Urban low-efficiency land intelligent identification method fusing social media data

Abstract

The invention discloses a social media data fused urban inefficient land intelligent identification method, which comprises the following steps of: realizing multi-source data space fusion by acquiring a social media geographic marking text, an urban updating policy file, a remote sensing image and basic geographic data; a natural language processing and topic modeling method is used for extracting keywords from the city update text, a topic-feature word mapping table is constructed, a RoBERTa-l arge model and a semantic rule are combined to calculate a text emotion score, and a land use negative emotion map is generated; and constructing a multi-dimensional evaluation system including indexes such as plot ratio and road network density, calculating a composite index by adopting an AMP algorithm, dynamically allocating weights through an entropy weight method, and judging the low-efficiency land. And a white list area is introduced for verification and parameter iterative optimization, and finally, an identification result is visually displayed in WebG IS and AR platforms, so that the accuracy and decision support capability of urban low-efficiency land identification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of urban planning and land use analysis, in particular to a city low-efficiency land intelligent identification method fusing social media data and multi-source spatial information, and is specifically applied to the related fields of city renewal, inventory planning, public participation intelligent decision-making and the like. BACKGROUND

[0002] With the city entering the inventory update stage, low-efficiency land identification as an important prelink of city space optimization and update planning is more and more concerned by the government and planning decision-making departments. The existing city low-efficiency land identification method mainly relies on remote sensing image interpretation, planning index calculation or expert manual evaluation and the like, although it has certain spatial scale adaptability, but there are obvious limitations in data update frequency, public perception expression, intelligent processing capacity and the like. On the one hand, the traditional method often ignores the subjective experience and social perception of the residents in the city space, and it is difficult to identify the area with low public space use satisfaction and poor environmental feeling; on the other hand, the existing method lacks a systematic technical path in multi-source heterogeneous data fusion, quantitative index calculation and dynamic weight adaptation, resulting in the deficiency of the identification result in the accuracy, explainability and decision support effect.

[0003] In recent years, as an important carrier of daily expression of city residents, the data of social media has the characteristics of strong timeliness, clear geographical attribute, rich emotional information and the like, which provides a new possibility for supplementing the traditional spatial data. At the same time, the development of intelligent technologies such as natural language processing, large language model, topic modeling and multi-index comprehensive evaluation also provides a method support for the identification and modeling of city space problems. However, at present, there is still a lack of effective integration of social media data, city update text information and remote sensing / planning data, and a new identification path fusing social media data is needed to be put forward to improve the city update efficiency and the intelligent level of governance. SUMMARY

[0004] In order to solve the deficiencies mentioned in the background art, the purpose of the present application is to provide a city low-efficiency land intelligent identification method fusing social media data, which can intelligently identify the city low-efficiency land.

[0005] The purpose of the present application can be realized by the following technical solutions:

[0006] A city low-efficiency land intelligent identification method fusing social media data, comprising the following steps:

[0007] Step one: multi-source data acquisition and fusion processing

[0008] Social media geotagged text data and low efficiency land policy documents were collected through the Weibo Open API and the Ministry of Natural Resources website, respectively. Urban renewal texts were scanned using a high-speed document laser scanner with a horizontal and vertical scanning speed greater than 140ppm / 210ipm. Geospatial image data with a resolution better than 1m were obtained through satellite remote sensing technology, and urban basic data, including building outline vector layers, building construction year attribute tables, and land use status maps, were also obtained simultaneously. After converting the GPS coordinates in the social media data to local coordinates through the WGS84 coordinate system, they were superimposed on the GIS base map to achieve spatial matching.

[0009] Step two: Urban renewal and low efficiency land keyword modeling

[0010] Urban renewal planning texts and low efficiency land policy documents were input, and NLP tools such as Jieba and spaCy were used for word segmentation and noise removal, eliminating stop words and non-spatial description words. Word frequency was calculated based on TF-IDF weighting, and word vectors were generated using the BERT model to identify urban renewal and low efficiency land keywords. LDA / dynamic topic models were pre-trained to establish a theme-feature word mapping table for four types of renewal areas: residential, industrial, commercial, and service.

