Mount Mansfield: Mapping the Alpine Tundra
Using Object-Based Image Analysis to Extract Alpine Tundra Features on Vermont’s Highest Mountain
Editor’s Preface: Cale is back, this time with a look at the alpine tundra of Mount Mansfield. Download and check out the data from this work at the VT Open Geodata Portal here (metadata here). It makes a nice compliment to lower-resolution NLCD land cover datasets, and an even better one to the high-resolution (0.5m) 2016 statewide land cover set soon available from VCGI:
Project Introduction
As part of Penn State’s Master of Geographic Information Systems (MGIS) program, students get to choose a capstone project on a topic and study area of interest. As a Vermonter with a love of the outdoors (specifically an affinity for the mountains), I knew from the start of the MGIS program my project would in some way involve mountainous terrain. Not until my capstone proposal course did I narrow the idea to Mount Mansfield’s alpine tundra. My intent with this project was to produce a tangible data set and make it available for download, viewing, and use in future GIS, remote sensing, and interdisciplinary studies.
Before getting into the specifics, I wanted to summarize the project and provide the reader some context. This project created a shapefile — a file type that a user can load into a geographic information system (GIS) software package (e.g. ArcGIS, QGIS) — and associated files that represent Mount Mansfield’s alpine tundra. In other words, the data I created provides a footprint of the alpine tundra land cover on and along the ridges of Mount Mansfield. In this project and for this shapefile, alpine tundra is defined by three (3) features: Bare Rock, Alpine Vegetation (non-evergreen), and Subalpine Krummholz (evergreen). The coverage of the alpine tundra spans the Sunset Ridge Trail, the Maple Ridge Trail, and the portion of the Long Trail from the Mount Mansfield Forehead to the Mount Mansfield Adam’s Apple.
I created the shapefile through an object-based image analysis (OBIA). Data sources included high-resolution 4-band imagery (Red, Green, Blue, Infrared) from the Vermont Imagery Program (0.2 meters, 0.5 meters) collected in 2013 during leaf-off conditions and the National Agriculture Imagery Program (NAIP) (1 meter) collected in 2014 during leaf-on conditions, and high-resolution elevation data (Normalized Digital Surface Model, Digital Elevation Model, Slope) from the Vermont Lidar Program (0.7 meters) collected in 2014 during leaf-off conditions.
This article covers the following parts of the project:
- Problem
- Background
- Land Cover Classification
- Project Workflow
- Data Sets
- Image Interpretation Key
- Data Analysis
- Results
Problem
The specific problem I tried to solve was that current low resolution, pixel-based national land cover data sets do not capture the size, shape, or spatial arrangement of alpine tundra features on Mount Mansfield. To solve this problem, I used object-based image analysis on Vermont’s high-resolution imagery and elevation data to map Mount Mansfield’s alpine tundra features for use on local scales and by local communities.
This solution created a baseline land cover map for the alpine tundra that is usable by local stakeholders and is cost and schedule-efficient to complete (all done via software vice field work). Applications of the output include estimating the area of the alpine tundra, establishing a baseline for change detection, and incorporating the data into future GIS, remote sensing, and interdisciplinary studies.
Background
Vermont’s highest peak, Mount Mansfield (Latitude: 44° 32' 38'' N
Longitude: 72° 48' 52'' W), tops out at 4,393 feet (1339 meters). The image below orients readers, for those who are unfamiliar with the location or appearance of the mountain.
National level land cover data sets, specifically the National Land Cover Database (NLCD) mapped by the Multi-Resolution Land Characteristics (MRLC) Consortium, provides 30 meter spatial resolution land cover data for the United States. However, this data set does not provide enough granularity to identify the size, shape, or spatial arrangement of features on more localized levels (e.g. Mount Mansfield, a single mountain in a single state). This is the crux of the problem I was trying to solve. The series of images below demonstrate that while the NLCD 2011 data does capture the overall shape of Mount Mansfield, it does not provide enough granularity for local scale (State/County level) analysis. As you follow through the three images (each one zoomed in farther than the previous), notice how you start to see the mountain’s varying terrain in the NAIP imagery (left pictures) and how the square, 30 meter pixels in the NLCD coverage (center pictures) do not fully or accurately capture the size, shape, or spatial arrangement of Mount Mansfield’s features.
Land Cover Classification
This analysis set out to map and delineate alpine tundra, as defined by three feature types: Bare Rock, Alpine Vegetation, and Subalpine Krummholz. The image below shows a closeup of an area along the Long Trail that samples each feature type of interest.
