CAMELS Dataset Research Articles

Enhancing runoff predictions in data-sparse regions through hybrid deep learning and hydrologic modeling

Amidst growing concerns over climate-induced extreme weather events, precise flood forecasting becomes imperative, especially in regions like the Chaersen Basin where data scarcity compounds the challenge. Traditional hydrologic models, while reliable, often fall short in areas with insufficient observational data. This study introduces a hybrid modeling approach that combines the deep learning capabilities of the Informer model with the robust hydrological simulation by the WRF-Hydro model to enhance runoff predictions in such data-sparse regions. Trained initially on the diverse and extensive CAMELS dataset in the United States, the Informer model successfully applied its learned insights to predict runoff in the Chaersen Basin, leveraging transfer learning to bridge data gaps. Concurrently, the WRF-Hydro model, when integrated with The Global Forecast System (GFS) data, provided a basis for comparison and further refinement of flood prediction accuracy. The integration of these models resulted in a significant improvement in prediction precision. The synergy between the Informer’s advanced pattern recognition and the physical modeling strength of the WRF-Hydro significantly enhanced the prediction accuracy. The final predictions for the years 2015 and 2016 demonstrated notable increases in the Nash–Sutcliffe Efficiency (NSE) and the Index of Agreement (IOA) metrics, confirming the effectiveness of the hybrid model in capturing complex hydrological dynamics during runoff predictions. Specifically, in 2015, the NSE improved from 0.5 with WRF-Hydro and 0.63 with the Informer model to 0.66 using the hybrid model, while in 2016, the NSE increased from 0.42 to 0.76. Similarly, the IOA in 2015 rose from 0.83 with WRF-Hydro and 0.84 with the Informer model to 0.87 using the hybrid approach, and in 2016, it increased from 0.78 to 0.92. Further investigation into the respective contributions of the WRF-Hydro and the Informer models revealed that the hybrid model achieved the optimal performance when the contribution of the Informer model was maintained between 60%-80%.

Integration of Generative-Adversarial-Network-Based Data Compaction and Spatial Attention Transductive Long Short-Term Memory for Improved Rainfall–Runoff Modeling

This work presents a novel approach to rainfall–runoff modeling. We incorporate GAN-based data compaction into a spatial-attention-enhanced transductive long short-term memory (TLSTM) network. The GAN component reduces data dimensions while retaining essential features. This compaction enables the TLSTM to capture complex temporal dependencies in rainfall–runoff patterns more effectively. When tested on the CAMELS dataset, the model significantly outperforms benchmark LSTM-based models. For 8-day runoff forecasts, our model achieves an NSE of 0.536, compared to 0.326 from the closest competitor. The integration of GAN-based feature extraction with spatial attention mechanisms improves predictive accuracy, particularly for peak-flow events. This method offers a powerful solution for addressing current challenges in water resource management and disaster planning under extreme climate conditions.

Streamflow prediction in ungauged catchments through use of catchment classification and deep learning

Streamflow prediction in ungauged catchments is a challenging task in hydrological studies. Recently, data-driven models have demonstrated their superiority over traditional hydrological models in predicting streamflow in ungauged catchments. However, previous studies have overlooked the similarities between the training and the target catchments. Therefore, this study explores the role of catchment similarity in regionalization modeling using the publicly available CAMELS dataset. We employed the dynamic time warping-based KMeans (DTW-KMeans) time-series clustering technique to cluster the streamflow data from gauged catchments. We utilized the long short-term memory (LSTM) neural network to construct regional models for different classes of gauged catchment. Additionally, the mapping relationship between gauged catchment classes and static attributes was established using the random forest (RF). By combining the trained RF model with the static attributes of an ungauged catchment, we determined its class and used the corresponding regional LSTM to predict streamflow. To evaluate the effectiveness of the framework, we applied the classification-based regionalization modeling (CRM) and non-classification-based regionalization modeling (NRM) approach for comparison. The results indicate that: (1) The DTW-KMeans-based catchment classification method is generally accurate and reasonable; (2) the complexity of the LSTM model and the number of training catchments should be appropriately matched to improve streamflow prediction; and (3) catchment similarity plays a crucial role in regionalization modeling, the proportion of training catchments with high similarity to ungauged catchments significantly affects prediction results.

