Intelligent identification method for urban inefficient residential land based on convolutional neural network

Abstract

The invention discloses an urban low-efficiency residential land intelligent identification method based on a convolutional neural network, and the method comprises the steps: S1, carrying out the classification of each residential land parcel, and enabling the classified residential land parcels to serve as a residential land parcel classification case library; s2, determining a neighborhood range of each living plot in the learning area according to the plot dominant type, constructing a neighborhood function combination data set of each living plot as a sample set, and training a convolutional neural network according to the sample set to obtain a classification prediction model; s3, classifying the living parcels of the identification area according to the living parcel classification case library to obtain an internal classification result, processing the living parcels of the identification area as the sample set in the S2, and inputting the classification prediction model in the S2 to obtain a neighborhood identification result; and S4, in combination with the internal classification result and the neighborhood identification result in the step S3, identifying and obtaining the low-efficiency residential land in the identification area. The urban inefficient residential land can be automatically identified, and the method has the advantages of dynamic identification, accuracy and objectivity, popularization of results, real-time adjustment and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of urban land management, and in particular to a method for intelligently identifying inefficient urban residential land based on a convolutional neural network. Background Art

[0002] Identifying inefficient urban residential land involves identifying residential land that is being used inefficiently and underutilized, using specific standards and methods. For various reasons, this land may not be optimally utilized, leading to resource waste or unbalanced urban development. Identifying inefficient urban residential land aims to optimize land resource allocation, improve land use efficiency, and promote sustainable urban development.

[0003] Currently, the method for identifying inefficient urban residential land mainly measures the spatial benefits of different plots by constructing an evaluation index system, while convolutional neural networks are mainly used for urban remote sensing imagery and street scene recognition and classification.

[0004] Among these, comprehensive spatial benefit evaluation utilizes indicator calculation methods such as network analysis, two-step mobile search, proportional modeling, gravity modeling, and indicator thresholding, along with static data such as statistical data and questionnaire data. It uses weighted scores from different dimensions of benefit evaluation to assess the efficiency of different plots, ultimately identifying inefficient plots. However, these methods lack the big data tools needed to accurately characterize the concentration and flow of factors and the real-time activities of land users to design and measure evaluation indicators. They also fail to consider the spatial connections created by the functional combinations surrounding the plots, making them incapable of meeting the demands of high-quality urban development in the future. Convolutional neural network image recognition, on the other hand, focuses solely on macro-level urban scene classification. Through unsupervised labeling, semantic segmentation, and feature extraction, it has been applied to remote sensing extraction, street scene recognition, and object detection. However, due to the accuracy of remote sensing imagery and street scene images, these methods focus less on the microscopic scale of cities, fail to further assess the efficiency of urban plots, and therefore fail to meet the requirements of identifying inefficient land use. Summary of the Invention

[0005] Purpose of the invention: In response to the above-mentioned shortcomings, the present invention proposes an intelligent identification method for urban inefficient residential land based on convolutional neural networks, which can automatically identify urban inefficient residential land. It has the advantages of dynamic identification, accuracy and objectivity, generalizable results and real-time adjustment, and supports large-scale automated identification that may appear in the future.

[0006] Technical solution:

[0007] The present invention provides a method for intelligently identifying inefficient urban residential land based on a convolutional neural network, comprising:

[0008] S1. Obtain the regional quality index and internal quality index of each residential plot in the learning area, classify each residential plot based on the regional quality index and internal quality index, and use the regional quality index and internal quality index as the residential plot classification case library;

[0009] S2. Determine the neighborhood range of each residential plot in the learning area based on the dominant plot type. Combine land use classification data, building land classification data, and urban geographic interest boundary data to obtain a raster for each residential plot. Construct a dataset of neighborhood function combinations for each residential plot as a sample set. Based on this dataset, a convolutional neural network is trained to obtain a classification prediction model.

[0010] S3, classify the residential blocks in the identified area according to the residential block classification case library obtained in S1, obtain internal classification results, perform the same processing on the residential blocks in the identified area as the sample set in S2, input it into the classification prediction model trained in S2, perform low-efficiency land identification, and obtain neighborhood identification results;

[0011] S4. Combining the internal classification results and neighborhood identification results obtained in S3, the inefficient residential land in the identification area is identified.

[0012] Specifically, in S1, after obtaining the regional quality index and internal quality index of the residential plots in the learning area, they are fused with the land use classification data, building land classification data and urban geographic interest boundary data to classify each residential plot accordingly.

[0013] Specifically, in S1, the regional quality indicators of each residential plot include six indicators: environmental ecology, efficient land use, convenient transportation, public livability, complete facilities, and functional flow;

[0014] The environmental ecological indicators include internal environmental compliance and green space service efficiency. The internal environmental compliance is calculated by dividing the green space area by the total land area. The green space service efficiency is calculated by converting the number of people within the residential area within the green space's reach within a set time into the service efficiency per hectare of green space.

[0015] The land use efficiency indicators include the level of land price realization, residential land efficiency and land use mix. The level of land price realization is calculated by dividing the actual transaction price of residential land into the benchmark land price of the residential land level. The residential land efficiency is calculated by dividing the population size of the residential area within the residential land into the building area. The land use mix is ​​calculated by the mix of different types of facilities and points of interest.

