Science Stories: Adventures in Bay-Delta Data

Protecting Outmigrating Juvenile Spring-run Chinook—Part 3

Pete Nelson, PhD
Department of Water Resources

First… it’s been a while! I wrote two previous posts (Science for Salmon Part 1 and Part 2) about developing our annual forecast for the number of juvenile Spring-run Chinook salmon (“spring-run”) entering the Delta. Since then, the team developing this Juvenile Production Estimate (JPE) has been focused mainly on the modeling that uses various fish counts and environmental conditions to predict the number and timing of outmigrating spring-run. The modeling component is crucial because this is how we finally arrive at a forecast, but it’s a challenge to explain because it requires a lot of advanced mathematics.

Sometimes, scientists will point to a model into which the data are dumped and treat it as a bit of a “black box” from which emerges, sans further explanation, the answer (Figure 1). Our question—how many juvenile spring-run enter the Delta each year?—is a tough one, and we want to make the process used to determine our answer less opaque. So, this article is going to be a balancing act between that “black box” approach and an attempt to explain the beauty behind the math.

Spring-run Chinook salmon have gone from being one of the most abundant Chinook runs in California to being one of the scarcest; consequently, they are listed as threatened under both the California and Federal Endangered Species Acts. The Sacramento-San Joaquin Delta is the Chicago O’Hare International for California Chinook, the hub for adult fish returning from the ocean to their natal streams to spawn, and for the juveniles migrating downstream back to the ocean. Salmon biologists, water managers, and fishermen worry about the gauntlet of threats these anxious travelers experience, including warm water, predators, and taking the “wrong turn” into water diversions. Managing the risk posed to the spring-run population by these diversions is the responsibility of the Department of Water Resources(DWR), and to do this we need to know when and how many salmon youngsters are entering the Delta each year.

The Modeling Architecture: More Than Just a Guess

A spring-run JPE isn't a single number pulled from thin air (or a fish pulled from a barrel); it is generated by a multi-model framework that combines several sophisticated statistical sub-models. To understand how we get to the final answer, we’re going to look at the four main challenges: counting out-migrating juvenile salmon, identifying which of those fish are actually spring-run, estimating their probability of surviving their journey to the Delta, and finally forecasting how many will actually reach that stage in their life-long journey.

Challenge 1: Counting the Invisible (BT-SPAS-X)

The first step is estimating the total number of outmigrating juvenile salmon passing by sampling locations at sites like Knights Landing on the mainstem Sacramento River and on various tributaries. Figure 2, above, shows the expected abundance of outmigrating juvenile Chinook and their weekly capture probability using an example from Lower Clear Creek, 2008. There’s a lot going on in these figures: The upper graph shows weekly abundance estimates, and the lower graph shows weekly capture probabilities in a rotary screw trap. Both figures show the flow or discharge rate in green, the abundance estimates from a modified Peterson model in blue circles and the predicted fish numbers from the BT-SPAS-X model using vertical bars, gray when the estimates are based on trap captures and pink where no data were available. The thin black bars show the range around each weekly estimate within which we’re quite certain the “true” number lies; where that span is narrow, we think we’re close to reality, but where that span is broad our confidence is much reduced. There’s a bit more to the figure, but I’m going to leave it there. If you’re curious about every last detail (nice!), get in touch and I’ll help you out.

Anyway, all of this depends on Rotary Screw Traps which are floating funnels placed in the current, facing the outmigrating salmon as they travel downstream to catch a portion of the passing fish (see Figure 5 for a simplified illustration). We mark a bunch of the fish we catch and then release them upstream of the trap so they have to pass the trap again. We expect, then, to catch a mixture of a few those marked fish and more unmarked fish (i.e., not previously caught). The ratio of marked to unmarked fish tells us the trap “efficiency.” But trap efficiency is strongly affected by both the trap placement and stream conditions, giving results that are sometimes distressingly poor, catching as few as 5 in 2000 marked fish moving down the big mainstem river. These methods traditionally only look at one rotary screw trap at a time, and rely on regular sampling and efficiency tests, but the accuracy can be really poor if there are any interruptions in sampling.

To improve our abundance estimates, we’ve started using a model with penalized splines—mathematical curves that act like a flexible ruler—to smooth out the trends in outmigration, allowing us to predict the trap efficiency when the trap wasn't even in the water, or when efficiency tests didn’t happen. The model "borrows" information from all the other years at that site when the traps were running, and across other sites, as well as using current environmental conditions like stream flow to help fill in the gaps. It also accounts for "process error," recognizing that there will always be some unexplained variation in efficiency.

Challenge 2: Who is Who? (The PLAD Model)

The next challenge is that multiple runs of Chinook salmon—fall, late-fall, winter, and spring—all use the Sacramento River watershed and, based on appearance, the juveniles are indistinguishable. Traditionally, biologists used Length-at-Date (LAD) criteria, which basically assumed that if a fish was a certain size on a certain day, it belonged to a specific run. However, using length-at-date for run assignment is notoriously inaccurate, often misidentifying half of the fish it tries to categorize.

To fix this, our colleague, Noble Hendrix, developed the Probabilistic Length-at-Date (PLAD) model. Instead of a "yes or no" assignment, PLAD uses genetic data from a subset of fish to build a "mixture model.” This model looks at the distribution of fish sizes from a genetically-determined sample of juvenile salmon and calculates the probability that a juvenile is a spring-run based on its length, when it was caught, and where it was caught (Figure 3). It’s like a more accurate version of a growth chart that acknowledges that some "late bloomers" or "early birds" from different runs might overlap in size, and allows us to assign each fish captured a probability regarding its run-type (e.g., 60% likelihood spring run and 40% winter run). By using this probabilistic approach, we can assign run types even to fish that weren't genetically tested, while being honest about how certain we are in that assignment.

