Scholarlink

<p>Rapid urbanization in semi-arid regions subjects metropolitan rivers to a distinctive form of hydrologic-physical impairment—herein designated as Type 4 degradation—characterized by connectivity fragmentation, baseflow depletion, and physical habitat homogenization, in which groundwater over-extraction elevates zero-flow days and concrete channelization eliminates substrate heterogeneity. Conventional assessment protocols, developed predominantly for humid-region perennial streams, inadequately capture the multidimensional connectivity dynamics critical to water-limited systems. To address this diagnostic gap, this study developed a hybrid Random Forest–Deep Neural Network (RF-DNN) framework that integrates objective feature selection with nonlinear modeling capacity and couples SHAP-based interpretability for mechanistic inference. A comprehensive indicator system encompassing 26 metrics across hydrology, hydrochemistry, physical habitat, biological organization, and socioeconomic pressure was constructed and evaluated using 96 observations collected from 24 monitoring sites across Xi’an’s Ba, Chan, Feng, and Hei Rivers over 2 years. Under leave-one-river-out (LORO) cross-validation, the framework achieved a relative deviation of 7.6 ± 1.3%, Cohen’s <i>κ</i> of 0.76 ± 0.05, and ecological validity ρ of 0.79 ± 0.07, outperforming AHP-Fuzzy (15.2 ± 2.8%) and AHP-TOPSIS (17.1 ± 3.3%) methods as well as stand-alone XGBoost, SVM, Full-RF, and Full-DNN models (all paired-test <i>p</i> &lt; 0.05). SHAP analysis identified urbanization rate (mean |SHAP|= 0.068), ammonia nitrogen (0.032), population density (0.028), and flow velocity (0.022) as dominant predictors of health variation, and revealed critical ecological transition zones—bootstrap-validated at 45% impervious cover (95% CI 38–52%) and 0.8&#xa0;mg/L NH₃-N (95% CI 0.65–0.94&#xa0;mg/L), with synergistic toxicity amplification of approximately 42% under co-occurring oxygen depletion (DO &lt; 4.0&#xa0;mg/L). Spatial assessment revealed a systematic longitudinal gradient (Hei &gt; Feng &gt; Chan &gt; Ba), with the downstream urban reaches of Ba River exhibiting a mean dry-season RHI of only 0.2482, while wet-season improvements produced grade-level transitions at 10 of 24 sites (41.67%). The framework advances mechanistic diagnosis for semi-arid urban watersheds and supports evidence-based prioritization of flow reallocation, coupled NH₃-N/DO management, and riparian habitat reconstruction for restoration planning.</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

Urban river health evaluation in semi-arid Xi’an, China: a hybrid RF-DNN framework integrating multi-source and SHAP-based interpretability

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Yajie Qu,
  • Tao Yang,
  • Haiyan Li,
  • Yilin Liao,
  • Xiangnan Lei,
  • Jiaqi Yuan

摘要

Rapid urbanization in semi-arid regions subjects metropolitan rivers to a distinctive form of hydrologic-physical impairment—herein designated as Type 4 degradation—characterized by connectivity fragmentation, baseflow depletion, and physical habitat homogenization, in which groundwater over-extraction elevates zero-flow days and concrete channelization eliminates substrate heterogeneity. Conventional assessment protocols, developed predominantly for humid-region perennial streams, inadequately capture the multidimensional connectivity dynamics critical to water-limited systems. To address this diagnostic gap, this study developed a hybrid Random Forest–Deep Neural Network (RF-DNN) framework that integrates objective feature selection with nonlinear modeling capacity and couples SHAP-based interpretability for mechanistic inference. A comprehensive indicator system encompassing 26 metrics across hydrology, hydrochemistry, physical habitat, biological organization, and socioeconomic pressure was constructed and evaluated using 96 observations collected from 24 monitoring sites across Xi’an’s Ba, Chan, Feng, and Hei Rivers over 2 years. Under leave-one-river-out (LORO) cross-validation, the framework achieved a relative deviation of 7.6 ± 1.3%, Cohen’s κ of 0.76 ± 0.05, and ecological validity ρ of 0.79 ± 0.07, outperforming AHP-Fuzzy (15.2 ± 2.8%) and AHP-TOPSIS (17.1 ± 3.3%) methods as well as stand-alone XGBoost, SVM, Full-RF, and Full-DNN models (all paired-test p < 0.05). SHAP analysis identified urbanization rate (mean |SHAP|= 0.068), ammonia nitrogen (0.032), population density (0.028), and flow velocity (0.022) as dominant predictors of health variation, and revealed critical ecological transition zones—bootstrap-validated at 45% impervious cover (95% CI 38–52%) and 0.8 mg/L NH₃-N (95% CI 0.65–0.94 mg/L), with synergistic toxicity amplification of approximately 42% under co-occurring oxygen depletion (DO < 4.0 mg/L). Spatial assessment revealed a systematic longitudinal gradient (Hei > Feng > Chan > Ba), with the downstream urban reaches of Ba River exhibiting a mean dry-season RHI of only 0.2482, while wet-season improvements produced grade-level transitions at 10 of 24 sites (41.67%). The framework advances mechanistic diagnosis for semi-arid urban watersheds and supports evidence-based prioritization of flow reallocation, coupled NH₃-N/DO management, and riparian habitat reconstruction for restoration planning.