[0016] The accessibility indicators include public transportation completeness, work commuting convenience, and slow traffic conditions. Public transportation completeness is calculated by the ratio of land covered by public transportation stations within a set distance to the total land area. Work commuting convenience is calculated by the effective commuting time / number of effective commutes of users within the residential area. Slow traffic conditions are calculated by the length of the slow road network / residential land area.

[0017] The public livability indicators include built environment quality, public space resilience, and age-friendly index. Built environment quality is derived through user evaluation scores on mapping software or real estate software. Public space resilience is calculated by weighting the accessibility and efficiency of emergency evacuation sites and passages. The age-friendly index is calculated by the proportion of land covered by elderly care and childcare institutions within the designated space of a residential plot to the total land area.

[0018] The facility improvement index includes facility completeness and facility intelligence level. The facility completeness is calculated by the proportion of public service and commercial service facilities covered within the set space of the residential plot to the total land area. The facility intelligence level is calculated by the number of smart charging piles obtained by statistics.

[0019] The functional flow indicators include job-housing balance and spatial flow vitality. The job-housing balance is calculated by dividing the employment land area by the total residential land area in the buffer zone within a set distance from the centroid of the residential plot. The spatial flow vitality is obtained by counting the average thermal arrival frequency and residence time of mobile phone signals within the residential plot.

[0020] Specifically, in S1, the internal quality indicators of the residential plot include population concentration, spatial vitality, development intensity, building density, housing quality, living environment, facility intelligence and spatial resilience;

[0021] Among them, population concentration is calculated by total population / plot area, spatial vitality is calculated by nighttime population heat, development intensity is calculated by floor area ratio, building density is calculated by building density, residential quality is obtained by building age and style scoring, living environment is calculated by greening rate, facility intelligence is obtained by the number and distribution data of smart charging pile points of interest, and spatial resilience is obtained by emergency shelter and channel data.

[0022] Specifically, in S1, the classification of each residential plot is as follows:

[0023] By combining subjective and objective weighting methods, the weights of regional quality indicators and internal quality indicators of residential plots are determined, and the comprehensive score of each residential plot is calculated, and the residential plots are classified accordingly.

[0024] Specifically, in S1, the k-means algorithm is used to classify each residential block, and the k value is determined by observing the change in the sum of squares of clustering errors under different k values.

[0025] Specifically, in S2, the neighborhood range of each residential block is calculated using the following model:

[0026]

[0027] Among them, R base is the base radius, which is determined by the urban spatial structure; is the influence coefficient of the dominant function of residential plot u, for The global average value, A u is the area of ​​residential plot u, A ref is the median area of ​​each field in the residential plot, β is the adjustment coefficient of the dominant function, and γ is the area impact attenuation factor.

[0028] Specifically, in S2, after determining the neighborhood range of each residential block, it is enhanced through fixed position images and then subjected to category-weighted random sampling processing.

[0029] Specifically, in S2, when constructing the neighborhood functional combination dataset of each residential plot, the color of each functional area in the residential plot is mapped to a preset efficiency category code by establishing RGB values ​​and functional category indexes. The coding scheme complies with the provisions of the "Guidelines for the Classification of Land and Sea Use for National Land Space Survey, Planning, and Use Control".

[0030] Specifically, in S2, the convolutional neural network adopts the ResNet-34 network model, and introduces a cosine annealing learning rate scheduling strategy with an initial learning rate of 0.0005, combined with a 0.5 Dropout layer and an Adam optimizer for optimization.

[0031] Specifically, S2 also includes preprocessing the input residential block neighborhood function combination dataset, including: using standardized input to adapt the model architecture, and performing data enhancement processing on the residential block neighborhood function combination dataset through random scaling, cropping and horizontal flipping strategies.

[0032] Specifically, in S4, the efficiency level set I in the internal classification result is defined as {L I ,M I ,H I}, in ascending order L I <M I < H I Set update priority I > M I > HI ;

[0033] Define the efficiency level set E={L E ,M E ,H E}, in descending order H E > M E > L E Set update priority H E > M E > L E ;

[0034] Then, the type set of residential land is G=I×E={L I L E ,L I M E ,L I H E ,M I L E ,M I M E ,M I H E ,H I L E ,H I M E ,H I H E}, based on the above sequence, the judgment matrix is ​​constructed using the relative importance scale;

[0035] Then, for any two residential land types G1=(I1,E1) And G2=(I2,E2), the priority relationship is:

[0036]

[0037] Finally determine L I H E Type of residential land is the residential land with the highest priority for renewal, H I L E Type of residential land is the last residential land to be considered for renewal.

