Fully distributed hydrological models take into account the spatial variability of a catchment, allowing for a more accurate representation of its heterogeneity, and assessing its hydrological response at multiple locations. However, physics-based fully distributed models can be time-consuming when it comes to model runtime and calibration, especially for large-scale catchments. On the other hand, deep learning models have shown great potential in the field of hydrological modelling, outperforming lumped rainfall-runoff conceptual models, and improving prediction in ungauged basins via catchment transferability. Despite these advances, the field still lacks a multivariable, fully distributed hydrological deep learning model capable of generalizing to unseen catchments. To address the aforementioned challenges associated with physics-based distributed models and deep learning models, we explore the possibility of developing a fully distributed deep learning model by using Graph Neural Networks (GNN), an extension of deep learning methods to non-Euclidean topologies including graphs and meshes.

We develop a surrogate model of wflow_sbm, a fully distributed, physics-based hydrological model, by exploiting the similarities between its underlying functioning and GNNs. The GNN uses the same input as wflow_sbm: distributed static parameters based on physical characteristics of the catchment and gridded dynamic forcings. The GNN is trained to produce the same output as wflow_sbm, predicting multiple gridded variables related to rainfall-runoff, such as streamflow, actual evapotranspiration, subsurface flow, saturated and unsaturated groundwater storage, snow storage, and runoff. We show that our GNN model achieves high performance in unseen catchments, indicating that GNNs are a viable option for fully distributed multivariable hydrological models capable of generalizing to unseen regions. Furthermore, the GNN model achieves a significant computational speedup compared to wflow_sbm. We will continue this research, using the GNN-based surrogate models as pre-trained backbones to be fine-tuned with measured data, ensuring accurate model adaptation, and enhancing their practical applicability in diverse hydrological scenarios.

How to cite: Nelemans, P., Bentivoglio, R., Buitink, J., Meshgi, A., Hrachowitz, M., Dahm, R., and Taormina, R.: Towards Fully Distributed Rainfall-Runoff Modelling with Graph Neural Networks, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6846, https://doi.org/10.5194/egusphere-egu24-6846, 2024.

The significance of probabilistic hydrological forecasting has grown in recent years, offering crucial insights for risk-based decision-making and effective flood management. This study explores generative deep learning models, specifically the conditional variational autoencoder (CVAE), for probabilistic streamflow forecasting. This innovative approach is applied for forecasting streamflow one to seven days (s) ahead in 75 Canadian basins included in the open-source Canadian model parameter experiment (CANOPEX) database. CVAE is compared against two benchmark quantile-based deep learning models: the quantile-based encoder-decoder (ED) and the quantile-based CVAE (QCVAE).

Over 9000 deep learning models are developed based on different input variables, basin characteristics, and model structures and evaluated regarding point forecast accuracy and forecast reliability. Results highlight CVAE‘s superior reliability, showing a median reliability of 92.49% compared to 87.35% for ED and 84.59% for QCVAE (considering a desired 90% confidence level). However, quantile-based forecast models exhibit marginally better point forecasts, as evidenced by Kling-Gupta efficiency (KGE), with a median KGE of 0.90 for ED and QCVAE (compared to 0.88 for CVAE). Notably, the CVAE model provides reliable probabilistic forecasts in basins with low point forecast accuracy.

The developed generative deep learning models can be used as a benchmark for probabilistic streamflow forecasting due to the use of the open-source CANOPEX dataset. Overall, the results of this study contribute to the expanding field of generative deep learning models in hydrological forecasting, offering a general framework that applies to forecasting other hydrological variables as well (precipitation and soil moisture).

How to cite: Jahangir, M. S. and Quilty, J.: Probabilistic streamflow forecasting using generative deep learning models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2939, https://doi.org/10.5194/egusphere-egu24-2939, 2024.

