A Comparison of Image Classifications using UAV Aerial Imagery for Mapping Phragmites australis in Goat Island Marsh Francis S. Hourigan Master of Science in Environmental Management May 19th 2016 University of San Francisco
Wetland Impacts from Invasive Species Wetlands provide a variety of functions or Phragmites australis (common reed grass) “ecosystem services”: Invades wetlands across the United States and particularly large areas in the Great Groundwater recharge Lakes 3 . Carbon storage Species richness or biodiversity 1 Non-native species are less desirable than Exotic species have been introduced those of our native ecosystem. and proliferated over the last 150 Crowd out native species years. 2 Alter specialized habitats Decrease the biodiversity 4 1 Mitsch, William J.; Gosselink, James G. Wetlands. Wiley. Kindle Edition. (2015) 2 Chambers, R. M., Meyerson, L. a. & Saltonstall, K. Expansion of Phragmites australis into tidal wetlands of North America. Aquat. Bot. 64, (1999): 261 – 273. 3 Hazelton, E. L. G., Mozdzer, T. J., Burdick, D. M., Kettenring, K. M. & Whigham, D. F. Phragmites australis management in the United States: 40 years of methods and outcomes. AoB Plants 6, (2014): 1 – 19. 4 (Steve Kohlman PhD, Pers. Comm. December 2015)
Common Reed Grass: Phragmites australis Phragmites is a member of the Poaceae family (grasses). It stands on average 2-4 meters (up to 13 feet) high. Good nesting habitat for marsh birds. Bank stabilization and sediment accretion 3 . Invades initially by seed and spreads by root rhizomes and stolons 4 . Treated by Grazing and Spraying Two genetic subspecies of Phragmites australis native to the greater San Francisco Bay and Delta. Phragmites australis subsp. Altered and degraded wetlands and low salinity berlandieri tidal marshes are more susceptible to invasion Phragmites australis subsp. by Phragmites americanus Low salinity marshes usually support greater species richness than their Indistinguishable without genetic testing higher salinity counterparts 2 . from the non-native invader. 1 1 (http://ucjeps.berkeley.edu/eflora/) 2 Chambers, R. M., Osgood, D. T., Bart, D. J. & Montalto, F. Phragmites australis Invasion and Expansion in Tidal Wetlands: Interactions among Salinity, Sulfide, and Hydrology. Estuaries 26, (2003): 398 – 406. 3 Philipp, K. R. & Field, R. T. Phragmites australis expansion in Delaware Bay salt marshes. Ecol. Eng. 25, (2005): 275 – 291. 4 Hazelton, E. L. G., Mozdzer, T. J., Burdick, D. M., Kettenring, K. M. & Whigham, D. F. Phragmites australis management in the United States: 40 years of methods and outcomes. AoB Plants 6, (2014): 1 – 19.
Goat Island Marsh, Rush Ranch Solano Land Trust National Estuarine Research Reserve (NERR) Rush Ranch is a 2,070-acre open space preserve that is owned and operated by • the Solano Land Trust. • It is a working cattle ranch and a protected tidal saltmarsh habitat, as well as a National Estuarine Research Reserve (NERR). It is situated within the Suisun Bay and part of the extensive marsh habitat of the • Sacramento-San Joaquin Delta (www.solanolandtrust.org). The purpose of the Goat Island Marsh Restoration Project is to reestablish tidal flows to the site and to reestablish characteristic marsh features and vegetation. Restoration Goals: Widen inlet channel Lower the perimeter levee Expand existing Submerged Aquatic Vegetation (SAV) ponds Active weed control and native species revegetation.
Methods Imagery Acquisition: National Agricultural Imagery Program (NAIP) Imagery (~ 1m) was acquired as an Esri Map Service from the Sonoma County Vegetation Mapping and Lidar Project (sonomavegmap.org) and clipped to the Area of Interest in Goat Island Marsh. The Unmanned Aerial Vehicle (UAV) Mission was flown by the DJI Phantom 4 quad-copter, using the PIX4D Capture App. on April 5 th 2016. An Orthomosaic Image of 81 photos was created along with a Digital Surface Model (DSM) with the PIX4D Mapper Software. Resolution was 3.4 cm / pixel
Green-up Signature of Phragmites in Early Spring (April 5 th 2016)
Methods Imagery Pre-Processing and Segmentation Segmented UAV Imagery of AOI: RAW Mosaic Comp- Segment PIX4D or Imagery Image or osite / OBIA Drone2 (.raw .tif Mosaic (Imagery pixel Map .jpg ) Dataset + DEM) groups UAV Imagery Pre-processing Workflow:
Methods Segmentation Object Based Image Analysis (OBIA) is used in complex landscapes where there is a lot of heterogeneity of texture and color. Segments homogeneous groups of pixels into objects Objects can be classified into types. The result is a smoother image classification with less salt and pepper appearance 1 . Fig. 2. (a) Aerial photograph of heterogeneous landscape (b) fine scale segmentation (c) coarse scale segmentation (d) object based classification of woody cover, resulting in 97% accuracy (originally from: Levick and Rogers, 2008). 1 1 T. Blaschke. Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, Vol. 65, 1 , 2010: 2 – 16.
