Long-term streamflow forecasting in data-scarce regions: Insightful investigation for leveraging satellite-derived data, Informer architecture, and concurrent fine-tuning transfer learning

Journal of Hydrology, Journal Year: 2024, Volume and Issue: 631, P. 130772 - 130772

Published: Feb. 2, 2024

Language: Английский

---

Shifted dominant flood drivers of an alpine glacierized catchment in the Tianshan region revealed through interpretable deep learning

npj Climate and Atmospheric Science, Journal Year: 2025, Volume and Issue: 8(1)

Published: Jan. 25, 2025

Language: Английский

---

Hydrological Projections in the Third Pole Using Artificial Intelligence and an Observation‐Constrained Cryosphere‐Hydrology Model

Earth s Future, Journal Year: 2024, Volume and Issue: 12(4)

Published: April 1, 2024

Abstract The water resources of the Third Pole (TP), highly sensitive to climate change and glacier melting, significantly impact food security millions in Asia. However, projecting future spatial‐temporal runoff changes for TP's mountainous basins remains a formidable challenge. Here, we've leveraged long short‐term memory model (LSTM) craft grid‐scale artificial intelligence (AI) named LSTM‐grid. This has enabled production hydrological projections seven major river TP. LSTM‐grid integrates monthly precipitation, air temperature, total mass (total_GMC) data at 0.25‐degree grid. Training employed gridded historical evapotranspiration sets generated by an observation‐constrained cryosphere‐hydrology headwaters TP during 2000–2017. Our results demonstrate LSTM grid's effectiveness usefulness, exhibiting Nash‐Sutcliffe Efficiency coefficient exceeding 0.92 verification periods (2013–2017). Moreover, monsoon region exhibited higher rate increase compared those westerlies region. Intra‐annual indicated notable increases spring runoff, especially where meltwater contributes runoff. Additionally, aptly captures before after turning points highlighting growing influence precipitation on reaching maximum total_GMC. Therefore, offers fresh perspective understanding spatiotemporal distribution high‐mountain glacial regions tapping into AI's potential drive scientific discovery provide reliable data.

Language: Английский

---

Deep learning algorithms and their fuzzy extensions for streamflow prediction in climate change framework

Journal of Water and Climate Change, Journal Year: 2024, Volume and Issue: 15(2), P. 832 - 848

Published: Feb. 1, 2024

Abstract The present study analyzes the capability of convolutional neural network (CNN), long short-term memory (LSTM), CNN-LSTM, fuzzy CNN, LSTM, and CNN-LSTM to mimic streamflow for Lower Godavari Basin, India. Kling–Gupta efficiency (KGE) was used evaluate these algorithms. Fuzzy-based deep learning algorithms have shown significant improvement over classical ones, among which is best. Thus, it further considered projections in a climate change context four-time horizons using four shared socioeconomic pathways (SSPs). Average 2041–2060, 2061–2080, 2081–2090 are compared that 2021–2040 changed by +3.59, +7.90, +12.36% SSP126; +3.62, +8.28, +12.96% SSP245; +0.65, −0.01, −0.02% SSP370; +0.02, +0.71, +0.06% SSP585. In addition, two non-parametric tests, namely, Mann–Kendall Pettitt were conducted ascertain trend point projected streamflow. Results indicate provides more precise prediction than others. identified variations across different SSPs facilitate valuable insights policymakers relevant stakeholders. It also paves way adaptive decision-making.

Language: Английский

---

Improved streamflow prediction accuracy in Boreal climate watershed using a LSTM model: A comparative study

PLOS Water, Journal Year: 2025, Volume and Issue: 4(4), P. e0000359 - e0000359

Published: April 21, 2025

Streamflow plays a vital role in water resource management and environmental impact assessment. This study is novel application of the Long Short-Term Memory (LSTM) model, type recurrent neural network, for real-time streamflow prediction Upper Humber River Watershed western Newfoundland. It also compares performance LSTM model with physically based SWAT model. The was optimized by tuning hyperparameters adjusting window size to balance capturing historical data ensuring stability. Using single input variables such as daily average temperature or precipitation, achieved high Nash-Sutcliffe Efficiency (NSE) 0.95. In comparison, results show that delivers more competitive performance, achieving an NSE 0.95 versus SWAT’s 0.77, percent bias (PBIAS) 0.62 compared 8.26. Unlike SWAT, does not overestimate flows excels predicting low flows. Additionally, successfully predicted using data. Despite challenges interpretability generalizability, demonstrated strong particularly during extreme events, making it valuable tool cold climates where accurate forecasts are crucial effective management. highlights potential model’s

Language: Английский

---

TSWS: An observation-based streamflow dataset of Tianshan Mountains watersheds (1901–2019)

Scientific Data, Journal Year: 2025, Volume and Issue: 12(1)

Published: April 29, 2025

Due to scarcity of data and complex hydrological conditions in the Tianshan region, long-term complete streamflow are lacking. This study produced a multi-basin dataset, named watershed (TSWS) by comparing results Hydrologiska Byråns Vattenavdelning Long Short-Term Memory models, analyzed spatiotemporal variation streamflow.TSWS dataset provides daily for 56 watersheds monthly 89 Mountains 1901-2019. The simulations 40 (daily scale) 70 (monthly passed S-tests (Nash-Sutcliffe efficiency ≥0.5, percent bias ≤25%, ratio root-mean-square error standard deviation measured ≤0.7). showed an overall increasing trend streamflow, especially from 1990 2019; spatially, it higher west south, lower east north. first comprehensive simulation its long time series will provide important reference climatic studies.