While the Bare Rock, Alpine Vegetation, and Subalpine Krummholz make up the features of interest, they are not the only features on the mountain. Other feature types include:
- Building
- Radio/TV Tower
- Auto Road/Parking Lot
- Gravel Construction Road
- Coniferous Tree (non-alpine)
- Deciduous Tree
- Ski Trail
- Car
In order to scope my analysis and focus on the features of interest, I grouped all other features into a class named Other.
Project Workflow
I completed this analysis in four stages:
- Data acquisition — downloading the relevant imagery and elevation data sets
- Data preprocessing — adjusting the data sets to fit the appropriate map projection and study area size
- Data processing — creating the shapefile containing the mapped and delineated features of interest
- Accuracy assessment — determining the quantitative and qualitative accuracy of the output shapefile
The image below shows the four stages and the major components of each stage.
Data Sets
I obtained all imagery and elevation data sets for this project from the Vermont Open Geodata Portal.
Imagery data sets (sampled in the image below) included:
Note that all of the imagery data sets contained four spectral bands (at the spatial resolutions listed above): Red, Green, Blue, and Infrared.
Elevation data sets (sampled in the image below) included:
- Normalized Digital Surface Model (2014, 0.7m)
- Hydro Enforced Digital Elevation Model (2014, 0.7m)
- Slope derived (in ArcGIS Pro) from the Digital Elevation Model (2014, 0.7m)
Image Interpretation Key
When conducting remote sensing studies where you are trying to extract features, it is useful to follow the elements of image interpretation and create an image interpretation key. The image interpretation key characterizes features according to how they appear in the the image relative to image tone, image texture, shadow, pattern, association, shape, size, and site. The results (the key) provide a comparison of feature characteristics that allow you to identify similarities and differences between the features. My image interpretation key highlighted differences between the alpine tundra features of interest (Bare Rock, Alpine Vegetation, Subalpine Krmmholz) in a way that helped to shape the eCognition rule set I used to map and delineate these features. More on the rule set later.
The following set of images shows the visual part of my image interpretation key (there is also a written part) for the features of interest in True Color imagery (Red/Green/Blue bands), Color Infrared imagery (Infrared/Red/Green bands), and the Normalized Difference Vegetation Index (NDVI) representations. For the NDVI representation (a measure of vegetation health in a sense), the important item to note is that the color ramp in the images below ranges from red (low NDVI values, barren areas) to yellow (medium NDVI values, shrub/grassland) to green (high NDVI values, dense/healthy vegetation). All features are highlighted in red boundaries, and each set of images shows four sample features per representation.
Data Analysis
With the imagery and elevation data sets downloaded and prepared, combined with the knowledge gained from developing the image interpretation key, I moved to the object-based image analysis portion of the project. This was the part where I used eCognition Developer software to map, delineate, and export the alpine tundra features of interest.
The general premise of object-based image analysis is to segment images — break images into groups of pixels (called objects)—and classify the objects as features (e.g. assign Group A to the Bare Rock class and assign Group B to the Alpine Vegetation class). Note that the terms Group A and Group B are used for exampled purposes only. Once you classify all objects, you can export the features from eCognition Developer to a shapefile for use in other GIS software packages.
To classify features within eCognition, you develop what is called a rule set (mentioned earlier). The rule set consists of rules (logical statements) that tell the software what to do with the layers loaded and the objects within eCognition. Examples rules could include:
- Assign all objects with nDSM (height) < 5 feet to the Short Tree class
or
- Assign all objects with NDVI < 0 to the Barren Land class
The above are summarized versions of rules. Within the eCognition software itself, the rules are not as readable as I have presented above.
The importance of introducing some background for the rule set is that it provided the method in which I mapped the alpine tundra features on Mount Mansfield. I created a rule set that, based on the feature characteristics identified in the image interpretation key, delineated the Bare Rock, Alpine Vegetation, and Subalpine Krummholz features and exported the result to a shapefile.
The following image provides an overview of the rule set I created for this project. Note that this overview highlights what I did, but not actually how (the actual rule set shows the how part).
As seen in the rule set overview, I listed other features I encountered during the analysis that I had to rectify, specifically shadow, snow, and mixed classes (either a mix of Bare Rock/Alpine Vegetation or a mix of Alpine Vegetation/Subalpine Krummholz). The image below shows a sample of the mountain containing both shadow and snow.
I rectified the snow and shadow by creating rules that used attributes from the other imagery layers to assign the objects to the appropriate classes.