Meta-LSTM in hydrology: Advancing runoff predictions through model-agnostic meta-learning

In the field of hydrology, deep learning has become a prevalent tool for runoff simulation. However, the limitations stem from their primary focus on normality, which fails to accurately capture rare events. This study introduces a novel Long Short-Term Memory (LSTM) model, termed Meta-LSTM, which is based on the Model-Agnostic Meta-Learning (MAML) framework. The Meta-LSTM model is capable of dynamically fine-tuning its parameters to ensure effective adaptation to various runoff scenarios. From this Meta-LSTM model, the Meta-enhanced paradigm is subsequently proposed, ensuring effective adaptation to diverse models and datasets. We conducted experiments with various models on the CAMELS and CAMELS-AUS datasets. Compared to conventional deep learning models for runoff simulation, Meta-enhanced models exhibit substantial improvements. Specifically, the Meta-LSTM model reduces the PBIAS from −13.86 % to −5.18 %, increasing the Kling-Gupta Efficiency (KGE) by 7 % on the CAMELS dataset. For the Meta-GR4J-CNN model, the PBIAS decreases by 21 %, and the Nash-Sutcliffe Efficiency (NSE) increases by 2 % on the CAMELS-AUS dataset. This demonstrates the ability of the Meta-enhanced model to more accurately represent observed data without significant additional time costs. Our approach overcomes the problem of uniform training in simulations of complex scenarios and promises to significantly improve the accuracy of future hydrologic studies.

A Novel Strategy for Automatic Selection of Cross‐Basin Data to Improve Local Machine Learning‐Based Runoff Models

AbstractPrevious studies have shown that regional deep learning (DL) models can improve runoff prediction by leveraging large hydrological datasets. However, training a DL regional model using all data without screening may degrade local performance. This study focuses on constructing enhanced local models through the utilization of cross‐basin data. To this end, we propose an approach that employs a novel training strategy to optimize DL model training for specific basins. The approach measures the impact of any one basin's gradient on the loss of the basin of interest, providing insights into the relationships between different basins. The approach was validated using 531 basins from the CAMELS dataset. Results suggest that local performance degradation is a common occurrence in regional models, and imbalanced data are likely to result in a specific pattern dominating the entire regional model. In comparison to a regional model simply trained with all basins, the median Nash‐Sutcliffe efficiency (NSE) for our models is 0.031 higher. In particular, the increase in NSE can exceed 0.2 for some dry basins. Our findings indicate that this novel DL strategy can significantly improve model performance in specific basins using large hydrological datasets, while mitigating local performance loss.

Data reformation – A novel data processing technique enhancing machine learning applicability for predicting streamflow extremes

Hydrologists have been actively exploring the utility of machine learning (ML) models for predicting streamflow. While ML methods have proven to be as accurate as conventional modeling techniques for streamflows well represented in the training set, they continue to lack satisfactory skills for extreme events. In this study, a novel ‘data reformation’ technique is proposed based on the Relative Strength Index (RSI) – a measure of speed and direction of changes in the time series. RSI homogenizes all observations to a constrained 0–100 range, and all ‘out-of-sample’ data in the testing set fall within the space of the training set. Long Short-Term Memory network with an attention mechanism is used to train three ML models using 55,055 events from the CAMELS dataset (670 basins, 1980–2014). Predictions are made for 12,424 events, of which 3,810 are significantly higher than streamflows in the training set. The ML model based on RSI-reformed data exhibits superior performance, as compared to other advanced ML models without data reformation. Peaks up to 15 times larger than those in the training events are accurately predicted, leading to an outperforming model skill for 433 out of 670 catchments. These findings indicate that incorporating a new data reformation technique into the data pre-processing step in ML modeling can enhance the utility of ML models for extreme events. This research encourages further exploration to identify better data reformation methods to enable confident ML predictions.

Enhancing process-based hydrological models with embedded neural networks: A hybrid approach

Deep learning (DL) models have demonstrated exceptional performance in hydrological modeling; however, they are limited by their inability to output untrained hydrological variables and lack of interpretability compared to process-based hydrological models. We propose a hybrid approach that combines the conceptual EXP-Hydro model with embedded neural networks (ENNs), replacing its internal modules while maintaining adherence to hydrological knowledge. The resulting hybrid model can predict untrained hydrological variables without requiring post-processing or pre-training procedures. We tested 15 hybrid models that replace different internal modules across 569 basins in the contiguous United States using the CAMELS dataset. Additional experiments were conducted to generalize hydrological relationships within ENNs and further use them to improve the EXP-Hydro model's performance. Results show that all hybrid scenarios outperform the ordinary EXP-Hydro model, with an optimal median Nash-Sutcliffe efficiency (NSE) of 0.701 in the evaluation period – comparable to state-of-the-art LSTM and conceptual hydrological model featuring an error-correcting post-processor. Reasonable patterns of runoff and snow-related processes are captured by ENNs in respective hybrid models. We further used the runoff (snow-related) pattern to improve the ordinary EXP-Hydro model with median NSE increasing from 0.496 to 0.567 (raising median NSE from 0.601 to 0.677 in snow-influenced region). Our study highlights the potential for using ENNs in enhancing process-based hydrological models' performance while maintaining interpretability within a novel hybrid framework.