Challenge 3: Estimating Survival (Smolt Survival)

The previous models help us to estimate the abundance and timing of juvenile spring-run produced each year in the Sacramento River watershed, but before these fish reach the Delta they must survive something of a gauntlet as they migrate from smaller, higher elevation streams to the confluence of the Sacramento and the San Joaquin rivers. This passage is fraught with danger, ranging from potentially poor water conditions to hungry predators, and estimating the probability of their survival is critical to developing our ultimate estimate (Figure 4).

Our survival model uses fish detection data from acoustic telemetry studies to estimate survival rates for juvenile salmon migrating from their natal tributaries, down the Sacramento River or the Feather River, and eventually to the Delta. Fish tagged with tiny sound-emitting devices are detected as they pass a series of underwater listening devices along the migration path, and each detection is logged as an individual fish of known origin at a specific date and time. So, yes, the government is listening. Just not to you. A sequence of detections provides a measure of downstream passage speed, and a failure of detection suggests the likely demise of that fish somewhere upstream. The primary determinants of survival in the model are fish size (bigger fish have an improved probability of success) and environmental conditions, such as flow or water temperature. For this reason, survival probabilities vary from year to year and from fish to fish, and the model integrates newly acquired data with the previous results each year.

Challenge 4: Predicting the Future (Stock-Recruit and In-Season Models)

Now that we have a decent idea of how many juvenile salmon are moving from their rearing habitat high in the watershed and we have an estimate of their survival probabilities for different flow conditions and fish sizes, we can use models to forecast how many of those are likely to make it to the Delta. We have two general ways of doing this:

First, we can use Spawner-Outmigrant Stock-Recruit models to predict how many outmigrating juveniles that today’s parents will produce over the coming year. This is like trying to guess how many apples an orchard will produce based on how many trees were planted and what the weather is expected to be like. We include environmental conditions in our forecast models, like water temperature and peak flows during spawning and rearing. Our findings show that cooler water and higher flows generally lead to more productive salmon populations.

As the migration season begins, we switch to the other way of forecasting the expected number of outmigrants, using an In-Season Outmigrant Model. This is similar to how a GPS updates your expected arrival time as you drive. Basically, at any given date during the outmigration season, the In-season Outmigrant Model counts how many outmigrants have passed a rotary screw trap and then uses the outmigration timing recorded from previous years to forecast how many more outmigrants we expect to pass by season’s end. Early in the season (say, late December), very few outmigrants have passed most of the rotary screw trap sites, so the forecast is broad and uncertain. But as more fish pass the traps, say by the end of February or March, the model "learns" more about the current season’s migration and environmental conditions, and the forecast becomes much more precise.

Bringing it All Together: The Integrated JPE

The final "answer" comes from the integrated JPE model, which stitches all these pieces together. It takes the spawner counts, expands the rotary screw trap catch counts to estimate outmigrant abundances with BT-SPAS-X, identifies the spring-run among those outmigrants with PLAD, forecasts total season outmigrant numbers with Stock-Recruit and In-season Outmigrant models, and then uses survival models to forecast how many of those expected outmigrants will make it as they swim from their natal creeks down the mainstem Sacramento River to the legal boundary of the Delta where the river flows under the I-Street bridge in the city of Sacramento (Figure 5).

 

This modeling process relies on a massive "field-to-cloud" effort that pipes data from over 40 different information streams overseen by dozens of data stewards to ensure that the data we put into the models are reliable and updated on a weekly basis for real time management decisions. Well…that’s the system that’s being built and should be in place by the end of 2026. Because each step—from counting a single egg to tracking a juvenile past a trap—has some uncertainty, our models propagate that error throughout the entire process. This means we won’t just give management teams a single JPE number wrought in stone (the answer!); we will give them a distribution of possible JPEs centered on a highest probability estimate that expresses our degree of confidence in the JPE forecast. And we can update that forecast through the season as more data roll in and our confidence in the forecast increases.

Patience truly does lead to beauty. By moving away from simple guesses and toward these sophisticated, probabilistic tools, we are creating a more transparent and robust way to protect our threatened spring-run Chinook salmon. For us, the answer to how many juvenile spring-run are entering the Delta is going to be different every year, each time generating additional, challenging questions: What was the effect of rainfall this time? How did ocean conditions affect the health of returning adults? And how should we ensure that there are sufficient water resources for this incredible fish every year?

Table 1. This table displays each of the icons used in Figure 5, a schematic of the spring run JPE, accompanied by text describing what each icon represents.

Image	Explanation
	Sacramento River and its tributaries
	Sierra Nevada mountains with snowpack
	Dams and other diversions give us the means to store and regulate water use
	Chinook salmon are an anadromous fish, juveniles hatching and rearing in freshwater but completing most of their growth in the ocean; adults return to freshwater to spawn
	Pacific Ocean offers salmon abundant food, vast space and rapid growth potential
	Data: numbers, size, genetics and more; this information is crucial to understanding salmon ecology, determining fisheries policy, and making water use decisions
	Rotary Screw Trap: these capture juvenile salmon as they migrate downstream, holding them so that we can count and measure them before releasing them back into the river
	Scientists collect data in the field, analyze samples in the lab, enter those data into the system, and develop and run the models to understand the patterns
	Cloud storage for the data provides a central repository accessible to all
	Models, including the response of salmon production to positive environmental conditions, help to make sense of lots of data
	State capital where policy is developed, such as that determining how water use decisions are made
	Decisions like a mandatory reduction in water exports to prevent drawing migrating juvenile salmon into pumping plants, for example

Note: Dr. Josh Korman, who led the technical team developing these models, died on 25 May 2026. Josh was one of the rare quantitative ecologists who also spent major amounts of time in the field—with a mask and snorkel, pulling nets and counting fish—and his keen enthusiasm for the challenge of understanding threatened fish populations was inspiring. We lost Josh before the completion of this project, but we were lucky to work with him. As we finish, I trust that we’d have made Josh proud.

By Pete Nelson

What’s this post about?

