A novel hybrid SCS-LSTM model for intelligent flood forecasting in data-scarce small mountainous catchments
Study Region: Ungauged small catchments in Dawukou District, Northwest China. Study Focus: To address the critical data scarcity in PUBs, we propose a hybrid SCS-LSTM framework that integrates the Soil Conservation Service (SCS) hydrological model with Long Short-Term Memory (LSTM) neural networks....
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-10-01
|
| Series: | Journal of Hydrology: Regional Studies |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2214581825004367 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Study Region: Ungauged small catchments in Dawukou District, Northwest China. Study Focus: To address the critical data scarcity in PUBs, we propose a hybrid SCS-LSTM framework that integrates the Soil Conservation Service (SCS) hydrological model with Long Short-Term Memory (LSTM) neural networks. The SCS model is innovatively employed to convert zero-dimensional rainfall point data (e.g., peak rainfall intensity) into physically consistent one-dimensional flood hydrographs, simulating key hydrological processes (e.g., infiltration, runoff generation, and routing). By generating 117 synthetic time-series hydrographs, this method provides process-informed training datasets for machine learning models. Three hybrid architectures (SCS-BPNN, SCS-GRU, SCS-LSTM) are rigorously compared to identify the optimal approach for mountainous PUBs. New Hydrological Insights for the Region: The SCS-LSTM model achieved superior performance, with mean Nash-Sutcliffe Efficiency (NSE) and correlation coefficient (R) values exceeding 0.83, outperforming SCS-BPNN and SCS-GRU by 6.1 % (NSE) and 11.0 % (R), respectively. Key advancements include:1.Data Scarcity Mitigation: SCS-generated hydrographs overcome the ''zero-dimensional data trap'' in PUBs, enabling robust training of machine learning models where traditional data-driven approaches fail.2.Process-Aware Forecasting: The hybrid framework captures both hydrological dynamics (e.g., flood peak timing) and nonlinear rainfall-runoff relationships, critical for flash flood prediction in mountainous regions.3.Regional Transferability: This methodology offers a scalable solution for ungauged catchments in monsoonal Asia, where > 60 % of basins lack observational data, yet face escalating flood risks under climate change. |
|---|---|
| ISSN: | 2214-5818 |