The limitations of conventional streamflow forecasting

The gap

Evidence profile

Sourced from the stated research gap and future work and future-work section and stated challenges of the source papers, classified as general, spanning 7 journals.

Supporting evidence — 8 representative gaps

• Understanding Long-Term Groundwater Storage Variability Using GRACE Data and Explainable Machine Learning (2026) · Hydrology · doi

  The lack of understanding of the temporal dynamics of groundwater storage and its long-term changes. The need for a novel hybrid modeling framework that integrates multi-source satellite and climate data with machine learning and explanatory artificial intelligence techniques. The limited application of sequence-based modeling with model interpretability in semi-arid closed basins.

  generalstated research gap
  Keywords: lack understanding temporal dynamics groundwater storage long-term changes
• A VMD–PSO hybrid framework with machine and deep learning for Aaccurate spatial–temporal daily rainfall occurrence prediction in Pakistan’s diverse climatic regions (2026) · Modeling Earth Systems and Environment · doi

  The current method can be improved in future studies by adding satellite observations and remote sensing, which will provide a more thorough spatial representation and enable the detection of exceptional rainfall occurrences. In order to better capture temporal correlations over large distances than recurrent networks, more complex deep learning archi- tectures, including Transformer models and attention, may be investigated. Additionally, the system can be tailored to incorporate real-time early warning in order to mitigate disaster risks in sensitive locations and encourage flood preparedness. Addi- tionally, multi-objective optimization would be useful in achieving a balance between interpretability of the model, computational efficiency, and forecast accuracy, making the framework more practical to use on a broad scale. Author contributions H.F. and S.H. designed and implemented the deep learning framework and performed experiments. M.A.H. contrib- uted to data preprocessing, model benchmarking, and statistical analy- sis. S.S. assisted with theoretical formulation and result interpretation. All authors reviewed and approved the final manuscript. Funding Open access funding provided by NTNU Norwegian Univer- sity of Science and Technology (incl St. Olavs Hospital - Trondheim University Hospital) Data availability No datasets were generated or analysed during the current study.

  generalfuture workevidence 5/5
  Keywords: current order deep learning model framework funding hospital improved future adding satellite observations remote sensing
• Machine and Deep Learning Approaches for Drought Characterization and Prediction: a Comprehensive Review (2026) · Environmental Modeling & Assessment · doi

  Drought remains one of the most complex and multifaceted natural hazards, shaped by nonlinear interactions among climatic, hydrological, and ecological systems. Over the past decades, drought research has evolved from simple statistical indices to advanced hybrid and deep learning frameworks, reflecting a concerted effort to improve accu- racy, reliability, and interpretability under changing climatic conditions. Traditional drought indices—such as SPI, SPEI, and PDSI—have long served as foundational tools for his- torical monitoring and trend assessment. However, their reliance on stationary climate assumptions and limited flex- ibility constrains their performance under evolving environ- mental conditions. Conversely, machine learning (ML) and deep learning (DL) techniques—including Random Forests, SVM, ANN, CNN, and LSTM—offer superior capabili- ties in capturing nonlinear patterns and temporal depen- dencies. By leveraging multi-source datasets from remote sensing, ground observations, and climate projections, these models enable more accurate and spatially resolved 1 3Machine and Deep Learning Approaches for Drought Characterization and Prediction: a Comprehensive…drought forecasting. Recent developments emphasize the rise of hybrid and ensemble frameworks, which integrate the strengths of process-based and data-driven methods. These models combine the physical realism of hydrological simulations with the adaptive learning power of AI, yield- ing improved predictive accuracy and interpretability. The integration of CMIP6-based climate projections with deep learning architectures represents a promising direction for future drought scenario modeling and adaptive water-man- agement planning. Recent developments have also focused on improving the operational accessibility of drought analysis through dedicated software platforms, which address limitations of existing tools such as restricted functionality, limited distri- butional flexibility, and dependence on advanced program- ming expertise (Terzi and Üçüncü) [256]. The emergence of cloud computing, IoT-enabled observation systems, and explainable AI has further transformed the drought moni- toring landscape. Platforms such as Google Earth Engine facilitate near-real-time drought mapping, while physics- informed and interpretable AI models promote transparent, trustworthy predictions. Integrating community feedback and participatory data systems enhances the social relevance of drought forecasts, bridging scientific insights and practi- cal decision-making. Looking forward, the future of drought research lies in interdisciplinary collaboration, adaptive modeling, and scalable AI systems that couple hydrologi- cal knowledge with advanced analytics. By uniting physical understanding, data-driven innovation, and policy integra- tion, researchers and decision-makers can transition from reactive drought response toward proactive, predictive, and sustainable management. Such integrated approaches will be critical for ensuring water security, supporting agricul- tural resilience, and building climate-adaptive systems for future generations. Acknowledgements The authors wish to thank the anonymous reviewers for their constructive comments to enhance the quality of the manuscript. Author Contributions C.N.: Conceptualization, Methodology, Data curation, Formal analysis, Investigation, Validation, Visualization, Writing – original draft editing. A.M: Methodology, Data curation, Formal analysis, Writing – review & editing, Supervision. Funding There was no funding for this project. Data Availability Data will be made available on request.