Accurate modeling of various hydrological variables is important for water resource management, flood forecasting, and pest control. Deep learning models, especially Long Short-Term Memory (LSTM) models based on Recurrent Neural Network (RNN) structures, have shown significant success in simulating streamflow, soil moisture, and model parameter assessment. With the development of large language models (LLMs) based on attention mechanisms, such as ChatGPT and Bard, we have observed significant advancements in fields like natural language processing (NLP), computer vision (CV), and time series prediction. Despite achieving advancements across various domains, the application of attention-based models in hydrology remains relatively limited, with LSTM models maintaining a dominant position in this field. This study evaluates the performance of 18 state-of-the-art attention-based models and their variants in hydrology. We focus on their performance in streamflow, soil moisture, snowmelt, and dissolved oxygen (DO) datasets, comparing them to LSTM models in both long-term and short-term regression and forecasting. We also examine these models' performance in spatial cross-validation. Our findings indicate that while LSTM models maintain strong competitiveness in various hydrological datasets, Attention models offer potential advantages in specific metrics and time lengths, providing valuable insights into applying attention-based models in hydrology. Finally, we discuss the potential applications of foundation models and how these methods can contribute to the sustainable use of water resources and the challenges of climate change.

How to cite: Liu, J., Shen, C., and Bindas, T.: Can Attention Models Surpass LSTM in Hydrology?, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20636, https://doi.org/10.5194/egusphere-egu24-20636, 2024.

Streamflow simulation or rainfall-runoff modelling has been a topic of research for the past few decades which has resulted in a plethora of modelling approaches ranging from physics models to empirical or data driven approaches. There are many physics-based (PB) models available to estimate streamflow, but still there exists uncertainty in model outputs due to incomplete representations of physical processes. Further, with advancements in machine learning (ML) concepts, there have been several attempts but with no/little physical consistency. As a result, models based on ML algorithms may be unreliable if applied to provide future hydroclimate projections where climates and land use patterns are outside the range of training data. 

Here we test blended models built by combining PB model state variables (specifically soil moisture) with ML algorithms on their ability to simulate streamflow in 671 catchments representing diverse conditions across the conterminous United States.

For this purpose, we develop a suite of blended hydrological models by pairing different PB models (Catchment Wetness Index, Catchment Moisture Deficit, GR4J, Australian Water Balance, Single-bucket Soil Moisture Accounting, and Sacramento Soil Moisture Accounting models) with different ML methods such as Long Short Term Memory network (LSTM), eXtreme Gradient Boosting (XGB).

The results indicate that the blended models provide significant improvement in catchments where PB models are underperforming. Furthermore, the accuracy of streamflow estimation is improved in catchments where the ML models failed to estimate streamflow accurately.

How to cite: Sourya, D. A. and Rathinasamy, M.: Can Blended Model Improve Streamflow Simulation In Diverse Catchments ?, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18154, https://doi.org/10.5194/egusphere-egu24-18154, 2024.

National scale hydrological models are required for many types of water sector applications, for example water resources planning. Existing UK national-scale model frameworks are based on conceptual numerical schemes, with an emerging trend towards incorporating deep learning models. Existing literature has shown that groundwater/surface water interactions are key for accurately representing future flows, and these processes are most accurately represented with physically-based hydrological models.

In response to this, our study undertakes a comparative analysis of three national model frameworks (Neural Hydrology, HBV, SHETRAN) to investigate the necessity for physically-based hydrological modelling. The models were run with the full ensemble of bias-corrected UKCP18 12km RCM data which enabled a direct comparison of future flow projections. We show that whilst many national frameworks perform well for the historical period, physically-based models can give substantially different projections of future flows, particularly low flows. Moreover, our study illustrates that the physically-based model exhibits a consistent trajectory in Budyko space between the baseline and future simulations, a characteristic not shared by conceptual and deep learning models. To provide context for these results, we incorporate insights from other national model frameworks, including the eFlag project.

How to cite: Lewis, E., Smith, B., Birkinshaw, S., He, H., and Pritchard, D.: Benchmarking hydrological models for national scale climate impact assessment, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18762, https://doi.org/10.5194/egusphere-egu24-18762, 2024.

In addressing the challenge of streamflow prediction in ungauged basins, this study leveraged deep learning (DL) models, especially long short-term memory (LSTM) networks, to predict streamflow for pseudo ungauged basins in Japan. The motivation stems from the recognized limitations of traditional hydrological models in transferring their performance beyond the calibrated basins. Recent research suggests that DL models, especially those trained on multiple catchments, demonstrate improved predictive capabilities utilizing the concept of streamflow regionalization. However, the majority of these studies were confined to geographic regions within the United States.