One thing to note is the the NDVI (introduced above) played the most important role in the feature classification (assigning objects to classes). The image below defines NDVI mathematically — a normalized ratio of the Infrared band reflectance and Red band reflectance, with a value between -1 and 1 — and shows the NDVI ranges discovered for the alpine tundra features on Mount Mansfield.
It is important to note that the NDVI ranges presented above are not universal values. The Alpine Vegetation class will not always be between -0.1 and 0.1; these are just the numbers I found for the particular set of imagery on the specific image acquisition date. The same applies for the Bare Rock and Subalpine Krummholz. Though not universal, the ranges observed are directionally accurate (i.e. for similar lighting conditions in early May, Bare Rock NDVI values will generally be less than Alpine Vegetation NDVI values, and Alpine Vegetation NDVI values will be less than Subalpine Krummholz NDVI values).
Results
OK, now for the results. I hope you agree that, even though I spent a lot of time building up to this point, it was necessary (or at least useful) to provide the background and context that give meaning to the results.
I mentioned earlier that eCognition Developer provides the function to export the result of the rule set to a shapefile for use in GIS software packages. The images below show the shapefile I produced with my rule set, as well as the accompanying legend. The specific colors in the legend were chosen to match the actual feature colors at the time of the imagery used in the analysis (early May).
When conducting remote sensing studies, it is necessary to assess the accuracy of the output, both quantitatively and qualitatively. The image below shows the error matrix, often called the confusion matrix. It is a quantitative assessment of 624 randomly sampled points to see if my eCognition rule set placed features in the correct class (Bare Rock, Alpine Vegetation, Subalpine Krummholz). The two metrics most useful are the User’s Accuracy column and the Producer’s Accuracy row.
User’s Accuracy is a measure of how accurately I mapped the features from a map user’s perspective (i.e. the map I produced contains 100 Bare Rock features, but how many of the 100 are actually Bare Rock in reality?) It is a measure of commission errors.
Producer’s Accuracy is a measure how accurately I mapped the features from a map producer’s perspective (i.e. the area I mapped contains 100 Bare Rock features, but how many of those 100 did I show as Bare Rock?) It is a measure of omission errors.
The quantitative accuracy does not show the entire picture of how well (or not well) I mapped the alpine tundra. A different sample of 624 points could have produced much different accuracy percentages for all of the accuracy metrics. This is why it is important to include the qualitative accuracy assessment as well. It is difficult to compare the output shapefile to an imagery basemap when looking at the entire mountain at once. To better show the qualitative comparison, I created a series of six images along different parts of Mount Mansfield that compare the imagery basemap to the output shapefile.
My personal opinion is that the qualitative comparison provides a better assessment of how well (or not well) I mapped the alpine tundra. Looking through the images, you can identify areas where I was more or less successful with the mapping, as opposed to interpreting the numbers in the error matrix as the final say for all sections of the study area.
In addition to viewing the output in this article, you can access the shapefile and associated files here (metadata here). The files included in the Mount Mansfield land cover shapefile are:
- Feature Geometry (.shp)
- Index File (.shx)
- dBASE Table (.dbf)
- Coordinate System (.prj)
- Metadata (.xml)
- Spatial Index (.sbn, .sbx)
- Character Encoding (.cpg)
There are a few notes about the files that you should be aware of. The feature geometry file is 110 MB, and the dBASE Table is 80 MB (it contains 321,726 records). In addition, the shapefile is projected into NAD83 State Plane Vermont FIPS 4400 Meters (EPSG: 32145). The metadata file contains a comprehensive documentation of the data, analysis, and results, beyond what I discussed in the article.
For those object-based image analysis enthusiasts out there, I am happy to provide you the eCognition rule set (.dcp file) that I created for this project. Let me know if you are interested, and I will send it to you.
Conclusion
I wanted to reiterate the problem I was trying to solve and the objective of my analysis. I found that national level data sets such as the National Land Cover Database do not capture the size, shape, or spatial arrangement of alpine tundra features on Mount Mansfield. My solution used object-based image analysis on Vermont’s high resolution imagery and elevation data to create a tangible land cover data set that captures the alpine tundra on a local scale and is usable by local stakeholders. The following image provides a visual summary of the problem and my solution.
I want to thank Jarlath O’Neil-Dunne, who advised me throughout this project. His time, expertise, and guidance played a significant role in the success of my analysis.
I hope that you can find use in this data set, whether you are viewing the data or using it in a future GIS, remote sensing, or interdisciplinary study.
Let me know if you have any questions or comments about this project.
Cale Kochenour