Methods Photointerpretation of Reference Data Reference Data: Random points were created and then buffered by one meter. The resulting random polygons were assigned to one of 4 classes using the high resolution UAV 2016 imagery. The number of Training Data polygons was approximately 60% of the original reference data sample. The remaining 40% of the reference data was used to create the Accuracy Assessment Data for the error matrices. 1 Congalton, R. G., Green, K. Assessing the Accuracy of Remotely Sensed Data: Principals and Practices. CRC Press Inc. (1999).
Methods Photointerpretation of Reference Data: Training Data (60%): Class # of Polygons Phragmites australis 48 Mixed Emergent 60 Low Marsh 57 Upland 47 Accuracy Assessment Data (40%): Class # of Polygons Phragmites australis 30 Mixed Emergent 40 Low Marsh 36 Upland 36
Methods Image Classification Image Classification Workflow • GIS Ground Truth Data • Photointerpreted Training Data Reference data • Orthomosaic Creation • DSM Creation Image Processing • Segmentation of Similar Pixels • Create Training Sample Polygons Training Data • Train Support Vector Machine Image • Classify Raster Classification • Error (AKA Confusion) Matrix Segmented NAIP 2014 Imagery using a Accuracy Assessment Segment Mean Shift (1 pixel ~ 1 meter minimum segment size). 1 Dronova, I. Object-Based Image Analysis in Wetland Research: A Review. Remote Sens. 7, (2015): 6380 – 6413.
Analysis Image Classification Results NAIP 2014 Imagery Classification UAV 2016 Imagery Classification Speckled / salt and pepper appearance Well defined vegetation /class boundaries
Analysis Error Matrix Error Matrix Creation in Model Builder: Extract Pixel Values Create a Frequency Create Pivot Table and from Accuracy Table of Truth vs. Export Error Matrix Assessment Points Predict Values Classified values are extracted from the Accuracy Assessment from the remaining 40% of the reference data. Two attribute fields are created: ‘Truth’ and ‘Predict’ The frequency for each class in the truth and predict fields are computed. A pivot table is generated with the class headings and their relative frequencies in an Error Matrix format.
Results Error Matrix NAIP Overall Accuracy = 49% NAIP Producer’s Accuracy NAIP 2014 Classified Imagery: Error Matrix 1 Upland = 94% Low Marsh = 47% Reference Data Low Mixed Phragmite Mixed Emergent = 13% Class Upland Row Total Marsh Emergent s Phragmites = 47% Classified Data Upland 34 2 0 3 39 Low 0 17 35 11 63 NAIP User’s Accuracy Marsh Mixed Upland = 87% 2 13 5 2 22 Emergent Low Marsh = 27% Phragmite 0 4 0 14 18 s Mixed Emergent = 23% Column Phragmites = 78% 36 36 40 30 142 Total UAV Overall Accuracy = 52% UAV Producer’s Accuracy UAV 2016 Classified Imagery: Error Matrix 1 Upland = 92% Low Marsh = 36% Reference Data Mixed Emergent = 10% Low Mixed Phragmite Class Upland Row Total Marsh Emergent s Phragmites = 80% Classified Data Upland 33 1 0 2 36 Low UAV User’s Accuracy 0 13 35 3 51 Marsh Upland = 92% Mixed 2 14 4 1 21 Emergent Low Marsh = 25% Phragmite 1 8 1 24 34 Mixed Emergent = 19% s Phragmites = 71% Column 36 36 40 30 142 Total 1 Congalton, R. G., Green, K. Assessing the Accuracy of Remotely Sensed Data: Principals and Practices. CRC Press Inc. (1999).
Recommend
More recommend
