Generating continuous streamflow information through integrated climate-hydrology modeling at fine spatial scales of the order of a few kilometers is often challenged by computational costs associated with running high-resolution (HR) climate models. To address this challenge, the present study explores deep learning approaches to generate HR streamflow information from that at low resolution (LR), based on runoff generated by climate models. Two sets of daily streamflow simulations spanning 10 years (2011–2020), at LR (50 km) and HR (5 km), for the Ottawa River basin, Canada, are employed. The proposed deep learning model is trained using upscaled features derived from LR streamflow simulation for the 2011–2018 period as input and the corresponding HR streamflow simulation as the target; data for 2019 are used for validation. The model estimates for the year 2020, when compared with unseen HR data for the same year, suggest good performance, with differences in monthly mean values for different accumulation area categories in the −0.7–5% range and correlation coefficients for streamflow values for the same accumulation area categories in the 0.92–0.96 range. The developed framework can be ported to other watersheds for generating similar information, which is required in climate change adaptation studies.

  • A physically consistent deep learning framework for generating high-resolution streamflow information from low-resolution streamflow sequences.

  • Enhancing the value of climate-hydrology integrated modeling outputs to support local-scale adaptation strategies.

  • An adaptable generalized framework that can be applied to other regions and climate-hydrology system outputs.

Information about continuous streamflow magnitudes within a watershed is often required to support several water resources development and management-related activities, including water supply planning, sediment and nutrient management, hydrokinetic resource assessment, waste load allocations, inundation mapping, erosion risk management, etc. As most of the world's watersheds are sparsely gauged, this information at various spatial resolutions is often generated by setting up watershed-level deterministic type lumped or distributed hydrologic models (Singh & Frevert 2002). Lumped models are easy to setup while distributed models require a considerable amount of data and experience and could be quite complex to setup. Streamflow generation within a watershed is a physical problem and when physically based integrated climate-hydrology models are opted to deal with this problem in the context of climate change assessment and adaptation planning, streamflow generation at fine spatial resolutions could become very challenging due to high computational costs involved (e.g., Sushama et al. 2004; Diro & Sushama 2019; Teufel & Sushama 2021, 2022). To address this challenge, rapidly emerging machine learning approaches could be quite useful (e.g., Kratzert et al. 2019; Stengel et al. 2020; Wu et al. 2021; Teufel et al. 2023). This study focuses on the development and testing of a deep learning framework, based on image super-resolution techniques, for generating high-resolution (HR) streamflow information for the Ottawa River basin in conjunction with the physically based regional climate model global environmental multiscale (GEM) of Environment and Climate Change Canada (Côté et al. 1998). The aim is to estimate streamflow at 5 km resolution (i.e., HR) for the Ottawa River basin from the corresponding streamflow information at 50 km resolution (i.e., low resolution (LR)), by combining physical understanding of runoff generation dynamics within a watershed with modern machine learning architectures.

Machine learning and hybrid modeling approaches, including several variants of deep learning, are becoming increasingly popular in solving applied problems in many disciplines of science and engineering. Among these are streamflow and flood forecasting (e.g., Badrzadeh et al. 2013; Ding et al. 2020; Pradhan et al. 2020; Burgan 2022; Girihagama et al. 2022; Tounsi et al. 2023), HR inundation mapping (Carreau & Naveau 2023), climate downscaling (e.g., Pan et al. 2019; Wang et al. 2021), weather forecasting (e.g., Chen et al. 2019; Zhang et al. 2022), water quality modeling (e.g., Zhi et al. 2021), remote sensing (e.g., Paul & Nagesh Kumar 2018; Tang et al. 2022; Yang et al. 2022) and others. Additional information on the range of studies and discussions on several aspects of machine learning, along with its advantages and limitations, can be found in an earlier review compiled by Govindaraju (2000) and in more recent reviews by Shen (2018), Sit et al. (2020) and Camps-Valls et al. (2021).

Some advancements along the lines of this paper that have appeared in the literature are discussed here. Recently, Sdraka et al. (2022) explored different deep learning models and evaluation metrics for downscaling remote sensing images, i.e., for increasing the resolution of remotely sensed images. These models were developed based on convolutional neural network (CNN) and generative adversarial network (GAN) architectures, with different upsampling techniques, such as preupsampling and postupsampling. In the preupsampling technique, the upsampling is done to generate an image in HR from LR before applying convolutional layers, while in the postupsampling technique, it is done after. Similarly, deep learning architectures were used by Wu et al. (2021) to emulate HR images from LR images by employing a GAN-based architecture. They emulated air temperature and dew point fields at a resolution of 250 m from 2.5 km resolution. Sajjadi et al. (2018) proposed a frame-recurrent model for generating HR images by combining two different deep learning models, of which the first is encoder-decoder CNN and the second one consisted of convolutional layers and residual blocks, operating in LR space with a postupsampling to emulate HR images. Teufel et al. (2023) modified this framework to study a physical climate-hydrology problem using preupsampling, making the model architecture independent of the scaling factor, and replacing the first deep learning model component with known information about the wind field, to simulate intense precipitation events at a resolution of 250 m from a simulation at 2.5 km. Integration of wind fields within the deep learning architecture provided a mechanism for bringing in physical bases within the machine learning model. Other recent super-resolution applications using deep residual network in the area of climate science include Wang et al. (2021) and Höhlein et al. (2020), where temperature and precipitation were downscaled from 50 km/36 km to 4 km and wind from 31 to 9 km, respectively.

