Wells Dam Fish Counting Validation Study
- Nate Bennett
- Feb 26
- 5 min read
Salmon are the backbone of the Pacific Northwest ecology, supporting tribal and recreational fisheries, commercial harvest, and providing necessary ecosystem function in both marine and freshwater life history phases. For decades, fish managers and hydroelectric dam operators have dedicated thousands of hours to monitoring and enumerating salmon returns. This critical data supports our understanding of run strength, population-level success, ESA listing status, and informs fishery openings and closures.
Until very recently, human counters have performed this important enumeration work. Today, a new technology is modernizing the fish-counting process. This technology is not only alleviating the need for human counters but also demonstrating improved accuracy and providing additional data not easily obtained with human counters. Importantly, this technology has the potential to decrease costs associated with arduous and expensive human labor.
Ark CV is the leader in this technology space. The Washington-based company’s goal is to leverage the power of computer vision to automate this important fish-counting process, delivering accurate counts at a lower cost to operators, fish managers, and users alike. Over the past year, Ark CV has been evaluated using real operational footage across winter, spring, peak summer migration, and fall conditions to determine whether automated monitoring can perform reliably under those conditions. Ark CV has partnered with Natural Resources staff at Douglas PUD (East Wenatchee, WA) at Wells Dam to demonstrate and develop enumeration capability. Tens of thousands of salmon pass two adult salmon fish ladders and viewing windows at Wells Dam annually. In one set of peak-migration validation windows alone, the system processed 32,087 sockeye across 40 high-volume hours, achieving 98.7% agreement with official human counts using raw model outputs prior to QA (Quality Assurance) correction.
“I think we were a little suspicious that computer vision would struggle to resolve edge cases that human fish counters can. Fish counting is difficult since fish passage isn’t one-directional. Fish are passing our count windows upstream and downstream. Therefore, the computer needs to correctly add and subtract these events. Importantly, edge cases where a fish is obscured in the background by a conspecific in the foreground were also expected to be very difficult for computers to enumerate. After development and tweaks to Ark CV’s model during the test period, I think we have arrived at a place where the model is capable of enumerating fish, identifying species, fish size, and adipose presence at a higher accuracy rate than human counters.”
Andrew Gingerich
Douglas County Public Utility District, Natural Resources Supervisor
Ark CV is built as a production-grade computer vision system capable of automating fish passage monitoring across the Pacific Northwest. The system is capable of automating data collection that goes beyond just fish counts. Ark CV’s computer vision model can detect species, measure length to discern jacks from multiple-marine-year fish, and identify adipose-fin presence, producing multiple data points for every detection that can support decisions made by fisheries management.
Each season presents a unique set of challenges for fish monitoring, and Ark CV is designed to perform in all of them. Below we have summarized seasonal validation results conducted at Wells Dam fish viewing windows. To accomplish this, recorded video was obtained from Douglas PUD and ingested into a computer vision model developed by Ark CV. Tests were performed blindly, without knowledge of what manual counts had been reported through the video duration. Results were then verified after the naive effort was completed by Ark CV’s model.
Winter
Performance remained stable in winter conditions. Across five winter validation days totaling 25 steelhead observations, Ark CV achieved 100% species agreement (25/25 correctly identified as steelhead) and 96% adipose agreement (24/25 correct adipose classifications).
Importantly, these results were drawn from 76 hours of reviewed footage spanning January through mid-April. Although overall winter fish volumes were low, the system was evaluated across a substantial amount of video under variable lighting and environmental conditions. Importantly, the model can quickly remove long periods (hours and days) of lack of activity, when fish are not passing through the field of view. This footage is otherwise reviewed by human counters at 8-16X speed, presenting an opportunity for error by human counters who are trying to quickly digest slow periods of fish migration.
Spring
Ark CV demonstrated sensitivity to early-run fish. In a May validation window encompassing 31 hours of footage, the system identified an early-arriving sockeye prior to the first official seasonal report on June 11 by human counters. Subsequent blind video review (reviewers were unaware of both the model prediction and official human count) confirmed the detection, illustrating the system’s ability to detect out-of-pattern fish without relying on seasonal assumptions and biases cast by human counters who are not expecting a given species during an unlikely seasonal period.
A core component of Ark CV’s architecture is its confidence-based QA workflow. Every detection receives a confidence score. Low-confidence detections are automatically flagged for short review clips. This allows human reviewers to focus exclusively on edge cases rather than manually reviewing hours of footage. Computer vision counts assigned high confidence scores (majority of cases) are automatically tallied into the database results without the need for review. This workflow contributed to improved results during spring validation.
During the May Chinook validation window (209 total Chinook), 197 fish were identified automatically at high confidence (94.3%). Twelve detections (5.7%) were flagged for QA review. Total review time was under five minutes, and all twelve were confirmed as valid detections through blind review. Final species agreement for that time window reached 100%, while adipose classification agreement was 98.1% (205/209).
This workflow transforms what would traditionally require hours of manual review into a focused review of only a small fraction of detections that warrant additional scrutiny.
Summer (Peak Migration)
During peak migration in July 2024, Ark CV was evaluated in three independent high-volume validation time windows totaling 40 peak hours and 32,087 sockeye observations. Because the 2025 Columbia River sockeye run was significantly lower than the 2024 run, archived 2024 video was used to pressure-test the model under high-density conditions (worst case scenarios).
Using raw model outputs prior to any QA correction, the system achieved 98.7% agreement with official human counts. These video clips included passage periods with sustained passage rates exceeding 1,000 fish per hour. Agreement remained consistent across all three periods, ranging from 98.2% to 98.9%, with no evidence of instability at high fish densities. The model showed no systematic overcount bias and maintained consistent performance across ladders and time windows.
Fall
Seasonal robustness continued into the fall, when fish coloration and morphology changed during upstream migration as many fish species transitioned into spawning condition.
Across two late-September validation periods (September 28 and September 30), Ark CV was evaluated on a combined multi-species dataset that included Chinook, Coho, and Steelhead. This window had 53 combined Chinook and Coho observations, and species agreement reached 98.1% (52/53 correct), with a single cross-species misclassification and no missed detections.
Adipose classification also remained strong across the combined dataset. Steelhead adipose agreement was 95.4% (104/109), while Chinook adipose classification achieved 100% agreement (36/36). These results indicate stable multi-species performance as fish coloration and morphology change during upstream migration.
While fall sample sizes remain smaller than peak summer migration volumes, these results indicate stable multi-species performance as fish appearance evolves throughout the season.
In summary, these validation periods demonstrate that Ark CV maintains high agreement across species, seasons, and migration intensity, with a workflow designed for real-world deployment rather than laboratory conditions.
If you are interested in evaluating Ark CV on your own video or passage data, we welcome the opportunity to run a validation test using your footage. Please contact Nate at Ark CV if you’d like to learn more.
Nate Bennett
801-688-5885
Comments