Urban river health evaluation in semi-arid Xi'an, China: a hybrid RF-DNN framework integrating multi-source and...

Area of Science:

• Environmental Science
• Hydrology
• Ecology

Background:

• Semi-arid metropolitan rivers face unique "Type 4 degradation" due to urbanization, including fragmented connectivity and habitat loss.
• Existing assessment methods, designed for humid regions, fail to capture the complexities of water-limited river systems.

Purpose of the Study:

• To develop and validate a hybrid machine learning framework for diagnosing Type 4 river degradation.
• To identify critical ecological thresholds and key drivers impacting river health in semi-arid urban environments.

Main Methods:

• A hybrid Random Forest-Deep Neural Network (RF-DNN) framework was developed, integrating feature selection and SHAP interpretability.
• A comprehensive indicator system with 26 metrics across hydrology, hydrochemistry, habitat, biology, and socioeconomic factors was used.
• The framework was evaluated using data from 24 monitoring sites across four rivers over two years.

Main Results:

• The RF-DNN framework achieved high accuracy (7.6% deviation, 0.76 Cohen's κ, 0.79 validity), outperforming traditional methods and standalone models.
• Urbanization rate, ammonia nitrogen, population density, and flow velocity were identified as dominant predictors of river health.
• Critical ecological thresholds were defined for impervious cover (45%) and ammonia nitrogen (0.8 mg/L), with synergistic effects under low dissolved oxygen.

Conclusions:

• The developed RF-DNN framework provides a robust tool for mechanistic diagnosis of river health in semi-arid urban watersheds.
• Findings support targeted restoration strategies, including flow reallocation, nutrient and oxygen management, and habitat reconstruction.
• The study highlights the need for tailored assessment protocols for water-limited urban river systems.