基于街景图像与深度学习的城市活力智能识别方法研究-以青岛为例

深度学习算法

语义分割

需积分: 1 33 浏览量 2024-10-13 12:37:54 上传评论 1 收藏 14.87MB PDF 举报

资源推荐

资源详情

资源评论

International Journal of Applied Earth Observation and Geoinformation 123 (2023) 103476

Available online 4 September 2023

Deep exploration of street view features for identifying urban vitality: A

case study of Qingdao city

Zepeng Ma

College of Geodesy and Geomatics, Shandong University of Science and Technology, Qingdao 266590, China

ARTICLE INFO

Keywords:

Street view image

Deep learning algorithms

Semantic segmentation

SHAP method

Social perception

Urban vitality

ABSTRACT

Urban vitality has signicant practical implications for urban management and planning. In this study, we

propose a comprehensive research framework that combines street scene images, point of interest (POI) data,

road network data, and residential land data, and employs deep learning algorithms to explore the characteristics

and inuencing factors of urban vitality from a social perception perspective. By designing multi-scale semantic

segmentation models, emotion perception models, and street perception models, we deeply explore the street

features of the city. At the block level, we use weighted calculation methods to quantify urban vitality by

combining POI data and residential land data, accurately characterizing the city. Finally, we analyze the driving

factors of urban vitality using random forest and SHAP methods. The research results show that Chengyang

District and Laoshan District have advantages in visual perception, while Shibei District and Shinan District

exhibit advantages in urban vitality. The overall urban vitality in the main urban area of Qingdao City is low,

with high scores in emotion perception and visual perception, but low scores in transportation accessibility and

facility convenience. Visual perception factors play a signicant role in urban vitality, highlighting the impor-

tance of urban street beautication and humanized design in economic development and environmental con-

struction. The analytical results of this study contribute to optimizing urban spatial features and provide

references for urban planning and construction.

1. Introduction

The competitiveness of cities is closely related to national strength

(Zhang, 2020). In the 21st century, China has rapidly entered the track

of urbanization, bringing not only opportunities but also challenges. In

recent years, “urban issues” have become increasingly serious, such as

inadequate urban planning, poor urban openness and inclusiveness, lack

of cooperation and communication between cities, and the emergence of

empty or ghost cities in various regions (Zhao and Zhao, 2022). The

main reason for these problems is the uneven expansion of cities (Dang,

2023). In order to ensure high-quality and stable development of cities,

it is crucial to enhance urban vitality (Zhang, 2018).

The denition and research focus of urban vitality have been

continuously evolving. Early studies on urban vitality mainly focused on

the economic and market vitality of cities (Della Spina, 2019), consid-

ering urban vitality as the self-development ability of cities in terms of

production relations, as well as prosperity and stability in economic,

environmental, and social aspects. However, with a shift in research

thinking, scholars have started to understand urban vitality from

different perspectives and dimensions. For instance, Sulis et al., (2018)

pointed out that urban vitality is related to people’s behaviors in a

certain place. Liang et al., (2022) and Zhu et al., (2020) argued that

urban vitality can be represented through human activities and the

interactive features generated within the city.

A comprehensive measurement of urban vitality is a focal issue in

urban studies, yet there is currently a lack of a xed indicator system

that can fully and accurately reect the level of urban vitality. Previous

studies have proposed different indicators and evaluation methods. For

example, Gülden Demet and Giritlio

glu, (2008) used the number of

residences, population, commercial areas, and the price per square

meter of land as indicators of urban vitality. Nia, (2021) argued that

urban aesthetics design has a direct impact on the level of urban vitality.

Fuentes et al., (2020) evaluated the vitality of communities in Santiago,

Chile, based on the diversity of urban form, architecture, and spatial

characteristics. These evaluation methods are subjective to a certain

extent and cannot comprehensively and accurately measure urban

vitality.

In recent years, scholars have closely associated urban vitality with

human behavior. They use data collection and analysis of people’s be-

haviors to understand the level of urban vitality. For example, Delcl

os-

Ali

o et al., (2019) analyzed and validated Jacobs’s theory of urban vi-

tality based on GIS data and data tracked on smartphones. Botta and

Guti

errez-Roig, (2021) argued that a vibrant urban environment refers

to areas in a city with high levels of human activity and interaction, and

Contents lists available at ScienceDirect

International Journal of Applied Earth

Observation and Geoinformation

journal homepage: www.elsevier.com/locate/jag

https://doi.org/10.1016/j.jag.2023.103476

Received 2 June 2023; Received in revised form 25 August 2023; Accepted 29 August 2023

International Journal of Applied Earth Observation and Geoinformation 123 (2023) 103476

they demonstrated that combining mobile phone data with crowd-

sourced urban features can lead to a better understanding of urban ac-

tivity. Mao and Zhong, (2020) explored the level of urban vitality and its

inuencing factors based on POI (Points of Interest) data and location

services data, emphasizing the subjectivity of individuals in urban vi-

tality. Yin et al., (2019) used mobile app location data to describe the

ow of people across towns, exploring the distribution patterns of urban

vitality. This approach is closer to the essence of urban vitality and has

better objectivity. However, the difculties in obtaining mobile

signaling and location data, as well as the high data costs, greatly limit

research on urban vitality.