In this study, a deep learning model based on the residual neural network (ResNet) architecture is proposed for generating HR streamflow information from that at LR for the Ottawa River basin in Canada. To the knowledge of the authors, such a problem has not been dealt with in the hydrologic literature and hence this contribution will advance our understanding of streamflow generation dynamics at HR based on physically consistent machine learning architecture and will therefore make a novel contribution to the literature. Machine learning approaches are data-driven and do not require solving complex analytical formulations as in the case of physically based climate-hydrology integrated modeling approaches and thus will be beneficial in reducing the computational costs considerably once the parameters of the architectures involved are properly trained and validated.

The rest of this paper is organized as follows. Section 2 describes methodological aspects of both the climate and routing models and the deep learning architecture considered in the study. Development of a baseline benchmark and deep learning experiments to help optimize model structure are described in Section 3, along with detailed results of the study. Relevant scientific discussions are provided in Section 4. Finally, Section 5 lists main conclusions, contributions and limitations of the study, as well as future research directions.

Streamflow data and study area

This study considers streamflow generated at 50 km resolution based on runoff from a regional climate model – GEM (Côté et al. 1998) – simulation at the same resolution for the 2011–2020 period, driven by European Centre for Medium-Range Weather Forecasts’ (ECMWF) ERA-Interim (Hersbach et al. 2020), at the lateral boundaries over a pan-Arctic domain (Figure 1). The land part of GEM is represented using the Canadian Land Surface Scheme (CLASS; Verseghy 2011) and the cell-to-cell routing scheme WATROUTE (Sushama et al. 2004; Poitra et al. 2011; Huziy & Sushama 2016) is used to generate streamflow from GEM/CLASS simulated runoff. The routing scheme solves the water balance equation at each grid cell and relates water storage to streamflow using Manning's equation. Flow directions, channel lengths and slopes required by the routing scheme are derived from the HydroSHEDS database (Lehner et al. 2008). Additional information about GEM and various parametrization schemes can be found in Diro & Sushama (2019).
Figure 1

(a) Left panel: Experimental domain of GEM simulation (in red), with every tenth grid point shown. In blue is the study region. Right panel: Zoomed view of the study region (in blue) with the Ottawa River basin shown in black. (b) Accumulated drainage area (AA) at 50 km LR and 5 km HR and location of selected study points (P1 to P6) in the river basin are respectively shown in left, center and right panels.

Figure 1

(a) Left panel: Experimental domain of GEM simulation (in red), with every tenth grid point shown. In blue is the study region. Right panel: Zoomed view of the study region (in blue) with the Ottawa River basin shown in black. (b) Accumulated drainage area (AA) at 50 km LR and 5 km HR and location of selected study points (P1 to P6) in the river basin are respectively shown in left, center and right panels.

Close modal

Although the streamflow simulations are available over the entire pan-Arctic domain at 50 km resolution, for the purposes of this study, streamflow simulations covering only the study region, i.e., the Ottawa River basin, are considered; the Ottawa River, which is about 1,200 km long, is one of the major tributaries of the Saint-Lawrence drainage system with a total drainage area of 146,000 km2, of which 65% (35%) is located in the province of Quebec (Ontario), Canada. Furthermore, streamflow at 5 km resolution for the Ottawa River basin are also generated using WATROUTE based on the same GEM simulation. The 50 and 5 km streamflow data will respectively be referred to as LR and HR data hereafter.

Given the very large range of streamflow values in the Ottawa River basin, the generated LR and HR daily streamflow values, transformed into specific flows by dividing the streamflow by accumulated drainage area (AA), are used for developing and evaluating the deep learning framework, discussed in the next section, considering HR values as the target for training, validation and testing purposes. A comparison of the drainage information in terms of AA at both LR and HR for the Ottawa River basin is also provided in Figure 1, which clearly demonstrates the added value of HR. Also shown in this figure are the locations of six selected points (i.e., P1, P2, P3, P4, P5 and P6) on the main stem and tributaries of the Ottawa River, which have been selected for detailed analysis to understand streamflow generation dynamics within the river basin.

Deep learning framework

The proposed deep learning framework used in this study is shown in Figure 2. The conventional neural network architecture typically comprises elements such as convolutional layers, activation functions, pooling functions, filters and fully connected neurons. However, as the depth of the neural network increases, a critical challenge known as the vanishing gradient issue arises during backpropagation. As the network depth increases, the gradients of the loss function corresponding to the network parameters become exceedingly small, making it challenging to update weights for previous layers, subsequently impacting model performance. To overcome this, He et al. (2016) introduced the ResNet architecture, which significantly advanced the field of deep learning and helped solve classification and regression problems more efficiently. This architecture is adopted in this study. The key concept underlying the ResNet involves organization of network layers into distinct blocks, allowing data to seamlessly flow within and also around each block, referred to as skip connection. The skip connections play a pivotal role by combining the input to a given residual block with the output derived from that same block through addition or concatenation, effectively amalgamating the input tensor and output tensor. However, it is noteworthy that opting for concatenation results in large tensors. The consequential increase in the number of parameters becomes particularly significant when a greater number of residual blocks is employed. Hence, in this study, element-wise addition is performed with the output from the residual block and the input flowing to a residual block. The ResNet architecture facilitates a more efficient flow of gradients during backpropogation. A similar architecture was also used by Teufel et al. (2023) to simulate HR extreme precipitation events from LR data. The hyperparameters of the proposed architecture to be tuned are the batch size, kernel size, number of residual blocks and filters used in each block. The learnable parameters are the weights and biases corresponding to each convolutional layer. For obtaining optimum weights, the error between the reconstructed and real images is minimized.
Figure 2

A schematic depiction of the framework developed for generating HR streamflow information from LR streamflow data using deep learning approach.