A Juvenile Production Estimate (JPE), as we discussed in Part 1 of this essay, is an estimate of the number and timing of outmigrating juvenile spring-run Chinook Salmon (“spring run”) as they enter the Sacramento-San Joaquin Delta. It is an important tool for protecting these fish because it helps water managers anticipate when these salmon may be at risk of becoming entrained in water diversions as well as serving as an important check on the status of this population. In this part, I’ll describe the cutting-edge genetic and modeling tools we’re using to distinguish spring run from the other Central Valley Chinook. This series will finish with a final installment full of the quantitative modeling we’re developing to pull in all the salmon and environmental data and actually produce a forecast of juvenile spring-run production.

Distinguishing spring-run Chinook from other salmon

Ronald Reagan once said, “A tree’s a tree: How many more do you need to look at?” I don’t know about Mr. Reagan, but most of us could probably tell that there’s a difference between a valley oak from a ponderosa pine (hint: one has pinecones and the other has acorns); however, even the best fish biologists can’t tell a spring-run from a fall-run Chinook based on looks alone.

Table 1. Central Valley Chinook salmon life cycle timing.

Until recently, the most practical way to separate the four different runs (remember? there are fall run, late fall run, winter run, and, of course, spring run) was to consider the size of the fish and the day of the year it was observed. Each run spawns during a different time of the year, and if those developing eggs and fry grow at approximately the same rate, you might expect that spring run, with peak spawning in September and October, are going to be a little larger than fall run, which spawn about two months later (Table 1). Similarly, late fall- and winter-run fish, which begin their spawning even later are likely to be smaller still. As you peruse Table 1, you’ll see that the four runs spawn during different times of the year (light blue bars), but there’s a lot of overlap, particularly between spring run and fall run. When you factor in variations in water temperature and food availability across the watershed, which directly affect the duration of egg incubation and larval and juvenile growth, there’s even more potential for overlap in size between the runs. To make run identification more challenging still, the juveniles from different spawning locations are typically moving downstream during roughly the same timeframe (November through July), and juveniles from some runs (especially late fall run and spring run) over-summer where they can find cold enough water and don’t migrate down until the following migration season, but typically earlier in the season and at a much larger size. Nonetheless, observing outmigrating juveniles with the first rains in November, you’d probably expect the largest of these fish to be late fall-run and spring-run fish that delayed migrating to the ocean and spent their first summer in their natal streams, with successively smaller fish being young-of-the-year winter run, spring run, and fall run.

Figure 1. River length-at-date (LAD) run identification curves (colored panels) and the genetic identity of fish (panels A-D) for juvenile Chinook collected at Chipps Island ( from Johnson et al. 2017). Word is, this figure is based on older genetic results and a lot of the bigger fall-run here are probably spring-run or late fall-run (thanks Brett Harvey!).

Counting cards, counting salmon

Now take a look at Figure 1. The background of each panel, made up of curved shapes, is colored to indicate the predicted size ranges on any given day of the year for the four run-types: Fall run as orange, late fall run as green, winter run as blue, and spring run as purple. These colors are generated using a mathematical formula that relates a fish’s length-at-date or LAD to a particular run-type. For example, looking at the blue winter-run curves, you’ll see that they (the fish and the shapes) start small, early in the time period—not even 50 mm long in September—but growing to 200 mm as early as about April. Therefore, if the models are correct, we would expect a small, 50 mm fish in October to be the progeny of parents who returned to fresh water in the winter (winter run). Similarly, a 75 mm fish in October is likely to be a late fall-run fish, and a juvenile 80 mm or larger in October is expected to be a fall-run fish. That’s if the fish play by our rules! Now look at the black dots on these graphs: Each dot represents the size of an actual individual, outmigrating juvenile salmon that was genetically identified as belonging to one of the four run-types, and the dots don’t all stay in the colored curve where we’d expect based on length. Still need convincing that these models aren’t perfect? Figure 2 compares the percentage of outmigrants belonging to each run as determined by LAD versus genetic test, and the success of the LAD approach at predicting run type could only be described as abysmal. Bottom line: LAD models are a less than dependable option for identifying the run of a juvenile salmon. Still, LAD does provide the foundation for another important element of our JPE program: probabilistic length-at-date or PLAD models.

Figure 2. Percent of total juvenile Chinook salmon by field year assigned to each run (rows), based on length-at-date (LAD) criteria versus genetic analysis (from Brandes et al. 2021).

The PLAD approach also uses the size of a juvenile salmon and the date captured to assign run type, but, employing a similar approach to counting cards when trying to beat your nephew during a competitive game of “Hi-Lo”^[1],; PLAD also uses probabilistic modeling to estimate the uncertainty associated with that run assignment. Figure 3 shows how two run types might be distinguished with varying degrees of confidence based on day of capture, size, and our history of genetically determined run types of that date and size combination. Here, the larger the fish, the more likely it is to be from that “blue” run type. Each contour line is defined by a probability measure. That measure of probability relies on our database of juvenile salmon captures where we’ve used genetic techniques to determine the run-type for juveniles sampled across a range of sizes, dates and locations of capture. We’re particularly careful to retain a tissue sample from fish with a low probability PLAD run assignment; later genetic analysis is expected to improve our model’s accuracy. As this database is expanded, our ability to include such factors as the tributary where the salmon was captured and environmental factors (e.g., flow and temperature) likely to affect growth rates will be improved. In other words, the PLAD model does not provide a run-type determination with complete certainty, but it should dramatically improve our ability to distinguish between these runs and it’s cheaper, quicker, and easier than doing the genetics on every single fish we sample.

Figure 3. Conceptual depiction of probabilistic length-at-date (PLAD) juvenile salmon size ranges for two runs (from Nelson et al. 2023).

