Ecological forecasting – i.e., making iterative predictions about the future state of ecosystems that account for uncertainty and are updated with new data as they become available – can be a critical tool for the management of aquatic ecosystems. Forecasts of the future state of aquatic ecosystems (e.g., water quantity, water quality, fisheries) are increasingly needed as global change precludes the use of historical conditions for guiding predictions of future aquatic conditions. Recent advances in data availability and ecosystem models have positioned aquatic researchers to increase their use of forecasting for prediction of stream discharge, fisheries, hypoxia, algal blooms, endangered and protected species, drinking water availability, and other metrics of ecosystem functioning. Forecasts benefit a variety of users, including managers, policymakers, and the public, for purposes ranging from preemptive mitigation of water quality and quantity concerns to adaptively protecting aquatic biota and planning recreational activities. Forecasting also allows researchers to refine ecological theory by repeatedly confronting models with data, thereby improving models over time. In this session, we solicit diverse presentations on both methodological and application-based research on forecasting hydrodynamics, biogeochemistry, and ecology in aquatic ecosystems. We invite submissions that examine the role of forecasting to support adaptation and mitigation by engaging management, industry, communities, and other decision-makers in the face of climatic extreme events and other environmental challenges. Our overarching goal is to illuminate the value of ecological forecasting as an approach for guiding climate change adaptation and mitigation in aquatic ecosystems.
Lead Organizer: Mary Lofton, Virginia Tech (melofton@vt.edu)
Co-organizers:
Rafael Marcé, Institut Català de Recerca de l'Aigua (rmarce@icra.cat)
Cayelan Carey, Virginia Tech (cayelan@vt.edu)
Presentations
02:00 PM
A Comparison of Neural Network Models for Water Quality Forecasting (8088)
Primary Presenter: Marcus Lapeyrolerie, UC Berkeley (marcuslapeyrolerie@me.com)
This project has the objective of making a comprehensive comparison of state-of-the-art machine learning methods for the prediction of water temperature, dissolved oxygen and chlorophyll-a concentration in limnological time series. The time series that are used to evaluate the models are taken from the National Ecological Observatory Network's (NEON) Ecological Forecasting Challenge, an open challenge where teams can submit forecasts to data that is collected and made publicly accessible by NEON. A common finding across the challenge is that a day of year historical mean model (also referred to as the climatology model) commonly produces top scoring forecasts; thus, a primary consideration in this project has been to compare the performance of machine learning models to the climatology model. Using Darts, a Python library that implements various machine learning (ML) models for time series forecasting like Long Short Term Networks (LSTM) to more contemporary models like Temporal Fusion Transformers and Temporal Convolutional Networks, we evaluate the probabilistic forecasts from 8 different neural network models for water temperature, dissolved oxygen and chlorophyll-a concentration at 34 different sites across North America over the course of a year. This research is unprecedented in that we use machine learning models that have not been examined previously for limnological time series forecasting, and few studies have had a similar geographical scale. We found that all except one of the machine learning models performed as well or better than the climatology model across all three target variables according to the continuous ranked probability score. From our early explorations, we learned that missing data imputation greatly affected the accuracy of the machine learning models; thus, we designed a bespoke method to fill in missing data using historical data to achieve our results. Our work supports that machine learning models can accurately forecast limnological time series and are deserving of further development.
02:15 PM
LEVERAGING SEASON-AHEAD FORECASTS OF LAKE WATER QUALITY TO PREDICT FISH STRESS AND VULNERABILITY (7699)
Primary Presenter: Emma Blackford, University of Wisconsin-Madison (eblackford@wisc.edu)
Recent advances in subseasonal-to-seasonal scale predictions of climate and hydrology variables provide prospects for sectoral management; however, little attention has been devoted to prediction of water quality factors that directly or indirectly impact aquatic habitat conditions. Concurrently, significant effort has been aimed at advancing routine lake monitoring and assembly of local scale datasets of water quality and aquatic abundance. Presently there is a gap in pairing predictions with lake-specific data to understand the ability of models to predict water quality and habitat conditions and subsequently inform preseason management opportunities. Sufficient summertime oxythermal habitat for fish is closely tied to water quality parameters, most notably water temperature, dissolved oxygen, and nutrient levels. Cold-water fish are at high risk from a changing climate, given the direct reduction of habitat availability under warming temperatures. In response, direct prediction (e.g. fish stress) and indirect prediction (e.g. air and water temperature, nutrient levels, DO, oxythermal stress indices, and combinations) models were developed, using global and local hydro-climate features as predictors. Existing long-term datasets monitoring Wisconsin lakes were leveraged to assess a variety of metrics to predict oxythermal stress conditions for cold-water fish species on a season-ahead scale. Modeling and predicting changes at this time scale can inform actionable resource decisions, and compliment short-term and long-term climate change adaptation strategies developed by Departments of Natural Resources and other organizations. Future research includes evaluating anticipatory management strategies, particularly for predictions of extreme conditions, including fish kills, and understanding how lake water conditions and predictability may evolve under a changing climate.
