Mapping and Characterizing Urban Forest with a Combination of LiDAR data and Color Digital Orthophotos

Eli Rodemaker, Remote Sensing Team Leader

John Colwell, Remote Sensing Expert

Norm Roller, Remote Sensing Program Manager

Marshall and Associates, Inc.

1603 Cooper Point Road NW, Olympia WA 98502

erodemaker@marshallgis.com

Abstract

Urban planners and foresters are increasingly interested in the amount, location, and condition of trees in the city.  Marshall has developed a procedure that uses digital color orthophotos and LiDAR data to map and evaluate the urban forest resource more accurately than is possible with either source of data, alone.  The LiDAR data is used to identify objects that may be trees, based on a height threshold.  Automated procedures are then applied to the photographic data to edit out objects that are not part of the tree canopy.  Additional information can then be extracted about the canopy, including height of individual trees, crown cover per unit area, and location of “holes” in the canopy, as well as estimates of bulk canopy volume.  An example of such information produced for the City of Seattle is described.

Introduction

Increasingly, foresters in charge of managing the trees in urban environments want more information than is generally available, in order to better manage the forest resource.  The urban forest conditions are of interest for a variety of reasons, including: 1) their role in watershed management; 2) their value for carbon sequestration and perhaps for ameliorating the urban climate; and 3) in some relatively unaffected forest preserves in parks, there is interest in saving these forests as examples of natural pre-urban ecosystems, as well as protecting them from encroachment of exotic species.

This paper reports on the efforts to use two sources of remote sensing data (LiDAR data, and high-resolution multispectral aircraft data) to provide some of the desired information.

Background

Prior to this work, information about the City of Seattle’s tree resources was sparse.  Costly field inventory was available for some locations but the majority of the City was not cataloged.  The available data consisted of limited windshield surveys and a few GPS-based censuses.  Availability of recent LiDAR data and nearly contemporaneous Color DOQQ data encouraged the generation of previously unknown resource information.  Most of the LiDAR data were acquired for the Puget Sound LiDAR Consortium by TerraPoint, LLC in early 2000 and early 2001.  TerraPoint flew a multiple-return scanning laser altimeter in a small fixed-wing aircraft with a circa 0.9 meter on-the-ground laser spot, nominal across- and along-track pulse spacing of 1.5 meters, and 50% overlap of adjacent flight lines, providing an average of circa 1 pulse/square meter. 

.  The resulting raster data has a nominal 6 foot spatial resolution with elevations recorded in floating-point feet measured to thousands of feet.

The contractor delivered LiDAR all-return points, a first-return surface, bare-earth points, and a gridded bare-earth DEM.  Bare-earth points were separated from points in the canopy and on structures by automatic geometric filtering (virtual deforestation, or 'VDF') (Haugerud and Harding, 2001).  The bare-earth DEM was produced by sampling a TIN Triangulated Irregular Network) constructed from these points, with few, if any, modifications.  The bare-earth DEM is known to have some inaccuracies.

Marshall was also given access to multispectral color aerial orthophotographs (DOQQ), in digital form.  This imagery was collected in the summer of 2000 at a nominal spatial resolution of 0.5 feet.  Orthophotographs were obtained for the City of Seattle and surrounding area (e.g., Mercer Island).  All data processing was performed in LEICA Geosystems’ (Atlanta, GA) ERDAS IMAGINE® software.  Conversion of the project products to ESRI shapefile format was performed with ArcGIS^TM 8.3 software from Environmental Systems Research Institute (Redlands, CA).

APPROACH

LiDAR data can be very useful as an indication of heights of objects above the terrain.  There are a variety of ways to deal with this issue.  The approach we used consisted of several steps, including distinguishing trees and their associated heights by stratifying the LiDAR heights file so that it only showed vegetation.  This vegetation information was produced by categorizing the color orthophoto data into vegetation and not-vegetation.  Figure 1 is an example of the appearance of a digital color image.  Other corrections were also made for errors in the bare earth terrain model, and buffering of the infrastructure.

Figure 1.  Typical Orthophoto.

Red, green, and blue color aerial photograph corrected for topography.