Figure 2

A schematic depiction of the framework developed for generating HR streamflow information from LR streamflow data using deep learning approach.

Close modal
The input to the deep learning model consists of (a) time-dependent and (b) time-independent features. The time-dependent features represent LR daily specific flows (i.e., LR streamflow divided by LR AA) for four different AA ranges (102–103 km2; 103–104 km2; 104–105 km2 and >105 km2) for the training, validation and testing periods, i.e., for 10 years (from January 1, 2011 to December 31, 2020), upscaled to HR using a bilinear interpolation. Thus, overall, the time-dependent features have up to four channels of 3,650 images (ignoring the impact of the leap year), of 190 × 230 pixel format at HR (5 km) obtained from upscaling the corresponding 19 × 23 pixel images at LR (50 km); examples of four channels after upscaling are shown in Figure 3 for clarity.
Figure 3

Time-dependent specific flow (i.e., streamflow/AA) features at LR (left column) and the corresponding upscaled features (center column). Binary values for the HR AA are shown in the right column. Rows 1 to 4 correspond to AA ranges 102–103 km2, 103–104 km2, 104–105 km2 and >105 km2.

Figure 3

Time-dependent specific flow (i.e., streamflow/AA) features at LR (left column) and the corresponding upscaled features (center column). Binary values for the HR AA are shown in the right column. Rows 1 to 4 correspond to AA ranges 102–103 km2, 103–104 km2, 104–105 km2 and >105 km2.

Close modal

Of the five time-independent inputs, four are obtained from HR AA values falling in the following ranges of AA: (1) 102–103 km2, (2) 103–104 km2, (3) 104–105 km2 and (4) >105 km2; the fifth one corresponds to the lake mask at HR, with 0 reflecting lakes and 1 otherwise. All inputs, both time-dependent and time-independent, are input to the deep learning model. The objective of the deep learning model (Figure 2) in the overall framework is to minimize the error between the predicted and the target HR images.

The Adam optimizer is used in the study. This optimizer, first proposed by Kingma & Ba (2014), is a stochastic gradient descent method which is based on adaptive estimation of first-order and second-order moments. The rectified linear unit is used as the activation function for all convolutional layers.

Deep learning experiments and optimization of the model structure

Prior to employing the deep learning model for estimating specific flow and streamflow in an HR setting, targeted experiments are conducted to optimize model structure, selection of hyperparameters and choice of an appropriate loss function and input features. For this purpose, upscaled LR-specific flows for the years 2011–2018 are used as input and the corresponding HR-specific flows as the target. Data for 2019 are used for validation purposes.

Ten different combinations of inputs and loss functions are experimented using a constant set of hyperparameters in order to identify the best input combination and a reasonable loss function. These combinations are shown in Table 1, labeled as config_1 to config_10. Among the four time-dependent inputs, a single range of AA encompassing all possible AA values is considered for config_1, while for the remaining nine configurations, three ranges of AA are used. For config_1 and config_4, no time-independent variable is used, while for config_8, config_9 and config_10, all independent variables are used, including information about lakes. For config_5, config_6 and config_7, only information related to AA is used. For the remaining two configurations (i.e., config_2 and config_3), three of the five time-independent variables are used. Regarding the loss function, three different functions are evaluated. One of these being just the mean squared error (MSE). The other two also involve MSE, but are weighted by the logarithm of AA. One of these options gives a weight of 0 to cells where there are lakes. It is worth noting that the difference between config_6 and config_7 and that between config_8, config_9 and config_10 is only the form of the loss function. Each of the above configurations is run for 1,000 epochs and the errors are found to be between 0.0002 and 0.0003 for most of the configurations. By analyzing the results after 1,000 epochs visually and also based on error measures, it is found that the performance of the first configuration is the least accurate, suggesting that multiple inputs and weighted loss functions are the most suitable choices. Configurations config_7 and config_10 are the ones that produced minimal errors, lower than 0.00005. These two configurations use the loss functions that assign zero weight to cells containing lakes. Visually, the images generated by these two configurations resemble better the real HR images. For this study, config_7 is selected for additional analyses. The loss function thus considered is the MSE weighted with log10(AA) values from HR. As a consequence of logarithmic transformation of AA values, the larger the AA value, the higher the weight. The weights thus roughly range from 0 to 6 since the maximum AA is of the order of 106 km2. Figure S1 (Supplementary Material) shows the map of weights for the Ottawa River basin. The cells with lakes are assigned zero weights as no streamflow generation is assumed for these cells.

