Fish counting,
validated in the field.

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 8–16× speed, presenting an opportunity for error by human counters who are trying to quickly digest slow periods of fish migration.

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—the majority of cases—are automatically tallied into the database results without the need for review.

During the May Chinook validation window, 197 of 209 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.

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.

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.

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).

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.