PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Review of Deep Learning Applications in Hydrology and Water Resources

Muhammed Sit, Bekir Z. Demiray, Zhongrun Xiang, Gregory J. Ewing, Yusuf Sermet, Ibrahim Demir

arXiv:2007.12269v1physics.geo-phcs.LGstat.ML

TL;DR

Hydrology and water resources face increasingly complex data and modeling challenges, while evidence on deep learning applications remains dispersed across tasks and domains. This systematic review analyzes 129 publications and finds rapidly increasing use across diverse water-sector applications, while highlighting dataset and ethical challenges.

  • Problem

    Hydrological systems are heterogeneous, and existing physical models do not always perform and scale well because of intrinsic limitations.

  • Method

    The paper systematically reviews 129 publications on deep neural networks for hydrological tasks, grouped by water-sector application area.

  • Results

    Deep learning use in the water sector increased exponentially, with CNNs and LSTMs most common while Transformers were absent from the reviewed studies.

  • Takeaways & Limitations

    The review shows deep learning has potential across diverse water-sector applications and identifies directions for future hydrology and water-resources research.

  • Takeaways & Limitations

    Few reviewed flooding studies used curated, labeled datasets, contributing to redundant methods and limiting cumulative progress in the field.

Abstract

from arXiv · show

The global volume of digital data is expected to reach 175 zettabytes by 2025. The volume, variety, and velocity of water-related data are increasing due to large-scale sensor networks and increased attention to topics such as disaster response, water resources management, and climate change. Combined with the growing availability of computational resources and popularity of deep learning, these data are transformed into actionable and practical knowledge, revolutionizing the water industry. In this article, a systematic review of literature is conducted to identify existing research which incorporates deep learning methods in the water sector, with regard to monitoring, management, governance and communication of water resources. The study provides a comprehensive review of state-of-the-art deep learning approaches used in the water industry for generation, prediction, enhancement, and classification tasks, and serves as a guide for how to utilize available deep learning methods for future water resources challenges. Key issues and challenges in the application of these techniques in the water domain are discussed, including the ethics of these technologies for decision-making in water resources management and governance. Finally, we provide recommendations and future directions for the application of deep learning models in hydrology and water resources.

Abstract … Deep Learning

The paper reviews deep learning applications across hydrology and water resources, while documenting review methods, neural-network concepts, task types, and reproducibility challenges. It positions deep learning as a scalable approach for learning from complex hydrological data but emphasizes that application practices remain insufficiently standardized.

  • Abstract: The paper is a pre-print submitted to arxiv.
  • Introduction: 175 zettabytes of global digital data are expected by 2025, alongside growing water-related data from sensors, climate concerns, management, and hazard monitoring.
  • Introduction: Physical models are deterministic and reliable but may not perform or scale well because hydrological systems are inherently heterogeneous.
  • Introduction: The review systematically examines deep learning use cases across flooding, weather, land use and soil, water quality, surface water, water resources, and groundwater.
  • Introduction: Deep learning methods in hydrology lack standardized practices for data quality, preparation, execution, validation, and documentation, limiting deterministic reproducibility.
  • Literature Review: The literature review describes its search methodology, extracted manuscript information, deep learning architectures, machine-learning task types, and summary figures.
  • Review Methodology: 129 publications remained for comprehensive review after screening an initial 1515-publication search and filtering 315 publications.
  • Deep Learning: Deep learning uses multilayer artificial neural networks to map inputs to outputs, learn complex representations, and support regression, classification, sequence prediction, matrix prediction, unsupervised learning, and reinforcement learning.Stacked layers form complex later representations from simpler earlier ones.

Architectures … Recurrent Neural Networks (RNNs)

The review introduces cornerstone neural-network architectures used in hydrology and water-resources research, emphasizing CNNs, GANs, and RNNs. These architectures extract features, generate or translate data, and model sequential information, while more complex RNN variants address training problems at greater computational cost.