[0011] Step three: Generation of land use negative sentiment map based on social media data

[0012] The four types of theme-feature word mapping tables constructed were used as a keyword table to search all the microblog text data obtained. All microblog texts and posting locations containing the keyword table were extracted, and the microblog texts and posting locations were classified according to the theme-feature word correspondence. A pre-trained sentiment analysis model using Chinese text as the training set, RoBERTa-large, was used in combination with semantic enhancement rules for degree adverbs and negative words to calculate the sentiment scores of all classified microblog texts. A sentiment score of 0.5 was set as the negative sentiment threshold, and a land use negative sentiment map was generated using kernel density estimation.

[0013] Step four: Calculation of renewal potential based on AMP algorithm and entropy weight method

[0014] An index system was constructed including plot ratio, road network density, rent vacancy rate, resident vitality, and spatial satisfaction. The AMP algorithm was used to calculate the composite index, and the entropy weight method was used to dynamically assign index weights. A composite index of 50 was set as the threshold for low efficiency land, and land parcels with a composite index below the threshold were marked.

[0015] Step five: White list verification and parameter iteration

[0016] Second, the white list area is determined by the age attribute of the building and the government purple line file. The deviation rate between the low efficiency land judgment result and the spatial distribution of the white list area is compared. If the deviation rate exceeds 10%, the weight parameters of the AMPA algorithm are adjusted until the deviation rate is less than or equal to 10%.

[0017] Step six: visualization and decision support

[0018] The generated negative emotion map and low efficiency land distribution heat map are superimposed and calculated in the WebGIS platform to identify the final urban low efficiency land. The urban road network, urban functional area, and other vector spatial information are combined with the urban low efficiency land recognition result to generate an urban low efficiency land distribution map. The urban spatial status map, resident negative emotion map, and urban low efficiency land distribution map are placed in an AR interactive touch screen with a resolution of not less than 7680x4320 pixels and a screen size of not less than 86 inches to realize interactive display of the three views.

[0019] Further, the urban low efficiency land intelligent identification method fusing social media data has the characteristics that the NLP tool is used for word segmentation and noise reduction processing in step two, which specifically includes the following sub-steps:

[0020] The input urban renewal planning text and low efficiency land policy file are segmented by using an NLP tool (such as Jieba or spaCy) to divide the continuous text sequence T = {t1, t2, …, tn} into a series of word sequences W = {w1, w2, …, wm}, where n is the total number of characters in the text, m is the total number of segmented words, and m ≤ n. Noise reduction processing is performed, including removing stop words and non-spatial description words. The stop word table S_stopwords contains common but meaningless words for topic analysis, such as "of", "is", "in", etc. The non-spatial description word table S_nonspace contains words that do not involve geographical location or spatial feature description. The cleaned word sequence W_clean is filtered by the following logic:

[0021]

[0022] Where "∈" means "belongs to", "≠" means "does not belong to", and "∧" means logical "and" operation. This step ensures that the subsequent analysis focuses on key information related to urban renewal and low efficiency land.

[0023] Further, the urban low efficiency land intelligent identification method fusing social media data has the characteristics that the keyword identification in step two specifically includes the following steps:

[0024] The weight w_t of each word t is calculated by using the TF-IDF algorithm, and the calculation formula is as follows: ​

[0025] w_t = tf(t) * log(N / (df(t) + 1))

[0026] where tf(t) denotes the term frequency of term t in the current document, N denotes the total number of documents, and df(t) denotes the number of documents containing term t. Then, each term t is converted into a semantic vector vt e Rd using a pre-trained BERT model to capture the semantic information of the term. The cosine similarity sim(vt, vc) = (vt · vc) / (||vt|| * ||vc||) between vt and a pre-defined city renewal and low utility land related concept vector vc e Rd is calculated. The TF-IDF weight wt of the term and the cosine similarity sim(vt, vc) thereof are combined, and the term with a score higher than a preset threshold Threshold is selected as the final recognized keyword.