For this study, a total number of 211 catchments were delineated and investigated, distributed across all four primary islands of Japan (Kyushu - 32, Shikoku - 13, Honshu - 127, and Hokkaido - 39) encompassing a comprehensive sample of hydrological systems within the region. The catchments were obtained corresponding to the streamflow observation points and their combined area represented more than 43% of Japan's total land area, after accounting for overlaps. After cleaning and refining the streamflow dataset, the remaining catchments (198) were divided into training (~70%), validation (~20%), and holdout test (~10%) sets. A combination of dynamic (time-varying) and static (constant) variables were obtained on a daily basis corresponding to the daily streamflow data and provided to the models as input features. However, the final model accorded higher significance to dynamic features in comparison to the static ones. Although the models were trained on daily time steps, the results were aggregated to monthly timescale. The main evaluation metrics included the Nash-Sutcliffe Efficiency (NSE) and Pearson’s correlation coefficient (r). The final model achieved a median NSE of 0.96, 0.83, & 0.78, and a median correlation of 0.98, 0.92, & 0.91 corresponding to the training, validation, and test catchments, respectively. For the validation catchments, 90% exhibited NSE values greater than 0.50, and 97% demonstrated a correlation surpassing 0.70. Correspondingly, these proportions were observed at 77% and 91% for the test catchments.

The results presented in this study demonstrate the feasibility and efficacy of developing a data-driven model for streamflow prediction in ungauged basins utilizing streamflow regionalization. The final model exhibits commendable performance, as evidenced by high NSE and correlation coefficients across the majority of the catchments. Importantly, the model's ability to generalize to unseen data is highlighted by its remarkable performance on the holdout test set, with only a few instances of lower NSE values (< 0.50) and correlation coefficients (< 0.70).

How to cite: Servia, H., Albrecht, F., Saxe, S., Bierti, N., Kawasaki, M., and Kurihara, S.: A Step Towards Global Hydrologic Modelling: Accurate Streamflow Predictions In Pseudo-Ungauged Basins of Japan, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6432, https://doi.org/10.5194/egusphere-egu24-6432, 2024.

In recent years, machine learning models, particularly Long Short-Term Memory (LSTM), have proven to be effective alternatives for rainfall-runoff modeling, surpassing traditional hydrological modeling approaches ¹. These models have predominantly been implemented and evaluated for rainfall-runoff simulations. However, operational hydrology often requires short- and mid-term forecasts. To be effective, such forecasts must consider past observed values of the predicted variables, requiring a data assimilation procedure ^2,3,4. This presentation will evaluate several approaches based on the combination of open-source machine learning tools and data assimilation strategies for short- and mid-term discharge forecasting of flood and/or drought events. The evaluation is based on the rich and well-documented CAMELS dataset ^5,6,7. The tested approaches include: (1) coupling pre-trained LSTMs on the CAMELS database with a Multilayer Perceptron (MLP) for prediction error corrections, (2) direct discharge MLP forecasting models specific for each lead time, including past observed discharges as input variables, and (3) option 2, including the LSTM-predicted discharges as input variables. In the absence of historical archives of weather forecasts (rainfall, temperatures, etc.), the different forecasting approaches will be tested in two configurations: (1) weather forecasts assumed to be perfect (using observed meteorological variables over the forecast horizon in place of predicted variables or ensembles) and (2) use of ensembles reflecting climatological variability over the forecast horizons for meteorological variables ensembles made up of time series randomly selected from the past. The forecast horizons considered range from 1 to 10 days, and the results are analyzed in light of the time of concentration of the watersheds.

