|
|
| |
|
|
| 本文已被：浏览 67次 下载 0次 |
|
|
|
| 基于多源数据的声环境改善关键住宅建筑识别技术研究 |
|
路晓东, 李政媛, 谢庄秀, 祝培生, 赵明辉
|
|
大连理工大学
|
|
| 摘要: |
| 交通噪声已成为威胁城市居民公共健康的重要环境问题。存量更新背景下，亟需快速识别建成区声环境质量需要改善的关键住宅建筑。研究立足中国城市特点，基于多源数据构建适用于我国的识别模型。以大连市典型中心城区作为研究样本，首先通过软件模拟获取建筑交通噪声评价，将立面噪声超过70dBA的住宅作为关键建筑进行识别，并整合GIS与多源数据，从道路特征、规划布局、建筑属性三个维度构建特征指标。其次，运用随机森林（RF）、极限梯度提升（XGBoost）和逻辑回归（Logistic）三种机器学习方法分别建立关键建筑识别模型，对比实效；最后引入SHAP分析方法以解释特征变量的作用机制。结果表明：在多模型对比中，RF模型表现最优，整体准确率75%。SHAP分析显示，规划布局要素在关键住宅建筑识别中占据主导地位。具体来看，路网“标准化角度选择度”、“公交兴趣点”、路网“标准化角度整合度”、“噪声暴露度”是判定建筑是否为关键住宅建筑的重要因素。该研究可以为城市声环境精细化治理提供决策依据。 |
| 关键词: 机器学习 交通噪声 多源数据 随机森林模型 SHAP分析 |
| DOI： |
| 分类号:TU112 |
| 基金项目:语言声房间和厅堂的语音传输指数（STI）评价阈值研究,国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Multi-source Data-based Identification Technology of Key Residential Buildings for Sound Environment Improvement |
|
LuXiaodong, LiZhengyuan, XieZhuangxiu, ZhuPeisheng, ZhaoMinghui
|
|
Dalian University of Technology
|
| Abstract: |
| Traffic noise has become a major environmental challenge threatening urban public health, and the problem is particularly severe in Chinese cities undergoing rapid urban renewal, where the need for accurate and efficient identification of residential buildings with poor acoustic conditions is urgent. To address this challenge, this study develops a multi-source data-based identification framework at the building scale, specifically adapted to China’s urban context, and applies it to a typical central urban area in Dalian. A total of 1,247 residential buildings were screened through GIS-based analysis, focusing on those within 50 meters of major roads, and traffic noise levels were simulated in SoundPLAN with input from UAV-monitored traffic flow data. Buildings with fa?ade noise exceeding 70 dBA, were labeled as key buildings in need of noise mitigation. Following the “emission–transmission–reception” framework, a comprehensive feature system was established: road-related indicators such as road area density, length index, total network length, Euclidean distances to 2/4/6/8-lane roads, and arterial adjacency formed the “emission” dimension; planning-related “transmission” features included space syntax metrics—standardized angular choice (NACH) and standardized angular integration (NAIN)—as well as building density, floor area ratio, compactness, and bus-stop points of interest (POIs); and “reception” attributes captured building morphology (height, area, convexity, fractal dimension, polygon edges, shape index) and Resident population. And a novel “noise exposure degree” indicator was introduced to quantify each building’s openness of sound transmissionto road noise sources by analyzing the proportion of unobstructed line-of-sight connections to evenly distributed point sources along roads. All road and planning indicators are calculated and summarized based on the buffer zone and assigned to the corresponding buildings. For the building attribute indicators. The dataset was preprocessed using Pearson correlation and variance inflation factor diagnostics to remove redundancy and multicollinearity, normalized through min–max scaling, and balanced via SMOTE oversampling. Stratified sampling produced an 80/20 train–test split, and three models—Logistic Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)—were trained using grid search with five-fold cross-validation. Model evaluation employed multiple metrics including accuracy, precision, recall, F1 score, confusion matrix, and ROC-AUC. Results demonstrate that the RF model achieved the most balanced and robust performance, with accuracy of 75.2%, precision of 0.598, recall of 0.716, F1 score of 0.652, and an AUC of 0.79, outperforming XGBoost (higher precision but much lower recall, 0.481) and LR (weakest across all metrics). After being verified in the research area, RF successfully identified 358 ultra-noisy residential buildings and 827 residential buildings that met the requirements, with a recall rate of 89%, achieving a recall rate of 89%. To enhance interpretability, SHAP (SHapley Additive exPlanations) analysis was used to quantify the contribution and directional effect of each variable. Results revealed a clear hierarchy: planning-related features dominated, followed by road characteristics, while building morphology contributed least. NACH was the strongest predictor, reflecting that areas with higher route choice and accessibility are more vulnerable to noise exceedance; bus-stop POI density also had a significant positive impact, indicating intensified transport activity; NAIN captured the effect of network connectivity in concentrating traffic flows and amplifying noise; and both noise exposure degree and population density further increased exceedance risks in highly exposed, densely inhabited areas. Meanwhile, distance to two-lane roads showed a strong negative effect consistent with attenuation laws, while intrinsic building attributes such as height and polygon edges exerted only minor influence. These findings indicate that external spatial and transport-related factors, rather than intrinsic building form, largely determine building-level noise risks. Methodologically, the study demonstrates the integration of GIS-based multi-source features, advanced ensemble learning models, and SHAP interpretability to bridge the gap between coarse noise mapping and fine-grained building-level assessments. Practically, the proposed framework offers a cost-effective, scalable, and transparent tool for municipalities lacking noise maps, enabling rapid identification of priority residential buildings requiring intervention under China’s stock-renewal agenda. The results also provide actionable guidance: urban planning strategies that focus on improving road-network accessibility and connectivity, regulating traffic around bus nodes, and reducing building exposure are likely to be more effective than modifying building morphology alone. While the framework is robust, it has limitations, such as the feature system has not yet taken into account the microscopic physical properties such as building materials and structures. Future research should incorporate multi-scale, dynamic datasets to further strengthen robustness and applicability. Overall, the Random Forest-based identification model with SHAP interpretation achieves reliable predictive performance and transparent feature insights, providing scientific evidence to support refined governance of the urban acoustic environment and promoting healthier, more sustainable residential neighborhoods in Chinese cities. |
| Key words: Machine learning Construction noise Multi-source data Random forest model SHAP Analysis |
|
|
|
|