Once the trees in an urban area are identified, some of their attributes can be assessed.  These attributes can include several tree height classes, a simple estimate of “tree volume”, tree canopy closure, and tree “Gap” analysis.  Once a data set has been prepared for an urban area, the data can be used with ancillary data to learn things about the value of the urban tree resource.  One way of doing this is to run CITYgreen analyses for current conditions, and for several scenarios of possible future conditions. 

Marshall Pre-Processing of the Data

There were several aspects of the data that required special attention. The first step employed by Marshall was to create two seamless files from the 21 quadrangles of LiDAR data that cover Seattle.  One file was of the bare earth surface model and another one was of the first returns surface model.  For each quadrangle, spurious values exist along the edges of the file extent and within the LiDAR.  Marshall applied our error flag and background removal algorithms to the 42 separate files (2 surface types x 21 quads).  Once false background data was removed, the quadrangles were joined together in ERDAS.  The quadrangles do not abut one another, but instead overlap at the edges, i.e., elevation values from each neighboring file.  Overlapping pixels were averaged together since the pixel values were often not the same.  Spurious elevations of less than sea level were present within many of the LiDAR files and were preserved in error field layers for reference during product analysis and QA/QC.

Once the bare earth and first returns mosaics were created, Marshall created a feature heights file by differencing the two mosaics, pixel by pixel.  The resulting initial feature height file is shown in Figure 2.  There are still errors in this product due to erroneous interpolation between “ground surface” sample points derived from the LiDAR data.  The result is that certain infrastructure (e.g., road overpasses) still show objects with heights.  There are also “negative” heights reported as shown in Figure 3.  The origin of some of these negative heights is due to inaccurate interpolation of the ground layer is shown in Figure 4.

Figure 2.  LiDAR Height Above “Bare Earth” Product.

Figure .  Bare Earth Model Problem – Interpolation Across Ground Surface Data Voids.

CIR Aerial Imagery on the left, LiDAR ‘Bare Earth’ digital elevation model on the right.  Non-ground features are removed from the LiDAR point data leaving a void of ground samples at buildings (Haugerud and Harding, 2001).  Ground elevation values are interpolated from remaining points.

Spurious negative feature heights were preserved in another error field.  In addition, anomalously high values were set to the highest “real” heights in the data set.  The feature heights file was then stratified into a potential tree canopy file by restricting the LiDAR height data to feature heights greater than 5 feet.   Buildings were removed from this layer by applying the City of Seattle’s Building Outlines GIS layer previously generated by hand from an older color orthophotographic image set.  This building polygon layer was converted to a 1 ft pixel spatial resolution raster file and then a 12 foot (3.66 meter) buffer was appended to each raster building outline.  The buffer was necessary to account for two effects: 1) the scale of the LiDAR signal at a six foot resolution versus the building outlines layer that was tightly constrained to the edges of features on a 1 foot photo, and 2) a limited amount of positional error in the buildings outline layer due to parallax effects.  An image showing this infrastructure mask is shown in Figure 5.  The buffered building outlines were used to mask the potential canopy layer to produce an initial LiDAR Tree Canopy Cover file.  This masked LiDAR file is shown in Figure 6.  A drawback to this approach is the possible elimination of some pixels belonging to trees next to buildings.

Figure 5.  Infrastructure Editing.

Building infrastructure GIS layer in Cyan with LiDAR feature heights layer as base image.

Figure 6.  Initial LiDAR Canopy Product.

LiDAR feature heights layer editied for building infrastructure.  Both elevated roadways and movable tall objects such as cargo containers, truck, railcars, and airplanes remain.

There are still errors in this product due to erroneous interpolation between “ground surface” sample points derived from the LiDAR data.  The result is that certain infrastructure (e.g., road overpasses) still show objects with heights that are not trees.  Both omission errors and commission errors occur in this product.  The commission errors are due to heights associated with non-vegetated terrain.  The omission errors are due to trees located near buildings, which were removed by the “buffering” of the infrastructure mask.

