This is the fourth in a multi-part series examining how artificial intelligence may impact California water.
California lacks information on how much water flows through the vast majority of the state’s rivers and streams—only 8% have stream gages, instruments that measure and record this water (also known as “streamflow”). This limits water managers’ ability to efficiently allocate water between competing users and priorities, including the environment.
AI, particularly machine learning, is beginning to help. Because AI is skilled at analyzing large datasets and identifying complex patterns, it has shown potential to fill in important gaps for California water. These include estimates of past streamflows—both in places that weren’t monitored and/or were altered by human actions—as well as forecasting future streamflows. However, AI’s ability to accurately predict streamflow hinges on having sufficient data. In California, these models are further challenged by the state’s complex, human-managed water system.
California agencies use natural flow estimates made by machine learning
The state relies on machine learning (AI models that learn from data without explicit instructions) to estimate what unimpaired flows used to be in California thanks to the Natural Flows Database, a public tool created by the US Geological Survey (USGS), The Nature Conservancy (TNC), and other organizations. This tool describes what seasonal and monthly streamflows would have been without human alterations across the state. The California Department of Fish and Wildlife uses these modeled flows for their Watershed Streamflow Criteria, which allow the State Water Board to estimate the amount of water needed by California streams. The tool has also been instrumental to achieving goals of the California Salmon Strategy as well as informing the implementation of the Cannabis Cultivation Policy.
Modeling streamflows is challenging, but needed
However, predicting flows in waterways affected by human actions—including water diversions, dam operations, and wastewater discharges—has proved to be more challenging. TNC, in partnership with a company called UpstreamTech, recently released new models to predict daily natural streamflow—that is, how much water would flow through a stream without dams and diversions—and estimate actual, altered flows at locations without stream gages. The new effort leverages “deep learning,” an advancement that emulates human learning, including adjusting its approach when it makes mistakes. However, predictions of both natural and altered daily flows are not as accurate as the monthly and seasonal natural streamflows in the Natural Flows Database. This is not surprising—daily streamflows vary over many orders of magnitude and are notoriously difficult to predict. Estimating uncertain human activities upstream adds another layer of complexity, but skillful models of actual flows would help inform more precise water management.
New AI-powered forecast tools are starting to inform decision-making
While the Natural Flows Database provides estimates of past streamflow, which is useful for setting management and policy goals, many water managers need forecasts of future streamflows in order to assess flood or drought risk and inform dam operations. UpstreamTech’s AI-based streamflow forecasting tool, HydroForecast, which forecasts future flows in gaged and ungaged basins, is already being used in California. USGS also recently released a tool—River DroughtCast—that uses deep learning to forecast droughts and give communities extra time to prepare for water shortages.
Novel AI approaches can provide information about streams at lower cost
AI can also be used to supplement stream gages, the gold standard for monitoring, when there is limited funding. The USGS-developed Flow Photo Explorer, for example, uses machine learning models (namely, RankNet) to rank photos of flow and estimate relative river flows. These estimates provide timing of extreme high- and low-flow events and—combined with other stream information—can estimate the flow of water, providing an affordable way to supplement river flow data from traditional stream gages at a fraction of the cost. While these data cannot replace high-quality stream gaging, this approach offers useful supplementary information. Organizations including the California Department of Fish and Wildlife, TNC, and UC San Diego Center for Western Weather and Water Extremes are using them across the state. The Flow Photo Explorer is also being applied in other states to address challenges such as predicting harmful algal blooms and estimating snow depth.
AI is data hungry—and it’s only as good as what it eats
While AI models are emerging as an important tool for understanding river flow and water availability, it’s important to recognize that developing these models requires training and validation with long-term, high-quality datasets. Unfortunately, with recent federal budget proposals including large cuts to science agencies, more data collection is likely to fall on state and local entities to collect the information they need. It’s thus more important than ever that California continues and increases monitoring of its rivers—because without it, even the best AI models won’t have reliable data to learn from.