A new criteria for determining the best decomposition level and filter for wavelet-based data-driven forecasting frameworks- validating using three case studies on the CAMELS dataset

A new criteria for determining the best decomposition level and filter for wavelet-based data-driven forecasting frameworks- validating using three case studies on the CAMELS dataset

Runoff Prediction Based on Deep Residual Shrinkage Long Short-term Memory Network

Aiming at the lack of sufficient meteorological data for runoff prediction in areas with few data, the accuracy of the data is improved by using soft thresholding under the deep attention mechanism, and a deep residual shrinkage long-short memory network prediction model is constructed combined with the watershed characteristic data. The model is validated by using 134 watersheds with spatial correlations in the CAMELS dataset. According to the experimental findings, the model is more accurate than the LSTM model, HIV model, and mHM model in predicting runoff in regions with little data, with an accuracy of 86%.

Large-sample assessment of varying spatial resolution on the streamflow estimates of the wflow_sbm hydrological model

Abstract. Distributed hydrological modelling moves into the realm of hyper-resolution modelling. This results in a plethora of scaling-related challenges that remain unsolved. To the user, in light of model result interpretation, finer-resolution output might imply an increase in understanding of the complex interplay of heterogeneity within the hydrological system. Here we investigate spatial scaling in the form of varying spatial resolution by evaluating the streamflow estimates of the distributed wflow_sbm hydrological model based on 454 basins from the large-sample CAMELS data set. Model instances are derived at three spatial resolutions, namely 3 km, 1 km, and 200 m. The results show that a finer spatial resolution does not necessarily lead to better streamflow estimates at the basin outlet. Statistical testing of the objective function distributions (Kling–Gupta efficiency (KGE) score) of the three model instances resulted in only a statistical difference between the 3 km and 200 m streamflow estimates. However, an assessment of sampling uncertainty shows high uncertainties surrounding the KGE score throughout the domain. This makes the conclusion based on the statistical testing inconclusive. The results do indicate strong locality in the differences between model instances expressed by differences in KGE scores of on average 0.22 with values larger than 0.5. The results of this study open up research paths that can investigate the changes in flux and state partitioning due to spatial scaling. This will help to further understand the challenges that need to be resolved for hyper-resolution hydrological modelling.

Use of streamflow indices to identify the catchment drivers of hydrographs

Abstract. Time irreversibility or temporal asymmetry refers to the steeper ascending and gradual descending parts of a streamflow hydrograph. The primary goal of this study is to bring out the distinction between streamflow indices directly linked with rising limbs and falling limbs and to explore their utility in uncovering processes associated with the steeper ascending and gradual descending limbs of the hydrograph within the time-irreversibility paradigm. Different streamflow indices are correlated with the rising and falling limbs and the catchment attributes. The key attributes governing rising and falling limbs are then identified. The contribution of the work is on differentiating hydrographs by their time irreversibility features and offering an alternative way to recognize primary drivers of streamflow hydrographs. A series of spatial maps describing the streamflow indices and their regional variability in the Contiguous United States (CONUS) is introduced here. These indices complement the catchment attributes provided earlier (Addor et al., 2017) for the CAMELS data set. The findings of the study revealed that the elevation, fraction of precipitation falling as snow and depth to bedrock mainly characterize the rising limb density, whereas the aridity and frequency of precipitation influence the rising limb scale parameter. Moreover, the rising limb shape parameter is primarily influenced by the forest fraction, the fraction of precipitation falling as snow, mean slope, mean elevation, sand fraction, and precipitation frequency. It is noted that falling limb density is mainly governed by climate indices, mean elevation, and the fraction of precipitation falling as snow; however, the recession coefficients are controlled by mean elevation, mean slope, clay, the fraction of precipitation falling as snow, forest fraction, and sand fraction.