  • Architectures: The architecture subsection summarizes cornerstone neural-network architectures used in the reviewed papers and briefly discusses related ANN concepts.
  • Convolutional Neural Networks (CNNs): CNNs use at least one convolutional layer that receives a 3D tensor and applies cascaded kernels to extract intricate knowledge from the input.For an RGB image tensor, separate convolution operations and kernels process the color-channel matrices.
  • Convolutional Neural Networks (CNNs): Activation layers, typically using ReLU after convolution, introduce nonlinearity, while pooling layers reduce input size while preserving positional knowledge.
  • Generative Adversarial Networks (GANs): GANs consist of generator and discriminator CNNs competing in a min-max game to improve fake-example generation and fake-versus-real discrimination.The generator produces fake examples, whereas the discriminator determines whether inputs are fake.
  • Generative Adversarial Networks (GANs): GANs generate new dataset-like samples by mapping random noise to real samples and can also perform translation tasks such as super-resolution.
  • Recurrent Neural Networks (RNNs): RNNs contain recurrent layers whose neurons connect to one another, conveying each output both to the next layer and to the next neuron within the same layer.This recurrent connection allows information learned by one neuron to move through the sequence.
  • Recurrent Neural Networks (RNNs): RNNs are suited to sequential data such as time series or text because they retain information from earlier parts and move it to later parts.Earlier text can affect the meaning extracted from later text.
  • Recurrent Neural Networks (RNNs): LSTM and GRU networks address the vanishing-gradient problem affecting RNN training, but they have greater computational complexity.

Long Short-Term Memory (LSTM) Networks … Elman Network (ENN)

The reviewed recurrent architectures address sequential data through memory retention, gated computation, autoregressive mapping, or contextual state reuse. LSTM improves long-term memory but is resource-intensive, GRU reduces complexity, NAR models predict time series using current and previous samples, and ENNs retain prior hidden states through context units.

  • Long Short-Term Memory (LSTM) Networks: LSTM networks preserve information over longer short-term memory periods, supporting more efficient training on sequential datasets but requiring greater resources.LSTM neurons produce two values through successive activations and operations, rather than one output.
  • Long Short-Term Memory (LSTM) Networks: Figure 5 compares the computational structure of RNN, LSTM, and GRU nodes.
  • Gated Recurrent Unit (GRU) Networks: LSTM networks generally address vanishing gradients and support advances in natural language processing and time-series prediction, but their time complexity is a drawback.GRU networks reduce this complexity while retaining efficacy and produce one output to convey learned features.
  • Nonlinear Autoregressive (NAR) Models: NAR models predict time series by mapping current and previous samples to outputs through a nonlinear function such as a polynomial or neural network.When a neural network performs the training, the NAR network is classified as an RNN because it processes sequential inputs.
  • Elman Network (ENN): An Elman Network is an RNN implementation with three layers, including one hidden layer connected to context units.Figure 6 depicts an Elman Network architecture with two input and two output neurons.
  • Elman Network (ENN): After the first iteration, an Elman Network copies the hidden-layer state into context units, preserving the previous sample’s network state for later iterations.

Autoencoders (AE)

Autoencoders are unsupervised neural networks that reduce dataset dimensionality into latent representations while learning to reconstruct the original data. They use the same values as network inputs and outputs.

  • Autoencoders (AE): Autoencoders reduce dataset dimensionality using hidden-layer neurons that form a latent vector representation.The latent vector represents the dataset within the hidden layer.
  • Autoencoders (AE): Implemented unsupervised, autoencoders learn representations without labeled outputs.They generate a dataset representation within their hidden-layer neurons.
  • Autoencoders (AE): Using the same values as inputs and outputs, autoencoders learn to reconstruct data samples into their original form.

Deep Q Networks (DQN)

Deep Q Networks (DQN) are reinforcement-learning algorithms that use a predetermined rewarding policy to help an agent maximize rewards through interactions with its environment. Unlike conventional Q-Learning, DQN differs in how it generalizes.

  • DQN is a reinforcement-learning algorithm with a predetermined rewarding policy.
  • DQN provides an agent-environment framework for selecting actions that maximize received rewards.
  • DQN differs from conventional Q-Learning in how it generalizes.

Extreme Learning Machines (ELM)

Extreme Learning Machines are three-layer neural networks that randomize input-to-second-layer weights while training second-to-third-layer weights. They have also been criticized as unoriginal.

  • Extreme Learning Machines (ELM): ELM networks randomize input-to-second-layer weights and train the weights connecting to the third layer.They are described as three-layer neural networks.
  • Extreme Learning Machines (ELM): ELM networks have been criticized for being unoriginal.

Summary of Articles

The review summarizes 129 papers, finding that CNNs and LSTMs dominate because of their success in matrix and sequence prediction, while publication activity and architecture use increased over time. Studies commonly relied on authority-acquired datasets, rarely released software, and most often used Keras, creating reproducibility challenges.