Genetics

Thus far, we’ve been blithely referring to the use of “genetics” for distinguishing amongst Chinook run types without a care in the world for the hard-working folks who extract the DNA from these fish and examine the sequence of molecules contained therein. So how do they do it? DNA sequences can be used for differentiating amongst groups at varying levels—blood found at a crime scene might be examined first at the species level: Is this blood from a human victim or is this merely evidence of poor kitchen practices where someone butchered that darned rooster who persisted in crowing at 4 AM every morning? DNA can be used to determine if this is poultry or human blood, and it can also be used to determine which human left that blood. As you might imagine, distinguishing between individual humans requires more detail (more genetic markers, more time and expense) than determining foul versus fowl. The laboratory techniques for making these determinations are continually being improved, and DWR is now using a method that allows field crews to take a tissue sample from a fish and right there, in as little as 30 minutes, learn whether that fish is a spring-run salmon or not. The fish is unharmed, the sampling is quick, and the process is relatively inexpensive.

Our team has several genetic tests that we employ, each giving results with varying levels of precision and each with different demands in terms of time, effort and cost. Our initial test uses a CRISPR-based technique known as SHERLOCK (Baerwald et al. in press). It’s fast and cheap—less than $2.26 per sample—and gives one of three results: Early/Early, Late/Late, and Early/Late. These three terms relate to Central Valley Chinook run timing and the knowledge that these behaviors have largely been attributed to a single genetic region. Like humans, each salmon has two copies of each gene; each juvenile salmon inherits either an Early or a Late gene from dad, and an Early or Late gene from mom. Thus, each juvenile salmon ends up with either two Early genes (Early/Early or “homozygous Early”) or two Late genes (Late/Late or “homozygous Late”), or the juvenile may end up with one of each gene (Early/Late—“heterozygous” as the geneticists would say). It turns out that both winter- and spring-run salmon are Early/Early, and Late/Late fish are either fall-run or late fall-run salmon. Additional testing can distinguish between winter and spring run. What about the Early/Late heterozygotes? Yet another genetic test may allow us to assign those fish to a specific run, but this test costs more and must be processed in the laboratory, which takes more time.

Not only are these genetic tests critical to the ostensibly simple counts (“three more spring run and one more winter run”) that go into calculating a JPE, but they also serve to improve our PLAD model, testing and refining our ability to predict how to make run assignments using only size and date. After all, the PLAD models are still the quickest and cheapest way to determine how many juvenile salmon are produced by each of the runs.

What next?

Maryam Mirzakhani, the amazing Iranian-born mathematician, said, “The beauty of mathematics only shows itself to more patient followers.” Similarly, our efforts to forecast the number of juvenile spring run each year requires a little patience...and requires what ecologists refer to as quantitative modeling. This is where we take the monitoring data, PLAD results and the output of genetic analyses, and join it with environmental information—principally water temperature and flow rates—to generate a final JPE number. But for this, you’ll have to wait for Part 3.

Further Reading

• Baerwald, M. R., E. C. Funk, A. M. Goodbla, M. A. Campbell, T. Thompson, M. H. Meek, and A. D. Schreier. Rapid CRISPR-Cas13a genetic identification enables new opportunities for listed Chinook salmon management. Molecular Ecology Resources. in press.
• Brandes, P. L., B. Pyper, M. Banks, D. Jacobson, T. Garrison, and S. Cramer. 2021. Comparison of Length-at-Date Criteria and Genetic Run Assignments for Juvenile Chinook Salmon Caught at Sacramento and Chipps Island in the Sacramento–San Joaquin Delta of California. San Francisco Estuary and Watershed Science. [accessed 2023 Jan 3]. 19 (3).
• Johnson, R., S. Windell, P. L. Brandes, J. L. Conrad, J. Ferguson, P. A. L. Goertler, B. N. Harvey, J. Heublein, J. A. Israel, D. W. Kratville, J. E. Kirsch, R. W. Perry, J. Pisciooto, W. R. Poytress, K. Reece, and B. G. Swart. 2017. Science Advancements Key to Increasing Management Value of Life Stage Monitoring Networks for Endangered Sacramento River Winter-Run Chinook Salmon in California. San Francisco Estuary and Watershed Science.  15 (3).
• Nelson, P. A., M. Baerwald, O. Burgess, E. Bush, A. Collins, F. Cordoleani, H. DeBey, D. Gille, P. A. L. Goertler, B. Harvey, R. C. Johnson, J. Kindopp, E. Meyers, J. Notch, C. C. Phillis, G. Singer, and T. Sommer. 2022. Considerations for the Development of a Juvenile Production Estimate for Central Valley Spring-Run Chinook Salmon. San Francisco Estuary and Watershed Science. 20 (2).

1. Hi-Lo is a simple card game in which the dealer turns over the top card in a full deck and the player then guesses whether the next card in the deck will be higher or lower than the first card. Guess right and the player wins; guess wrong and the dealer wins (if the cards are equal, then neither player wins). The process continues as each card is revealed in succession. To give a very simple example, if you count the face cards (Jacks, Queens, Kings—12 in the deck, total), and the dealer has revealed 10, chances are very good that a Queen is going to be followed by a lower card.

Untangling the estuary’s food web with data analysis and synthesis

Blog by Rosemary Hartman. Paper and analysis by Tanya L Rogers, Samuel M Bashevkin , Christina E Burdi, Denise D Colombano, Peter N Dudley, Brian Mahardja, Lara Mitchell, Sarah Perry, and Parsa Saffarinia.

Did you know that one female threadfin shad can lay 5,000 to 20,000 eggs at once?! The California Department of Fish and Wildlife’s Fall Midwater Trawl, which surveys the Sacramento-San Joaquin Delta and upper San Francisco Estuary, caught 1,551 threadfin shad in 2022. If half of those were female, they could have produced three to fifteen MILLION baby fish. But in 2023 the Fall Midwater Trawl only caught 1,922 threadfin shad. What happened to all the baby fish?

Apart from the number of parents, fish populations (and other populations) are usually controlled by a combination of three ecological processes: 1) Individuals cannot find enough food or nutrients to survive and reproduce (what’s known as a ‘bottom-up’ process), 2) they are eaten by predators (what’s known as a ‘top-down’ process), or 3) they encounter unfavorable environmental conditions such as high water temperatures that lower survival. It is frequently very hard to tell which process is dominating, but understanding whether lack of food or too many predators is controlling a population can be very helpful in figuring out how to recover populations (Figure 1).