The effect of calibration data length on the performance of a conceptual hydrological model versus LSTM and GRU: A case study for six basins from the CAMELS dataset

We systematically explore the effect of calibration data length on the performance of a conceptual hydrological model, GR4H, in comparison to two Artificial Neural Network (ANN) architectures: Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRU), which have just recently been introduced to the field of hydrology. We implemented a case study for six river basins across the contiguous United States, with 25 years of meteorological and discharge data. Nine years were reserved for independent validation; two years were used as a warm-up period, one year for each of the calibration and validation periods, respectively; from the remaining 14 years, we sampled increasing amounts of data for model calibration, and found pronounced differences in model performance. While GR4H required less data to converge, LSTM and GRU caught up at a remarkable rate, considering their number of parameters. Also, LSTM and GRU exhibited the higher calibration instability in comparison to GR4H. These findings confirm the potential of modern deep-learning architectures in rainfall-runoff modelling, but also highlight the noticeable differences between them in regard to the effect of calibration data length.

Maximizing the Potential of Artificial Intelligence to Perform Evaluations in Ungauged Washbowls

Long short-term memory networks (LSTM) offer precision in the prediction that has never been seen before in ungauged basins. Using k-fold validation, we trained and evaluated several LSTMs in this study on 531 basins from the CAMELS data set. This allowed us to make predictions in basins for which we did not have any training data. The implication is that there is usually sufficient information in available catchment attributes data about similarities and differences between catchment-level rainfall-runoff behaviors to generate out-of-sample simulations that are generally more accurate than current models when operating under ideal (i.e., calibrated) conditions, i.e., when using under idealized conditions. In other words, existing models are generally less accurate when working under idealized conditions than out-of-sample simulations. We found evidence that including physical limits in LSTM models improves simulations, which we believe should be the primary focus of future research on physics-guided artificial intelligence. Putting in place additional physical constraints on the LSTM models.

The proper care and feeding of CAMELS: How limited training data affects streamflow prediction

Accurate streamflow prediction largely relies on historical meteorological records and streamflow measurements. For many regions, however, such data are only scarcely available. Facing this problem, many studies simply trained their machine learning models on the region's available data, leaving possible repercussions of this strategy unclear. In this study, we evaluate the sensitivity of tree- and LSTM-based models to limited training data, both in terms of geographic diversity and different time spans. We feed the models meteorological observations disseminated with the CAMELS dataset, and individually restrict the training period length, number of training basins, and input sequence length. We quantify how additional training data improve predictions and how many previous days of forcings we should feed the models to obtain best predictions for each training set size. Further, our findings show that tree- and LSTM-based models provide similarly accurate predictions on small datasets, while LSTMs are superior given more training data.

Impact of training data size on the LSTM performances for rainfall–runoff modeling

For predicting catchment runoff with data-driven methods, a long historical database of measurements is required. The current study focuses on the assessment of a deep learning model named long short-term memory (LSTM) for rainfall–runoff relationship with different training data size. The developed model has been evaluated on twenty catchments with diverse hydrological conditions obtained from the freely available CAMELS dataset. In order to prove the efficiency of the proposed model for runoff prediction, we test its performances against the traditional feed-forward neural network model. The studied models have been trained with the same input parameters and different size of training data to show the effect of data length on the prediction performances. To this end, the length of training data was varied from 3 to 15 years, while the model was tested on 10 years of data. The results show that the deep LSTM outperforms the traditional model in terms of statistical indicators over different size of training sets. The proposed deep LSTM model can predict runoff with acceptable performances using 9 years of data length in the training procedure, a result that improves when using 12 years. In addition, it has been proven that the deep LSTM model may be efficient even when using small data size (3 years) compared to its benchmarked model which require 9 years for similar results. Thus, the LSTM network is a powerful deep learning model able to learn the behavior of rainfall–runoff relationship with a minimum data length.

Using hydrological and climatic catchment clusters to explore drivers of catchment behavior

