University of Washington

Loss and Salvage Predictor

For model details and calibration methods see Tillotson et al. (2022).

Tillotson MD, Hassrick J, Collins AL, Corey P. 2022. Machine Learning Forecasts to Reduce Risk of Entrainment Loss of Endangered Salmonids at Large-Scale Water Diversions in the Sacramento-San Joaquin Delta, California. San Francisco Estuary and Watershed Science. https://doi.org/10.15447/sfews.2022v20iss2art3

Select Week Date

Unavailable dates silently adjusted.
Select Graphic Scaling

Set Annual Limits
Select Calibration
Loss for both species



Steelhead salvage & sum limits

Loss and Salvage Predictor Online Tool

This tool provides a graphical interface to predictive loss and salvage of endangered winter-run Chinook Salmon and Steelhead at the CVP and SWP pumping facilities at the southern end of the Sacramento-San Joaquin Delta, based on the model by Tillotson et al. (2022).

Loss and Salvage are related terms for the disposition of juvenile fish at the pumping facilities. Salvage refers to fish that are removed from exported water and eventually released back to the Delta (for more details, see Fish Salvage in State Water Project and Central Valley Project, CA Open Data). Loss refers to fish that are entrained in the exported water, but not salvaged and are therefore lost due to mortality. Salvaged fish are identified by their species and age to calculate loss (for more details, see Chinook Salmon Loss Estimation for Skinner Delta Fish Protective Facility and Tracy Fish Collection Facility, 7/9/2018). For monitoring purposes, loss is important for regulatory compliance so recent calibrations of the process are based on predicting loss.

Whether predicting loss, or alternatively salvage, realtime forecasts are depicted over the water year (WY), which begins October 1. The juvenile migration season here is typically from December to mid-June. A prediction for the week is made with weekly aggregated environmental conditions and the prior week's observations. Also available is the following week's prediction, which is made with the current week's observations and projected environmental conditions for the following week.

  1. The "Most Recent" run, which uses current environmental conditions for the current week's forecast of loss/salvage and projects those conditions one week into the future for the next week's forecast of loss.
  2. A historical run, which depicts observed environmental conditions and observed loss/salvage.
  3. A hindcast of a historical year that was not included in the selected calibration data set.

There are several model calibrations to choose from:

Loss for both species
  • All data, WY 2009-2025: Version 3, the current model calibration, and the default, includes all data available in these water years.
  • Non-0 filtered data, WY 2009-2025: Version 3, is an alternative model calibration and includes all data except for the days with 0 counts of fish.
  • WY 2009-2020: Version 2 model calibration predicting loss for both winter-run and steelhead.
  • WY 1999-2019: Version 1 model calibration predicting loss for both winter-run and steelhead
Steelhead salvage & sum limits
  • WY 2009-2020: Version 2 model calibration predicting loss for winter-run and salvage for steelhead
  • WY 1999-2019: Version 1 model calibration predicting loss for winter-run and salvage for steelhead

Over the years, management practices have changed and data (e.g., OMR flow) may have missing historical values. Users may want to select the calibration that reflects years of more current operations and more complete environmental data. Older calibrations are still made available in case users want to access these. Additionally, although loss is always computed for winter-run Chinook salmon, the user may select the legacy management target: "salvage" for Steelhead.

Environmental variables used for prediction include weekly averages of Old and Middle River flow (OMR), combined CVP and SWP exports, San Joaquin River flow at Vernalis, Sacramento River flow at Freeport, and water temperature at Mallard Island, as well as cumulative rainfall (inches) over the Delta area and the Delta Cross Channel gate status. The default calibration ("All data, WY: 2009-2025") uses the 7-day sum of precipitation for consistency with the one week aggregation of all the other variables, whereas previous calibration used a 5-day sum. The changes in results are negligible and the 7-day sum of precipitation is more aligned in temporal resolution with the other weekly variables.

Results from the Loss and Salvage Predictor tool for historical years draw from existing model calibration runs for each species. Thus, historical results will be the same point estimates and confidence bounds in the query outputs. This is different from the outputs in Track-a-Cohort, where each time the model is rerun with near-real-time data, the values for historical point estimates and confidence bounds will differ due to randomization in the randomForest R function, but are statistically equivalent.

Graphics include: Observations for the current water year (small circles). The forecast window (green triangle) with high, median and low projections (short lines). The historical accumulated take (grey line) and 95% confidence bounds on the observed cumulative fraction of the total for the week (grey area).

Version Release History

Version 3: recalibration using data from 2009–2025, including improved QCed OMR data. For consistency, the "All data, WY 2009-2025" data set is also used in the Track-a-Cohort prediction of Weekly Loss of Juveniles in the Delta. Released 2026-10-20.

Version 2: recalibration using only data from 2009–2020, reflecting years of more current operations and more complete environmental data. Released in 2021-Jan.

Version 1: original release of the tool in 2020-Jun with calibration provided by Michael Tillotson (personal communication).

Data Sources

Data sources for data used in Predictions