Table 1

Experimental model configurations: different inputs and loss functions considered in the study

ConfigurationTime-dependent features (flow/AA) for four ranges of AA (km2)
Time-independent features for four ranges of AA (km2) and presence/absence of lakes
Loss functions (MSE and AA weighted MSE, wMSE)
0–106103–104104–105> 105102–103103–104104–105> 105LakeswMSEwMSE + LakesMSE
config_1 ✓           ✓ 
config_2  ✓ ✓ ✓  ✓ ✓ ✓    ✓ 
config_3  ✓ ✓ ✓  ✓ ✓ ✓  ✓   
config_4  ✓ ✓ ✓      ✓   
config_5  ✓ ✓ ✓ ✓ ✓ ✓ ✓  ✓   
config_6  ✓ ✓ ✓ ✓ ✓ ✓ ✓  ✓   
config_7  ✓ ✓ ✓ ✓ ✓ ✓ ✓   ✓  
config_8  ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓   ✓ 
config_9  ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓   
config_10  ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓  ✓  
ConfigurationTime-dependent features (flow/AA) for four ranges of AA (km2)
Time-independent features for four ranges of AA (km2) and presence/absence of lakes
Loss functions (MSE and AA weighted MSE, wMSE)
0–106103–104104–105> 105102–103103–104104–105> 105LakeswMSEwMSE + LakesMSE
config_1 ✓           ✓ 
config_2  ✓ ✓ ✓  ✓ ✓ ✓    ✓ 
config_3  ✓ ✓ ✓  ✓ ✓ ✓  ✓   
config_4  ✓ ✓ ✓      ✓   
config_5  ✓ ✓ ✓ ✓ ✓ ✓ ✓  ✓   
config_6  ✓ ✓ ✓ ✓ ✓ ✓ ✓  ✓   
config_7  ✓ ✓ ✓ ✓ ✓ ✓ ✓   ✓  
config_8  ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓   ✓ 
config_9  ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓   
config_10  ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓  ✓  

wMSE stands for weighted mean squared error

For the analysis of hyperparameters, eight different choices are evaluated. These are indicated as config_101 to config_108 in Table 2 and the results of this investigation are given in Table S1, along with the time required to train the model, and Figure S2 (Supplementary Material). Here, the goal is to analyze, which combination(s) of hyperparameters would considerably improve the accuracy of the model. That is, whether increasing the number of channels in the inner layers or increasing the number of layers of the model would improve the accuracy of the modeled images. With the increase in the number of channels in the inner layers, the number of trainable parameters also increases and, consequently, the time to train the model increases as well. Although the time to train the model is much longer, the quality of results obtained with config_108 is much better than that obtained with other configurations; config_102 differs from config_101 only by batch size. By increasing the batch size, the tensors become larger resulting in only a small improvement in time, but with almost similar accuracy in terms of loss function. Configurations config_104 and config_105 show a small improvement in the model accuracy compared to config_101, i.e., increasing the number of channels in the inner layers lead to better accuracy. The modeling results presented in the remainder of this paper are based on config_108. It must be noted that a post-processing step is necessary to convert the model output to streamflow by multiplying it by AA values from the HR domain.

Table 2

Experimental configurations evaluated for optimizing the deep learning model structure

ConfigurationNumber of layersNumber of channelsBatch sizeEstimated size (Megabyte)
config_101 12 16 (10 layers) 10 146 
config_102 12 16 (10 layers) 50 146 
config_103 12 32 (10 layers) 10 288 
config_104 24 16 (10 layers)
32 (10 layers) 
10 455 
config_105 46 8 (10 layers)
16 (22 layers)
32 (12 layers) 
10 670 
config_106 12 64 (10 layers) 10 579 
config_107 12 128 (10 layers) 10 1,159 
config_108 12 256 (10 layers) 10 2,329 
ConfigurationNumber of layersNumber of channelsBatch sizeEstimated size (Megabyte)
config_101 12 16 (10 layers) 10 146 
config_102 12 16 (10 layers) 50 146 
config_103 12 32 (10 layers) 10 288 
config_104 24 16 (10 layers)
32 (10 layers) 
10 455 
config_105 46 8 (10 layers)
16 (22 layers)
32 (12 layers) 
10 670 
config_106 12 64 (10 layers) 10 579 
config_107 12 128 (10 layers) 10 1,159 
config_108 12 256 (10 layers) 10 2,329 

Model performance assessment

To assess the value of the deep learning model in simulating HR streamflow, a baseline reference is generated from LR data, schematically shown in Figure 4. Streamflow values from LR are divided by the corresponding LR AA values to obtain LR-specific flows, which are then upscaled to HR using a bilinear interpolation in a similar manner as shown in Figure 3. The upscaled specific flows are then multiplied by HR AA values derived from the HR domain. Finally, the streamflow image is filtered depending on the range of HR AA. This baseline is used for benchmarking model performance and uncovering the value of the deep learning model as presented below.
Figure 4

Schematic diagram showing generation of baseline HR streamflow sequences.

Figure 4

Schematic diagram showing generation of baseline HR streamflow sequences.

Close modal

To test the model on unseen data, LR-specific streamflow for the year 2020 is used as time-dependent input to generate the corresponding HR streamflow, which is used in a variety of assessments presented below, i.e., the behavior of overall streamflow distributions, monthly averaged flows and monthly maximum flows. From a hydrological viewpoint, streamflow dynamics, including the magnitude and seasonal distributions, are very much tied with accumulated drainage areas, starting from the headwaters to the downstream points. Given this intuition, the performance of the model is assessed with respect to each of the four HR ranges/categories of AA discussed before in the methodology section. In all assessments and comparisons, the information and statistics derived from the baseline act as the reference.

Overall streamflow distributions

Frequency distributions of streamflow for each of the four categories of AA for the baseline, deep learning model predictions and the target HR data are compared in Figure 5. The streamflow values predicted by the model for the AA categories 1 to 3 are closer to the target values than the baseline as reflected from the respective distribution comparisons. Regarding the AA category 4, the overall distribution of predicted values is relatively closer to that of the target values for larger streamflow magnitudes, but with lower frequencies for certain magnitudes. However, the overall distribution is better simulated than the baseline. Quantitative performance statistics, provided in Table S2 and Figure S4 (Supplementary Material), further support the above-noted observation, i.e., predicted streamflow values are more similar to the target HR values than the baseline.
Figure 5

Comparison of frequency distributions derived from the baseline, model predictions and the target HR data for four AA categories: 102103 km2 (top left panel), 103104 km2 (top right panel), 104105 km2 (bottom left panel) and >105 km2 (bottom right panel).