[0038] Beneficial Effects: This invention identifies inefficient urban residential land by comprehensively considering the internal classification results and neighborhood identification results of residential land. Furthermore, based on existing indicator evaluation and identification, it introduces a convolutional neural network to learn the neighborhood function combination patterns of residential plots with different efficiency levels to identify whether new residential plots are inefficient. This invention can automatically identify inefficient urban residential land, offering advantages such as dynamic identification, accuracy and objectivity, and scalable and real-time adjustment of results, supporting the potential large-scale automated identification in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 is a flow chart of the method of the present invention;

[0041] Figure 2 Schematic diagram of the area to be identified in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of clustering results of a learning area in an embodiment of the present invention;

[0043] Figure 4 Schematic diagram of constructing a residential block neighborhood function combination dataset in an embodiment of the present invention;

[0044] Figure 5 This is a graph showing the changes in loss and accuracy during model training in an embodiment of the present invention;

[0045] Figure 6 This is a flow chart of model prediction and recognition in an embodiment of the present invention;

[0046] Figure 7 This is a schematic diagram of the internal classification results of low-efficiency residential land in an embodiment of the present invention;

[0047] Figure 8 Schematic diagram of the neighborhood identification result of inefficient residential land in an embodiment of the present invention;

[0048] Figure 9 This is a schematic diagram of the comprehensive identification results of inefficient residential land in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present application is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of the present invention should have the same general meaning as those understood by persons of ordinary skill in the field to which the present invention belongs.

[0050] The present invention is based on the convolutional neural network to identify the urban inefficient residential land intelligent method Figure 1 As shown, including:

[0051] S1. Obtain the regional quality index and internal quality index of each residential plot in the learning area, classify each residential plot based on the regional quality index and internal quality index, and use the regional quality index and internal quality index as the residential plot classification case library;

[0052] In the present invention, the learning area can be set as a set area of ​​a city, such as a central urban area.

[0053] In the present invention, after obtaining the regional quality index and internal quality index of the residential plots in the learning area, they can be combined with land use classification data, building land classification data and urban geographic interest boundary data for fusion processing, and each residential plot can be classified accordingly.

[0054] In this invention, the consideration of high-quality development and factor mobility is increased, and six major indicators, namely, environmental ecology, efficient land use, convenient travel, public livability, complete facilities, and functional mobility, can be used as regional quality indicators of residential plots.

[0055] Specifically, environmental ecological indicators include internal environmental compliance and green space service efficiency. Internal environmental compliance generally uses the green space ratio to reflect the greening level of residential land and the future ecological and environmental development potential of the land. This can be calculated by dividing the green space area by the total land area. Data can be obtained through existing mapping software such as AutoNavi and Baidu Maps, or through real estate software such as Beike and Anjuke. Green space service efficiency refers to the per capita service efficiency of park green space. This can be calculated by converting the number of people within the residential area within the green space's reach within a set timeframe, which can be 15 minutes, into the service efficiency per hectare of green space. Data can be obtained through existing mapping software such as AutoNavi and Baidu Maps, or through real estate software such as Beike and Anjuke.

[0056] Land use efficiency indicators include land price realization, residential land efficiency, and land use mix. Land price realization can be calculated by dividing the actual transaction price of residential land parcels by the benchmark land price for the residential land parcel's corresponding level. Generally speaking, higher land price realization indicates greater economic vitality for the residential parcel. Data sources are available through land transaction market websites. Residential land efficiency uses the population size within a defined spatial range, which can be 100 square meters. Residential land efficiency can be calculated by dividing the population size of the residential area within the residential parcel by the building area. Data sources are available from the Tencent location-based big data platform or mapping software such as AutoNavi. Land use mix can be calculated by calculating the mix of points of interest (POIs) of different types of facilities. This reflects the vitality and sustainability of residential land use. Data sources are available from mapping software such as AutoNavi.

[0057] Accessibility indicators include public transportation availability, work commute convenience, and slow traffic. Public transportation availability can be measured by public transportation station coverage, reflecting the objective level of transportation access within a residential area. This is calculated as the ratio of land covered by public transportation stations to the total land area within a set distance range, typically 1000 meters. Data sources include mapping software such as AutoNavi. Work commute convenience can be measured by average weekday commuting time, reflecting the ease of commuting for users within the residential area. This is calculated by dividing the effective commuting time by the number of effective commutes within the residential area. Data sources include mobile phone signaling data from residents within the residential area. Slow traffic conditions can be measured by the slow traffic network density, reflecting the health and safety of residents within the future community. This is calculated as the slow traffic network length (km) divided by the residential land area (km²). Data sources include OSM road network data or mapping software such as AutoNavi.

[0058] Public livability indicators include built environment quality, public space resilience, and an age-friendly index. Built environment quality can be measured by streetscape and building age to reflect the quality of public space and community quality. This can be obtained through community ratings on mapping software such as Baidu Street View or real estate apps such as Beike or Anjuke, with the average of these ratings being used. Public space resilience can be measured by the accessibility and efficiency of emergency evacuation sites and passages to reflect a space's disaster response capacity. This can be calculated through a weighted calculation of the accessibility and efficiency of emergency evacuation sites and passages, with data sourced from databases of mapping software such as AutoNavi. The age-friendly index can be measured by the coverage rate of elderly care and childcare facilities within a designated area to reflect the inclusiveness and age-friendly nature of a residential plot. This is calculated as the proportion of land covered by elderly care and childcare facilities within a designated area within a residential plot. The designated area can be within 500 meters of the residential plot, with data sourced from mobile phone signaling data from residents within the residential plot or mapping software such as AutoNavi.