References

1. Kratzert F, Klotz D, Brenner C, Schulz K, Herrnegger M. Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks. Hydrol Earth Syst Sci. 2018;22(11):6005-6022. doi:10.5194/hess-22-6005-2018

2. Bourgin F, Ramos MH, Thirel G, Andréassian V. Investigating the interactions between data assimilation and post-processing in hydrological ensemble forecasting. J Hydrol (Amst). 2014;519:2775-2784. doi:10.1016/j.jhydrol.2014.07.054

3. Boucher M ‐A., Quilty J, Adamowski J. Data Assimilation for Streamflow Forecasting Using Extreme Learning Machines and Multilayer Perceptrons. Water Resour Res. 2020;56(6). doi:10.1029/2019WR026226

4. Piazzi G, Thirel G, Perrin C, Delaigue O. Sequential Data Assimilation for Streamflow Forecasting: Assessing the Sensitivity to Uncertainties and Updated Variables of a Conceptual Hydrological Model at Basin Scale. Water Resour Res. 2021;57(4). doi:10.1029/2020WR028390

5. Newman AJ, Clark MP, Sampson K, et al. Development of a large-sample watershed-scale hydrometeorological data set for the contiguous USA: data set characteristics and assessment of regional variability in hydrologic model performance. Hydrol Earth Syst Sci. 2015;19(1):209-223. doi:10.5194/hess-19-209-2015

6. Kratzert, F. (2019). Pretrained models + simulations for our HESSD submission "Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets", HydroShare, https://doi.org/10.4211/hs.83ea5312635e44dc824eeb99eda12f06

7. Kratzert, F. (2019). CAMELS Extended Maurer Forcing Data, HydroShare, https://doi.org/10.4211/hs.17c896843cf940339c3c3496d0c1c077

How to cite: Saint Fleur, B. E., Gaume, E., Savary, M., Akil, N., and Theriez, D.: Short- and mid-term discharge forecasts combining machine learning and data assimilation for operational purpose, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16474, https://doi.org/10.5194/egusphere-egu24-16474, 2024.

By employing Machine Learning techniques on the US-CAMELS dataset, we discern a minimal number of streamflow features. Together with meteorological forcing, these features enable an approximate reconstruction of the entire streamflow time-series. This task is achieved through the application of an explicit noise conditional autoencoder, wherein the meteorological forcing is inputted to the decoder to encourage the encoder to learn streamflow features exclusively related to landscape properties. The optimal number of encoded features is determined with an intrinsic dimension estimator. The accuracy of reconstruction is then compared with models that take a subset of static catchment attributes (both climate and landscape attributes) in addition to meteorological forcing variables. Our findings suggest that attributes gathered by experts encompass nearly all pertinent information regarding the input/output relationship. This information can be succinctly summarized with merely three independent streamflow features. These features exhibit a strong correlation with the baseflow index and aridity indicators, aligning with the observation that predicting streamflow in dry catchments or with a high baseflow index is more challenging. Furthermore, correlation analysis underscores the significance of soil-related and vegetation attributes. These learned features can also be associated with parameters in conceptual hydrological models such as the GR model family.

How to cite: Bassi, A., Mira, A., Höge, M., Fenicia, F., and Albert, C.: Learning Catchment Features with Autoencoders, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9190, https://doi.org/10.5194/egusphere-egu24-9190, 2024.

Among the non-structural strategies for mitigating the huge economic losses and casualties caused by floods, the implementation of early-warning systems based on real-time forecasting of flood maps is one of the most effective. The high computational cost associated with two-dimensional (2D) hydrodynamic models, however, prevents their practical application in this context. To overcome this drawback, “data-driven” models are gaining considerable popularity due to their high computational efficiency for predictions. In this work, we introduce a novel surrogate model based on the Transformer architecture, named FloodSformer (FS), that efficiently predicts the temporal evolution of inundation maps, with the aim of providing real-time flood forecasts. The FS model combines an encoder-decoder (2D Convolutional Neural Network) with a Transformer block that handles temporal information. This architecture extracts the spatiotemporal information from a sequence of consecutive water depth maps and predicts the water depth map at one subsequent instant. An autoregressive procedure, based on the trained surrogate model, is employed to forecast tens of future maps.

As a case study, we investigated the hypothetical inundation due to the collapse of the flood-control dam on the Parma River (Italy). Due to the absence of real inundation maps, the training/testing dataset for the FS model was generated from numerical simulations performed through a 2D shallow‐water code (PARFLOOD). Results show that the FS model is able to recursively forecast the next 90 water depth maps (corresponding to 3 hours for this case study, in which maps are sampled at 2-minute intervals) in less than 1 minute. This is achieved while maintaining an accuracy deemed entirely acceptable for real-time applications: the average Root Mean Square Error (RMSE) is about 10 cm, and the differences between ground-truth and predicted maps are generally lower than 25 cm in the floodable area for the first 60 predicted frames. In conclusion, the short computational time and the good accuracy ensured by the autoregressive procedure make the FS model suitable for early-warning systems.