Figure 5

Comparison of frequency distributions derived from the baseline, model predictions and the target HR data for four AA categories: 102103 km2 (top left panel), 103104 km2 (top right panel), 104105 km2 (bottom left panel) and >105 km2 (bottom right panel).

Close modal

Monthly mean daily flows

Monthly mean daily streamflow values for each month of the year are calculated from the baseline, model predictions and the target HR data and compared in Figure 6. For AA categories 1 and 2, mean values are better captured by the model than the baseline and are in close agreement with the target values nearly for all months. Regarding the AA category 3, the model predicted mean daily streamflow values are similar to the target values for most of the months. However, for certain months (e.g., October, November and December), baseline values are more similar to the target values than the model predictions. Regarding the AA category 4, clearly, the model predicted mean values are more similar to the target values when compared with the baseline.
Figure 6

Comparisons of monthly mean daily flows derived from the baseline, model predictions and the target HR data for four AA categories: 102103 km2 (top left panel), 103104 km2 (top right panel), 104105 km2 (bottom left panel) and >105 km2 (bottom right panel).

Figure 6

Comparisons of monthly mean daily flows derived from the baseline, model predictions and the target HR data for four AA categories: 102103 km2 (top left panel), 103104 km2 (top right panel), 104105 km2 (bottom left panel) and >105 km2 (bottom right panel).

Close modal
Spatial comparisons of the deep learning model predicted and HR target values of monthly mean daily flows for all 12 months of the year are provided in Figure 7, along with the respective differences. It is interesting to note that apart from the main stem and major tributaries of the Ottawa River, the differences are small and the spatial patterns of streamflow are well captured by the model. For larger river segments, the disparity is due to the accumulation of smaller differences. Thus, the deep learning model has demonstrated high degree of fidelity in reproducing spatial patterns of monthly mean daily flows for the entire basin.
Figure 7

Spatial patterns of monthly mean daily flows (m3/s) derived from model predictions and HR target data and their differences (m3/s) for all months for the year 2020 for the Ottawa River basin.

Figure 7

Spatial patterns of monthly mean daily flows (m3/s) derived from model predictions and HR target data and their differences (m3/s) for all months for the year 2020 for the Ottawa River basin.

Close modal

Monthly high flows

Similar to the monthly mean daily flow comparisons, monthly maximum flows for each month of the year are obtained from the baseline, deep learning model predictions and the target HR data and are compared in Figure 8. For the AA category 1, model predicted flows are in better agreement with the target flows than the baseline for two-third of the months. For the AA categories 2 and 3, mixed behavior can be noticed, i.e., model predicted monthly maximum flows are larger/smaller than the target values for certain months of the year. For the AA category 4, predicted monthly maximum flows are closer to the target values than the baseline, except December, but are slightly over-predicted. Over-predicted high flows noted for certain months are potentially due to high inter-annual variability (Figure S3 in the Supplementary Material). It is worth noting that the absolute maximum values may not occur at the same spatial location in the baseline, model predictions and HR data, however, this aspect is not verified in the study. Additionally, it is also important to note that much longer datasets than just one year data used in this proof-of-concept type study are required to arrive at definitive conclusions, especially for high flows.
Figure 8

Comparisons of monthly maximum daily flows derived from the baseline, model predictions and the HR target data for four AA categories: 102103 km2 (top left panel), 103104 km2 (top right panel), 104105 km2 (bottom left panel) and >105 km2 (bottom right panel).

Figure 8

Comparisons of monthly maximum daily flows derived from the baseline, model predictions and the HR target data for four AA categories: 102103 km2 (top left panel), 103104 km2 (top right panel), 104105 km2 (bottom left panel) and >105 km2 (bottom right panel).

Close modal

Comparisons of monthly mean daily flows at selected points within the river basin

To assess how well the model is able to predict flows at selected tributaries within the Ottawa River basin, monthly mean daily flows for each month of the year are obtained for the baseline, deep learning model predictions and the target HR information at six selected points, which are shown in Figure 1. The comparisons of these estimates are shown in Figure 9. Overall, the results shown in this figure suggest that the simulated monthly mean daily flows are in good agreement with the target HR flows, compared to the baseline, despite the smaller deviations for certain points, e.g., Point 1 and Point 6. Such deviations are expected since only one year of simulated data has been used in these comparisons to elucidate this proof-of-concept study. Although computationally very expensive, future studies should endeavor to consider longer time periods.
Figure 9

Comparison of monthly mean daily flows for six selected points located in the Ottawa River basin.

Figure 9

Comparison of monthly mean daily flows for six selected points located in the Ottawa River basin.

Close modal

This proof-of-concept study aimed at developing a deep learning modeling framework to simulate HR streamflow information from LR streamflow data produced within a climate-hydrology integrated setting using a regional climate model. For training and testing the model, targeted HR streamflow information is also generated using the same climate-hydrology integrated setting. The model development is undertaken following a systematic approach comprising of training, validation and testing phases. The results (cf. Section 3) suggest that the proposed framework is a reasonable choice for generating HR information from coarse resolution data.

