using UAV Aerial Imagery for Mapping Phragmites australis in Goat - - PowerPoint PPT Presentation

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 “ecosystem services”:

• Groundwater recharge
• Carbon storage
• Species richness or biodiversity 1

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)

Phragmites australis (common reed grass) Invades wetlands across the United States and particularly large areas in the Great Lakes 3. Exotic species have been introduced and proliferated over the last 150

• years. 2

Non-native species are less desirable than those of our native ecosystem.

• Crowd out native species
• Alter specialized habitats
• Decrease the biodiversity4

Common Reed Grass: Phragmites australis

Two genetic subspecies of Phragmites australis native to the greater San Francisco Bay and Delta.

• Phragmites australis subsp.

berlandieri

• Phragmites australis subsp.

americanus

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.

Altered and degraded wetlands and low salinity tidal marshes are more susceptible to invasion by Phragmites

• Low salinity marshes usually support

greater species richness than their higher salinity counterparts 2 . 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

accretion3.

• Invades initially by seed and spreads

by root rhizomes and stolons4.

• Treated by Grazing and Spraying

Indistinguishable without genetic testing from the non-native invader. 1

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 5th 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 5th 2016)

Methods

Imagery Pre-Processing and Segmentation

UAV Imagery Pre-processing Workflow:

RAW Imagery (.raw .tif .jpg) PIX4D or Drone2 Map Mosaic Image or Mosaic Dataset Comp-

• site

(Imagery + DEM) Segment / OBIA pixel groups

Segmented UAV Imagery of AOI:

• 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

• 1T. Blaschke. Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, Vol. 65, 1, 2010: 2–16.

Object Based Image Analysis (OBIA) is used in complex landscapes where there is a lot of heterogeneity of texture and color.

• Segments homogeneous groups
• f pixels into objects
• Objects can be classified into

types. The result is a smoother image classification with less salt and pepper appearance 1.

Methods

Segmentation

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

Reference data

• GIS Ground Truth Data
• Photointerpreted Training Data

Image Processing

• Orthomosaic Creation
• DSM Creation

Training Data

• Segmentation of Similar Pixels
• Create Training Sample Polygons

Image Classification

• Train Support Vector Machine
• Classify Raster

Accuracy Assessment

• Error (AKA Confusion) Matrix

1 Dronova, I. Object-Based Image Analysis in Wetland Research: A Review. Remote Sens. 7, (2015): 6380–6413.

Segmented NAIP 2014 Imagery using a Segment Mean Shift (1 pixel ~ 1 meter minimum segment size).

Analysis

Image Classification Results

UAV 2016 Imagery Classification NAIP 2014 Imagery Classification

Speckled / salt and pepper appearance Well defined vegetation /class boundaries

Extract Pixel Values from Accuracy Assessment Points Create a Frequency Table of Truth vs. Predict Values Create Pivot Table and Export Error Matrix

Analysis

Error Matrix Error Matrix Creation in Model Builder:

• 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

Classified Data Reference Data

Class Upland Low Marsh Mixed Emergent Phragmite s Row Total Upland 34 2 3 39 Low Marsh 17 35 11 63 Mixed Emergent 2 13 5 2 22 Phragmite s 4 14 18 Column Total 36 36 40 30 142

Classified Data Reference Data

Class Upland Low Marsh Mixed Emergent Phragmite s Row Total Upland 33 1 2 36 Low Marsh 13 35 3 51 Mixed Emergent 2 14 4 1 21 Phragmite s 1 8 1 24 34 Column Total 36 36 40 30 142

NAIP 2014 Classified Imagery: Error Matrix 1 UAV 2016 Classified Imagery: Error Matrix 1

NAIP Producer’s Accuracy Upland = 94% Low Marsh = 47% Mixed Emergent = 13% Phragmites = 47% NAIP User’s Accuracy Upland = 87% Low Marsh = 27% Mixed Emergent = 23% Phragmites = 78%

1 Congalton, R. G., Green, K. Assessing the Accuracy of Remotely Sensed Data: Principals and Practices. CRC Press Inc. (1999).

UAV Producer’s Accuracy Upland = 92% Low Marsh = 36% Mixed Emergent = 10% Phragmites = 80% UAV User’s Accuracy Upland = 92% Low Marsh = 25% Mixed Emergent = 19% Phragmites = 71%

NAIP Overall Accuracy = 49% UAV Overall Accuracy = 52%

Conclusion

Error Matrix Accuracy Assessment

The Overall Accuracy of the UAV classification (52%) was greater than the NAIP classification (49%).

