wide range of questions in basic and applied science, and are a valuable resource from a well-studied tropical rain forest for teaching and education. They also provide a quantitative baseline against which future conditions can be assessed.

D. Key words: biomass; carbon; Costa Rica; forest structure; landscape; La Selva Biological Station; LiDAR; light detection and ranging; Neotropics; remote sensing; sustainability.

CLASS II. RESEARCH ORIGIN DESCRIPTORS

A. Overall project description: Not applicable, this is a stand-alone project.

1. Site description: The study site is 128 km² of Tropical Wet Forest in the Atlantic lowlands of Costa Rica (Figs. 1, 2, 3). Embedded within the study site is La Selva Biological station, which covers 16 km². A detailed description of La Selva is in given in McDade et al. (1994).

a. Site type: The original site vegetation is classified as Tropical Wet Forest in the Holdridge life-zone system (Hartshorn and Hammel 1994).

b. Geography: Atlantic lowlands of Costa Rica (Figs. 2, 3). The southwest corner of the study area is at easting 820129 and northing 1141996 and the northeast corner is at easting 830629 and northing 1161996 in UTM zone 16 N (Sanford et al. 1994).

c. Habitat: The site has a mixed-land-use history, including cleared or developed areas, selectively-logged forest, terra-firme old-growth forest, current and abandoned plantations and pastures, second-growth forests of varying ages, and old-growth swamp forest (McDade et al. 1994).

d. Geology, landform: Relief is from 37 to 504 m (WGS 1984) on highly weathered Oxisols, and most of the landscape is on undulating upland plateaus (Kleber et al. 2007). Alluvial terraces near the Puerto Viejo and Sarapiquí rivers support higher nutrient availability through alluvial and colluvial deposition. In the rest of the landscape, nutrient availability is lower, although erosion can rejuvenate nutrient supply on deeply weathered soils (Porder et al. 2006).

e. Watersheds, hydrology: The entire study area is at the foot of Barva Volcano in Braulio Carrillo National Park and adjacent lowlands. All major watersheds originate outside the study area above the upper elevation boundary of the data set.

f. Site history: A detailed site history is given in McDade et al. (1994).

g. Climate: At La Selva mean annual temperature at 45 m elevation is 26 C, and annual rainfall is c. 4 m. There is a moderately drier season from February to May, but all months receive > 100 mm of precipitation (long-term station records at La Selva). Rainfall increases steeply with elevation, reaching c. 8 m at a 600 m elevation site (El Plástico) 2 km from the upper boundary of the data set area.

2. Experimental or sampling design

a. Design characteristics: The study area was selected to represent mixed-land-use lowland rain forest, including all 16 km² of La Selva Biological Station and additional areas outside the La Selva property including plantations and areas within Braulio Carrillo National Park up to 504 m elevation.

b. Permanent plots: Not applicable.

c. Data collection: Remote sensing data were collected on 13–14 March 2006.

3. Research methods

a. Field/laboratory: Data were collected for 128 km² of mixed-land-use tropical rain forest on 13–14 March 2006 (Kellner 2008, Kellner et al. 2009). The total number of discrete elevation estimates is 127,849,839 and mean sampling density was 1.99 observations/m². In the configuration used here, the Leica ALS50 was capable of recording the locations of up to three reflecting surfaces within the returned laser signal for every emitted pulse. These correspond to three discrete elevation estimates for each laser footprint and are termed ‘returns’. The minimum number of returns for a given laser pulse is one. In forested terrain, this may correspond to a laser pulse that passes through the forest canopy and is reflected by the ground surface. The maximum number of returns is three, corresponding to three distinct reflecting surfaces within the canopy. Some pulses do not reach the ground surface, and may be reflected entirely by canopy vegetation. In our data, the variables return and nreturn distinguish the return number, and number of returns for each laser pulse. Each observation was classified as ‘ground’ (class = 2) or ‘vegetation’ (class = 5) using local topographic modeling and visual examination of the three-dimensional distribution of LiDAR observations (Cognocarta GIS LLC, San Jose, Costa Rica). Because LiDAR estimates the elevation of vegetation or ground surfaces relative to a reference ellipsoid (in this case, WGS 1984), quantifying height above ground requires interpolation of a ground surface and subtraction of ground from vegetation elevation (Fig. 1). We processed LiDAR data to generate a raster ground surface (digital terrain model, DTM) using natural neighbor interpolation of LiDAR elevation estimates that were classified as ground. Field studies demonstrate that LiDAR DTM estimates are precise and accurate within old-growth forest and the wider landscape at La Selva (see section III.A.4 and Figs., 4, 5). Canopy height can therefore be calculated by subtracting ground from elevation variables (but see V.G.4.A).