The strategy adopted for generating HR streamflow sequences from LR data based on image processing techniques has its origin in video super-resolution, which has been applied previously in remote sensing, medical sciences and earth sciences (e.g., Trinh et al. 2014; Dong et al. 2016; Caballero et al. 2017; Liu et al. 2017; Sajjadi et al. 2018). The image processing super-resolution technique for HR streamflow data generation is still a new concept, which is expected to mature in the coming years. To our knowledge, this is the first study to apply the super-resolution technique for HR streamflow generation from LR streamflow data.

From a hydrological perspective, streamflow generation within a watershed is a nonlinear physical process, with complex streamflow dynamics that are highly tied with the drainage patterns when the whole watershed is viewed as a single entity from the headwaters to the downstream points. To bring in this physical intuition in the machine learning model structure, accumulated drainage area and specific flow concepts, which are well established in the field of hydrology, are introduced in the model structure. The relatively smaller range of specific flows compared to that of streamflow enabled efficient training of the model and ensured physical realism too. From a machine learning perspective, it is found that, increasing the number of channels helped improve the accuracy of predicted flows compared to increasing the batch size. Although the number of trainable parameters increased in the former case, a structure containing relatively higher number of channels is retained in the study.

It must be noted that generating HR streamflow information from regional and global climate models is quite challenging and sometimes infeasible due to high computational costs involved. The deep learning modeling framework developed in this study is quite flexible and therefore can be ported to other watersheds and regions for generating HR streamflow information from LR data originating from climate models and hence can be an efficient approach for enhancing the utility of climate model outputs. The main limitation of this study is that a single year is considered in the testing phase, which limits assessing model ability in capturing high flows particularly. Due to computational resource constraints, longer simulations are not attempted for this proof-of-concept study.

Future research could consider other deep learning architectures and additional physical features in the model structure and additional super-resolution applications to gain further insights. It must be noted that the goal here is to transform streamflow generated at LR to HR. The accuracy of HR data generated will therefore rely on the quality of the LR data as streamflow observation data is not integrated in this framework as this was outside the scope of the current study, but is worth exploring in the future.

The evaluation of the deep learning model on unseen independent data for generating HR streamflow information from that of LR for the Ottawa River basin is performed considering (1) domain-specific overall streamflow distributions, (2) monthly mean daily flows, (3) monthly high flows and (4) comparisons of monthly mean daily flows at six selected points within the river basin. These evaluations are conducted with reference to four categories of AA (i.e., 1: 102–103 km2; 2: 103–104 km2; 3: 104–105 km2 and 4: >105 km2), representing increasing values of drainage area, and also based on the entire river basin. From the various results and modeling aspects discussed in this paper, the following main conclusions can be drawn.

  • Overall distribution functions of streamflow values for the entire domain, categorized with respect to various ranges of AA, are in good agreement with those of the target data compared to the baseline. Quantitative performance statistics in terms of mean, coefficient of variation and coefficient of skewness for the same AA categories further confirm this assertion.

  • The deep learning model is able to better simulate the monthly mean daily streamflow values for each month of the year for AA categories 1, 2 and 4, while for the AA category 3, mixed performance is noticed, with baseline values more similar to the target values than the model predictions for certain months of the year. With respect to spatial patterns of monthly mean daily flows, differences between model predictions and the target values are small, except for the major tributaries of the Ottawa River. Thus, given these results, the image super-resolution deep learning techniques are worthwhile to consider for hydrological and climatological applications.

  • It is found that for the AA category 1, model predicted monthly maximum flows are in better agreement with the target flows than the baseline for two-third of the months. For the AA categories 2 and 3, model predicted monthly maximum flows are larger/smaller than the target values for certain months of the year. For the AA category 4, predicted monthly maximum flows are closer to the target values than the baseline. However, to enhance confidence in these observations, results from longer data sets will be required, which are not attempted in this study.

  • Assessment of model performance at six selected points along the various tributaries within the Ottawa River basin, representing a range of drainage areas, suggests that the simulated monthly mean daily flows are in good agreement with the target HR flows, compared to the baseline, despite smaller deviations for certain points. Since only one year of data has been used in these comparisons, such deviations are not surprising.

Finally, as mentioned earlier, this study is the first attempt for generating HR streamflow information from LR streamflow based on image super-resolution, driven by machine learning architectures and employing physically consistent climate-hydrology integrated modeling, often required to support climate change assessment and adaptation-related projects. From that perspective, this study makes a unique contribution to the broad literature on hydrology and climate change science. Given the limited evaluations conducted in this study, more in-depth assessment will be required using longer data sets than the one year of unseen data used in this proof-of-concept type study.

This research was funded by the Canadian Space Agency (Grant 21SUESDFIM), Natural Sciences and Engineering Research Council of Canada, Trottier Institute for Sustainability in Engineering and Design and McGill Sustainability Systems Initiative. The GEM simulations and deep learning experiments considered in this study were performed on the supercomputer managed by the Digital Research Alliance of Canada and Calcul Québec. The helpful comments from three anonymous reviewers are much appreciated.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.