Language: Английский

---

Monthly Streamflow Forecasting for the Irtysh River Based on a Deep Learning Model Combined with Runoff Decomposition

Water, Journal Year: 2025, Volume and Issue: 17(9), P. 1375 - 1375

Published: May 2, 2025

The mid- and long-term hydrological forecast is important for water resource management disaster prevention. Moreover, forecasts in the region with poorly observed field meteorological data are a great challenge traditional models due to complexity of processes. To address this challenge, machine learning model, particularly deep model (DL), provides new tool improving accuracy runoff prediction. In study, we took Irtysh River, one longest rivers Central Asia well-known trans-boundary river basin poor observations, as an example develop based on LSTM combined decomposition by Maximal Overlap Discrete Wavelet Transform (MODWT) process target variables predicting monthly streamflow. We also proposed XGBoost-SHAP (Extreme Gradient Boost-SHapley Additive Explanations) method identification predictors from large-scale indices streamflow forecast. results suggest that MODWT shows robustness between training test period. better performance than benchmark without illustrated increased NSE. well identified nonlinear relationship streamflow, determined can be physically explained. Compared mutual information method, improves study illustrate ability forecast, methods developed provide effective approach improve prediction scarcely catchments.

Language: Английский

---

Enhancing the streamflow simulation of a process-based hydrological model using machine learning and multi-source data

Ecological Informatics, Journal Year: 2024, Volume and Issue: 82, P. 102755 - 102755

Published: Aug. 3, 2024

Streamflow simulation is crucial for flood mitigation, ecological protection, and water resource planning. Process-based hydrological models machine learning algorithms are the mainstream tools streamflow simulation. However, their inherent limitations, such as time-consuming large data requirements, make achieving high-precision simulations challenging. This study developed a hybrid approach to simultaneously improve accuracy computational efficiency of simulation, which integrates Block-wise use TOPMODEL (BTOP) model into eXtreme Gradient Boosting (XGBoost), i.e., BTOP_XGB. In this approach, BTOP generates simulated using Latin hypercube sampling algorithm instead calibration reduce costs. Then, XGBoost combines with multi-source errors. which, serval input variable selection employed choose relevant inputs remove redundant information model. The validated compared standalone at three stations in Jialing River basin, China. results show that performance BTOP_XGB significantly better than models. NSE Beibei, Xiaoheba, Luoduxi increases by 54%, 21%, 83%, respectively. Meanwhile, time saved >90% original calibrated BTOP. less affected parameter sample sizes amounts, demonstrating robustness simplifies complexity enhances stability learning, jointly improving reliability provides potential shortcut over basins areas or limited observed data.

Language: Английский

---

Streamflow Prediction with Time-Lag-Informed Random Forest and Its Performance Compared to SWAT in Diverse Catchments

Water, Journal Year: 2024, Volume and Issue: 16(19), P. 2805 - 2805

Published: Oct. 2, 2024

This study introduces a time-lag-informed Random Forest (RF) framework for streamflow time-series prediction across diverse catchments and compares its results against SWAT predictions. We found strong evidence of RF’s better performance by adding historical flows time-lags meteorological values over using only actual values. On daily scale, RF demonstrated robust (Nash–Sutcliffe efficiency [NSE] > 0.5), whereas generally yielded unsatisfactory (NSE < 0.5) tended to overestimate up 27% (PBIAS). However, provided monthly predictions, particularly in with irregular flow patterns. Although both models faced challenges predicting peak snow-influenced catchments, outperformed an arid catchment. also exhibited notable advantage terms computational efficiency. Overall, is good choice predictions limited data, preferable understanding hydrological processes depth.

Language: Английский

---

Performance of LSTM over SWAT in Rainfall-Runoff Modeling in a Small, Forested Watershed: A Case Study of Cork Brook, RI

Water, Journal Year: 2023, Volume and Issue: 15(23), P. 4194 - 4194

Published: Dec. 4, 2023

The general practice of rainfall-runoff model development towards physically based and spatially explicit representations hydrological processes is data-intensive computationally expensive. Physically models such as the Soil Water Assessment tool (SWAT) demand spatio-temporal data expert knowledge. Also, difficulty complexity compounded in smaller watershed due to constraint models’ inability generalize hydrologic processes. Data-driven can bridge this gap with their mathematical formulation. Long Short-Term Memory (LSTM) a data-driven Recurrent Neural Network (RNN) architecture, which better suited solve time series problems. Studies have shown that LSTM competitive performance hydrology studies. In study, comparative analysis SWAT Cork Brook shows results from were flow prediction NSE 0.6 against 0.63, respectively, given limited availability data. do not overestimate high flows like SWAT. However, both these struggle low values estimation. Although interpretability, explainability, use across different datasets or events outside training may be challenging, are robust efficient.

Language: Английский

---