[0059] Facility improvement indicators include facility completeness and facility intelligence. Facility completeness can be measured by the coverage rate of public and commercial service facilities, which is used to evaluate the external social impact of a plot. This is calculated by measuring the proportion of land covered by public and commercial service facilities within a designated residential area, which can be within 1,000 meters of the residential area. Data can be obtained from databases provided by mapping software such as AutoNavi. The facility intelligence level can be measured by the number of smart charging stations, a key indicator of the intelligent and networked nature of future communities. This can be obtained through statistics, and the data can be obtained from mapping software such as AutoNavi or WeChat mini-programs.

[0060] Functional mobility indicators include job-housing balance and spatial mobility vitality. Job-housing balance, also known as the job-housing balance degree, reflects the likelihood of nearby employment in land use arrangements within a residential area. This can be calculated by dividing the area of ​​employment land by the total area of ​​residential land within a buffer zone within a set distance from the centroid of the residential area. The set distance can be 1000 meters, and data can be obtained from mapping software such as AutoNavi. Spatial mobility vitality can be measured by the length of stay and frequency of use in a space. This can be obtained by statistically analyzing the average frequency of mobile phone signaling arrivals and the length of stay within the residential area. This data can be obtained from mobile phone signaling data from residents within the residential area or mapping software such as AutoNavi.

[0061] In the present invention, the internal quality index of the residential plot focuses on the interior of the residential plot, which is more microscopic than the regional quality index and more in line with the actual situation of the residential plot.

[0062] Specifically, the internal quality indicators of residential plots include population concentration, spatial vitality, development intensity, building density, housing quality, living environment, facility intelligence and spatial resilience.

[0063] Among them, population concentration can be calculated by dividing the total population by the area of ​​the plot, and the data source can be obtained from the mobile phone signaling data of the internal population of the residential plot; spatial vitality can be calculated by night-time population heat, and the data source can be obtained from the mobile phone signaling data of the internal population of the residential plot; development intensity can be calculated by floor area ratio, and the data source can be obtained from remote sensing data or map software such as Baidu Map; building density can be calculated by building density, and the data source can be obtained from remote sensing data or map software such as Baidu Map; residential quality can be obtained by scoring the age and style of the building, and the data source can be obtained from street view data or real estate software such as Beike; human settlement environment can be calculated by green space ratio, and the data source can be obtained from the geographic interest boundary data (AOI) of the green space; facility intelligence can be obtained by the number and distribution data of smart charging pile points of interest (POI), and the data source can be obtained from facility POI data; spatial resilience can be obtained by emergency shelter and channel data, and the data source can be obtained from remote sensing data or map software such as Baidu Map.

[0064] In the present invention, land use classification data can be derived from some existing land use classification databases, such as the European Space Agency 10m precision data in 2020; construction land classification data can be derived from some existing construction land classification databases, such as the paper by Gong Peng et al. (Mapping Essential Urban Land Use Categories in China (EULUC-China): Preliminary Results for 2018.), and geographic interest boundary data can be obtained from map software databases of previous years, such as obtained from Baidu Maps in 2024.

[0065] In the present invention, the classification of each residential plot is specifically as follows:

[0066] By combining subjective and objective weighting methods, the weights of regional quality indicators and internal quality indicators of residential plots are comprehensively determined to avoid the entropy weight bias problem caused by data interpolation and fitting. The comprehensive score of each residential plot is then calculated, and the residential plots are classified accordingly, with each plot being assigned a classification label of inefficient, general or efficient.

[0067] Specifically, the subjective weighting can be done by using the analytic hierarchy process (AHP), and the objective weighting can be done by using the entropy weight method, and the final weight W of a certain quality indicator of the residential plot can be obtained. j ,as follows:

[0068] W j =α*W sj +(1-α)W oj ;

[0069] Among them, α represents the combination coefficient, which reflects the preference degree of subjective weight; W sj 、W oj They represent the weight of the j-th quality indicator determined by subjective and objective weighting respectively.

[0070] In this invention, the advantage of the entropy weight method is that it is very objective. However, its disadvantage is that if the data itself has some errors, the weights of the indicators may not be consistent with actual cognition. Therefore, this application incorporates the hierarchical analysis method, which determines the weights of different indicators by comparing the importance of each pair of indicators. This corrects the weight results of the entropy weight method and ensures that the final weights are consistent with actual cognition.

[0071] In the present invention, the k-means algorithm can be used to classify each residential block, and the k value is determined by observing the change of the sum of squared clustering errors (SSE) under different k values; specifically, the k value can be determined by the "elbow rule".

[0072] S2. Determine the neighborhood range of each residential plot in the learning area based on the dominant plot type. Combine land use classification data, building land classification data, and geographic interest boundary data to obtain a raster for each residential plot. Construct a dataset of neighborhood function combinations for each residential plot as a sample set. Based on this dataset, a convolutional neural network is trained to obtain a classification prediction model.