How to cite: Pianforini, M., Dazzi, S., Pilzer, A., and Vacondio, R.: A Transformer-Based Data-Driven Model for Real-Time Spatio-Temporal Flood Prediction, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-566, https://doi.org/10.5194/egusphere-egu24-566, 2024.

Predicting future flood hazards in a changing climate requires adopting a stochastic framework due to the multiple sources of uncertainties (e.g., from climate change scenarios, climate models, or natural variability). This requires performing multiple flood inundation simulations which are computationally costly. Data-driven models can help overcome this issue as they can emulate urban flood maps considerably faster than traditional flood simulation models. However, their lack of generalizability to both terrain and rainfall events still limits their application. Additionally, these models face the challenge of not having sufficient training data. This led state-of-the-art models to adopt a patch-based framework, where the study area is first divided into local patches (i.e., broken into smaller terrain images) that are subsequently merged to reconstruct the whole study area prediction. The main drawback of this method is that the model is blind to the surroundings of the local patch. To overcome this bottleneck, we developed a new deep learning model that includes patches' contextual information while keeping high-resolution information of the local patch. We trained and tested the model in the city of Zurich, at spatial resolution of 1 m. The evaluation focused on 1-hour rainfall events at 5 min temporal resolution and encompassing extreme precipitation return periods from 2- to 100-year. The results show that the proposed CNN-attention model outperforms the state-of-the-art patch-based urban flood emulator. First, our model can faithfully represent flood depths for a wide range of extreme rainfall events (peak rainfall intensities ranging from 42.5 mm h^-1 to 161.4 mm h^-1). Second, the model's terrain generalizability was assessed in distinct urban settings, namely Luzern and Singapore. Our model accurately identifies water accumulation locations, which constitutes an improvement compared to current models. Using transfer learning, the model was successfully retrained in the new cities, requiring only a single rainfall event to adapt the model to new terrains while preserving adaptability across diverse rainfall conditions. Our results suggest that by integrating contextual terrain information with local terrain patches, our proposed model effectively generates high-resolution urban pluvial flood maps, demonstrating applicability across varied terrains and rainfall events.

How to cite: Cache, T., Gomez, M. S., Blagojević, J., Beucler, T., Leitão, J. P., and Peleg, N.: Improving the Generalizability of Urban Pluvial Flood Emulators by Contextualizing High-Resolution Patches, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8102, https://doi.org/10.5194/egusphere-egu24-8102, 2024.

The domain of spatial flood prediction is dominated by hydrodynamic models, which, while robust and adaptable, are often constrained by computational requirements and slow processing times. To address these limitations, the integration of Deep Learning (DL) models has emerged as a promising solution, offering the potential for rapid prediction capabilities, while maintaining a high output quality. However, a critical challenge with DL models lies in their requirement for retraining for each new domain area, based on the outputs of hydrodynamic simulations generated for that specific region. This need for domain-specific retraining hampers the scalability and quick deployment of DL models in diverse settings. Our research focuses on bridging this gap by developing a fully generalized DL model for flood prediction.

FloodWaive's approach pivots on creating a DL model that can predict flood events rapidly and accurately across various regions without requiring retraining for each new domain area. The model is trained on a rich dataset derived from numerous hydrodynamic simulations, encompassing a wide spectrum of topographical conditions. This training is designed to enable the model to generalize its predictive capabilities across different domains and weather patterns, thus overcoming the traditional limitation of DL models in this field.

Initial findings from the development phase are promising, showcasing the model's capability to process complex data and provide quick, accurate flood predictions. The success of this fully generalized DL modeling approach could revolutionize applications of flood predictions such as flood forecasting and risk analysis. Regarding the later, real-time evaluation of flood protection measures could become a reality. This would empower urban planners, emergency response teams, and environmental agencies with the ability to make informed decisions quickly, potentially saving lives and reducing economic losses.