  generalfuture workevidence 5/5
  Keywords: drought learning systems deep climate adaptive advanced models future nonlinear climatic hydrological indices hybrid frameworks
• BiasCast: learning and adjusting real time biases from meteorological forecasts to enhance runoff predictions (2026) · Hydrology and Earth System Sciences · doi

  The operational deployment of deep learning based flood forecasting models faces fundamental challenges when tran- sitioning from high-quality reanalysis to meteorological forecast data, with domain shift between these data sources leading to model performance degradation. While previous approaches have focused on bias-correcting meteorological inputs through statistical methods (Lenderink et al., 2007) or machine learning techniques applied to precipitation fore- casts (Ko et al., 2020; Zhang et al., 2020), these methods rely on comparing forecasts with meteorological observa- tions that themselves contain uncertainties (Bárdossy et al., 2022). This study addressed the challenge through system- atic evaluation of Long Short-Term Memory architectures and training techniques that learn bias correction directly from the more reliable discharge observations, following the paradigm suggested by Kirchner (2009) of using river dis- charge as the primary constraint. Our experiments across 451 Central European catchments demonstrated that appropriate neural network designs can transform the domain shift prob- lem from a major obstacle into a learnable pattern correction task. Sequential Forecast LSTM architectures, when com- bining meteorological hindcast data with past discharge ob- servations, provided the most effective framework for miti- gating forecast-induced biases. This configuration achieved a median NSE of 0.71, surpassing even the reanalysis base- line simulation and establishing discharge integration, if data are available in near real time as in the LamaH domain, as a critical component for operational forecast accuracy. To quantify the bias propagation caused by the domain shift, we conducted cross-domain evaluation revealing per- formance deterioration when reanalysis-trained models were applied to forecast inputs. In this setting, the median Nash- Sutcliffe Efficiency decreased from 0.58 to 0.33, represent- ing a 0.25 reduction in model skill. This performance degra- dation stems from fundamental differences in data distri- butions between reanalysis and forecast datasets, violating the assumption of identically distributed training and test- ing data that underlies machine learning model generaliza- tion (Goodfellow et al., 2016). Analysis of the spatial pat- terns of this input data domain shift, quantified using the 1- Wasserstein distance across the 451 basins, reveals that to- pographically complex, high-elevation, and snow-dominated catchments experience the largest distributional differences between reanalysis and ECMWF-HRES inputs, potentially driven by the limited ability of numerical weather prediction models to resolve orographic effects at their operational res- olution (Haiden et al., 2024; Lavers et al., 2021). Among the tested neural network architectures, the Se- quential Forecast LSTM demonstrated superior performance for operational forecasting applications. This architecture achieved a median NSE of 0.63 with notable st

  generalfuture workevidence 5/5
  Keywords: forecast domain reanalysis operational meteorological shift learning models model performance bias inputs architectures discharge median
• A Diffusion-Based Framework for High-Resolution Precipitation Forecasting over CONUS (2026) · Artificial Intelligence for the Earth Systems · doi