[0073] In this paper, due to the problems of inconsistent data years, inconsistent classification standards for each functional area, and inconsistent spatial coverage among multi-level land use classification data, construction land classification data, and geographic interest boundary (AOI) data, spatial overlay and data fusion processing is required for the geographic interest boundary (AOI), multi-level land use classification data, and construction land classification data. Specifically, raster algebra operations are used to process the geographic interest boundary (AOI), multi-level land use classification data, and construction land classification data to generate a global land use classification base map with consistent years, classification standards, and spatial coverage. Based on the spatial distribution characteristics of residential land, a buffer tool is used to clip the neighborhood boundaries of each plot, and output a dataset of residential plot neighborhood functional combinations. The spatial distribution characteristics of residential land can be obtained from relevant local planning documents.

[0074] Specifically, the land use classification data and the building land classification data were merged in ArcGIS, and updated using the latest geographic interest boundary data to generate a land classification base map. Then, based on the spatial distribution characteristics of residential land, the neighborhood range was determined, and the buffer analysis tool in the spatial analysis tool of ArcGIS software was used to clip the neighborhood range boundary of each plot to obtain a raster of each residential plot. Python was used to process the raster of each residential plot and export tiff images, and each plot was assigned a corresponding category label. Finally, a residential plot neighborhood function combination dataset was constructed.

[0075] In the present invention, the neighborhood range of each residential plot in the learning area is determined according to the dominant type of the plot, which can be calculated using the following model:

[0076]

[0077] Among them, R base is the base radius, which is determined by the urban spatial structure; is the influence coefficient of the dominant function of residential plot u, for The global average value, A u is the area of ​​residential plot u, A ref is the median area of ​​each field in the residential plot, β is the adjustment coefficient of the dominant function, and γ is the area impact attenuation factor.

[0078] The above model can be used to adaptively calculate the neighborhood radius of each residential block, and then determine its appropriate neighborhood range.

[0079] In the present invention, after determining the neighborhood range of each residential block in the learning area, it can also be enhanced through fixed-position images to retain its spatial position regularity to the greatest extent.

[0080] Specifically, fixed position image enhancement is as follows:

[0081] Setting the neighborhood radius, we can get:

[0082] l´(x)=l(T θ (x)),stФ(T θ )≤τ;

[0083] Among them, l´(x) represents the image at the data point x after enhancement; τ represents the transformation limit parameter, Ф(T θ )≤τ means to maintain the constraint on the defined position and control the image transformation amplitude; l(T θ (x)) represents the image at data point x through T θ (x) Transformed image; T θ (x) describes the spatial transformation at data point x, including translation and rotation and scaling. Specifically, the OpenCV library-based mesh-locked translation, random scaling and cropping, and elastic mesh deformation can be used as follows:

[0084] ;

[0085] Among them, ∆x and ∆y represent the translation of data point x in the corresponding direction, θ rot Indicates the rotation angle of data point x, s x 、s y They respectively represent the scaling amount of the data point x in the corresponding direction.

[0086] Furthermore, the classification labels for each region generally include three classification labels: low efficiency, general, or high efficiency. Among them, the majority are general labels, which are far higher than low efficiency and high efficiency labels. In order to prevent the image categories obtained by subsequent classification from being unbalanced, the present invention can use category-weighted random sampling processing on the aforementioned data set to dynamically balance the proportion of each category in the data and suppress the model's preferential learning of high-frequency categories. The details are as follows:

[0087] Use W v =n total / n v Perform inverse frequency weighted calculation on different label categories and assign weights, where n total is the total number of samples, n v is the number of samples in the vth category.

[0088] In this invention, when constructing a dataset of neighborhood functional combinations for each residential plot, the colors of each functional area within the residential plot can be mapped to a preset performance category code by establishing RGB values ​​and functional category indexes. Specifically, the coding scheme complies with the provisions of the "Guidelines for the Classification of Land and Sea Use for National Land and Spatial Survey, Planning, and Usage Control."

[0089] Specifically, the color mapping of each functional area is the preset performance category code, as follows:

[0090] Assume there are w functional categories, each category corresponds to a unique RGB color value C i , i represents the category number. For each data point x, its functional category is c(x), then the mapping is expressed as Color(x)=C c(x) , where Color(x) represents the RGB color value corresponding to the data point x.

[0091] Specifically, the process of mapping the colors of each functional area to the preset performance category codes is as follows:

[0092] Obtain the neighborhood function combination dataset of each residential block, perform synchronous spatial clipping on each residential block and its neighborhood according to the boundaries, construct a white mask for each residential block, color its neighborhood according to its function and the aforementioned coding scheme, and perform color fusion calculation, finally obtaining the neighborhood range function color image of each residential block, and then construct the neighborhood function combination dataset of each residential block.

[0093] In the present invention, 70% of the samples in the sample set are randomly selected as the training set, 20% of the samples are used as the validation set, and the remaining 10% are used as the test set.

[0094] In the present invention, the convolutional neural network can adopt the ResNet-34 network model, which first performs spatial grid encoding on the images in the training set to generate a multi-channel feature matrix containing the functional identification of the plot, the road network density and the spatial correlation strength, and then adaptively learns the spatial coupling law of the functional combination in the neighborhood window through the convolution kernel, and uses the nonlinear activation function to capture the functional compatibility threshold. Finally, the multi-scale spatial dependency is aggregated through the global pooling layer to output a probability distribution map representing the relationship between the plot and the surrounding functional combination. It can simultaneously analyze the spatial distribution of the white area of ​​the residential plot, the surrounding colored functional area, and the connection form of the road network. Compared with models such as VGG, the ResNet-34 network model can accurately identify the image features of the road network and functional mixture through hierarchical feature fusion under fewer parameters, and maintain a strict correspondence between color and function.