b. Instrumentation: Data were collected by the Leica ALS50 aboard a fixed-wing aircraft on 13–14 March 2006. The Leica ALS50 is a discrete-return scanning laser altimeter.

c. Taxonomy and systematics: Not applicable.

d. Permit history: Permits for study at the site from the Organization for Tropical Studies and the Costa Rican government were maintained over the life of the project.

e. Legal / organizational requirements: See II.B.3.d.

1. Latest update: 12 February 2009

2. Latest archive date: 12 February 2009

3. Metadata status: The metadata are complete and up to date.

4. Data verification: Data quality has been assessed using field-validation studies and exploratory data analysis. We conducted visual and statistical examinations, and quantitatively assessed precision and accuracy in the field. Based on a visual examination, 38 points in the original LiDAR files were omitted from the current version. Each omitted point was an obvious outlier, occurring above an acceptable elevation range. We assessed the accuracy of LiDAR elevation estimates by comparing elevation values from the LiDAR DTM to 4184 ground-surveyed control points that were well distributed within the old-growth forest, and 8798 control points in the entire 16 km² of La Selva Biological Station. At the location of each control point, elevation was measured in the field using optical leveling techniques. These values were subsequently corrected by adding 11.44 m to account for a known underestimation bias between field measurements and the WGS 1984 reference ellipsoid (Hofton et al. 2002, Clark et al. 2004). Within the 16 km² landscape at La Selva the relationship between field and LiDAR estimates of ground elevation is: field measured elevation (m) = 0.999 ± (0.051) × LiDAR predicted elevation (m) - 1.289 ± (0.001), P_intercept < 0.001 P_slope < 0.001, r² = 0.996, RMSE = 1.66 m, n = 8798 (Fig. 4). P values and standard errors for regression parameters (in parentheses) were estimated using a non-parametric bootstrap with 2,000 iterations. The non-parametric bootstrap is appropriate for spatial data, where the assumption of independence among observations is unlikely to hold (Clark 2007). This indicates that LiDAR elevation estimates were precise and accurate throughout the topographic and structural conditions at the site.

Several sources of error can influence LiDAR elevation estimates, including the diameter of the laser pulse relative to the size of openings in the forest canopy, laser power, and how the inbound laser pulse was processed to identify reflecting surfaces. Exogenous to the LiDAR system, topographic slope can introduce bias in LiDAR elevation estimates because steep terrain increases the elevation range over which a given laser pulse is reflected to the sensor (Clark et al. 2004). We reexamined the relationship between field and LiDAR estimates of ground elevation in old-growth forest using the subset of field measurements that were collected within old-growth forest. The relationship between field and LiDAR estimates of ground elevation in old-growth forest is: field measured elevation (m) = 0.999 ± (0.001) × LiDAR predicted elevation (m) - 1.406 ± (0.126) m, P_intercept < 0.001 P_slope < 0.001, r² = 0.994, RMSE = 1.85 m, n = 4184 (Fig. 4). To determine whether there is slope-induced bias in LiDAR elevation estimates, we regressed the difference between each field and LiDAR estimate of ground elevation on topographic slope calculated using the DTM and standard methods. There is no evidence of slope-induced bias in LiDAR elevation estimates in old-growth forest: field measured elevation - LiDAR predicted elevation (m) = -0.004 ± (0.004) × slope (degrees) -1.381 ± (0.049) m, P_intercept < 0.001 P_slope = 0.146, r² = 0.000, RMSE = 1.84 m, n = 4184 (Fig. 5). In the 16 km² La Selva landscape, the relationship is: field measured elevation - LiDAR predicted elevation (m) = -0.004 ± (0.002) × slope (degrees) - 1.351 ± (0.031) m, P_intercept < 0.001 P_slope = 0.058, r² = 0.000, RMSE = 1.66 m, n = 8798 (Fig. 5). Although the latter relationship is marginally significant, this is a function of the large sample size, as indicated by the small coefficient of determination and effect size, and we conclude that it is not biologically significant for most analyses. Additional field studies comparing forest structure and dynamics quantified using LiDAR and field measurements are in Kellner et al. (2009).