[0027] Further, the city low utility land intelligent identification method fusing social media data, characterized in that the step two of pre-training the LDA dynamic topic model to establish a topic-feature word mapping table specifically comprises the following steps:

[0028] The pre-trained LDA topic model is used to analyze the keywords and context information identified in step two. Let K topics be identified by the model, and each topic θk (k 1, 2, …, K) is a probability distribution about the words in the vocabulary V, denoted as θk p(w|θk) (p(w1|θk), p(w2|θk), …, p(w|θk)). By setting a topic-word probability threshold ε, the words with a probability higher than ε in each topic are screened out to form the feature word set Wk of the topic. The screened topics and their corresponding feature word sets Wk are mapped to the four pre-set types of renewal areas (residential, industrial and mining, commercial, and service), and a topic-feature word mapping table T {(θk, Wk, Type) | k 1, 2, …, K, Type e {residential, industrial and mining, commercial, service}} is established, where Type represents the type of the renewal area to which it belongs, and an example is as follows:

[0029]

[0030] Further, the city low utility land intelligent identification method fusing social media data, characterized in that the step three of constructing the RoBERTa-large model to calculate the text sentiment score specifically comprises the following steps:

[0031] The NLPCC2014 micro-blog sentiment classification corpus is used as the training set, 80% of the data is randomly extracted as the training set, and 20% is used as the test set. The cross-entropy loss function is used for model fine-tuning training. During the training process, the batch size is set to 16, the learning rate is 2e-5, and the training rounds are 3 rounds (epoch). The fine-tuned RoBERTa-large model is used for vector encoding and forward inference of the text to be analyzed. The probability distribution of the sentiment classification label is output, and the probability corresponding to the positive label is taken as the original sentiment score of the text. The preset semantic rules are used for post-processing of the score. The text containing negative words is identified and the sentiment polarity is reversed. The text containing degree adverbs is identified and the score amplitude is added or subtracted according to the weight coefficient set in the dictionary. The final sentiment score after fusing the semantic structure is output. A fixed threshold (0.5) is used to classify the sentiment score. The text with a score lower than the threshold is identified as a negative sentiment text, and the sentiment recognition result output based on deep semantic understanding and semantic rule fusion is realized.

[0032] Further, the urban low-efficiency land intelligent identification method fusing social media data has the characteristics that the step three of generating the land use negative sentiment map by the kernel density method specifically includes the following steps:

[0033] The kernel density difference parameter is set in the GIS platform, the Gaussian kernel function is used for spatial smoothing processing of the negative sentiment points, and continuous spatial modeling of the sentiment intensity is realized. The bandwidth parameter is used to control the heat influence range, and the city scale is set to 300-500 meters to balance the local sensitivity and overall readability and generate a rasterized heat output.

[0034] Further, the urban low-efficiency land intelligent identification method fusing social media data has the characteristics that the step four of constructing the index system of volume rate, road network density, rent vacancy rate, resident vitality, and space satisfaction specifically includes the following steps:

[0035] The volume rate is calculated by the computer through the formula volume rate The road network density is calculated by the computer through the formula road network density The rent vacancy rate is calculated by the computer through the formula rent vacancy rate , wherein R is the average rent of the research unit, Sv is the comprehensive average intensity of the building floor area with low nighttime light, and Sb is the sum of the building floor area in the research unit. The resident vitality is calculated by the entropy method by weighting the population dynamic density, walking activity intensity, business activity, public space utilization, and social interaction intensity. The specific formula is as follows: population dynamic density walking activity intensity business activity Se is the area of the top 20% of the night light intensity of the commercial land of the study unit, S is the area of the study unit, and public space utilization rate Social interaction intensity Resident vitality where wj is the index weighting weight, and Xj is the calculated index; the spatial satisfaction is calculated by the spatial alienation degree index formula The calculation is as follows, where Ni is the frequency of positive identification type keywords in the social media in the study unit, and N is the total frequency of keywords in the social media in the study unit.