[0095] In the present invention, when training the convolutional neural network, a cosine annealing learning rate scheduling strategy with an initial learning rate of 0.0005 can be introduced, combined with a 0.5 Dropout layer and an Adam optimizer for optimization to prevent the model from overfitting the local detail features of non-critical areas.

[0096] In the present invention, in order to expand data diversity while maintaining the topological invariance of the road space and the spatial correspondence between functional color blocks, the input residential block neighborhood function combination dataset needs to be preprocessed, including: using standardized input to adapt the model architecture, and performing data enhancement processing on the residential block neighborhood function combination dataset through random scaling, cropping and horizontal flipping strategies.

[0097] S3, classify the residential blocks in the identified area according to the residential block classification case library obtained in S1, obtain internal classification results, perform the same processing on the residential blocks in the identified area as the sample set in S2, input it into the classification prediction model trained in S2, perform low-efficiency land identification, and obtain neighborhood identification results;

[0098] S4. Combining the internal classification results and neighborhood identification results obtained in S3, the inefficient residential land in the identification area is identified.

[0099] In this invention, the principle that the lower the efficiency evaluation in the internal classification results, the stronger the renewal demand, and the higher the efficiency evaluation in the neighborhood identification results, the higher the renewal potential is adopted to construct a comprehensive judgment matrix for inefficient residential land, and the following is obtained:

[0100] Define the effectiveness level set I in the internal classification result = {L I ,M I ,H I}, in ascending order L I < M I < H I Set update priority I > M I > H I ;

[0101] Define the efficiency level set E={L E ,M E ,H E}, in descending order H E > M E > L E Set update priority H E > M E > L E ;

[0102] Then, the type set of residential land is G=I×E={L I LE ,L I M E ,L I H E ,M I L E ,M I M E ,M I H E ,H I L E ,H I M E ,H I H E}, based on the above sequence, the judgment matrix is ​​constructed using the relative importance scale.

[0103] Then, for any two residential land types G1=(I1,E1) And G2=(I2,E2), the priority relationship is:

[0104]

[0105] Finally, we can determine L I H E The residential land of type (low inside and high outside) is the residential land with the highest priority for renewal. I L E Residential land of this type (high inside and low outside) is the last residential land to be considered for renewal.

[0106] The present invention provides a specific embodiment, which is carried out in Nanjing. The Huayuan Road area of ​​Xuanwu District is selected as the identification area. The specific range is enclosed by "Xuanwu Avenue - Longpan Road - Bancang Lane - Jiangwangmiao Street", covering 10 communities in Suojin Village Street and Xuanwu Lake Street, with a total area of ​​more than 360 hectares. Figure 2 As shown in the figure, based on the administrative divisions of the "Nanjing Land and Spatial Master Plan (2021-2035)", the central urban area of ​​Nanjing was selected as the learning area (the area where the learning dataset will be constructed), including the Jiangnan main city and the Jiangbei new main city, covering an area of ​​808 square kilometers. Because the identification area belongs to the central urban area of ​​Nanjing, the actual dataset production excluded the land parcel data within the identification area.

[0107] In this embodiment, residential plot data is interpreted from Nanjing urban data, and the fields include boundary point coordinates, name, address and construction year, which can be obtained through Anjuke or Baidu Maps; mobile phone signaling data comes from China Unicom operators, and the fields include thermal value, residence time, visit frequency and grid ID; POI data is obtained through Baidu Maps, and the fields include name, facility type and UID; OSM road network data is obtained from OpenStreetMap, and the fields include name, road grade and UID; land use classification data comes from the 2020 European Space Agency 10m precision data, and the fields include name, land use type and UID; construction land classification data comes from the paper by Gong Peng et al. (Mapping Essential UrbanLand Use Categories in China (EULUC-China): Preliminary Results for 2018.), and the fields include name, land use type and UID; urban AOI data is obtained from Baidu Maps in 2024, and the fields include name, land use type and UID.

[0108] In this embodiment, during the clustering process of the k-means algorithm, according to the elbow plot results, it can be seen that the optimal k value for clustering is around 2 or 3. Returning to the specific problem, it is not reasonable to classify residential land into two categories (i.e., "low efficiency" and "high efficiency"), so the final k value is 3, and the residential plots are divided into three categories: "low efficiency", "general" and "high efficiency". Finally, the clustering results of residential plots in the central urban area of ​​Nanjing are obtained, as shown in Figure 2. Figure 3 shown.

[0109] In this embodiment, when the convolutional neural network is trained to obtain a classification prediction model, the neighborhood range is determined to be a radius of 1 km according to the "Nanjing 15-minute Community Living Circle Planning Guidelines".