Street view images, as a new type of data source, has advantages of

wide coverage, providing street-level landscape information, and low

data collection costs. Street view imagery contains a wealth of visual

information about urban spaces, including static urban built environ-

ments and dynamic pedestrians or vehicles on streets, which offers new

perspectives and dimensions for urban research (Zhang, et al., 2019;

Voordeckers et al., 2021). Previous studies have utilized street view

imagery data in various research areas, such as investigating the impact

of greenery on mental health, exploring comfortable urban travel ex-

periences, and extracting semantic scenes. With the continuous expan-

sion of street view imagery data sources and advancements in the eld of

articial intelligence, it brings both new opportunities and challenges

for research in geographic spatial-related domains (Gao, 2020).

Deep convolutional neural networks (CNNs) have emerged as a

pivotal and effective tool in the advancement of computer vision,

playing an increasingly vital role in image recognition and classication

(Berman et al., 2014). Furthermore, they offer new opportunities for

modeling and perceiving geographic locations (Wang et al., 2020). In

order to quantitatively analyze the semantic aspects of urban scenes and

extract their deeper connotations, researchers have started employing

deep learning methods for scene classication, semantic segmentation,

and object detection in street-view imagery(Kruse et al., 2021; Yu and

Wang, 2016; Yin et al., 2015; Liang et al., 2017; Markevych et al., 2017).

However, challenges such as category confusion continue to plague

scene classication and semantic segmentation techniques in the pro-

cessing of street-view images (Tian et al., 2019a; Tian et al., 2019b). The

introduction of complex semantic scene factors has necessitated higher

demands for scene classication and semantic segmentation tasks

(Zhang, 2018).

This study presents a comprehensive research framework to over-

come the limitations of previous studies. Specically, we employ ne-

scale street view imagery and deep dive into urban street features

from the perspective of social perception by utilizing a designed multi-

scale semantic segmentation model, emotional perception model, and

street perception model. Subsequently, at the plot scale, we quantify

urban vitality through weighted calculations, combining point of in-

terest data and residential land data, to accurately depict urban char-

acteristics. Lastly, we analyze the factors driving urban vitality using

random forest and SHAP methods. This work makes the following

contributions: (1) It designs three deep learning models that fully exploit

feature information in street view data; (2) It provides a method to

measure urban vitality features at a ne scale; (3) By quantifying human

perception of locations and uncovering regional characteristics, it pro-

motes tailored approaches to balance urban development; (4) It eluci-

dates the non-linear relationship between urban street view features and

urban vitality. This work enables a more comprehensive and accurate

evaluation of urban vitality levels, providing effective guidance for

urban development. It is of great signicance in addressing current

urban issues and promoting healthy urban growth.

2. Study area and data

2.1. Study area

Qingdao is located in the southeastern part of Shandong Province. It

is the economic center of Shandong Province and one of the earliest

coastal cities in China. With the process of urbanization, Qingdao’s

streetscapes have exhibited more distinctive features. The main urban

area of Qingdao city is chosen as the study area to analyze the urban

vitality at the residential block level. The study area is shown in Fig. 1.

2.2. Data source and preprocessing

2.2.1. Street view data

Street view images are a novel source of data that provides users with

real-life visual representations of streets, buildings, and other urban

physical environments. In this study, Baidu Maps’ street view imagery

was selected as the data source. The street view imagery was obtained

using Python web scraping with the Baidu Street View Map API.

During the data preprocessing phase, road network data of the study

area was downloaded from OpenStreetMap and integrated into ArcMap

for road network simplication. To focus on the internal features of the

city, highways and roads where street view vehicles cannot capture

imagery were excluded. Road centerlines were extracted and compared

with available street view imagery on Baidu Maps . Road segments

without street view coverage were removed. Following the research

approach of (Yang et al., 2021), sampling points were obtained at in-

tervals of 50 m along the road centerlines to ensure comprehensive

photo coverage and balanced data size. The latitude and longitude of

each sampling point were computed, and the nearest geocoded location

for panoramic imagery was retrieved using the Baidu API. For each

location, the angle between the road and the front-facing street view was

calculated, and street view images were collected from four specic

camera angles, with a xed pitch angle of 20

◦

. Images were crawled for

sampling points at positions

+ 0

◦

+ 90

◦

+ 180

◦

, and

+ 270

◦

, as

shown in Fig. 2.