  • Summary of Articles: 129 papers were comprehensively reviewed, although shared-figure data points vary because columns may contain multiple values or omit relevant information.The subsection presents visual summaries and a table of extracted information from each paper.
  • Architectures: CNNs and LSTMs were the most frequently used architectures, reflecting their success in matrix prediction and sequence prediction tasks important to hydrologic modelling.Despite extensive LSTM use, GRU networks with similar performance saw little significant usage.
  • Trends over time: Deep-learning utilization in the water field increased over time, with annual usage of neural-network architectures rising through March 2020.The review also projects continued growth for architectures such as DQNs that were widely used elsewhere but less used in the water field.
  • Data and code accessibility: Datasets were typically acquired from authorities or governmental agencies, while open-sourcing study software was unusual, making published outcomes difficult to reproduce.The review contrasts water research with broader deep-learning practice, where previously existing datasets and open-source models are more common.
  • Frameworks: Keras was inferred to be the most-used deep-learning framework in the water field, although TensorFlow appeared to be the first choice and PyTorch had smaller usage.Much TensorFlow usage came from Keras, which typically operates on top of TensorFlow through a higher-level interface.

Results

The Results section briefly summarizes the papers introduced earlier and groups them according to their use cases.

  • Results: The papers are summarized and interpreted by use case, with the results presented in Table 1.The section provides brief summaries of papers presented in the previous section.

Streamflow and Flood

Deep learning is widely applied to streamflow prediction, flood forecasting, susceptibility mapping, and monitoring, using recurrent, convolutional, optimized, decomposed, hybrid, and encoder-decoder architectures. Studies report improvements over physical or standalone deep learning models and enable image-based water-level and flooding assessments.

  • Streamflow and Flood: RNN, LSTM, NAR, and ENN architectures are commonly used for runoff prediction and flood forecasting because rainfall and runoff are time-series data.These applications target major rainfall-runoff modeling tasks.
  • Streamflow and Flood: LSTM models improved daily runoff prediction over SAC-SMA+Snow-17 and outperformed calibrated SAC-SMA and the National Water Model across 531 U.S. watersheds, including ungaged basins.The 2019 study used k-fold cross validation.
  • Streamflow and Flood: LSTM-PSO outperformed both standalone LSTM and LSTM-ALO after particle swarm optimization and ant lion optimization were used to tune LSTM parameters and hidden layers.The models were proposed to address deep-learning hyper-parameter optimization.
  • Streamflow and Flood: Runoff and flood models were enhanced through preprocessing with variational mode decomposition and intrinsic mode functions, or proper orthogonal and singular value decomposition.These approaches decomposed rainfall or runoff data before deep learning.
  • Streamflow and Flood: Stack autoencoder LSTM outperformed LSTM alone for daily discharge prediction, while encoder-decoder LSTMs addressed multiple-time-step flood forecasting.The SAE-LSTM used one-week discharge values as inputs.
  • Streamflow and Flood: LSTM networks were coupled with physical models to improve streamflow forecast accuracy, using meteorological inputs and simulated discharge from GHMs+CaMa Flood during 1971-2020.The hybrid approach improved flood simulations produced by a physical model.
  • Streamflow and Flood: DNNs and CNNs mapped area- or pixel-level flood susceptibility from 11 or 13 variables, while CNNs estimated real-time hourly water levels and flooding indicators from satellite and CCTV images.CCTV applications included water-level estimation, flooded-area calculation, SOFI, and stream-shape segmentation.

Land and Soil

Deep learning supports land-use, land-cover, and soil-related applications ranging from object-wise image segmentation to snow-depth estimation and time-series simulation. CNN, Mask R-CNN, DBN, and LSTM models are applied to classify land features, measure snow cover, and model land-surface temperature.

  • Land-use and land-cover mapping: Object-based CNNs perform object-wise rather than pixel-wise segmentation for very fine spatial resolution remote-sensing images and land-use classification.Researchers developed objective-based CNNs and blocks-based object-based CNNs to identify land use and land types.
  • Land-use and land-cover mapping: CNN models such as SegNet and VGG16 are used to segment and classify agricultural fields, while other studies identify impervious surfaces, vegetation, and bare soil.Related tasks include identifying agricultural fields, impervious surfaces, wetland types, water body types, and crop types.
  • Snow cover: Mask R-CNN segments a measuring rod in surveillance-camera images, enabling OpenCV to estimate snow depth from the rod’s snow-covered portion.Other work applies DBN and CNN models to estimate snow depth.
  • Land and soil time series: An ensemble composition-based LSTM simulates daily land-surface temperature after decomposing the original series into Intrinsic Mode Functions and a residue item.Other land- and soil-related applications include soil salinity and vegetation dynamics over time.

