Research: Lidar for Science and Resource Management

Lidar-based vegetation delineation in NPS units

Smoothed and un-smoothed classification results provided to the National Park Service. Examples: Jean Lafitte National Park. [Enlargement]

Maps depicting the spatial distribution of vegetation communities are used extensively for conservation, land management applications, and decision support systems as aid-tools in decision–making, sustainable development and adaptive management practices. Vegetation maps can also be used in change analysis and assessment of extreme storm impact on different vegetation communities.

The vegetation classification is based on the kmeans clustering (unsupervised classification) of the Experimental Advanced Airborne Lidar (EAARL) vegetation metrics. Five-meter-resolution grids of bare-Earth (BE), canopy heights (CH), canopy-reflection ratio (CRR), and height of median energy (HOME) were used as input data. The data comprise tiles of 2 square kilometers for each metric. The unsupervised classification was carried out within the statistical open-source software R using the rgdal (Keitt et al., 2008) and proj4 (Urbanek, 2008) packages to read and write geotiff files, and the cluster (Maechler et al., 2005) package to do k-means classifications.

First, the metrics geotiff tiles are read in the R environment and are saved as ASCII comma delimited tables (*.csv) with the four metrics values recorded for each pair of XY coordinates in the tile . Further the data is merged and all "no-data" points with values equal to -1000 are extracted. A principal component analysis on the remainder of the data is performed, and each principal component is multiplied by its proportion of variance explained. A high performance k-means algorithm (Kaufman and Rousseeuw, 1990, Struyf et al., 1997) is used to do unsupervised classification followed by a re-classification in which class 1 data has the lowest mean of canopy heights and class n (where n is the number of classes) has the highest mean of canopy heights. The last step is to transform the results back in geotiff tiles of 2 square kilometers.

The classification results can be opened in any software that reads geotiff projected files such as Global Mapper, ENVI, ERDAS or ArcGIS. In order to mitigate the high variance observed in some park's vegetation, the merged classification result can be smoothed by applying a majority filter (circular neighborhood of 10 m radius) followed by a low-pass filter in ArcGIS 9.x. If no majority value was computed for a certain neighborhood, then the original value was preserved. This resulted in more homogeneous patches of vegetation but concealed park's smaller canals and roads. Both smoothed and un-smoothed classification results were provided to the park service.

References

Kaufman, L. and Rousseeuw, P.J. (1990). Finding Groups in Data: An Introduction to Cluster Analysis: Wiley, New York.

Keitt, T.H., Bivand, R., Pebesma, E., Rowlingson, B. (2008). rgdal: Bindings for the Geospatial Data Abstraction Library: R package version 0.5-25. http://www.gdal.org, http://rgdal.sourceforge.net/, http://sourceforge.net/projects/rgdal/

Maechler, M., Rousseeuw, P., Struyf, A., Hubert, M. (2005). Cluster Analysis Basics and Extensions: cluster R package version 1.11.11

Struyf, A., Hubert, M. and Rousseeuw, P.J. (1997). Integrating Robust Clustering Techniques in S-PLUS: Computational Statistics and Data Analysis, 26, 17–37.

Urbanek, S. (2008). proj4: A simple interface to the PROJ.4 cartographic projections library: R package version 1.0-4.