02:30 PM
FUNCTION-AS-A-SERVICE CLOUD COMPUTING FOR LAKE WATER QUALITY FORECASTING: PAST, PRESENT AND FUTURE (8319)
Primary Presenter: Renato Figueiredo, University of Florida (rjofig@gmail.com)
Iterative, near-term water quality forecasting requires flexible cyberinfrastructure (CI) that meets several requirements. First, the CI must be reliable and support decision-making. Second, the CI must scale to support forecasting across a diverse range of waterbodies. Third, it must support complex workflows with various modules (e.g. data collection, QA/QC, ensemble execution) that are orchestrated to run autonomously. Fourth, it must be affordable and accessible to the growing community of aquatic forecasters. The Function-as-a-Service (FaaS) cloud computing model supported by commercial and open-source platforms (e.g. Amazon Lambda, GitHub Actions, OpenWhisk) meets all of these needs. Specifically, FaaS 1) offers automation, scalability and fault-tolerance by leveraging large-scale distributed computing resources, 2) provides a high degree of flexibility and customization of computing environments through containerization (e.g. Docker), and 3) is pay-as-you-go, where charges (e.g. bills, allocation credits) are proportional to the time actually used for computing (not idle time). This presentation will: 1) overview experiences and lessons learned from a successful use case of FaaS in lake forecasting , 2) describe the development of a recently developed novel open-source software (FaaSr) which lowers the barrier to entry to use FaaS for workflows written in R, and 3) describe future directions and use cases of FaaS in aquatic ecosystem forecasting workflows.
02:45 PM
FORECAST SKILL OF DISSOLVED ORGANIC MATTER CONCENTRATIONS DECREASES DURING STRATIFIED CONDITIONS (8061)
Primary Presenter: Dexter Howard, Virginia Tech (dwh1998@vt.edu)
Dissolved organic matter (DOM) is a valuable metric of ecosystem functioning and water quality in freshwater systems. Despite its importance for biogeochemical cycling, ecosystem metabolism, as well as drinking water management, we are unaware of any forecasts (i.e., predictions of future conditions with specified uncertainty) that have been developed for DOM dynamics. To improve our understanding of DOM cycling and water quality, we developed the first 1-16-day ahead forecasts of epilimnetic fluorescent DOM (fDOM) in a small eutrophic, dimictic reservoir in Vinton, Virginia, USA. Our forecasts use a time series model based on forecasted water temperature and meteorology that is updated daily from high-frequency fDOM sensor data that are wirelessly transmitted to the cloud for data assimilation. We found that our forecasts were able to capture observed dynamics over a year. Forecast skill varied across seasons, with a higher RMSE in the summer stratified period (RMSE=4.2 quinine sulfate units (QSU)) compared to other seasons (RMSE=3.2 QSU). Over the 1-16-day forecast horizon, forecast RMSE increased from 1.3 to 5.6 QSU. Uncertainty (contributed by model parameters, driver data, model structure, and initial conditions) increased over the forecast horizon (standard deviation increasing from 3.2 to 5.3 QSU over the 16-day horizon), and was greatest during summer. These forecasts can help guide water managers as well as improve our understanding of the predictability of DOM in freshwater ecosystems.
03:00 PM
INTRODUCTION TO VERA, THE VIRGINIA ECOFORECAST RESERVOIR ANALYSIS: A NEW FRESHWATER ECOSYSTEM FORECASTING CHALLENGE (7865)
Primary Presenter: Cayelan Carey, Virginia Tech (cayelan@vt.edu)
As part of the new Virginia Reservoirs Long-Term Research in Environmental Biology (LTREB) program, our team has launched VERA, the Virginia Ecoforecast Reservoir Analysis. VERA is an open forecasting challenge to predict a suite of physical, chemical, and biological variables at two drinking water supply reservoirs before the data are collected. VERA is integrated with the Virginia Reservoirs LTREB state-of-the-art field monitoring program, which serves as a testbed for developing new approaches for data collection, access, and publishing. Our goal is to use forecasts to identify the fundamental predictability of a range of reservoir variables, spanning A (algae) to Z (zooplankton). The forecast guidelines and cyberinfrastructure are based on the National Ecological Observatory Network (NEON) Ecological Forecasting Challenge, with the inclusion of many additional freshwater variables. Participants are welcome to use any modeling approach to forecast any (or all!) of the variables in the VERA Challenge. To date, we have collected >1800 forecasts submitted to VERA for 25 variables, which has 10+ years of data available for model development in the Environmental Data Initiative repository, with more data made available every day in near-real time. At our website, we provide the challenge’s guidelines and forecast submission templates, a dashboard that shows the performance of forecast models in near-real time, and additional resources to learn more about VERA. We invite you to explore our datasets and submit forecasts to the VERA Challenge!
03:15 PM
WHAT CAN WE LEARN FROM 100,000 FORECASTS? A SYNTHESIS OF SUBMISSIONS TO THE AQUATICS THEME OF THE NEON FORECASTING CHALLENGE (8173)
Primary Presenter: Freya Olsson, Virginia Tech (freyao@vt.edu)
The NEON Forecasting Challenge (hereafter Challenge), hosted by the Ecological Forecasting Initiative Research Coordination Network, provides an open platform for participants to engage in ecological forecasting. The goal of the Challenge is to create a community of practice that builds capacity for ecological forecasting of NEON data products and improves our understanding of predictability across ecological variables. Here, we present results of a synthesis of a year of submissions to the aquatics theme of the Challenge, in which participants generated 1-30-day ahead forecasts of surface water temperature (SWT), dissolved oxygen, and chlorophyll-a at 34 stream and lake NEON sites across the USA. Aggregated across the year, 37 models were used to submit >120000 forecasts across all sites and variables. Forecasts were generated using process, machine learning, and statistical models that used a variety of weather covariates and represented uncertainty using different approaches. In general, SWT forecasts that included air temperature as a covariate and were process-based produced the most skillful predictions. Submitted SWT forecasts exhibited greater skill in lakes with more variable seasonal dynamics. Process models were able to capture the important coupling between air temperature and SWT, although empirical and machine-learning models offered valuable alternative methods. Altogether, our synthesis of thousands of forecasts highlights the value of using process-based models for skillful forecasting of lake water temperature.
SS42A - Ecological Forecasting as a Tool for Adaptation and Mitigation in Aquatic Ecosystems
Description
Time: 2:00 PM
Date: 5/6/2024
Room: Meeting Room KL