Figure 1. Fish populations depend on the number of fish you start with (parents), amount of food available, number of predators, and suitable environmental conditions such as temperature and salinity. Diagram by Rosemary Hartman, DWR.

To try and figure out which processes might be controlling threadfin shad (and other estuarine fishes), a team of scientists recently applied structural equation models to a long-term dataset of physical variables, phytoplankton, zooplankton, clams, and fishes. The resulting paper “Evaluating top-down, bottom-up, and environmental drivers of pelagic food web dynamics along an estuarine gradient” was recently published in the journal Ecology. It is a great example of how long-term monitoring data, data integration, cutting-edge statistics, and diverse teams can work together to provide new insights about our environment.

The team started by drawing out their hypothesized relationships between ecosystem components in a food-web diagram (Figure 2). They wanted to create a mathematical model that described the relationships between each component in their diagram, but could only do so if there were adequate data available. They compiled data from six different long-term monitoring programs and one model dataset for a combined package of forty years of data across the estuary. Unfortunately, some of the key variables in their conceptual model, such as specific types of phytoplankton, aquatic vegetation, and large, predatory fishes, haven’t been monitored as well as others, so couldn’t be used in the model, but what was available was impressive.

Figure 2. Conceptual model of the estuarine food web used as a basis for the NCEAS synthesis team's food web model. Diagram adapted by Rosemary Hartman, DWR, from Rogers et al, 2024. 

They applied a modeling technique that allowed them to quantify each of the connections in their food web diagram to determine whether changes in one population were due to top-down effects of their predators, bottom-up effects of their food supply, or environmental variables like temperature and flow, and this was not an easy task! As Tanya Rogers, one of the team members and co-lead author on the final paper said, “It was sometimes hard to find the right balance between detail and interpretability”. Really complicated models have lots of detail, but can be hard to understand, but if the model is too simple it doesn’t describe reality. They eventually settled on a model that was somewhere in between. They found that fish and zooplankton trends were more driven by food supply in freshwater areas of the estuary, but the same populations were more driven by predation in the brackish water areas of the estuary. Abiotic drivers (temperature, turbidity, and flow) were frequently important in all regions and at all levels of the food web and had similar or greater effects than food supply or predation.

This was a great example of using an integrated dataset to address big questions about the ecology of the estuary, but the coolest part of it may have been the team that put it together. This project was part of a synthesis program sponsored by the Delta Science Program and led by the National Center for Ecological Analysis and Synthesis (NCEAS). The Delta Science Program recruited a diverse team of researchers from State and federal resource agencies and local universities with complementary areas of expertise. The team then participated in training workshops run by NCEAS on open science practices, data synthesis, reproducible workflows, data publications, and statistical techniques. Once the team had these new skills, the NCEAS trainers helped them assemble the integrated datasets and analyses used for the project. This workshop resulted in a really cool paper, but more importantly the team gained the skills to do more research like this in the future. Plus, they had fun getting to work with other early career researchers from different organizations who bring different perspectives.

Most exciting of all, the team made several recommendations for future research, such as exploring nutrient dynamics in more detail, testing changes to the dynamics in different salinity zones, exploring different time scales, and filling monitoring gaps such as large predatory fishes. With the team’s new skills and the available data, there are a lot more possibilities to explore.

Further reading

• Tanya L Rogers, Samuel M Bashevkin, Christina E Burdi, Denise D Colombano, Peter N Dudley, Brian Mahardja, Lara Mitchell, Sarah Perry, Parsa Saffarinia. 2024. Evaluating top-down, bottom-up, and environmental drivers of pelagic food web dynamics along an estuarine gradient. Ecology. e4274.
• Brown, L. R., W. Kimmerer, J. L. Conrad, S. Lesmeister, and A. Mueller–Solger. 2016. Food webs of the Delta, Suisun Bay, and Suisun Marsh: an update on current understanding and possibilities for management. San Francisco Estuary and Watershed Science. 14 (3).
• Durand, J. R. 2015. A conceptual model of the aquatic food web of the upper San Francisco Estuary. San Francisco Estuary and Watershed Science.13 (3).

By Peter Nelson

The Challenge

Spring-run Chinook salmon (“spring-run”) are listed as threatened under both the California Endangered Species Act and the Federal Endangered Species Act. Like most salmon, these fish are anadromous: The adults, having grown and matured in the ocean, return to their natal stream to spawn, and the juveniles, after rearing in freshwater, eventually migrate downstream to the ocean (see Figure 1). During that downstream migration, juvenile salmon are exposed to a gauntlet of threats, including warm water temperatures, predators of all sorts, and “taking the wrong turn” through water diversions and getting lost on their way to the ocean. Managing or reducing the risk posed by water diversions is a responsibility of the Department of Water Resources, and to do that water managers need to know the number and timing of those outmigrating juvenile spring-run as they enter the Delta. Coming up with an accurate prediction of this—what’s termed a Juvenile Production Estimate (PDF) or JPE—is not simple. This is the first of a two part piece about our efforts to develop a JPE, both what’s been accomplished and what’s planned, as well as a timeline.

Figure 1. Spring-run chinook have a complex life cycle. The adults migrate upstream in January through March, but instead of spawning right away like most salmon they hold in coldwater pools all summer and spawn in the fall. Diagram by Rosemary Hartman, Department of Water Resources. Click to enlarge.

The Approach

We know when to expect adult spring-run to return to their natal streams to spawn based on past experience: Humans, beginning with the indigenous peoples of the West Coast, have been observing these runs for generations, and we might reasonably expect that the numbers of returning adult salmon are a decent predictor of the juvenile fish those returning salmon will eventually produce. Observations by multiple teams of biologists of adult salmon throughout the Central Valley allow us to predict the likely numbers of juvenile spring-run expected to migrate downstream and enter the Delta on the way to the Pacific Ocean each year. “Hold on a minute,” you might say, “What about the water in those streams? If the creeks are low and the water is warm, surely those baby salmon won’t do as well as they might when conditions are good.” You’d be right! The number of reproducing salmon—the parents—isn’t a perfect predictor of the number of offspring: There are many environmental factors that affect juvenile production, but, based on past studies of salmon ecology, we can include factors like flow in our analysis of the likely number of juveniles that will be produced by the annual return of adult salmon (for example, see Michel 2019; Singer et al. 2020).

