Monitoring Fish Passage with an Automated Imaging System Steve R. Brink, Senior Fisheries Biologist Northwest Hydro Annual Meeting 2014, Seattle.

• Slide 1

Monitoring Fish Passage with an Automated Imaging System Steve R. Brink, Senior Fisheries Biologist Northwest Hydro Annual Meeting 2014, Seattle Slide 2 ` Malad River Hydroelectric Project Slide 3 ` Malad Fish Passage Plan FERC requirement to construct, operate, and maintain two fishways on the Malad River that permit rainbow trout (> 25 cm) to pass upstream and downstream of diversion dams. Required to continuously monitor use of each fishway. Monitoring of fishway replaced annual population sampling in river. Slide 4 Upstream and Downstream Passage Initially fish movement monitored past viewing window with motion detection camera and DVR Resident trout movement occurs year round and many fish exhibit false passes. Annual upstream movements average ~2,700 fish and downstream movements have been as high as 8,300 Slide 5 ` Monthly Trout Passage 2008 - 2012 Slide 6 ` Total Length (cm) Number of Rainbow Trout Length Distribution of Passed Trout Slide 7 ` Goals Track objects moving past viewing window Differentiate objects as fish or non-fish Record movement as upstream or downstream Measure objects identified as fish within 2 cm accuracy Use measurements, direction of movement, and time of passing to filter false pass movements Automated Fish Imaging System Slide 8 System Topology Embedded Computer Host Client Camera User Interface Computer System uses two computers to identify the active area in which objects travel in. Objects are then detected, measured, and classified as fish or non-fish Host software runs on an embedded computer and detects / tracks fish sized objects from the camera. This data is passed to Client software on a networked PC Client computer is responsible for less time-critical tasks such as measurement and classification of the object Client computer supports the user interface and resides on IPC network Slide 9 Challenges to Tracking Objects in the Field Multiple Objects Fish Location in the Window Viewing Window Conditions Changing Water Levels Water Clarity Slide 10 Challenges to Measuring Fish in the Field Lighting Multiple Fish Shadows Slide 11 Measurement Accuracy / False Passes False pass criteria Upstream and Downstream pass within 2 minutes Trout measured within 2 cm total length Object measurement must be accurate in order to correct for False Passes 2010 - 15,503 total passes 5,116 false 2011 - 14,364 total passes 3,095 false +/- 2 cm Slide 12 Four Components to the Software Water Line Detection Object Detection and Tracking Object Measurement Object Classification Automated Fish Imaging System Slide 13 ` Water Line Detection Detects changes in water level Excludes water ripple & region above it Algorithm defines yellow line and active area below Slide 14 Background is modeled & subtracted from each frame to obtain a foreground Foreground filtered by size, aspect and ratio bounds If object persists for set number of frames a tracker-object is expressed Object Detection and Tracking Slide 15 ` Object Measurement Client computer performs length measurement Length calculated using active contour model (Snake) which essentially forms a rubber band around object contour Many trout move along bottom of window producing pattern, texture, and color similar to background Length algorithm restricts measurements to frames centered in-frame by calculating distance of object from the center Slide 16 ` Object Classification Object classifier uses a support vector machine (SVM) to find distinct areas of separation between fish/non-fish objects Object shape (image moment), object color, and textural information are inputs to SVM Classifier trained with non-fish images (vegetation, water surface waves, & shadows) as well as known fish images Training set includes approximately 250 objects Slide 17 ` How is it working.? Iterative validation process made with manual motion detection system Compared objects passing the viewing window across three days in September 2013 Detection and tracking 590 objects tracked and labeled as fish or non-fish All trout that passed the window were detected and tracked Object Classification 36 out 43 trout passing the window were correctly classified (fish) 7 trout classified as non-fish Object Measurement 35 trout measured 22 within 1 cm accuracy additional 4 within 2 cm accuracy Slide 18 ` Continued improvement Current image quality is not adequate for accurate classification and measurement Update camera to 8 megapixels @ 21 FPS (current camera is 1 megapixel @ 9 FPS) Additional camera improvements will include automatic lighting enhancement More powerful embedded computer to handle tracking, classification, and measurement tasks. Reduce client computer responsibilities. Object Classification taken further to species identification