In the context of other studies, results presented here are comparable in precision and accuracy to other LiDAR data for the same site. Hofton et al. (2002) reported a 2.52 m overestimation bias for the Laser Vegetation Imaging Sensor, a large-footprint waveform LiDAR (RMSE = 5.64 m). And Clark et al. (2004) reported 0.97 m overestimation bias using the small-footprint FLI-MAP system (RMSE = 2.29 m). In old-growth forest < 10 °, overestimation bias was slightly greater (1.01 m) and RMSE was 1.95 m (Clark et al. 2004). In our data, overestimation bias was 1.29 m throughout the entire landscape, and 1.41 m in old-growth forest. Precision of the relationships (RMSE) was 1.66 and 1.85 m respectively. The offsets and precision reported here and in Clark et al. (2004) are based on comparison of DTM values to field measurements, rather than direct comparison based on laser pulses. Use of the DTM introduces additional assumptions and uncertainty. Nonetheless, the slope of the relationship between field measurements and DTM elevation is very close to 1.000 (Fig. 4), and the results presented here constrain the contribution of vertical and horizontal uncertainty and interpolation errors to a narrow range.

1. The data are available from the Ecological Society of America’s data archives. Duplicate copies of the data are being stored at La Selva Biological Station.

2. Contact persons: (1) James R. Kellner, 260 Panama Street, Stanford, California, 94305. Telephone: +1 808-933-8121. Fax: +1 808-933-8120. E-mail: jkellner@stanford.edu (2) David B. Clark, La Selva Biological Station, Puerto Viejo de Sarapiquí, Costa Rica. Telephone: + 506 2766-6565. Fax: +506 2766-6535  E-mail: dbclark@sloth.ots.ac.cr (3) Michelle A. Hofton, Department of Geography, University of Maryland, College Park, Maryland, 20742, USA. E-mail: mhofton@umd.edu

3. Copyright restrictions: None.

4. Proprietary restrictions: None, although we would like to hear how the data are being used (e.g., for what research questions or teaching exercises).

5. Costs: None

1. Identity: AllTiles.zip -- The data are in 511 comma-delimited ASCII text files. Each file is a 500 × 500 m square plot. Plot locations are shown graphically in Fig. 3. To download selected data tiles, rather than the entire data set, please click here to go to a page listing all 511 files with links.

2. Size: File sizes are in Table 1.

3. Format and storage mode: Each file is comma-delimited ASCII text that was compressed using WinZip. The first line of each file contains variable names, subsequent lines contain data. In each file, the variables are: easting (UTM zone 16 North), northing (UTM zone 16 North), elevation (m), intensity, return number, number of returns, class, scan angle rank, GPS time. 

4. Header information: The first line of each file contains variable names.

1. Data forms or acquisition methods: The data were collected using airborne remote sensing. Acquisition methods are described in section II.B.3.a.

2. Location of completed data forms: Not applicable.

3. Data entry and verification procedures: Quality control and field studies are described in section III.A.4.

1. GIS data layer: land use history for La Selva Biological Station (lu00.shp). Download LandUseHistoryShapefile.zip

2. GIS data layer: mapped locations of LiDAR data tiles (500_m_tiles.shp). Download 500mTilesShapefile.zip

1. Generation of digital terrain model to estimate ground elevation

To estimate ground elevation and canopy height above ground, we generated a digital terrain model (DTM). The DTM was based on a natural neighbor interpolation applied to LiDAR elevation estimates that had been classified as ground returns (i.e., those laser pulses that were likely to have penetrated the canopy and been reflected by underlying topography). We then subtracted the predicted DTM elevation at the location of each laser pulse from its elevation estimate to quantify canopy height above ground. Field validation studies quantifying the precision and accuracy of ground topography and canopy height are in section III.A.4, Kellner (2008) and Kellner et al. (2009). Also see V.G.4.a.

