Species of conservation concern are often negatively impacted by human infrastructure whose use can nevertheless be modulated (e.g. power generation and water transport) to lessen consequences. Such adaptive management, however, requires on-the-ground workers having access to near real-time data, analyses, and results, all of which are predicated on a robust computational workflow. Here, we take the example of two pumping plants on the California Delta that supply significant proportions of water used by the Central Valley and Southern California, but which negatively impact endangered Pacific salmon, who migrate through the Delta. Fish samples and water export volumes are collected 24 hours-a-day, 7 days-a-week, year-round at the two plants, and data since 1993 are available as a single file on-line. Given the seasonality of both water use and salmon migrations, our first statistical focus was to develop Generalized Additive Mixed Models (GAMMs) to describe the historical patterning in the data. We then established simple data summaries and near-term forecasting models that could be updated with each new day of data. We built our analyses into a robust computational pipeline for continuous (daily) deployment of results to a live website with a simple interface for navigation.
Results/Conclusions
There is full coverage since 1993, producing a dataset of (currently) 19,808 daily-integrated samples that had an average of 7.06 Chinook (sd: 25.11, max: 518, 63.9% 0s) and 0.91 Steelhead (sd: 3.09, max: 92, 80.9% 0s). Our GAMMs showed the significant seasonality, long-term trends, and year-to-year variability of both water export and salmon counts, indicating the utility of combining historical-based analysis with contemporary status to provide the most fruitful forecasting frameworks. Initial forecasting models produced prediction intervals that were nearly as wide as the range of counts observed, but forecast skill (measured using log score and rank probability score) appears to improve rapidly. We have connected the necessary data retrieval, analysis, summary, and presentation pieces into a computational pipeline whose workflow is documented and automated via Travis-CI, with the end product being daily updates to the salvage.fish website. Generalizing from this specific case, we develop a generalized computational workflow made up of constituents that can work with a wide range of applications. Combining robust statistical methods, flexible computational pipeline components, and continuous deployment provides a powerful framework for analyzing regularly updated data to inform on-the-ground managers with the most relevant results.