Hydroelectric dams in the U.S. must prove to the Federal Energy Regulatory Commission that they are not killing endangered fish. That compliance requirement generates fish count datasets, collected by trained biologists watching fish move through ladders and spillways, logging species, injury status, and fin clips by hand. The process is error-prone, inconsistent across operators, and expensive. Some dams count hourly. Others count monthly. The data is a mess, and regulators still accept it.
The article traces one consultant's work building a human-in-the-loop computer vision system to automate fish counting at large hydroelectric dams, primarily in the Pacific Northwest where Bonneville Power Administration hydropower dominates the grid and native salmonid species are already listed as threatened or endangered. The interesting material is not the conclusion that CV can replace manual counting. It is the granular account of how you coordinate domain experts, fish biologists, as annotators inside a machine learning pipeline, how you handle labeling disputes over subtle attributes like clipped adipose fins, and how you communicate model progress to stakeholders who do not think in F1 scores.
The piece is worth reading in full for its treatment of dataset governance: what happens when your ground truth is produced by people working in poor lighting, in remote locations, under physically demanding conditions, and how you build a system that accounts for that without dismissing the expertise those people carry. If you work on annotation pipelines, environmental ML, or regulated-industry AI, this is a concrete case study, not a pitch deck.
[READ ORIGINAL →]