Badrzadeh
H.
,
Sarukkalige
R.
&
Jayawardena
A. W.
2013
Impact of multi-resolution analysis of artificial intelligence models inputs on multi-step ahead river flow forecasting
.
Journal of Hydrology
507
,
75
85
.
doi:10.1016/j.jhydrol.2013.10.017
.
Burgan
H. I.
2022
Comparison of different ANN (FFBP, GRNN, RBF) algorithms and multiple linear regression for daily streamflow prediction in Kocasu River, Turkey
.
Fresenius Environmental Bulletin
31
(
5
),
4699
4708
.
Camps-Valls
G.
,
Tuia
D.
,
Zhu
X. X.
&
Reichstein
M.
2021
Deep Learning for the Earth Sciences: A Comprehensive Approach to Remote Sensing
.
John Wiley & Sons, Climate Science and Geosciences
,
New York
.
Caballero
J.
,
Ledig
C.
,
Aitken
A.
,
Acosta
A.
,
Totz
J.
,
Wang
Z.
&
Shi
W.
2017
Real-time video super-resolution with spatio-temporal networks and motion compensation
.
 In Proceedings of the IEEE conference on computer vision and pattern recognition
,
4778
-
4787
.
Chen
R.
,
Wang
X.
,
Zhang
W.
,
Zhu
X.
,
Li
A.
&
Yang
C.
2019
A hybrid CNN-LSTM model for typhoon formation forecasting
.
GeoInformatica
23
,
375
396
.
Côté
J.
,
Gravel
S.
,
Méthot
A.
,
Patoine
A.
,
Roch
M.
&
Staniforth
A.
1998
The operational CMC–MRB global environmental multiscale (GEM) model. Part I: Design considerations and formulation
.
Monthly Weather Review
126
(
6
),
1373
1395
.
Ding
Y.
,
Zhu
Y.
,
Feng
J.
,
Zhang
P.
&
Cheng
Z.
2020
Interpretable spatio-temporal attention LSTM model for flood forecasting
.
Neurocomputing
403
,
348
359
.
Dong
C.
,
Loy
C. C.
,
He
K.
&
Tang
X.
2016
Image super-resolution using deep convolutional networks
.
IEEE Transactions on Pattern Analysis and Machine Intelligence
38
(
2
),
295
307
.
Girihagama
L.
,
Khlaiq
M. N.
,
Lamontagne
P.
,
Perdikaris
J.
,
Roy
R.
,
Sushama
L.
&
Elshorbagy
A.
2022
Streamflow modeling and forecasting for Canadian watersheds using LSTM networks with attention mechanism
.
Neural Computing and Applications
34
,
19995
20015
.
doi:10.1007/s00521-022-07523-8
.
Govindaraju
R. S.
2000
Artificial neural networks in hydrology. II: Hydrologic applications
.
Journal of Hydrologic Engineering
5
(
2
),
124
137
.
He
K. M.
,
Zhang
X. Y.
,
Ren
S. Q.
&
Sun
J.
2016
Deep residual learning for image recognition. In: IEE Conference on Computer Vision and Pattern Recognition, pp. 770–778. doi:10.1109/Cvpr.2016.90
.
Hersbach
H.
,
Bell
B.
,
Berrisford
P.
,
Hirahara
S.
,
Horányi
A.
,
Muñoz-Sabater
J.
,
Nicolas
J.
,
Peubey
C.
,
Radu
R.
,
Schepers
D.
,
Simmons
A.
,
Soci
C.
,
Abdalla
S.
,
Abellan
X.
,
Balsamo
G.
,
Bechtold
P.
,
Biavati
G.
,
Bidlot
J.
,
Bonavita
M.
,
De Chiara
G.
,
Dahlgren
P.
,
Dee
D.
,
Diamantakis
M.
,
Dragani
R.
,
Flemming
J.
,
Forbes
R.
,
Fuentes
M.
,
Geer
A.
,
Haimberger
L.
,
Healy
S.
,
Hogan
R.J.
,
Hólm
E.
,
Janisková
M.
,
Keeley
S.
,
Laloyaux
P.
,
Lopez
P.
,
Lupu
C.
,
Radnoti
G.
,
de Rosnay
P.
,
Rozum
I.
,
Vamborg
F.
,
Villaume
S.
&
Thépaut
J.-N.
2020
The ERA5 global reanalysis
.
Quarterly Journal of Royal Meteorological Society
146
(
730
),
1999
2049
.
Höhlein
K.
,
Michael
K.
,
Hewson
T.
&
Westermann
R.
2020
A comparative study of convolutional neural network models for wind field downscaling
.
Meterological Applications
27
,
1961
.
Kingma
D. P.
&
Ba
J.
2014
Adam: A method for stochastic optimization. In 3rd International Conference for Learning Representations. doi:10.48550/arXiv.1412.6980
.
Kratzert
F.
,
Klotz
D.
,
Shalev
G.
,
Klambauer
G.
,
Hochreiter
S.
&
Nearing
G.
2019
Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets
.
Hydrology and Earth System Science
23
,
5089
5110
.
doi:10.5194/hess-23-5089-2019
.
Lehner
B.
,
Verdin
K.
&
Jarvis
A.
2008
New global hydrography derived from spaceborne elevation data
.
EOS Transactions of American Geophysical Union
89
(
10
),
93
-
94
.
Liu
D.
,
Wang
Z.
,
Fan
Y.
,
Liu
X.
,
Wang
Z.
,
Chang
S.
&
Huang
T.
2017
Robust video super-resolution with learned temporal dynamics
.
In Proceedings of the IEEE International Conference on Computer Vision
,
2507
2515
Pan
B.
,
Hsu
K.
,
AghaKouchak
A.
&
Sorooshian
S.
2019
Improving precipitation estimation using convolutional neural network
.
Water Resources Research
55
,
2301
2321
.
Poitras
V.
,
Sushama
L.
,
Seglenieks
F.
,
Khaliq
M. N.
&
Soulis
E.
2011
