Automated substrate characterization using low-cost sidescan sonar

American Fisheries Society Meeting, Tampa, Florida Aug 2017

Transcript

1. Automated substrate characterization using low-cost sidescan sonar. Daniel Buscombe Sch. Earth Sci. & Environ. Sustainability, Northern Arizona University. American Fisheries Society Annual Meeting, Tampa, Florida, August 2017.
2. Outline • PyHum program • Sidescan processing and rectification • Unsupervised substrate classification • Supervised substrate classification • Challenges & future directions
3. Big boats ……………………………………. little boats Yellowstone River, Montana Colorado River, Arizona Penobscot River, Maine
4. Different substrate types have unique textures. Opportunity for fully automated substrate classification
5. Colorado River in Grand Canyon. Wide. Deep. Muddy
6. • Python program for working with Humminbird data 1. Reads data 2. Applies corrections 3. Remove shadows (manual or automated) 4. Rectifies imagery to make planform georeferenced imagery (“mapping”) 5. Substrate classification and mapping • Technical details: Buscombe (2017) Environmental Modelling and Software • https://github.com/dbuscombe-usgs/PyHum • http://dbuscombe-usgs.github.io/PyHum/ “PyHum”
7. “61-mile” data set
8. Reported sonar depth
9. Raw data Aerated water
10. Geometric & radiometric correction
11. Shadow removal
13. Rectification
14. Merged sidescan
15. Texture Lengthscale [m] Unsupervised substrate classification
16. Buscombe et al. (2016) Automated Riverbed Sediment Classification Using Low-Cost Sidescan Sonar. Journal of Hydraulic Engineering 142 (2) = න 2 Mean length of texture elements, in meters: = න 2 cos Normalized variance spectrum Vector of scales [m] Object of given height, length of shadow proportional to grazing angle, θ:
17. Merged sidescan
18. Unsupervised substrate classification Texture Lengthscale [m]
20. map Reach – aggregated substrate distributions
21. Buscombe (2017) Environmental Modelling & Software 89, 1 - 18
22. Supervised substrate classification
23. Conditional Random Field: • Sparse labels (my doodles) + image (greyscale scan) • Builds a probabilistic model for each label based on 1. Relative locations in the image and 2. Distribution of greyscales • Applies model to fill in gaps. Each iteration, it updates the probabilistic model
24. Minimal supervision!
27. • Challenge: Separation of bedform and grain roughness (also other highly discrete textures e.g. wood) • Possible future direction: more sophisticated texture identification ‘beyond variance’ (machine learning) Challenges & Future Directions 1
28. Challenges & Future Directions 2 • Challenge: sidescan imagery is poor quality because acoustic information is lacking • Therefore radiometric correction, better sidescan merging (data fusion) are very rudimentary • Possible future direction: Zhao et al (2017), Remote Sensing 9 (6) • Use a substrate classification to model backscattering by bed = better radiometric corrections = better quality scans
29. Last word. An “open-source” hardware system? • Black-box sonar: SL? TVG? Directivity? • So far, we’ve been lucky that manufacturers make data available and hack the data structure • How long will that last? • Interface with autonomous survey vehicles, ROVs – real-time data Source: https://www.asvglobal.com/ StarFish 453 OEM Sidescan System
30. • Buscombe (2017) Shallow water benthic imaging and substrate characterization using recreational-grade sidescan-sonar. Environmental Modelling and Software 89, 1-18 • Buscombe et al. (2016) Automated Riverbed Sediment Classification Using Low-Cost Sidescan Sonar. Journal of Hydraulic Engineering 142 (2) • PyHum • https://github.com/dbuscombe-usgs/PyHum Thanks! • Funding = USGS, BoR • Daniel Hamill (USU) • Adam Kaeser (FWS) • Paul Grams (USGS) • Ted Melis (USGS) • Sean Smith (UMaine) • Joe Wheaton (USU) • PyHum users!
33. PyHum vs. SonarTRX Point cloud Rubbersheeting (?)