A Comparative Analysis of Urban Tree Canopy Assessment Methods in Minnesota
Remote Sensing of Natural Resources and Environment | FR 5262 | University of Minnesota
Philip J Potyondy
Remote Sensing of Natural Resources and Environment | FR 5262 | University of MinnesotaA Comparative Analysis of Urban Tree Canopy Assessment Methods in Minnesota
Philip J Potyondy
12/12/2011
Project Description / Objectives
The primary objective of this study is to compare a number of Urban Forest Canopy analysis methods. This study will inform and potentially validate canopy assessment methods used by the Emerald Ash Borer Rapid Response Community Preparedness Project that is being conducted throughout Greater Minnesota by a team of researchers at the University of Minnesota in collaboration with community leaders and volunteers.
Urban Forest Canopy Assessment Methods
§ Digitize urban forest canopy cover using image classification remote sensing techniques and software over study areas.
§ Digitize urban forest canopy cover via technician photo interpretation over study area.
§ Digitize urban forest canopy cover via technician photo interpretation within stratified random sampled blocks.
§ Calculate urban forest canopy using field collected tree canopy width measurements within stratified random sampled blocks.
§ Calculate urban forest canopy using randomly generated points within study area interpreted by a technician.
Each method will result in an estimate of urban forest canopy cover within the included communities. Results will be compared. The results of this study will help shape future urban forest canopy assessment. The timing and financial cost of each method will also be discussed. The field collected tree data is considered existing data since it has already been collected by tree inventory teams within the communities. The primary reason for collecting this data is to determine the composition of the urban forest as communities prepare for the arrival of the emerald ash borer (Johnson, 2011). Urban forest canopy can be used to estimate a myriad of benefits that trees provide to communities (McPherson, 1995, 1996, 1998a, 1998b, 2010; McPherson, Nowak, & Rowntree, 1994; Minnesota Department of Natural Resources, 2000; Piego & Breuste, n.d.; USDA Forest Service, n.d.). The resulting canopy data will also be used to estimate how removal of all ash trees from a community will impact winter home energy consumption.
Materials / Tools / Concepts
The Minnesota communities of Hibbing, Hutchinson, and Rochester were chosen for this study. Communities selected for this study were chosen from communities that were already participating in the Emerald Ash Borer Rapid Response Community Preparedness Project. The three selected communities range in population size from 13,835 in Hutchinson, 16,237 in Hibbing, and 103,486 in Rochester. The communities are geographically spread across Minnesota north, central, and southern – in order with latitude and longitude: Hibbing (47.4273,-92.9376), Hutchinson (44.8884,-94.380), and Rochester (44.024,-92.470). Communities were also selected by virtue of having municipal utility service. We have found it is easier to gather utility data from municipal providers than from commercial energy providers.
Tree inventory data was utilized from data collected by trained volunteer community tree inventory teams from each community as part of the Emerald Ash Borer Rapid Response Community Preparedness Project. Both publicly and privately owned trees were included. Trees over maintained lawn were measured, but trees in natural or densely vegetated areas were not included. Besides gathering data on species, diameter at breast height, and condition rating, the volunteers also collect the average crown width of trees within the sample blocks. Crown width measurements were utilized to estimate the overall canopy within the study area. The tree inventory study was set up as a stratified random sampling method with blocks as the sampling unit stratified by zones. Zones were stratified based in canopy cover, building type, and street type (curvilinear or straight). The tree inventory design is based on A Statistical Method for the Accurate and Rapid Sampling of Urban Street Tree Populations (Jaenson, Bassuk, & Schwager, 1992).
National Agricultural Imagery Program (NAIP) Digital Orthorectified Images from 2010 were used (U.S. Department of Agriculture, Farm Service Agency, 2010). Images of each community were downloaded using the MN Northstar Mapper (Minnesota Geospatial Information Office, 2011). The imagery is summer “leaf-on“ 3-band natural color imagery with 1 meter pixel resolution and is formatted to the UTM coordinate system using NAD83.
Google and Bing imagery was also utilized to guide technician interpretations. Bing Bird’s Eye imagery was especially beneficial, where available. Other leaf off imagery was also referenced where available, but was not used as a primary source of interpretation.
ERDAS Imagine 2010 was utilized for the raster image classification and accuracy assessment. ESRI ArcGIS 10 was used for vector operations including area calculations, classification clean-up, clipping raster images to the study areas, digitizing tree canopies, and all related vector processing. iTree Canopy was utilized for point based classification (USDA Forest Service, 2011).
Procedures
Digitize urban forest canopy cover using image classification remote sensing techniques and software over study areas
ERDAS Imagine Objective was utilized to conduct an Object Oriented Classification to classify tree canopy from the 2010 NAIP imagery. There are only two classifications as a result of this procedure, tree canopy and not tree canopy. The classification was generated by using the following model processes. A spectral single feature probability (SFP) was used as a Raster Pixel Processor. Threshold and Clump with a probability threshold setting of 0.90 was used as a Raster Object Creator. Dilate with a square Kernel Size of 3 and Size Filter with a Minimum Object Size of 30 pixels and no Maximum Object Size were used as Raster Object Operators. Polygon Trace was used for the Raster To Vector Conversion. Island Filter with a Maximum Object Size of 1,000,000,000 square meters, Generalize with a Tolerance of 0.50 meters, and Split with a Probability Threshold of 0.45 and 5 Recursions were used as Vector Object Operators. Geometry and Template Match were used as Vector Object Processors. The Geometry Object Classifier was set to Multi Bayesian Network with a Pixel Probability Weight of 50 and set to Enforce Distribution Bounds. Geometry:Area was set to Type Gaussian and Locked with setting of Minimum 100.90 square feet, Maximum 6035.29 square feet, Mean 2549.07 square feet, and a Standard Deviation of 2710.31 square feet. Geometry:Circularity was set to Type Linear with setting of Minimum 0.60, Maximum 1.00, Mean 0.00, and a Standard Deviation of 0.10. There were also three Vector Cleanup Operators that were used: a Probability Filter with a Minimum Probability of 0.45, a Convex Hull, and a Template Match with Level of Detail set to 16.
