skip to main content

Title: Predicting Water Temperature Dynamics of Unmonitored Lakes With Meta‐Transfer Learning

Most environmental data come from a minority of well‐monitored sites. An ongoing challenge in the environmental sciences is transferring knowledge from monitored sites to unmonitored sites. Here, we demonstrate a novel transfer‐learning framework that accurately predicts depth‐specific temperature in unmonitored lakes (targets) by borrowing models from well‐monitored lakes (sources). This method, meta‐transfer learning (MTL), builds a meta‐learning model to predict transfer performance from candidate source models to targets using lake attributes and candidates' past performance. We constructed source models at 145 well‐monitored lakes using calibrated process‐based (PB) modeling and a recently developed approach called process‐guided deep learning (PGDL). We applied MTL to either PB or PGDL source models (PB‐MTL or PGDL‐MTL, respectively) to predict temperatures in 305 target lakes treated as unmonitored in the Upper Midwestern United States. We show significantly improved performance relative to the uncalibrated PB General Lake Model, where the median root mean squared error (RMSE) for the target lakes is 2.52°C. PB‐MTL yielded a median RMSE of 2.43°C; PGDL‐MTL yielded 2.16°C; and a PGDL‐MTL ensemble of nine sources per target yielded 1.88°C. For sparsely monitored target lakes, PGDL‐MTL often outperformed PGDL models trained on the target lakes themselves. Differences in maximum depth between the source and target were consistently the most important predictors. Our approach readily scales to thousands of lakes in the Midwestern United States, demonstrating that MTL with meaningful predictor variables and high‐quality source models is a promising approach for many kinds of unmonitored systems and environmental variables.

more » « less
Award ID(s):
Author(s) / Creator(s):
 ;  ;  ;  ;  ;  
Publisher / Repository:
DOI PREFIX: 10.1029
Date Published:
Journal Name:
Water Resources Research
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    Ecosystems around the globe are experiencing changes in both the magnitude and fluctuations of environmental conditions due to land use and climate change. In response, ecologists are increasingly using near‐term, iterative ecological forecasts to predict how ecosystems will change in the future. To date, many near‐term, iterative forecasting systems have been developed using high temporal frequency (minute to hourly resolution) data streams for assimilation. However, this approach may be cost‐prohibitive or impossible for forecasting ecological variables that lack high‐frequency sensors or have high data latency (i.e., a delay before data are available for modeling after collection). To explore the effects of data assimilation frequency on forecast skill, we developed water temperature forecasts for a eutrophic drinking water reservoir and conducted data assimilation experiments by selectively withholding observations to examine the effect of data availability on forecast accuracy. We used in situ sensors, manually collected data, and a calibrated water quality ecosystem model driven by forecasted weather data to generate future water temperature forecasts using Forecasting Lake and Reservoir Ecosystems (FLARE), an open source water quality forecasting system. We tested the effect of daily, weekly, fortnightly, and monthly data assimilation on the skill of 1‐ to 35‐day‐ahead water temperature forecasts. We found that forecast skill varied depending on the season, forecast horizon, depth, and data assimilation frequency, but overall forecast performance was high, with a mean 1‐day‐ahead forecast root mean square error (RMSE) of 0.81°C, mean 7‐day RMSE of 1.15°C, and mean 35‐day RMSE of 1.94°C. Aggregated across the year, daily data assimilation yielded the most skillful forecasts at 1‐ to 7‐day‐ahead horizons, but weekly data assimilation resulted in the most skillful forecasts at 8‐ to 35‐day‐ahead horizons. Within a year, forecasts with weekly data assimilation consistently outperformed forecasts with daily data assimilation after the 8‐day forecast horizon during mixed spring/autumn periods and 5‐ to 14‐day‐ahead horizons during the summer‐stratified period, depending on depth. Our results suggest that lower frequency data (i.e., weekly) may be adequate for developing accurate forecasts in some applications, further enabling the development of forecasts broadly across ecosystems and ecological variables without high‐frequency sensor data.

