Scholarlink

Purpose <p>Suspended Sediment Concentration (SSC) plays a key role in river management, water quality, and ecosystem health. However, accurate estimation remains challenging, particularly in data-scarce rivers.</p> Materials and methods <p>This study presents different methodologies to predict SSC for four tributaries of the Rhône River using only streamflow inputs. Three models were developed: a Simplified Rating Curve Approach (SiRCA), a Random Forest (RF), and a Convolutional Long Short-Term Memory (CNN-LSTM) network. To address data scarcity, a novel Transfer Learning (TL) framework was introduced in which CNN-LSTM models were pretrained on different data sources and evaluated against the model trained from scratch on data-scarce conditions.</p> Results and discussion <p>The Deep Learning (DL) approach outperformed the other methods, followed by the RF and then the proposed rating curve model. The CNN-LSTM achieved KGE<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\('\)</EquationSource> </InlineEquation> values ranging from 0.48 to 0.85, while the RF model reached values between 0.42 and 0.79. In comparison, SiRCA yielded lower performance, with KGE<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\('\)</EquationSource> </InlineEquation> values ranging from 0.38 to 0.79. The TL framework further improved CNN-LSTM performance in data-scarce conditions, with gains of up to 0.13 in R<sup>2</sup>, and 0.47 in KGE<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\('\)</EquationSource> </InlineEquation> and reductions of up to 17% in mean absolute error and 7% in root mean squared error.</p> Conclusion <p>The findings indicate that ML and DL models significantly enhance SSC prediction performance compared with the SiRCA method. The proposed TL framework can improve model performance in data-scarce rivers when using an adequate data source, offering a practical solution for SSC prediction in poorly monitored areas.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A transfer learning approach to improve suspended sediment concentration prediction in data-scarce rivers

  • 期刊论文

  • 发表日期：

    2026年5月1日

查看全文
  • Taha Hamadene,
  • Valérie Nicoulaud-Gouin,
  • Hugo Lepage,
  • Mitra Fouladirad

摘要

Purpose

Suspended Sediment Concentration (SSC) plays a key role in river management, water quality, and ecosystem health. However, accurate estimation remains challenging, particularly in data-scarce rivers.

Materials and methods

This study presents different methodologies to predict SSC for four tributaries of the Rhône River using only streamflow inputs. Three models were developed: a Simplified Rating Curve Approach (SiRCA), a Random Forest (RF), and a Convolutional Long Short-Term Memory (CNN-LSTM) network. To address data scarcity, a novel Transfer Learning (TL) framework was introduced in which CNN-LSTM models were pretrained on different data sources and evaluated against the model trained from scratch on data-scarce conditions.

Results and discussion

The Deep Learning (DL) approach outperformed the other methods, followed by the RF and then the proposed rating curve model. The CNN-LSTM achieved KGE \('\) values ranging from 0.48 to 0.85, while the RF model reached values between 0.42 and 0.79. In comparison, SiRCA yielded lower performance, with KGE \('\) values ranging from 0.38 to 0.79. The TL framework further improved CNN-LSTM performance in data-scarce conditions, with gains of up to 0.13 in R2, and 0.47 in KGE \('\) and reductions of up to 17% in mean absolute error and 7% in root mean squared error.

Conclusion

The findings indicate that ML and DL models significantly enhance SSC prediction performance compared with the SiRCA method. The proposed TL framework can improve model performance in data-scarce rivers when using an adequate data source, offering a practical solution for SSC prediction in poorly monitored areas.