  This manuscript presents a comprehensive evaluation of three AI-based precipitation forecasting models across the CONUS: a data-driven model, a hybrid model, and an HRRR-corrective model. These models are benchmarked against the operational HRRR forecasts to assess their performance at both regular and extreme rainfall intensities, with 1 km spatial resolution and lead times ranging from 1 to 12 hours. A central novelty of this work lies in the direct comparison across three DL models each with distinct advantages and limitations. To the best of our knowledge, this is the first study to conduct such a large-scale and structured comparison for CONUS-wide extreme precipitation, providing critical insight into how different inputs in DL models enhance learning under real-world operational constraints. Our results demonstrate that the hybrid model achieves the best performance at short lead times (1h), while the HRRR-corrective model consistently outperforms both the HRRR and other DL models at longer lead times (up to 12h). This indicates the value of incorporating physics-based forecasts as a conditional prior when predicting complex spatiotemporal rainfall patterns. Another contribution of this work is the detailed monthly and regional evaluation of model performance. We show that the DL model behaviors are not only statistically robust but also physically consistent with regional precipitation regimes. For example, model skill drops in regions and months characterized by convective or monsoonal activity, which are historically more challenging to predict. This physics-informed analysis framework adds interpretability and reliability to AI-based weather forecasting. We also introduce a UQ pipeline with enhanced interpretability for operational use. A new custom loss function is developed to handle spatial error structures and false positive/negative asymmetries in rainfall prediction. Furthermore, we propose a novel UQ framework that quantifies and visualizes spatiotemporal uncertainty to support stakeholder decision-making, bridging the gap between raw model outputs and actionable insights. 33 Future work will explore the development of a globally generalizable DL model for rainfall prediction, with scalability to diverse climates and input data sources. We also aim to refine our UQ framework with conformal prediction, which provides statistically valid uncertainty inter- vals for each forecast. Additionally, we plan to enhance model interpretability by incorporating explainability metrics (e.g., saliency maps, input-attribution analyses) to identify key predictors influencing rainfall intensity. These improvements will make the model’s outputs more transparent and actionable for end users such as emergency managers and forecasters. 34 Acknowledgments. This material is based upon work supported by the National Science Founda- tion under Grant No. RISE-2019758 within the NSF AI Institute for Research on Trustwort

  generalfuture workevidence 5/5
  Keywords: model models rainfall based hrrr precipitation operational performance lead times framework interpretability prediction evaluation three
• Machine Learning and Deep Learning for Snowmelt Prediction: Long-Term Trends and Implications for Water Resource Management (2026) · Water Resources Management · doi

  Future research should focus on integrating long-term hydrometeorological datasets with Machine Learning and Deep Learning approaches. Future research should explore the application of the study's findings to other regions and contexts. Future research should investigate the use of other Machine Learning and Deep Learning algorithms for snowmelt prediction.

  generalfuture-work sectionevidence 5/5
  Keywords: future research focus integrating long-term hydrometeorological datasets machine
• Machine Learning and Deep Learning for Snowmelt Prediction: Long-Term Trends and Implications for Water Resource Management (2026) · Water Resources Management · doi

  The study identifies the challenge of accelerating snowmelt rates due to global warming, altering water availability and escalating flood risks. The study highlights the need to understand the dynamics of snowmelt and its drivers. The study notes the challenge of integrating long-term hydrometeorological datasets with Machine Learning and Deep Learning approaches.

  generalstated challengesevidence 5/5
  Keywords: study identifies challenge accelerating snowmelt rates due global
• Comparative evaluation of attention-based and transformer deep learning models for multivariate daily streamflow prediction in rivers of West Azerbaijan, Iran (2026) · Earth Science Informatics · doi

  The limitations of conventional streamflow forecasting approaches. The need for advanced deep learning models that can learn nonlinear relationships and long-term temporal dependencies in hydrological time series. The gap between practical needs of water resources management and the limitations of conventional streamflow forecasting approaches.

  generalstated research gapevidence 5/5
  Keywords: limitations conventional streamflow forecasting approaches need advanced deep

Questions about this gap