A New Version of CANAPI for Mapping Changes in Tall Shrub Canopies in Arctic

Tundra

Mark James Chopping, Montclair State University, choppingm@mail.montclair.edu (Presenter)
Rocio Duchesne-Onoro, University of Wisconsin - Whitewater, duchesnr@uww.edu
Angela Erb, University of Massachusetts Boston, angela.erb001@umb.edu
Zhuosen Wang, ESSIC, University of Maryland, zhuosen.wang@nasa.gov
Crystal Schaaf, University of Massachusetts Boston, crystal.schaaf@umb.edu
Christopher Chopping, West Orange High School, tchopp00@gmail.com

A new version of the Canopy ANalysis with Panchromatic Imagery (CANAPI) code was developed following tall shrub mapping tests with WorldView-2 imagery over Alaskan

Arctic tundra. CANAPI results have some dependence on user-determined settings, so the project team performed iterative tests on the impact of analyst subjectivity by having

four team members plus one naïve user perform multiple runs with different settings (each time attempting to find an optimal set) and subsequently labeling the set that

produced the subjectively "best" result. Suitable test imagery was located using SQL queries in ArcGIS under the ABoVE Science Cloud Windows VM. QuickBird (QB02)

panchromatic and multispectral imagery from June 20, 2003 and WorldView-2 (WV02) panchromatic and multispectral imagery from July 14, 2015, and a smaller QuickBird image

subset previously used were used in a series of CANAPI runs. Only tall shrub canopy measurement results from the "best" CANAPI runs from each user were considered. The

relative uncertainty in the estimates of mean crown radius was lowest at 4.3%, 3.2%, and 4.5% for the test, QB02, and WV02 image subsets, respectively, while the

corresponding values for %tall shrub cover are 25.6%, 58.0%, and 35.2%; and for mean shrub height 25.2%, 24.7%, and 30.5%. The estimated absolute and % changes over

2003–2015 (again considering only the "best" CANAPI runs) showed increases in the number of crowns detected, though there was a wide disparity from user to user: from 122%

through 630% (though the latter was from the naïve user). Changes in mean crown radius were far less variable but showed changes in both directions, from -5% through +8%.

Changes in estimated tall shrub cover were highly divergent, ranging from 7% to 105% (ignoring the 349% result from the naïve user), while changes in mean shrub height varied

from no change through 129% (again, ignoring the result from the naïve user). Visual inspection of the results for the QB02 and WV02 imagery from the most experienced user

(Duchesne) indicated unambiguous increases in shrub number, size, cover, and height over the 2003-2015 period, consistent with the quantitative results (increase of 196%, 5%,

59%, and 15%, respectively). These values are outside the uncertainty range, with the exception of the change estimate for mean height (15% vs 25% and 31% uncertainty for

QB02 and WV02, respectively). In order to reduce the impact of analyst subjectivity and improve the results, a new version of the code was developed to exploit the multispectral

bands as well as the panchromatic imagery (Canopy ANalysis with Multispectral and Panchromatic Imagery, CANAMPI). To this end, orthorectification was performed for the

WV02 July 14, 2015 scene on the NCCS ADAPT linux VM using the Polar Geospatial Center ortho processing and the Alaska NED mosaic digital elevation model (approximate

example syntax: python pgc_ortho.py --epsg 32606 --resolution 0.5 --dem /DEM/alaskaned_mosaic_wgs84.tif --format GTiff --stretch ns --outtype UInt16 ~/nrIvishak_River/

/att/nobackup/ascuser/orthoout/). This produced accurately-geolocated panchromatic and multispectral image files on a 0.5 m grid; a Normalized Difference Vegetation Index

image was also produced using the NIR and Red bands. Tests were performed on CANAPI-like code that first identifies candidate tall shrubs in the usual way (i.e., by locating

crescent-shaped areas of bright pixels arising from shrub crown illumination), then evaluates the mean NDVI of the pixels of each candidate crown, and finally flags objects

unlikely to be shrubs using a fixed threshold value. Although this does improve accuracy by removing false positives, it was found that some areas with tall shrubs have

unexpectedly low NDVI values, potentially limiting the utility of this approach. An alternative strategy that uses image spectral information will be investigated: spectral vectors

may be extracted automatically from crowns initially identified using the panchromatic imagery and a distance metric (e.g., Mahalonobis distance) subsequently used as the

criterion.

Presentation: ASTM4_Poster_Chopping_43_72.pdf (25236k)

Associated Project(s): 

• Chopping-03

Poster Location ID: 43

Session Assigned: Vegetation Dynamics and Distribution