    more » « less
  2. null (Ed.)
    Basin-centric long short-term memory (LSTM) network models have recently been shown to be an exceptionally powerful tool for stream temperature (Ts) temporal prediction (training in one period and making predictions for another period at the same sites). However, spatial extrapolation is a well-known challenge to modeling Ts and it is uncertain how an LSTM-based daily Ts model will perform in unmonitored or dammed basins. Here we compiled a new benchmark dataset consisting of >400 basins across the contiguous United States in different data availability groups (DAG, meaning the daily sampling frequency) with or without major dams and studied how to assemble suitable training datasets for predictions in basins with or without temperature monitoring. For prediction in unmonitored basins (PUB), LSTM produced an RMSE of 1.129 °C and R2 of 0.983. While these metrics declined from LSTM's temporal prediction performance, they far surpassed traditional models' PUB values, and were competitive with traditional models' temporal prediction on calibrated sites. Even for unmonitored basins with major reservoirs, we obtained a median RMSE of 1.202°C and an R2 of 0.984. For temporal prediction, the most suitable training set was the matching DAG that the basin could be grouped into, e.g., the 60% DAG for a basin with 61% data availability. However, for PUB, a training dataset including all basins with data is consistently preferred. An input-selection ensemble moderately mitigated attribute overfitting. Our results indicate there are influential latent processes not sufficiently described by the inputs (e.g., geology, wetland covers), but temporal fluctuations are well predictable, and LSTM appears to be a highly accurate Ts modeling tool even for spatial extrapolation. 
    more » « less
  3. We conducted a macroscale study of 2,210 shallow lakes (mean depth ≤ 3m or a maximum depth ≤ 5m) in the Upper Midwestern and Northeastern U.S. We asked: What are the patterns and drivers of shallow lake total phosphorus (TP), chlorophyll a (CHLa), and TP–CHLa relationships at the macroscale, how do these differ from those for 4,360 non-shallow lakes, and do results differ by hydrologic connectivity class? To answer this question, we assembled the LAGOS-NE Shallow Lakes dataset described herein, a dataset derived from existing LAGOS-NE, LAGOS-DEPTH, and LAGOS-CLIMATE datasets. Response data variables were the median of available summer (e.g., 15 June to 15 September) values of total phosphorus (TP) and chlorophyll a (CHLa). Predictor variables were assembled at two spatial scales for incorporation into hierarchical models. At the local or lake-specific scale (including the individual lake, its inter-lake watershed [iws] or corresponding HU12 watershed), variables included those representing land use/cover, hydrology, climate, morphometry, and acid deposition. At the regional scale (e.g., HU4 watershed), variables included a smaller set of predictor variables for hydrology and land use/cover. The dataset also includes the unique identifier assigned by LAGOS-NE(lagoslakeid); the latitude and longitude of the study lakes; their maximum and mean depths along with a depth classification of Shallow or non-Shallow; connectivity class (i.e., whether a lake was classified as connected (with inlets and outlets) or unconnected (lacking inlets); and the zone id for the HU4 to which each lake belongs. Along with the database, we provide the R scripts for the hierarchical models predicting TP or CHLa (TPorCHL_predictive_model.R), and the TP—CHLa relationship (TP_CHL_CSI_Model.R) for depth and connectivity subsets of the study lakes. 
    more » « less
  4. Lake depth is an important characteristic for understanding many lake processes, yet it is unknown for the vast majority of lakes globally. Our objective was to develop a model that predicts lake depth using map-derived metrics of lake and terrestrial geomorphic features. Building on previous models that use local topography to predict lake depth, we hypothesized that regional differences in topography, lake shape, or sedimentation processes could lead to region-specific relationships between lake depth and the mapped features. We therefore used a mixed modeling approach that included region-specific model parameters. We built models using lake and map data from LAGOS, which includes 8164 lakes with maximum depth (Zmax) observations. The model was used to predict depth for all lakes ≥4 ha (n = 42 443) in the study extent. Lake surface area and maximum slope in a 100 m buffer were the best predictors of Zmax. Interactions between surface area and topography occurred at both the local and regional scale; surface area had a larger effect in steep terrain, so large lakes embedded in steep terrain were much deeper than those in flat terrain. Despite a large sample size and inclusion of regional variability, model performance (R2 = 0.29, RMSE = 7.1 m) was similar to other published models. The relative error varied by region, however, highlighting the importance of taking a regional approach to lake depth modeling. Additionally, we provide the largest known collection of observed and predicted lake depth values in the United States. 
    more » « less
  5. Accurate predictions of water temperature are the foundation for many decisions and regulations, with direct impacts on water quality, fishery yields, and power production. Building accurate broad-scale models for lake temperature prediction remains challenging in practice due to the variability in the data distribution across different lake systems monitored by static and time-series data. In this paper, to tackle the above challenges, we propose a novel machine learning based approach for integrating static and time-series data in deep recurrent models, which we call Invertibility-Aware-Long Short-Term Memory(IA-LSTM), and demonstrate its effectiveness in predicting lake temperature. Our proposed method integrates components of the Invertible Network and LSTM to better predict temperature profiles (forward modeling) and infer the static features (i.e., inverse modeling) that can eventually enhance the prediction when static variables are missing. We evaluate our method on predicting the temperature profile of 450 lakes in the Midwestern U.S. and report a relative improvement of 4\% to capture data heterogeneity and simultaneously outperform baseline predictions by 12\% when static features are unavailable. 
    more » « less