These estimates, however, are just that—we can’t know exactly how the varying amount of water will affect the survival of juvenile salmon as they grow and migrate, but we should get reasonably close, and we have another source of information to improve our estimates, the number of outmigrating juveniles that we observe directly as they swim towards the Delta: The streams where spring-run spawn regularly have rotary screw traps (Video) (RSTs, Figure 2) on them. These devices divert migrating juveniles into a holding pen where biologists count and measure them each day before releasing them back into the stream to continue their journey to the sea. Data from these RSTs give us another check on our estimates based on spawner production, and are themselves an alternative means for estimating spring -run juvenile production.

Figure 2. A rotary screw trap floating in the Yolo Bypass Toe Drain with its cone out of the water (not sampling). Photo courtesy of the Department of Water resources.

One last point: In order for water managers to use these predictions for how many (and when) spring-run are expected to reach the Delta, these estimates need to happen each year before spring-run are expected to enter the Delta when water managers need to make decisions about their operations. This is especially tricky for estimates that rely on that RST data because it only takes a few weeks for juvenile salmon to travel from the RSTs to the Delta. This means that the process of counting adult salmon and (especially) juvenile salmon in the RSTs, entering those data into a shared database, and crunching the numbers to produce a JPE must be fast, efficient and accurate.

Gathering Information

This is a collaborative, interagency effort, which we began by holding a broad-based, public workshop in September 2020 with the Department of Fish and Wildlife (see Nelson et al. 2022 for details) and writing a science plan (PDF) with our agency partners to determine what monitoring data were needed to develop a spring-run JPE. Estimating an annual spring-run JPE is complicated by (1) the broad geographic and geologic range of Central Valley streams that support spring-run, (2) the challenge of developing a holistic, coordinated multi-agency monitoring framework for generating quantitative estimates of juvenile spring-run across their range, (3) the variable life history displayed across the spring-run streams, and (4) the difficulty of distinguishing juvenile spring-run from other run types (fall run, late-fall run, and winter run) found in the same streams (we will talk more about distinguishing salmon run type in our next blog post).

Monitoring

Most of the monitoring in spring-run streams is conducted by the staff of several governmental agencies (e.g., Deer Creek), gathering data on the numbers and timing of returning adults and of migrating juveniles, and tracking the changes in these metrics from year-to-year, but monitoring historically was designed to focus on local management needs, employed multiple methods and focused on different life stages across the watershed. Some work has been done to integrate data on number of returning adults (CDFW's GrandTab dataset, which produced the graph of returning adults, Figure 3 below). However, a spring-run JPE will require more a coordinated approach with the means of combining data from more than 40 monitoring programs from eight regions, several governmental agencies, and nearly two dozen data stewards and managers, using diverse methods and having large discrepancies in monitoring histories. These are significant challenges, but they can be met as long as we’re aware of the limitations (see below).

Figure 3. Total escapement (number of returning adults) by tributary for 2000-2022. Click for an enlarged version broken out by tributary.

In addition to gathering data on the number and timing of returning adults and departing juveniles, we’ll also need data on year-to-year salmon spawning success and on the survival of those outmigrating juveniles as they move from higher elevation habitats through lower, slower and warmer tributaries, and as they migrate down the mainstem of the Sacramento River to finally reach the Delta (streams with major spring-run spawning are shown in Figure 4).

Environmental conditions too are crucial: Preeminent are the quantity of water in the system and water temperature; we know that these have strong effects on salmon survivorship and behavior. The number and location of predators also vary from year to year and can affect the number of juvenile spring-run reaching the Delta.

Figure 4. Map of the Sacramento River watershed highlighting the rivers and streams where data is being collected for the spring-run JPE. Some spring-run also spawn in the San Joaquin watershed, but they have not been added to the spring-run JPE dataset yet. Click to enlarge.

Data Management

You may have heard the expression, “garbage in, garbage out”? Wherever the phrase originated, it certainly applies to ecology! Quality data and metadata (how, when, where, and by whom the data are collected) are critical to an accurate spring-run JPE and its application to salmon conservation and water management. DWR led the formation of a team to design a data management system. This team conducted extensive outreach to the various monitoring programs for the seven spring-run spawning streams identified as most important to the JPE.

This data management system is now a reality, and is designed to provide timely access to machine-readable monitoring data and metadata. To meet the annual deadlines for calculating a spring-run JPE, new RST data must be compatible across programs and reported rapidly. Building the initial dataset took over a year because of historical inconsistencies in data reporting across monitoring programs, but state and federal agencies are collaborating to make newly collected data compatible from the moment of data entry. Data from some monitoring programs are now acquired automatically from digital entry and uploads are occurring directly from the field daily; the rest of the monitoring programs will move to this “field-to-cloud” data entry system over the next several years, improving data quality and the greatly facilitating the ease of access. All historical RST data are now publicly available from the Environmental Data Initiative (use search term “JPE”), and new RST data will be added to this repository on a weekly basis. Indeed, one of the most exciting and novel aspects of the spring-run JPE effort is that it has unified much of the existing data reporting from multiple agencies monitoring along with new monitoring under a common goal and purpose.

The spring-run JPE data management program

• has now standardized data collection methodologies, schemas, encodings, and processing protocols;
• produces machine-readable data for all RST monitoring programs (adult data will follow soon);
• uploads data in near real-time to a shared data management system; and
• makes data publicly accessible in a simple format.