2.  An R function to read the LiDAR data ( RFunction.txt). R is an open-source programming language and environment for statistical computation (R Development Core Team, 2009).

1. Archival procedures: The data are stored for long-term access in the GIS laboratory at La Selva Biological Station.

2. Redundant archival sites: Redundant copies of the data are stored by each author at their respective institutions.

1.  Kellner, J. R. 2008. Population and community dynamics of tropical rain forest canopy trees. The University of Georgia, Athens, Georgia, USA.

2.  Kellner J. R., D. B. Clark, and S. P. Hubbell. 2009. Pervasive canopy dynamics produce short-term stability in a tropical rain forest landscape. Ecology Letters 12:155–164.

3.  Kellner J. R., and G. P. Asner. 2009. Convergent structural responses of tropical forests to diverse disturbance regimes. Ecology Letters 12:887–897.

a. Question number 1. I subtracted ground from elevation to estimate canopy height, but the number is negative. How is this possible?

Answer: Although vegetation height must be ≥ 0 m aboveground, it is possible for vegetation height estimates to be negative. This occurs when interpolation errors in the DTM are propagated through the analysis, or when assumptions of the DTM are violated. For example, the DTM assumes that ground elevation is constant within a given pixel. On steep terrain where this assumption could be seriously violated, some laser pulses may fall beneath the interpolated DTM value. In addition, when the number of laser pulses that reach the ground surface is locally sparse, the DTM can have poor precision or accuracy (for example, see outlying points in Fig. 4 as potential candidates). Elevation estimates for individual laser pulses (i.e., the elevation variable) are probably made with greater precision and accuracy than derived estimates of ground elevation (i.e., the ground variable). This is because ground estimates are based on classification of vegetation and ground returns (class variable = 5 or 2, respectively), and interpolation of a gridded surface for points with class = 2. Both procedures introduce assumptions and error.

b. Question number 2. The number of returns should vary from 1 to 3, but there are > 3 returns associated with the same value of the time variable.

Answer: Although the Leica ALS50 was capable of digitizing very small differences in time associated with laser ranges, the time variable is not sufficiently precise to distinguish between returns. Here is an example. The time variable is recorded with two decimal places of accuracy. This indicates that the smallest recordable difference in time (0.01 s) is associated with the distance traveled by a laser pulse in 0.01 s. Using the speed of light, 299,752,458 m · s^-1, this distance is 2,997,524.58 m. Thus, the time variable is useful for determining when a given laser pulse was acquired (i.e., for distinguishing between flight lines, or dates of acquisition), but it is not useful for distinguishing between returns.

c. Question number 3. I have never used LiDAR data. What software do I need to get started?

Answer: In our experience, there are three types of tasks typically performed by ecologists that will dictate software requirements: (1) data processing to generate images, (2) visualization and basic summaries, and (3) statistical computation. Software needs vary widely depending on the objective of a study.

Data processing to interpolate images requires classification of LiDAR data points and interpolation of images from point subsets (e.g., creation of a DTM from points classified as ground). In practice, this could involve elimination of points over buildings, water or outside an acceptable elevation range using a combination of automated and manual editing, followed by interpolation using several algorithms (e.g., Clark et al. 2004). The current standard for LiDAR data processing is TerraScan (TerraSolid Ltd.). It can handle very large numbers of points and allows users to interact with and edit data on the fly. QuickTerrain Modeler (Applied Imagery Inc.) performs many of the same tasks as TerraScan and is excellent for rapid visualization of LiDAR point clouds and interpolated surfaces. Figure 1 was generated using QuickTerrain Modeler. Both pieces of software can perform basic statistical summaries of LiDAR data.

To visualize raster images made using LiDAR data we use ENVI (ITT Visual Information Solutions), which also contains sophisticated image processing routines that are accessible through IDL. For users who want to locate a field site or plots on a LiDAR elevation model, imagery can be visualized using ArcGIS.

For statistical computations, any programming environment that can handle large volumes of data will suffice. We use R, which is an open-source language and environment for statistical computation (R Development Core Team 2009).

R Development Core Team. 2009. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http://www.R-project.org.