The resulting vector shapefile was cleaned up in ArcGIS. Clean-up mostly entailed removing tree canopy polygons from areas that were actually grass. The cleaned-up shapefile was then used to Erase the canopy from the study area shapefile. The cleaned-up canopy shapefile was then Merged with the study area shapefile that had the canopy shaped holes. This erase and then merge method was used to avoid overlap. This resulted in a single shapefile that covers the study area and contains the tree canopy and the space without tree canopy. An attribute was added to the table denoting tree canopy as 1 and areas that are not tree canopy as 2. Zero was avoided because in the next step the shapefile was converted to a raster .img using ERDAS Imagine which uses a minimum bounding rectangle to generate a rectangular raster file. The space outside the study area or area of interest but inside the bounding rectangle by default receives a value of zero; thus we did not want areas without tree canopy that are within the study area to be lumped with the area outside the study area.
The resulting raster file was used to conduct an accuracy assessment in ERDAS Imagine. There was not any available suitable reference imagery other than the imagery that was used for the classification. Acquiring additional imagery was cost prohibitive and not within the budget for this project. The accuracy assessment was conducted without viewing the classification and with the assistance of an additional technician. Reference locations were randomly generated across the image. A total of 112 points were referenced to conduct the accuracy assessment with the study area.
Digitize urban forest canopy cover via technician photo interpretation over study area
Technicians used ArcGIS to view, interpret, and digitize canopy by hand from the same 2010 NAIP imagery as above across the entire study area.
To prepare for comparison and analysis with the other canopy assessment methods the resulting shapefile was then processed using multiple tools in ArcGIS. First the many polygons that made up the tree canopy were Dissolved into a single polygon with no attributes. The dissolved shapefile was then processed with the Identity tool by the study area boundary to ensure canopy outside the study area was not included and attributes of the study area boundary file were applied to the result. The Canopy of Study Area file was retained, canopy area was calculated, and canopy area was divided by area of the Study Area resulting in Percent Canopy of Study Area. The Canopy of Study Area file is then processed with the Identity tool by the Zones that were generated for the tree inventory. Canopy of Zones file is retained, canopy area is calculated by zone, and canopy area by zone is divided by the area of each given zone resulting in Percent Canopy by Zone. Canopy of Zones file is then processed with the Identity tool by the Sampled Blocks that were generated for the tree inventory, where field data was collected. Canopy of Blocks file is retained, canopy area is calculated by block, and canopy area by block is divided by the area of each given block resulting in Percent Canopy by Block.
Digitize urban forest canopy cover via technician photo interpretation within stratified random sampled blocks
Essentially, technicians used ArcGIS to view, interpret, and digitize canopy by hand from the same 2010 NAIP imagery as above only within the sampled blocks that were generated for the tree inventory. In actuality we clipped this data from the dataset that was digitized over the entire study area (as described above). Percent canopy by block was calculated by dividing canopy area per block by the area of each given block. Block level canopy data was used to estimate zone and community scale tree canopy by using an area based geographically weighted extrapolation equation (Appendix 1 – Area Based Geographically Weighted Extrapolation Equation).
Calculate urban forest canopy using field collected tree canopy width measurements within stratified random sampled blocks
Community tree inventory teams collected two canopy width measurements and averaged them to report average canopy width per tree (see above section Materials / Tools / Concepts). These average canopy widths were equated to area by dividing the average canopy widths by 2 to get the radius of the canopy and then the radius was squared and multiplied by π (Area of Circle = πr²). The sum of the areas of all the trees within a given block was used as an estimation of percent canopy for the block. This block level canopy data was used to estimate zone and community scale tree canopy by using an area based geographically weighted extrapolation equation (Appendix 1 – Area Based Geographically Weighted Extrapolation Equation).
Calculate urban forest canopy using randomly generated points within study area interpreted by a technician
iTree Canopy was utilized for this point based classification (USDA Forest Service, 2011). iTree Canopy is a web application that guides a user through an interpretation process to determine a cover classification for a given area. The user sets a study area by uploading a shapefile to the software. The user then determined which classes they would like to classify and how many randomly generated points they would like to classify. We chose the following cover classes: tree, non-tree vegetation, impervious building, and impervious ground. iTree Canopy uses Google Maps API to display the points within the study area. The Google provided imagery at the time of running this analysis was very coarse and was nearly impossible to interpret from. Bing however had much better imagery especially the Bird’s Eye View. X, Y coordinates were copied from the Google Map to a Bing Map for classification and the resulting data was collected within iTree Canopy. 100 points were interpreted with the study area. iTree Canopy provides a table and calculates percent cover along with standard error on the fly.
Results
Percent canopy varies across the various methods (See Appendix 2 – Summation of all Hibbing Canopy Assessment Methods)
Digitize urban forest canopy cover using image classification remote sensing techniques and software over study areas
This method resulted in a Urban Tree Canopy of 24.71%.
Accuracy Assessment on this method resulted in an overall classification accuracy of 88.39% (See Appendix 3 – Object Oriented Classification Accuracy Assessment)
Digitize urban forest canopy cover via technician photo interpretation over study area
This method resulted in a Urban Tree Canopy of 18.03%.