This system allows us to look at all the different data sources at once to learn new things! For example, if we plot the catch of salmon from the rotary screw traps at Mill Creek, the Feather River, Knights Landing, and Delta Entry from upstream to downstream (Figure 5) we see that the most upstream site (Mill Creek) catches salmon earlier than the downstream sites and catches a lot more of them. Moving downstream the catch gets smaller and smaller as juvenile salmon get lost, eaten, or die along the way. Mill Creek also has juvenile salmon leaving the stream as late as May or June, but very few of these fish make it all the way down to the Delta, indicating that later migrants might have a harder time surviving.

Figure 5. Plot of rotary screw trap catch over time for the spring of 2023 at several locations in the Central Valley. Click to enlarge.

In our next post on the spring-run JPE, we’ll describe the cutting-edge genetic tools we’re using to distinguish spring-run from the other Central Valley Chinook, the quantitative modeling we’re developing that pulls in all of the salmon and environmental data and actually produced a juvenile production estimate along with an indication of our confidence in that estimate, the peer-review process that will critique our program and recommend improvements, and where we expect to take this spring -run JPE program next.

Further Reading

• Atlas WI, Ban NC, Moore JW, Tuohy AM, Greening S, Reid AJ, Morven N, White E, Housty WG, Housty JA et al. 2020. Indigenous Systems of Management for Culturally and Ecologically Resilient Pacific Salmon (Oncorhynchus spp.) Fisheries. BioScience. 71(2):186-204.
• Nelson PA, Baerwald M, Burgess O, Bush E, Collins A, Cordoleani F, DeBey H, Gille D, Goertler PAL, Harvey B et al. 2022. Considerations for the Development of a Juvenile Production Estimate for Central Valley Spring-Run Chinook Salmon. San Francisco Estuary and Watershed Science. 20(2).
• Cordoleani F, Satterthwaite WH, Daniels ME, Johnson MR. 2020. Using Life-Cycle Models to Identify Monitoring Gaps for Central Valley Spring-Run Chinook Salmon. San Francisco Estuary and Watershed Science. 18(4).
• Harvey BN, Jacobson DP, Banks MA. 2014. Quantifying the Uncertainty of a Juvenile Chinook Salmon Race Identification Method for a Mixed-Race Stock. North American Journal of Fisheries Management. 34(6):1177-1186.
• Michel CJ. 2019. Decoupling outmigration from marine survival indicates outsized influence of streamflow on cohort success for California’s Chinook salmon populations. Canadian Journal of Fisheries and Aquatic Sciences. 76(8):1398-1410.
• Satterthwaite WH, Cordoleani F, O'Farrell MR, Kormos B, Mohr MS. 2018. Central Valley Spring-Run Chinook Salmon and Ocean Fisheries: Data Availability and Management Possibilities. San Francisco Estuary and Watershed Science. 16(1).
• Singer GP, Chapman ED, Ammann AJ, Klimley AP, Rypel AL, Fangue NA. 2020. Historic drought influences outmigration dynamics of juvenile fall and spring-run Chinook Salmon. Environmental Biology of Fishes. 103(5):543-559.

Blog by Rosemary Hartman, Data collated by Nick Rasmussen

Our last post gave you an introduction to the water weeds of the Delta. Invasive submerged aquatic vegetation is taking over the waterways, making it difficult for boaters, fish, water project operations, and scientific researchers (Khanna et al. 2019). As we described in the blog “Getting into the weeds”, they are hard to control too. But how do we collect data on aquatic weeds and what do those data look like?

How do we collect data?

There are two main types of data that we can work with (see IEP Technical Report 92 (PDF) for details). The first, is collected with areal photographs, and is known as ‘remote sensing’ (Figure 1). A specialized camera (sensor) is mounted on a platform (a drone, airplane, or satellite), and it can collect a regular-old photograph, or a hyperspectral photograph that records a lot of wavelengths of light that our eyes can’t see. Because different types of plants reflect different spectra of light (they are different colors) these photographs can be used to map location and extent of vegetation (Figure 2).

Figure 1. Diagram showing how remote sensing works. Light from the sun is reflected by objects on the ground. Different wavelengths of light are reflected differently. The sensor (camera) on the platform (airplane, satellite, drone, etc) registers the different wavelengths and stores them as an image file. Later, experts can classify the images based on which wavelengths of light were reflected.

The UC Davis Center for Spatial Technologies and Remote Sensing (CSTARS) has been mapping vegetation in the Delta using hyperspectral imagery collected with airplanes for most of the last 20 years (like the map in Figure 2). They create maps every year and share their data online via the Knowledge Network for Biocomplexity.

Figure 2. Hyperspectral image map of Suisun Marsh collected by CSTARS. Colors are used to represent different vegetation types (they aren't really that color).

Hyperspectral imagery is very good for identifying floating aquatic vegetation and terrestrial vegetation, but the water makes it hard to identify submersed vegetation. We can map where submersed vegetation is, but not what kind of vegetation is there. To look at community composition, we need to actually get out in the field and check on the weeds directly. To survey submersed weeds, researchers use a thatch rake (Figure 3) – an evil looking tool with sharp tines on one end and a long handle.

Figure 3. A thatch rake being used to sample aquatic vegetation. Photo from the Department of Water Resources.

To measure weeds, researchers either lower the rake into the water and twist it around to pick up the weeds, or they drag the rake behind the boat. When they pull the rake in, they rank the coverage of weeds on the rake head and identify the weeds to species. Several different groups have been collecting these data over the past twenty years– the State Parks Division of Boating and Waterways (DBW), UC Davis (including CSTARS), SePRO Corporation, and the Department of Water Resources. Dr. Nick Rasmussen (DWR) recently integrated all these datasets into one data publication available on the Environmental Data Initiative. He developed the data set because he was helping to write a series of reports about the environmental impact of drought and drought-related management actions. Those reports required rounding up aquatic vegetation data quickly, but at the time, virtually none of it was readily available. He wanted to fix that.