[0110] In this example, the coding scheme of the residential plot neighborhood function combination dataset follows the "Guidelines for Classification of Land and Sea Use for National Land Space Survey, Planning, and Use Control" (hereinafter referred to as the "Guidelines"), and refers to Figure 4 , where: ① The white block in the center of the image (RGB (255,255,255)) represents residential land (0701); ② The adjacent colored areas are matched with color value functions according to the secondary category coding in the "Guidelines" (e.g., commercial land red series: 0901→RGB (230,50,45)); ③ The road network is expressed with a black broken line (RGB (0,0,0)), and the line width is positively correlated with the road grade.

[0111] In this embodiment, when the convolutional neural network is trained, the loss and accuracy changes are shown in the figure below: Figure 5As shown in the figure, training was terminated when the loss function converged to a stable state and no overfitting was observed. The model effectively discriminated the efficiency of residential plots within the morphological constraints of core residential plots, achieving a validation accuracy of 85.1%. Finally, generalization performance evaluation based on the test set demonstrated that the model possessed good feature discrimination capabilities and transfer adaptability.

[0112] In this example, the weight of the validation set with the highest accuracy in the model training (epoch=84) is saved, and independent data that has not been trained on the model is selected as the test set. The weight file and the test set are input into the Python file to obtain the recognition result of the model prediction classification, as shown in the following example: Figure 6 shown.

[0113] In this embodiment, the land parcels to be identified in the Huayuan Road area are classified according to the residential land parcel classification case library obtained in S1, and the internal classification results are obtained, such as Figure 7 The same processing as the sample set in S2 is performed on the plots to be identified in the Huayuan Road area, and the results of neighborhood identification of inefficient residential land in the Huayuan Road area are obtained by inputting them into the classification prediction model trained in S2. Figure 8 It can be seen that the results show that inefficient residential plots are concentrated in Suojin Village and Bancang Street Community on the south side of the Huayuan Road area, with a small number in Jiangwangmiao Community and Huayuan Road Community in the central part.

[0114] In this embodiment, the aforementioned internal classification results and neighborhood identification results are combined, and the plot types with the median performance level of the two are eliminated to obtain four types of decision results: "low inside and high outside", "low inside and low outside", "high inside and high outside", and "high inside and low outside". Since the "high inside and low outside" type was not detected in the study area and formed an empirical empty set, three types of valid classification results were finally determined. The performance level evaluation of the internal classification results characterizes the intensity of the plot renewal demand and the degree of material environment decay. The performance level evaluation of the neighborhood identification results reveals the rationality of the spatial functional pattern and the gradient of renewal potential. Based on this, the "low inside and high outside" type plots are established as the priority renewal objects. The results show that the priority renewal plots in the Huayuan Road area include Yinghai Apartment, the Skin Research Institute Residential Area, the Tianyi Company Dormitory, Suojin Village 2, Suojin Village No. 5, and Gangzi Village No. 63. Figure 9 By comparing the list of old residential areas to be renovated in Xuanwu District, Nanjing, with the list of urban construction projects in the past five years, it can be seen that most of the above-mentioned plots have been included in the renovation list, which shows that this method has a high accuracy.

[0115] This method identifies inefficient urban residential land by comprehensively considering both the internal classification results of residential land and the neighborhood identification results. Building on existing indicator-based evaluation and identification, it introduces a convolutional neural network to learn the neighborhood functional combination patterns of residential plots of varying efficiency levels to identify whether new residential plots are inefficient. This method can automatically identify inefficient urban residential land, offering advantages such as dynamic identification, accuracy and objectivity, and scalable and real-time adjustment. This method supports the potential for large-scale automated identification in the future.

[0116] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention (including the claims) is limited to these examples. Within the scope of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present invention as described above, which are not provided in detail for the sake of simplicity.

[0117] The embodiments of the present invention are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for intelligently identifying inefficient urban residential land based on convolutional neural networks, characterized in that: include: S1. Obtain the regional quality index and internal quality index of each residential plot in the learning area, classify each residential plot based on the regional quality index and internal quality index, and use the regional quality index and internal quality index as the residential plot classification case library; S2. Determine the neighborhood range of each residential plot in the learning area based on the dominant plot type. Combine land use classification data, building land classification data, and urban geographic interest boundary data to obtain a raster for each residential plot. Construct a dataset of neighborhood function combinations for each residential plot as a sample set. Based on this dataset, a convolutional neural network is trained to obtain a classification prediction model. S3, classify the residential blocks in the identified area according to the residential block classification case library obtained in S1, obtain internal classification results, perform the same processing on the residential blocks in the identified area as the sample set in S2, input it into the classification prediction model trained in S2, perform low-efficiency land identification, and obtain neighborhood identification results; S4. Combining the internal classification results and neighborhood identification results obtained in S3, the inefficient residential land in the identification area is identified.

2. The intelligent identification method for inefficient urban residential land according to claim 1 is characterized in that: In said S1, after obtaining the regional quality index and internal quality index of the residential plots in the learning area, they are fused with the land use classification data, the building land classification data and the urban geographic interest boundary data, and each residential plot is classified accordingly.

