*Ecological Archives* A025-109-A1

Geoffrey A. Fricker, Jeffrey A. Wolf, Sassan S. Saatchi, and Thomas W. Gillespie. 2015. Predicting spatial variations of tree species richness in tropical forests from high-resolution remote sensing. *Ecological Applications* 25:1776–1789. http://dx.doi.org/10.1890/14-1593.1

Appendix A. A full description of remote sensing pre-processing.

*Imagery preprocessing*

All imagery pre-processing was performed in ArcGIS 10.2 and ENVI 4.8/5. A single Quickbird (QB) satellite image scene was used to calculate canopy reflectance indices for all vegetation plots on BCI. Atmospheric correction was applied using the FLAASH correction tool in ENVI-5.0. The QB satellite imagery was used to calculate the Ratio Vegetation Index (RVI), the Normalized Difference Vegetation Index (NDVI), and Enhanced Vegetation Index (EVI) in ENVI and spectral heterogeneity was calculated using a principle components analysis in ArcGIS. To calculate spectral heterogeneity we clipped the image to the vegetation plots so no water or clouds were included in the imagery. Using the clipped imagery, we applied a principal components analysis (PCA) to the satellite imagery to compute a 4-band principle component image. We found that >96 % of the spectral variation was composed in the first principle component, and focused the spectral heterogeneity calculation on only that first principle component. We also used ‘two-dimensional’ spectral variability using the second principal component; however the results were qualitatively similar and were not reported in the final results. The standard deviation of the pixels in the first band of the principle component, compose the spectral heterogeneity for each vegetation sub-plot and were highly correlated to the standard deviation of the EVI (R = 0.91). The Quickbird variance in imagery texture was calculated from the panchromatic (0.6 m resolution) using a 3 × 3 window in ENVI. We investigated additional spectral indices which were not included in our final analysis. For instance we calculated mean values of pairwise Euclidean distance (MED) and distances from the mass centroid (MCD) as indicators of spectral heterogeneity (in two spectral dimensions) (Rocchini et al. 2004).

*Lidar preprocessing *

The remote sensing data with the high spatial accuracy and precision is the DRL data collected by Blom Corporation and Northrup Grumman over BCNM. These data were collected during the wet season between 15 August 2009 and 10 September 2009 during eleven flights by a fixed-wing aircraft equipped with an Optech 3100 (Optech, Vaughan, ON, Canada) system capable of four returns per pulse. The mean flying height was 457.2 m and mean flight speed was 66.9 m/second. The system operated at a scan angle of 17°, using a scanning frequency of 48 Hz and laser frequency of 70 KHz. All flights produced a total of > 233 million laser shots and > 528 million individual data points, resulting in a point density of 5.6 points per square meter (ppm²) and 8.1 returns per square meter (rpm²). We used a 1-m digital surface model (DSM) and 1-m digital canopy surface model (DCSM) (Fig. 1). The DSM and DCSM were created by Blom Corp. who calibrated and filtered unprocessed lidar data using Bentley’s Microstation (Bently, Exton, PA) and then manually edited the product to make a bare-earth DSM. Blom Corp. verified vertical accuracy of the DSM using 36 Differential GPS survey points in flat open areas around the island. The average error in height between the ground survey points and the DSM was 6.9 cm with RMSE value of 7.6 cm. Blom Corp. produced the DCSM from the point of highest return above each cell on a 1-m grid. We compared the DCSM to ground-based canopy height measurements from 2009 that were collected every 5 m cross the entire 50-ha plot (Hubbell et al. 1999), which were in general agreement with the 1-m DSM. No additional efforts were made to minimize artifacts where understory vegetation may present commission errors in the ground point classification. In addition to a classified point cloud the lidar data also provides an un-calibrated intensity value as well as the return type. Possible return types include only return, first of many, intermediate of many or last of many returns. The return type and intensity information is used in the calculation of the canopy/gap fraction statistics.

In addition to the mean canopy height and the standard deviation of canopy height we also tried additional methods to quantify vertical forest structure. We computed relative height metrics (rh25/rh50/rh75) which measured relative canopy distribution throughout the vertical canopy profile. We also computed lidar intensity, canopy closure/canopy fraction with four different methods approximating light penetration using lidar return ratios (first/last/intermediate), lidar classification (canopy/ground/all) and lidar intensity following a study by (Hopkinson and Chasmer 2009). We found these measurements of forest structure did not yield stronger associations with our response variables and were slightly more difficult to interpret from an ecological point of view. The mean canopy height and standard deviation of canopy height within a singular forest plot are much easier to interpret and likely to be repeated. Three dimensional lidar point cloud profiles from four 10 × 100 m horizontal transects to illustrate the coefficient of variation (CV) in the standard deviation of maximum canopy height (main text Fig. 3). Transects were selected based on the CV for the lidar derived maximum canopy height variable. CV = Standard deviation of max canopy height/average canopy height. CV values were selected to represent the statistical distribution of values (1 sigma and 2 sigma+) on both sides of the distribution. Each grid cell represents 10 m of vertical and horizontal change.

Literature cited

Hopkinson, C., and L. Chasmer. 2009. Testing LiDAR models of fractional cover across multiple forest ecozones. Remote Sensing of Environment 113:275–288.

Hubbell, S. P., R. B. Foster, S. T. O'Brien, K. E. Harms, R. Condit, B. Wechsler, S. J. Wright, and S. L. de Lao. 1999. Light-gap disturbances, recruitment limitation, and tree diversity in a neotropical forest. Science 283:554–557.

Rocchini, D., A. Chiarucci, and S. A. Loiselle. 2004. Testing the spectral variation hypothesis by using satellite multispectral images. Acta Oecologica 26:117–120.