While this project is still in its developmental stages, the preliminary results point towards a significant leap in flood forecasting technology. The ultimate goal is to offer a universally deployable, real-time flood prediction tool, significantly enhancing our ability to mitigate the impact of floods worldwide.

How to cite: Hofmann, J. and Holt, A.: Revolutionizing Flood Forecasting with a Generalized Deep Learning Model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20907, https://doi.org/10.5194/egusphere-egu24-20907, 2024.

Stream water temperatures influence aquatic biodiversity, agriculture, tourism, electricity production, and water quality. Therefore, stakeholders would benefit from an operational forecasting service that would support timely action. Deep Learning methods are well-suited for this task as they can provide probabilistic forecasts at individual stations of a monitoring network. Here we train and evaluate several state-of-the-art models using 10 years of data from 55 stations across Switzerland. Static features (e.g. station coordinates, catchment mean elevation, area, and glacierized fraction), time indices, meteorological and/or hydrological observations from the past 64 days, and their ensemble forecasts for the following 32 days are included as predictors in the models to estimate daily maximum water temperature for the next 32 days. We find that the Temporal Fusion Transformer (TFT) model performs best for all lead times with a cumulative rank probability score (CRPS) of 0.73 ºC averaged over all stations, lead times and 90 forecasts distributed over 1 full year. The TFT is followed by the Recurrent Neural Network (CRPS = 0.77 ºC), Neural Hierarchical Interpolation for Time Series (CRPS = 0.80 ºC), and Multi-layer Perceptron (CRPS = 0.85 ºC). All models outperform the benchmark ARX model. When factoring out the uncertainty stemming from the meteorological ensemble forecasts by using observations instead, the TFT improves to a CRPS of 0.43 ºC, and it remains the best of all models. In addition, the TFT model identifies air temperature and time of the year as the most relevant predictors. Furthermore, its attention feature suggests a dominant response to more recent information in the summer, and to information from the previous month during spring and autumn. Currently, daily maximum water temperature probabilistic forecasts are produced twice per week and made available at https://drought.ch/de/allgemeine-lage/wassertemperatur/fliessgewaesser-1-1.html. 

How to cite: Padrón, R. S., Zappa, M., and Bogner, K.: Operational stream water temperature forecasting with a temporal fusion transformer model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18073, https://doi.org/10.5194/egusphere-egu24-18073, 2024.

Sea surface temperature (SST) is a critical parameter in the global ocean-atmospheric system, exerting a substantial impact on climate change and extreme weather events like droughts and floods. The precise forecasting of future SSTs is thus vital for identifying such weather anomalies. Here we present a novel three-dimensional (3D) neural network model based on self-attention mechanisms and Swin-Transformer for mid-term SST predictions. This model, integrating both climatic and temporal features, employs self-attention to proficiently capture the temporal dynamics and global patterns in SST. This approach significantly enhances the model's capability to detect and analyze spatiotemporal changes, offering a more nuanced understanding of SST variations. Trained on 59 years of global monthly ERA5-Land reanalysis data, our model demonstrates strong deterministic forecast capabilities in the test period. It employs a convolution strategy and global attention mechanism, resulting in faster and more accurate training compared to traditional methods, such as Convolutional Neural Network with Long short-term memory (CNN-LSTM). The effectiveness of this SST prediction model highlights its potential for extensive multidimensional modelling applications in geosciences.

How to cite: Wang, L., Zhang, Y., and Zhang, X.: Skilful prediction of mid-term sea surface temperature using 3D self-attention-based neural network, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9446, https://doi.org/10.5194/egusphere-egu24-9446, 2024.

Surface water floods are responsible for a substantial amount of damage to buildings, yet they have received less attention than fluvial floods. Nowadays, both research and insurance companies are increasingly focusing on these phenomena to enhance knowledge and prevention efforts. This study builds upon pluvial-related damage data provided by the Swiss Mobiliar Insurance Company and the Building Insurance of Canton Zurich (GVZ) with the goal of developing a data-driven model for predicting potential damages in future precipitation events.

