| Q&A
“Humans cannot count fish 24/7, seven days a week at multiple different locations,” Quirion said. “They can only count a small statistical sample, and they have to interpolate in between.” This uncertainty affects not only operators, but
also regulators who rely on monitoring data to assess compliance and environmental performance. According to Quirion, the resulting data can be difficult to fully trust. “The resulting data is sometimes challenging for
regulators to believe or to really trust,” he said. Automated monitoring using AI addresses this limitation by removing the need for sampling-based interpolation. AI systems can operate continuously, recording and analysing every fish that passes through a monitored section of infrastructure. According to Innovasea, this leads to significantly higher accuracy and consistency compared with manual approaches.
The added value of species-level data While automated fish counting alone represents a
substantial improvement over traditional methods, the ability to identify fish species adds a further layer of value. The importance of species-level data varies by river system. In some cases, fish populations may be dominated by a single species, and aggregate counts may therefore be sufficient to assess overall passage performance. In other systems, however, fish populations may be distributed relatively evenly across species, or specific species may be designated as threatened, endangered or otherwise protected. “There are river systems where there are species that
are deemed at risk,” Quirion said. “It’s very important to understand how hydro operations are affecting those specific species.”
Without species-specific information, operators may have limited ability to demonstrate how their facilities interact with protected populations. They may also struggle to assess whether mitigation measures are effective or proportionate.
“If you don’t know, then you can’t take action to effectively protect them,” Quirion said. Species identification also has practical relevance
for hydropower engineering and equipment design. Different species vary in size, body shape and swimming behaviour, which can influence how they interact with ladders, bypasses and other passage structures. Accurate data on species composition can therefore support both operational management and longer-term infrastructure planning.
Integrating species identification into
HydroAI The Species Aware capability will be integrated directly into Innovasea’s existing HydroAI platform. HydroAI is already commercially available and deployed at multiple hydropower sites in Canada, where it provides continuous, automated fish counts. In its current configuration, HydroAI uses a camera positioned above fish ladders, capturing a top-down view of fish as they pass. This perspective is well suited to counting, as it allows individual fish to be tracked and recorded as they move through the passageway. However, reliable species identification requires additional visual information. “For species classification, we will require a side-view underwater camera,” Quirion said. “That will be a second camera that will just be part of that same insert.”
The side-view camera captures the fish’s profile as it swims past, enabling the AI system to analyse visual characteristics such as size, shape and colour. These features are used by the classification models to distinguish between species. Both cameras are housed within a single insert designed to fit into existing fish ladders. This approach allows species identification to be added without replacing or reconfiguring the core monitoring infrastructure.
Training AI for real hydropower conditions
A central focus of the Species Aware project is ensuring that AI models perform reliably under real-world hydropower conditions. Unlike controlled laboratory environments, hydropower passageways present a range of challenges, including variable lighting, turbulence, debris and changing flow conditions. “We are training an AI system specifically for hydro conditions,” Quirion said. “We want it to excel in the type of environment that you find in hydropower operations.” Innovasea’s approach involves both model design and data selection. The company has developed proprietary techniques within its machine-learning framework that focus on identifying visual features of fish that are less sensitive to low light and turbulence. Equally important is how the models are trained. “All of our data sets are collected in hydro
environments,” Quirion said. “They contain the type of challenges that you’d see in the real world.” Innovasea has accumulated a large archive of video
data from hydropower installations, which it considers a key asset. From this archive, curated datasets are created for initial model training. These datasets provide a baseline level of performance when the model is first deployed. Once in operation, the system continues to collect field data. This data is then used to fine-tune the model over time, allowing performance to improve as the system is exposed to a wider range of conditions. This iterative process ensures that model development
remains closely aligned with operational reality rather than theoretical performance benchmarks.
Measuring and validating accuracy Accuracy validation is critical when introducing AI-based
monitoring into regulated environments. Innovasea validates the performance of its systems by comparing AI outputs directly with human observations, which remain the historically accepted reference point. After a model is deployed, Innovasea selects a small statistical sample of the data generated by the system. Human observers are then asked to independently count fish and identify species within that sample. The results are compared with the AI’s outputs. Each dataset produced by HydroAI is accompanied by a performance metric expressed as a margin of error. This margin reflects the difference between AI and human results and provides transparency about data reliability. HydroAI’s margin of error varies depending on
site conditions. At many installations, Innovasea has observed margins of errors of approximately 5% or less. At more challenging sites, performance can deteriorate, sometimes with margins of error as low as 20%. In all cases, results are provided alongside the data so that operators and regulators can understand how much confidence to place in the results. By comparison, purely manual monitoring methods often involve margins
www.waterpowermagazine.com | January 2026 | 13 Below: HydroAi at a fish ladder
Above: HydroAI fish counter
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45