2.2.2. Semantic segmentation dataset

The semantic segmentation dataset used in this study was obtained

from the open-source Cityscapes dataset. The dataset includes a total of

3,475 nely annotated images, consisting of 2,975 training images and

500 validation images. Each image in the dataset has a resolution of

1024 × 2048 pixels.

Due to the fact that the Cityscapes dataset contains a total of 34

categories, this study focuses on 19 specic categories. Considering the

image size of the dataset, which is 1024 × 2048 pixels, in order to reduce

the GPU memory usage during model training, the image size was

resized to 768 × 768 pixels. Additionally, data augmentation techniques

were applied to the dataset. The performance of the model was evalu-

ated using the validation set for accuracy assessment. The sample data

and labels of the dataset are shown in Fig. 3.

2.2.3. MIT places pulse 2.0 dataset

The MIT Places Pulse 2.0 dataset is a crowdsourced dataset for urban

landscape perception in street view images, publicly released by the MIT

Media Lab. It consists of 110,988 images captured between 2007 and

2012, sampled from 56 major cities in 28 countries across six continents.

The global coverage of the dataset makes it an ideal choice for training

urban perception models. The dataset is formed by collecting ratings

from different participants through information collection platforms

(Fig. 4) who rated pairs of random street view images from six di-

mensions: “beautiful,” “depressing,” “lively,” “wealthy,” and “boring.”

Zhang et al., (2018a) and Zhang et al., (2018b) employed a statistical

method similar to the concept of “schedule intensity” in sports compe-

titions (Park and Newman, 2005) to calculate the perception scores for

each dimension, resulting in scores distributed within the range of [0,

10]. In our research, we sampled this dataset to specify our dataset for

predicting urban perception.

2.2.4. COCO Captions datasets

The COCO Captions dataset, also known as Microsoft COCO

Z. Ma

剩余13页未读，继续阅读

评论收藏

内容反馈

pk_xz123456

粉丝: 2750
资源: 3914

基于街景图像与深度学习的城市活力智能识别方法研究-以青岛为例

深度学习项目-街景字符识别.zip

基于街景影像和深度学习技术的城市流动商贩空间分布制图(毕设&课设论文参考).pdf

基于深度学习的上海城市街景与景观美学公众认知研究.pdf

基于大规模街景图片人工智能分析的精细化城市修补方法研究.pdf

多源数据与深度学习支持下的人本城市设计：以上海苏州河两岸城市绿道规划研究为例.pdf

城市街景数据集图像分割

基于深度卷积神经网络的街景门牌号识别方法.pdf

深度学习数据集之图像分割数据集：城市街景图像分割数据集

基于深度学习的城市街道景观意向识别.pdf

基于街景图的中国省会城市天空开阔度研究

天池学习赛——街景数字识别

深度学习-行人重识别实战(2020)

densenet模型-深度学习CNN训练识别街景分类-不含数据集图片-含逐行注释和说明文档.zip

街道建成环境中的城市非正规性 基于北京老城街景图片的人工打分与机器学习相结合的识别探索.pdf

基于深度学习的图像美观度评价.pdf

基于深度学习街景影像解译和景感生态学的视域环境定量解读.pdf

深度学习图像识别数据集：城市街道和城市雾霾2分类数据集

城市尺度下的建筑色彩定量化测度——基于街景数据与机器学习的人本视角分析.pdf

基于深度学习的交通街景识别_牛雪磊 (1).caj

pspnet街景分割及mask-rxnn目标识别

基于注意力机制的街景图像语义分割方法.docx

电信设备-一种基于街景属性信息的定位图像显示方法及系统.zip

阿姆斯特丹市的城市街景图像数据集

基于深度学习的图像实例分割.pdf

YOLOv8-deepsort 实现智能车辆目标检测+车辆跟踪+车辆计数

YOLOv8网络结构图，自制visio文件，yolov8.vsds，需要的自取，在原有的基础上直接改就行了

Transformer模型实现长期预测并可视化结果（附代码+数据集+原理介绍）

最新资源

街道建成环境中的城市非正规性基于北京老城街景图片的人工打分与机器学习相结合的识别探索.pdf