Projected changes to streamflow characteristics over western Canada as simulated by the Canadian RCM
.
Journal of Hydrometeorology
12
,
1395
1413
.
Sajjadi
M. S. M.
,
Vemulapalli
R.
&
Brown
M.
2018
Frame-recurrent video super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 6626–6634. doi:10.1109/Cvpr.2018.00693
.
Sdraka
M.
,
Papoutsis
I.
,
Psomas
B.
,
Vlachos
K.
,
Ioannidis
K.
,
Karantzalos
K.
,
Gialampoukidis
I.
&
Vrochidis
S.
2022
Deep learning for downscaling remote sensing images: fusion and super-resolution
.
IEEE Geoscience and Remote Sensing Magazine
10
(
3
),
202
255
.
Singh
V. P.
&
Frevert
D. K.
2002
Mathematical Models of Small Watershed Hydrology and Applications
.
Water Resources Publications
,
Highlands Ranch, CO
.
Sit
M.
,
Demiray
B. Z.
,
Xiang
Z.
,
Ewing
G. J.
,
Sermet
Y.
&
Demir
I.
2020
A comprehensive review of deep learning applications in hydrology and water resources
.
Water Science and Technology
82
,
2635
2670
.
Stengel
K.
,
Glaws
A.
,
Hettinger
D.
&
King
R. N.
2020
Adversarial super-resolution of climatological wind and solar data
.
Proceedings of the National Academy of Sciences
117
(
29
),
16805
16815
.
doi:10.1073/pnas.1918964117
.
Sushama
L.
,
Laprise
R.
,
Caya
D.
,
Larocque
M.
&
Slivitzky
M.
2004
On the variable-lag and variable-velocity cell-to-cell routing schemes for climate models
.
Atmosphere-Ocean
42
(
4
),
221
233
.
doi:10.3137/ao.420401
.
Tang
X.
,
Tu
Z.
,
Wang
Y.
,
Liu
M.
,
Li
D.
&
Fan
X.
2022
Automatic detection of coseismic landslides using a new transformer method
.
Remote Sensing
14
,
2884
.
Teufel
B.
&
Sushama
L.
2021
2 °C vs high warming: Transitions to flood generating mechanisms across Canada
.
Water
13
(
11
),
1494
.
doi:10.3390/w13111494
.
Teufel
B.
&
Sushama
L.
2022
High-resolution modeling of climatic hazards relevant for the northern transportation sector
.
Climate Dynamics
59
,
3135
3151
.
doi:10.1007/s00382-022-06265-6
.
Teufel
B.
,
Fernanda
C.
,
Sushama
L.
,
Sun
L.
,
Khaliq
M. N.
,
Belair
S.
,
Shamseldin
A.
,
Nagesh Kumar
D.
&
Vaze
J.
2023
Physics-informed deep learning framework to model intense precipitation events at super resolution
.
Geoscience Letters
10
,
19
.
doi:10.1186/s40562-023-00272-z
.
Tounsi
A.
,
Abdelkader
M.
&
Temimi
M.
2023
Assessing the simulation of streamflow with the LSTM model across the continental United States using the MOPEX dataset
.
Neural Computing and Applications
35
,
22469
22486
.
doi.org/10.1007/s00521-023-08922-1.
Trinh
D.
,
Luong
M.
,
Dibos
F.
,
Rocchisani
J.
,
Pham
C.
&
Nguyen
T. Q.
2014
Novel example-based method for super-resolution and denoising of medical images
.
IEEE Transactions of Image Processing
23
(
4
),
1882
1895
.
Verseghy
D. L
.
2011
CLASS–the Canadian Land Surface Scheme (Version 3.5), Technical Documentation (Version 1)
.
Climate Research Division, Science and Technology Branch, Environment and Climate Change Canada
.
Wang
F.
,
Tian
D.
,
Lowe
L.
,
Kalin
L.
&
Lehrter
J.
2021
Deep learning for daily precipitation and temperature downscaling
.
Water Resources Research
57
,
e2020WR029308
.
Wu
Y. K.
,
Teufel
B.
,
Sushama
L.
,
Belair
S.
&
Sun
L. J.
2021
Deep learning-based super-resolution climate simulator-emulator framework for urban heat studies
.
Geophysical Research Letters
48
(
19
).
doi:10.1029/2021GL094737
.
Yang
R.
,
Luo
F.
,
Ren
F.
,
Huang
W.
,
Li
Q.
,
Du
K.
&
Yuan
D.
2022
Identifying urban wetlands through remote sensing scene classification using deep learning: A case study of Shenzhen, China
.
ISPRS International Journal of Geo-Informatics
11
,
131
.
Zhang
C.
,
Brodeur
Z. P.
,
Steinschneider
S.
&
Herman
J. D.
2022
Leveraging spatial patterns in precipitation forecasts using deep learning to support regional water management
.
Water Resources Research
58
,
e2021WR031910
.
Zhi
W.
,
Feng
D.
,
Tsai
W.-P.
,
Sterle
G.
,
Harpold
A.
,
Shen
C.
&
Li
L.
2021
From hydrometeorology to river water quality: Can a deep learning model predict dissolved oxygen at the continental scale?
Environmental Science and Technology
55
,
2357
2368
.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY-NC-ND 4.0), which permits copying and redistribution for non-commercial purposes with no derivatives, provided the original work is properly cited (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Supplementary data