• The Producer’s Accuracy of 80% for Phragmites is good
• (significant according to Congalton and Green (1999) 1 is 85%).
• Producer’s Accuracy = % of Phragmites mapped without errors (commission / omission)
• The User’s Accuracy of 71% is still fairly high (if not quite 85%).
• User’s Accuracy = % of Phragmites mapped that is actually Phragmites on the ground.

If the overall goal was to accurately map Phragmites australis, than a Producer’s Accuracy of 80% for the UAV imagery classification and User’s Accuracy of 71% are favorable values.

• The NAIP Producer’s (47%) and User’s (78%) accuracies are fairly good as well.
• The major difference is the usability of the image classification map:
• The UAV 2016 classification is more accurate for capturing Phragmites in the Map
• Both classifications are useful as maps of Phragmites actually on the ground.

1 Congalton, R. G., Green, K. Assessing the Accuracy of Remotely Sensed Data: Principals and Practices. CRC Press Inc. (1999).

Conclusion

Discussion & Next Steps…

• More GPS ground-truth or photo-interpreted points

should be used.1

• A hierarchical OBIA classification has been shown

in some cases to give better classification results.

– First classify upland from wetland  then low marsh vs. high marsh  finally individual species or alliances.

• Smaller macro-pixel size when segmenting the UAV

Imagery (or no segmentation?)

• Incorporating elevation (plant height) data from a

Digital Surface Model (DSM) and/or texture information (enhance the ability to identify Phragmites from the surrounding vegetation.3)

1 Congalton, R. G., Green, K. Assessing the Accuracy of Remotely Sensed Data: Principals and Practices. CRC Press Inc. (1999). 2Moffett, K. B. & Gorelick, S. M. Distinguishing wetland vegetation and channel features with object-based image segmentation. Int. J. Remote Sens. 34,

2013: 1332–1354.

3 Susan Ustin, Personal Communication February, 2nd 2016

In the above DSM: red is high elevation and blue is low elevation.

Conclusion

Further Research

Composite Raster Stack: RGB + DSM + Texture

Class Upland Low Marsh Mixed Emergent Phragmites Row Total Upland 33 1 2 36 Low Marsh 2 14 4 1 21 Mixed Emergent 13 35 3 51 Phragmites 1 8 1 24 34 Column Total 36 36 40 30 142

UAV 2016 & DSM Composite Image Error Matrix Overall Accuracy = 75%

UAV Composite Producer’s Accuracy Upland = 92% Low Marsh = 39% Mixed Emergent = 88% Phragmites = 80% UAV Composite User’s Accuracy Upland = 92% Low Marsh = 67% Mixed Emergent = 69% Phragmites = 71%

Management Recommendations

Unmanned Aerial Vehicles (UAV) in Environmental Management

• Early detection of Invasive species can improve management success by making their removal

more efficient and less expensive.1

• Remote Sensing can be a faster more economical monitoring tool to track vegetation changes
• ver time. 2
• Though UAVs are not a complete replacement for traditional field work, they do allow more

access to remote or challenging terrestrial and aquatic environments and can inform research with more complete visual assessments.

• The use of UAVs will likely continue to increase in all sectors. 2
• On May 4th 2016 the FAA Administrator announced plans to make the use of UAVs by students and researchers easier.

Currently, this is one of the major hurdles to the implementation of this technology.

• This week more restrictions on commercial drone operation were lifted.
• High resolution imagery is a powerful tool in environmental research, conservation, and

management.

1 Dvořák, P., J. Müllerová , T. Bartaloš , J. Brůna. Unmanned Aerial Vehicles for Alien Plant Species Detection and Monitoring. Int. Arc. of Photogrammetry and Rem. Sens. and Spatial Info. Sci., Vol XL- 1/W4, (2015): 83-90.

2Whitehead K & Hugenholtz C. Remote sensing of the environment with small unmanned aircraft systems (UASs), part 1: a review of progress and challenges. J. Unmanned Veh. Sys., vol: 216 , (2014):

69-8514 .

Acknowledgements:

• Dr. Tracy Benning,
• Dr. David Saah, and Dr. Gretchen Coffman, University of San Francisco

Steve Kohlman, Solano Land Trust Jared Lewis, San Francisco State University The GIS Department, WRA and Susan Ustin, University of California Davis