Both types of errors were corrected with the aid of the categorized DOQQ data.  Figure 1 above is an example of the appearance of a digital color image.  Some spectral confusion between vegetated and non-vegetated features occurs in a color DOQ mosaic.  However the goal of the use of the DOQ is to maximize LiDAR error reduction, not map trees with the DOQ.  To this end high accuracy spectral signatures were collected for vegetation and non-vegetation.  Further the DOQ data was averaged to 6 foot pixel resolution to reduce intercanopy variability.  The vegetation mask that was produced from this layer is shown in Figure 7.  When this mask was applied, only vegetated terrain with positive LiDAR heights remained, thus substantially reducing commission errors (Figure 8).  A limited amount of omission error was also corrected.  The building infrastructure layer without buffering was also masked by the vegetation categorization.  This produced some additional trees, especially along the edges of buildings, because they were green and had significant height.  The resulting recovery of omission errors as well as the commission-reduced areas are shown in Figure 9.  A final product showing all of the detectable trees with LiDAR and DOQQ data is shown in Figure 10.  These “tree” pixels were then used to generate the information of interest to the sponsor.

Figure 7.  Vegetated/Non-Vegetated Product.

Example of vegetation masks layer produced from color DOQ.  Red areas are representative of the non-vegetation mask (commission) and green of the vegetation omission reduction layer.

Figure 8.  Vegetation Editing.

LiDAR canopy product with DOQ mask for commission errors applied.  Cargo containers in the lower left and the elevated roadways are removed.
Figure 9.  Corrected Tree Canopy.

LiDAR canopy with both types of DOQ derived improvements.  Note the row of trees in the upper left added back to the LiDAR product initially masked by the building outline mask (including canopy hanging over the top of the buildings).

Figure 10.  Final Tree Canopy Product.

LiDAR based tree canopy layer in gold with DOQ base image with substantial improvement from a GIS layer masked approach. 

RESULTS

The information of interest to urban foresters included tree heights, classes, and other parameters.  For certain urban forest preserves, a greater degree of analysis may be desired.  One such place is Pioneer Park, on Mercer Island (City of Mercer Island), east of Seattle.  These preserves are examples of “undisturbed “ecosystems.  For Pioneer Park, Marshall developed a special set of information.  This information consisted of tree height classes, tree canopy closure (“variability”), and tree gap analyses.

Figure 11 shows several tree height classes.  This product was prepared by “binning” certain height classes from the continuous height file.  In addition, an image showing “gaps” in the canopy (vegetation less than 15 feet high) was produced.  One of the reasons a product like this is of interest is because these gaps are likely locations of establishment of noxious and/or foreign species.  An example product is shown in Figure 12.  A canopy closure product was also produced, and is shown in Figure 13.  Canopy closure was developed in a two-step process.  We used a spatial focal analysis window to generate a canopy height variability layer (variance within the window).  The window was passed pixel by pixel across the vegetation heights layer returning focal variance.  Marshall then indexed the canopy variability strata by canopy height strata to separate closure strata life forms.  In fact, open canopy areas need to be found during this indexing phase, as areas of non-forest types may have low variability, as would closed canopies.  Marshall created an expert decision rule algorithm to model the combined strata index (e.g. pixels with high variability and tree heights area mixed closure).  The indexed strata represent Canopy Closure (by lifeform) of: 1) Open; 2) Mixed; or 3) Closed.

Figure 11.    Canopy Height Stratification Product.

Figure 12.  Gap Detection Product.

Figure 13.  Canopy Closure Product.

These types of products are useful not only in their own right, but also as inputs to forest canopy analysis models.  We are currently using some of these information products and ancillary data to assess the current condition of the forest resource and its value in protecting watersheds, sequestering carbon, and other factors.  This is being done by using the remote sensing derived information in conjunction with other information to run the CITYgreen analysis program.  This program also allows urban planners to assess the consequences of selected future scenarios of land management.

SUMMARY AND CONCLUSIONS

This paper illustrates ways in which accurate and useful information on a city’s forest resource can be created from a combination of LiDAR data and high resolution color DOQQs.  These sources of data, both collected at about the same point in time, have not been available in most cities, where anecdotal and spot surveys have been all that is available.  The information that can be produced from the combination of these two sources of data is of value in its own right.  The information is more useful, and takes on greater significance, when its attributes and locational information make it possible to use powerful software analysis tools, such as CITYgreen.

SELECTED REFERENCES

Haugerud, R.A. and D.J. Harding.  Some Algorithms for Virtual Deforestation (VDF) of LiDAR Topographic Survery Data.  ISPRS Workshop, CommisionIII, Working Group3, Annapolis, MD, October 2001, 7pgs.