Abstract. The behavior of every catchment is unique. Still, we seek for ways to classify them as this helps to improve hydrological theories. In this study, we use hydrological signatures that were recently identified as those with the highest spatial predictability to cluster 643 catchments from the CAMELS dataset. We describe the resulting clusters concerning their behavior, location and attributes. We then analyze the connections between the resulting clusters and the catchment attributes and relate this to the co-variability of the catchment attributes in the eastern and western US. To explore whether the observed differences result from clustering catchments by either climate or hydrological behavior, we compare the hydrological clusters to climatic ones. We find that for the overall dataset climate is the most important factor for the hydrological behavior. However, depending on the location, either aridity, snow or seasonality has the largest influence. The clusters derived from the hydrological signatures partly follow ecoregions in the US and can be grouped into four main behavior trends. In addition, the clusters show consistent low flow behavior, even though the hydrological signatures used describe high and mean flows only. We can also show that most of the catchments in the CAMELS dataset have a low range of hydrological behaviors, while some more extreme catchments deviate from that trend. In the comparison of climatic and hydrological clusters, we see that the widely used Köppen–Geiger climate classification is not suitable to find hydrologically similar catchments. However, in comparison with novel, hydrologically based continuous climate classifications, some clusters follow the climate classification very directly, while others do not. From those results, we conclude that the signal of the climatic forcing can be found more explicitly in the behavior of some catchments than in others. It remains unclear if this is caused by a higher intra-catchment variability of the climate or a higher influence of other catchment attributes, overlaying the climate signal. Our findings suggest that very different sets of catchment attributes and climate can cause very similar hydrological behavior of catchments – a sort of equifinality of the catchment response.

Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets

Abstract. Regional rainfall–runoff modeling is an old but still mostly outstanding problem in the hydrological sciences. The problem currently is that traditional hydrological models degrade significantly in performance when calibrated for multiple basins together instead of for a single basin alone. In this paper, we propose a novel, data-driven approach using Long Short-Term Memory networks (LSTMs) and demonstrate that under a “big data” paradigm, this is not necessarily the case. By training a single LSTM model on 531 basins from the CAMELS dataset using meteorological time series data and static catchment attributes, we were able to significantly improve performance compared to a set of several different hydrological benchmark models. Our proposed approach not only significantly outperforms hydrological models that were calibrated regionally, but also achieves better performance than hydrological models that were calibrated for each basin individually. Furthermore, we propose an adaption to the standard LSTM architecture, which we call an Entity-Aware-LSTM (EA-LSTM), that allows for learning catchment similarities as a feature layer in a deep learning model. We show that these learned catchment similarities correspond well to what we would expect from prior hydrological understanding.

Enhancing Predictions in Ungauged Basins Using Machine Learning to Its Full Potential

In ungauged basins, long short-term memory (LSTM) networks provide unparalleled precision in prediction. Using k-fold validation, we trained and tested various LSTMs on 531 basins from the CAMELS data set, allowing us to make predictions in basins with no training data. The training and test data set contained 30 years of daily rainfall-runoff data from US catchments ranging in size from 4 to 2,000 km2, with aridity indexes ranging from 0.22 to 5.20, and 12 of the 13 IGPB vegetated land cover classes. Over a 15-year validation period, this effectively "ungauged" model was compared to the Sacramento Soil Moisture Accounting (SAC-SMA) model as well as the NOAA National Water Model reanalysis. Each basin's SAC-SMA was calibrated separately using 15 years of daily data. Across the 531 basins, the out-of-sample LSTM exhibited greater median Nash-Sutcliffe Efficiencies (0.69) than either the calibrated SAC-SMA (0.64) or the National Water Model (0.64). (0.58). This means that there is usually enough information in available catchment attributes data about similarities and differences between catchment-level rainfall-runoff behaviors to generate out-of-sample simulations that are generally more accurate than current models under ideal (i.e., calibrated) conditions. We discovered evidence that adding physical restrictions to the LSTM models improves simulations, which we believe should be the focus of future physics-guided machine learning research.

The CAMELS data set: catchment attributes and meteorology for large-sample studies

Abstract. We present a new data set of attributes for 671 catchments in the contiguous United States (CONUS) minimally impacted by human activities. This complements the daily time series of meteorological forcing and streamflow provided by Newman et al. (2015b). To produce this extension, we synthesized diverse and complementary data sets to describe six main classes of attributes at the catchment scale: topography, climate, streamflow, land cover, soil, and geology. The spatial variations among basins over the CONUS are discussed and compared using a series of maps. The large number of catchments, combined with the diversity of the attributes we extracted, makes this new data set well suited for large-sample studies and comparative hydrology. In comparison to the similar Model Parameter Estimation Experiment (MOPEX) data set, this data set relies on more recent data, it covers a wider range of attributes, and its catchments are more evenly distributed across the CONUS. This study also involves assessments of the limitations of the source data sets used to compute catchment attributes, as well as detailed descriptions of how the attributes were computed. The hydrometeorological time series provided by Newman et al. (2015b, https://doi.org/10.5065/D6MW2F4D) together with the catchment attributes introduced in this paper (https://doi.org/10.5065/D6G73C3Q) constitute the freely available CAMELS data set, which stands for Catchment Attributes and MEteorology for Large-sample Studies.