Unclassified Studies

Unclassified studies apply deep learning across coastal hydraulics, geothermal systems, terrain modeling, and basin-scale decision support. These works include image-based wave classification and tracking, geothermal temperature prediction, digital elevation model estimation, and GIS-based basin-scale decision support.

  • Coastal hydraulics: Deep learning studies in coastal hydraulics classify, monitor, and track nearshore waves using imagery.Kang (2019) used an improved CNN on images, while Kim et al. (2020) used multiple deep neural networks with coastal video imagery.
  • Geothermal systems: A DNN predicts geothermal spring and well temperatures from hydrogeochemical data, including chemical concentrations.The model uses hydrogeochemical inputs to estimate temperatures.
  • Terrain modeling: A very deep CNN estimates digital elevation models from single airborne or spaceborne images.The approach performs terrain-model estimation from individual remote-sensing images.
  • Decision support: A DNN was embedded into a GIS-based basin-scale decision support system.Rohmat et al. (2019) developed and embedded the model for basin-scale decision support.

Key Issues and Challenges

The review identifies data, conceptual, methodological, and reproducibility problems that constrain deep learning research in water resources. Key challenges include infeasible network sizes, inadequate benchmarking datasets, fragmented authority data, misunderstandings of deep learning, and insufficiently reported or justified technical choices.

  • Key Issues and Challenges: Universal approximation may require hidden layers that are too large to train and execute feasibly.Artificial neural networks with one hidden layer could theoretically represent any function, but feasibility can limit this capability.
  • Key Issues and Challenges: The water field lacks high-quality, curated, labeled, and published datasets for benchmarking and method development.This shortage slows the state of deep learning applications in the water domain.
  • Key Issues and Challenges: Authority-provided data are dispersed across agencies and can mismatch in temporal or spatial coverage, complicating dataset construction.Different agencies and areas may also have different periods of record.
  • Key Issues and Challenges: Studies often misunderstand deep learning as a specific technique rather than a broad class of neural-network-centered machine learning algorithms.The review identifies fundamental conceptual understanding as a problematic issue.
  • Key Issues and Challenges: Insufficient reporting of model and architecture details, together with dataset problems, creates persistent reproducibility barriers.These barriers overshadow reported accuracy and slow advancement in deep learning applications for water.
  • Key Issues and Challenges: Repeated trial-and-error discussions of layers, hyperparameters, or batch size without theoretical intuition weaken papers’ technical justification.The review states that not sharing intuition regarding technical choices prevents papers from reaching their goal.

Ethics in DL Applications

Deep learning in water resources raises ethical questions because it automates prediction and decision-making while increasing their reach and speed. Existing water-sector ethical dilemmas remain primarily distributional, but the literature rarely examines DL decision-making applications.

  • Ethics in DL Applications: Deep learning ethics center on what an application does when it automates prediction and decision-making in water resources.The review illustrates this concern with DL streamflow prediction in an urban catchment.
  • Ethics in DL Applications: DL expands decision-making’s reach and speed while potentially removing contextual information relevant to human judgment.This creates ethical concerns even when the underlying water-sector dilemmas remain familiar.
  • Ethics in DL Applications: Water-sector ethical decisions remain primarily distributional, involving who receives services, service quality, and whose risks are acceptable.These persistent dilemmas require resolution within water resources management and governance.
  • Ethics in DL Applications: Few reviewed water studies include DL decision-making components, creating an opportunity for research that integrates ethical work from other fields.Rohmat et al. (2019) is identified as an exception because its DL tool is integrated into a decision support system.
  • Ethics in DL Applications: DL’s new scale and speed require longstanding ethical questions about communities, the environment, and priorities to be examined in a new light.These questions are not exclusive to artificial intelligence or deep learning.

Recommendations and Conclusions

The paper systematically reviews 129 publications on deep neural networks for hydrological tasks and challenges. It highlights limited reuse of curated datasets and identifies applications supporting flood simulation, first responders, decision-makers, public education, and field personnel.

  • Conclusions: The review systematically selected 129 publications on recent deep neural-network applications addressing hydrological tasks and challenges.The publications were grouped by application area and analyzed using journal literature from January 2018 to March 2020.
  • Recommendations: Among 30 reviewed flooding papers, only a few used previously curated, labeled datasets, limiting architectural progress and producing repetitive studies with nearly identical methods.The passage connects dataset scarcity with limited cumulative development toward improved neural-network architectures and state-of-the-art advancement.
  • Recommendations: Deep learning can support realistic flood simulations, first-responder training, decision-making, public education, and heads-up recognition and decision support for field personnel.Potential field applications include sensor maintenance, structural renovation, and field experiments.
Loading 2007.12269v1…