It ended up being a fun challenge for Nick because he got to learn how to create a fully reproducible process for integrating the dataset – including some hard decisions about how best to combine data that was collected in very different ways. He did all the cleaning and formatting in R scripts that are made available in a public GitHub repository.

What do the data look like?

Well, because these data were collected by different programs for different purposes, there are pretty big differences in number of samples and distribution of samples over the years (Figure 4). So it’s a little difficult to detect trends.

Figure 4. Graph of number of submerged vegetation rake samples per year in the integrated vegetation dataset.

However, when we look at the relative abundance of different species that have shown up in the rake sample data (Figure 5), we can really see an expansion in Potamogeton richardsonii (Richardson’s pondweed, black bars in figure 5) and Najas guadalupensis (southern naiad, brown bars in figure 5) after 2014. It’s not clear whether these species, both native to California, were recent invaders in the Delta or whether early surveys didn’t know how to tell the difference between them and similar looking species such as Potamogeton crispus (curlyleaf pondweed, red bars in Figure 5).

Figure5. Graph of relative abundance of each species across all rake sample surveys by year. No sampling occurred 2011-2013. Click for a version that shows separate plots for each species.

To make the story a little more complicated, these species aren’t found evenly throughout the Delta. Both Potamogeton richardsonii and Najas guadalupensis are found chiefly in Franks Tract (Figure 6) and not in any of the other regions of the Delta. SePRO and DBW conduct extensive sampling every year within Franks Tract (see Caudill et al. 2019 (PDF)), but not as high an intensity in other areas, so high abundance there throws off the Delta-wide data if it is not weighted by location.

Figure 6. Average relative abundance of different types of submerged vegetation by region of the Delta for the entire dataset. Click for a version that shows separate plots for each species.

We can also look at the hyperspectral maps to give us a record of total coverage of submerged weeds by year (Figure 7). We can see that total coverage of weeds really increased between 2008 and 2017, then remained about the same from 2017 to 2022.

Figure 7. Graph of percent of waterways in the Delta covered with submerged vegetation by year. Triangles indicate missing years.

However, it is often more interesting to look at a smaller area of the Delta and see how vegetation shifts from year to year (Figure 8, 9). For example, distribution of weeds in Franks Tract – a large, open-water area of the Delta – changed dramatically when a barrier was installed in West False River in 2015 and 2021-2022. The open-water area in the middle of the tract filled in during 2015, but the area on the eastern side of the tract started to clear out during 2021-2022 (see Hartman et al, 2022 (PDF) for more information).

Figure 8. Map showing the location of Franks Tract - a large, open-water area with extensive vegetation and the site of many vegetation surveys by DBW and SePRO. Maps of vegetation in Franks Tract through the years are in Figure 9, below.

Figure 9. Hyperspectral image of Franks Tract showing installations of a barrier in West False River during 2015, 2021, and 2022, and resulting changes in distribution of submerged aquatic vegetation. Click to enlarge.

Between all these surveys, we’ve collected a lot of data on weeds, but we haven’t done an extensive analysis. There are lots more questions waiting to be asked! How does the distribution of weeds change with floods and droughts? Which species of weeds grow in shallow versus deep water? Are any species expanding in range? Download the dataset yourself and take a look!

Further reading

• Borgnis E, Boyer KE. 2015. Salinity tolerance and competition drive distributions of native and invasive submerged aquatic vegetation in the upper San Francisco Estuary. Estuaries and Coasts.39(3)1-11.
• Boyer KE, Safran SM, Khanna S, Patten MV. 2023. Landscape Transformation and Variation in Invasive Species Abundance Drive Change in Primary Production of Aquatic Vegetation in the Sacramento–San Joaquin Delta. San Francisco Estuary and Watershed Science. 20(4).
• Caudill J, Jones AR, Anderson L, Madsen JD, Gilbert P, Shuler S, Heilman MA. 2019. Aquatic plant community restoration following the long-term management of invasive Egeria densa with fluridone treatments. Management of Biological Invasions. 10(3):473-485.
• Christman MA, Khanna S, Drexler JZ, Young MJ. 2023. Ecology and Ecosystem Effects of Submerged and Floating Aquatic Vegetation in the Sacramento–San Joaquin Delta. San Francisco Estuary and Watershed Science. 20(4) 
• Hartman R, Rasmussen N, Bosworth D, Berg M, Ateljevich E, Flynn T, Wolf B, Pennington T, Khanna S. 2022. Temporary Urgency Change Petition of 2021 and Emergency Drought Salinity Barrier; Impact on Harmful Algal Blooms and Aquatic Weeds in the Delta (PDF). Report to the State Water Resources Control Board. Sacramento, CA. 
• IEP Aquatic Vegetation PWT, Khanna S, Conrad JL, Caudill J, Christman M, Darin G, Ellis D, Gilbert P, Hartman R, Kayfetz K et al. 2018. Framework For Aquatic Vegetation Monitoring in the Delta (PDF). Sacramento, CA: Interagency Ecological Program. Technical Report 92.
• Khanna S, Acuna S, Contreras D, Griffiths WK, Lesmeister S, Reyes RC, Schreier B, Wu BJ. 2019. Invasive Aquatic Vegetation Impacts on Delta Operations, Monitoring, and Ecosystem and Human Health (PDF). IEP Newsletter. 36(1):8-19.
• Khanna S, Ustin SL, Hestir EL, Santos MJ, Andrew M. 2022. The Sacramento-San Joaquin Delta genus and community level classification maps derived from airborne spectroscopy data. Knowledge Network for Biocomplexity. 
• Rasmussen NL, Conrad JL, Green H, Khanna S, Wright H, Hoffmann K, Caudill J, Gilbert P. 2022. Efficacy and fate of fluridone applications for control of invasive submersed aquatic vegetation in the estuarine environment of the Sacramento-San Joaquin Delta. Estuaries and Coasts. 45:1842-1860.
• Rasmussen N. 2023. Submersed aquatic vegetation community composition in the Sacramento-San Joaquin Delta integrated across four surveys ver 1. Environmental Data Initiative.