This work is a continuation of a previous method applied to Swiss data, relying on thresholds based on the quantiles of precipitation intensity and event volume, which, however, resulted in an excessive number of false alarms. First, a logistic regression has been assessed using different characteristics of the precipitation event. Subsequently, a random forest was established, incorporating terrain attributes to better characterize local conditions. Finally, a deep learning model was developed to account for the spatio-temporal properties of the precipitation fields on a domain larger than the targeted 1 km cell. The deep learning model comprises a convolutional neural network (CNN) for 4D precipitation data and subsequent dense layers, incorporating static attributes. The model has been applied to predict the probability of damage occurrence, as well as the damage degree quantified by the number of claims relative to the number of insured buildings.

How to cite: Horton, P., Mosimann, M., Kaderli, S., Martius, O., Zischg, A. P., and Steinfeld, D.: Deep-learning-based prediction of damages related to surface water floods for impact-based warning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17543, https://doi.org/10.5194/egusphere-egu24-17543, 2024.

Unmanned aerial vehicle (UAV) photogrammetry allows the generation of orthophoto and digital surface model (DSM) rasters of a terrain. However, DSMs of water bodies mapped using this technique often reveal distortions in the water surface, thereby impeding the accurate sampling of water surface elevation (WSE) from DSMs. This study investigates the capability of deep neural networks to accommodate the aforementioned perturbations and effectively estimate WSE from photogrammetric rasters. Convolutional neural networks (CNNs) were employed for this purpose. Three regression approaches utilizing CNNs were explored: i) direct regression employing an encoder, ii) prediction of the weight mask using an encoder-decoder architecture, subsequently used to sample values from the photogrammetric DSM, and iii) a solution based on the fusion of the two approaches. The dataset employed in this study comprises data collected from five case studies of small lowland streams in Poland and Denmark, consisting of 322 DSM and orthophoto raster samples. Each sample corresponds to a 10 by 10 meter area of the stream channel and adjacent land. A grid search was employed to identify the optimal combination of encoder, mask generation architecture, and batch size among multiple candidates. Solutions were evaluated using two cross-validation methods: stratified k-fold cross-validation, where validation subsets maintained the same proportion of samples from all case studies, and leave-one-case-out cross-validation, where the validation dataset originates entirely from a single case study, and the training set consists of samples from other case studies. The proposed solution was compared with existing methods for measuring water levels in small streams using a drone. The results indicate that the solution outperforms previous photogrammetry-based methods and is second only to the radar-based method, which is considered the most accurate method available.

This research was funded by National Science Centre, Poland, project WATERLINE (2020/02/Y/ST10/00065), under the CHISTERA IV programme of the EU Horizon 2020 (Grant no 857925).

How to cite: Szostak, R., Zimnoch, M., Wachniew, P., Pietroń, M., and Ćwiąkała, P.: Estimation of Small Stream Water Surface Elevation Using UAV Photogrammetry and Deep Learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15248, https://doi.org/10.5194/egusphere-egu24-15248, 2024.

Water flows and related mixing dynamics in the unsaturated zone are difficult to measure directly, so stable water isotope tracers have been used successfully to quantify flux and storage dynamics and to constrain process-based hydrological models as proxy data. In this study, a data-driven model based on deep learning was adapted to interpolate and extrapolate spatio-temporal isotope signals of δ¹⁸O and δ²H in soil water in three dimensions. Further, this was also used to help quantify evapotranspiration and groundwater recharge processes in the unsaturated zone. To consider both spatial and temporal dependencies of water isotope signals in the model design, the output space was decomposed into temporal basis functions and spatial coefficients using singular value decomposition. Then, temporal functions and spatial coefficients were predicted separately by specialized deep learning models in interdependencies among target data, such as the LSTM model and convolutional neural network. Finally, the predictions by the models were integrated and analyzed post-hoc using XAI tools.

Such an integrated framework has the potential to improve understanding of model behavior based on features (e.g., climate, hydrological component) connected to either temporal or spatial information. Furthermore, the model can serve as a surrogate model for process-based hydrological models, improving the use of process-based models as learning tools.

How to cite: Jung, H., Soulsby, C., and Tetzlaff, D.: A deep learning approach for spatio-temporal prediction of stable water isotopes in soil moisture, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-811, https://doi.org/10.5194/egusphere-egu24-811, 2024.