3. The intelligent identification method for inefficient urban residential land according to claim 1 is characterized in that: In S1, the regional quality indicators of each residential plot include six indicators: environmental ecology, efficient land use, convenient transportation, public livability, complete facilities, and functional flow; The environmental ecological indicators include internal environmental compliance and green space service efficiency. The internal environmental compliance is calculated by dividing the green space area by the total land area. The green space service efficiency is calculated by converting the number of people within the residential area within the green space's reach within a set time into the service efficiency per hectare of green space. The land use efficiency indicators include the level of land price realization, residential land efficiency and land use mix. The level of land price realization is calculated by dividing the actual transaction price of residential land into the benchmark land price of the residential land level. The residential land efficiency is calculated by dividing the population size of the residential area within the residential land into the building area. The land use mix is ​​calculated by the mix of different types of facilities and points of interest. The accessibility indicators include public transportation completeness, work commuting convenience, and slow traffic conditions. Public transportation completeness is calculated by the ratio of land covered by public transportation stations within a set distance to the total land area. Work commuting convenience is calculated by the effective commuting time / number of effective commutes of users within the residential area. Slow traffic conditions are calculated by the length of the slow road network / residential land area. The public livability indicators include built environment quality, public space resilience, and age-friendly index. Built environment quality is derived through user evaluation scores on mapping software or real estate software. Public space resilience is calculated by weighting the accessibility and efficiency of emergency evacuation sites and passages. The age-friendly index is calculated by the proportion of land covered by elderly care and childcare institutions within the designated space of a residential plot to the total land area. The facility improvement index includes facility completeness and facility intelligence level. The facility completeness is calculated by the proportion of public service and commercial service facilities covered within the set space of the residential plot to the total land area. The facility intelligence level is calculated by the number of smart charging piles obtained by statistics. The functional flow indicators include job-housing balance and spatial flow vitality. The job-housing balance is calculated by dividing the employment land area by the total residential land area in the buffer zone within a set distance from the centroid of the residential plot. The spatial flow vitality is obtained by counting the average thermal arrival frequency and residence time of mobile phone signals within the residential plot.

4. The intelligent identification method for inefficient urban residential land according to claim 1 or 3, characterized in that: In S1, the internal quality indicators of the residential plot include population concentration, spatial vitality, development intensity, building density, housing quality, living environment, facility intelligence and spatial resilience; Among them, population concentration is calculated by total population / plot area, spatial vitality is calculated by nighttime population heat, development intensity is calculated by floor area ratio, building density is calculated by building density, residential quality is obtained by building age and style scoring, living environment is calculated by greening rate, facility intelligence is obtained by the number and distribution data of smart charging pile points of interest, and spatial resilience is obtained by emergency shelter and channel data.

5. The intelligent identification method for inefficient urban residential land according to claim 1 is characterized in that: In S1, the classification of each residential plot is specifically as follows: By combining subjective and objective weighting methods, the weights of regional quality indicators and internal quality indicators of residential plots are determined, and the comprehensive score of each residential plot is calculated, and the residential plots are classified accordingly.

6. The intelligent identification method for inefficient urban residential land according to claim 1 is characterized in that: In S2, the neighborhood range of each residential block is calculated using the following model: Among them, R base is the base radius, which is determined by the urban spatial structure; is the influence coefficient of the dominant function of residential plot u, for The global average value, A u is the area of ​​residential plot u, A ref is the median area of ​​each field in the residential plot, β is the adjustment coefficient of the dominant function, and γ is the area impact attenuation factor.

7. The intelligent identification method for inefficient urban residential land according to claim 1 is characterized in that: In S2, after determining the neighborhood range of each residential block, it is enhanced through fixed-position images, and then subjected to category-weighted random sampling processing. After that, standardized input is used to adapt the model architecture, and the residential block neighborhood function combination dataset is enhanced through random scaling, cropping and horizontal flipping strategies.

8. The intelligent identification method for inefficient urban residential land according to claim 1 is characterized in that: In S2, when constructing the neighborhood functional combination dataset of each residential plot, the color of each functional area in the residential plot is mapped to a preset efficiency category code by establishing RGB values ​​and functional category indexes. The coding scheme complies with the provisions of the "Guidelines for the Classification of Land and Sea Use for National Land Space Survey, Planning, and Use Control".

9. The intelligent identification method for inefficient urban residential land according to claim 1 is characterized in that: In S2, the convolutional neural network adopts the ResNet-34 network model, and introduces a cosine annealing learning rate scheduling strategy with an initial learning rate of 0.0005, combined with a 0.5 Dropout layer and an Adam optimizer for optimization.

10. The intelligent identification method for inefficient urban residential land according to claim 1, characterized in that: In said S4, the efficiency level set I in the internal classification result is defined as {L I ,M I ,H I }, in ascending order L I < M I < H I Set update priority I >M I > H I ; Define the efficiency level set E={L E ,M E ,H E }, in descending order H E > M E > L E Set update priority H E > M E > L E ; Then, the type set of residential land is G=I×E={L I L E ,L I M E ,L I H E ,M I L E ,M I M E ,M I H E ,H I L E ,H I M E ,H I H E }, based on the above sequence, the judgment matrix is ​​constructed using the relative importance scale; Then, for any two residential land types G1=(I1,E1) and G2=(I2,E2), their priority relationship is: Finally determine L I H E Type of residential land is the residential land with the highest priority for renewal, H I L E This type of residential land is the last residential land to be considered for renewal.

Images