[0036] Further, the urban low-efficiency land intelligent identification method based on social media data has the characteristics that the step four of calculating the composite index by using the AMP algorithm and dynamically allocating the index weight by using the entropy weight method specifically comprises the following steps.

[0037] After the Adjusted Mazziotta-Pareto (AMP) algorithm is used to normalize each index, non-compensatory weighted integration is performed. The normalization of the original index is performed. Given the original index matrix X = {x ij}, it needs to be converted into a normalized matrix R = {r ij}, and the formula is as follows:

[0038]

[0039] where Min xj and Max xj are the minimum and maximum values of the index j, respectively, and Ref xj is the reference value. The weighted mean of each unit is calculated. If the index weight w j satisfies 0 < w j < 1 and , then the weighted mean of the unit i is The calculation formula is as follows:

[0040]

[0041] The weighted standard deviation of the unit i is calculated Combined with the weighted mean, the weighted coefficient of variation is obtained The formula for calculating the relative imbalance of the index distribution is as follows:

[0042]

[0043] The AMPI index is synthesized, and the calculation formula is as follows:

[0044]

[0045] Finally, the weights of the indicators are calculated based on the differences reflected by information entropy, using the following formula:

[0046]

[0047] Where m is the total number of evaluation indicators, and all weights satisfy ∑w j =1.

[0048] Furthermore, the method for intelligent identification of inefficient urban land use that integrates social media data is characterized in that step five, which involves filtering whitelist areas based on construction year and function, specifically includes the following steps:

[0049] Based on building age, areas with buildings ≤ 5 years old and areas with buildings 5 ​​years < 15 years old and renovation / renewal rates ≥ 30% are selected. Based on function, areas with commercial area > 50% are first identified, and then further filtered using the High-End Commercial Index (HCI). The formula is as follows: Luxury brand density = number of luxury brand stores / area; international brand proportion = number of international brand stores / total number of commercial stores; nighttime economy intensity = 22.00-02.00 passenger flow / average daily passenger flow; areas with HCI > 0.7 are selected; the spatial intersection of building age and function is obtained by overlaying the dimensions, and whitelist areas are selected.

[0050] Furthermore, the aforementioned method for intelligent identification of inefficient urban land use by integrating social media data is characterized in that, in step five, the deviation from the actual data exceeds 10%, specifically including the following steps:

[0051] Spatial distribution deviation rate using formula The calculation shows that A = inefficient output area of ​​the model, and B = whitelist area.

[0052] The beneficial effects of this invention are:

[0053] 1. The speed and breadth of resident opinion collection have been significantly improved. By utilizing the Weibo Open API in conjunction with geotagged text retrieval and topic-feature mapping tables, millions of residents' social media texts can be automatically captured and processed within 24 hours. Compared to the average cycle of 30 to 60 days for traditional questionnaires and field surveys, the information collection speed has been increased by more than 40 times. The social media coverage spans across regions and age groups, and the spatial distribution density of the data sample has increased by about 3.5 times. This enables a comprehensive reflection of the opinions and emotional state of residents in various functional areas of the city in a short period of time, meeting the real-time needs of urban renewal decisions for broad public opinion.

[0054] 2. By combining the AMP algorithm and the entropy weight method to dynamically allocate weights, and introducing government purple lines and the year of construction as a whitelist verification mechanism, the spatial deviation rate between the inefficient land use judgment results and authoritative data is improved by more than 50% compared with the deviation rate of the traditional static weighting method (about 20% to 25%). Through multi-source data fusion (social media, remote sensing, GIS basic data, etc.), the spatial positioning accuracy of inefficient land use is improved to within ±5 meters, significantly enhancing the spatial matching degree and policy adaptability of the judgment results.

[0055] 3. Based on WebGIS and AR interactive touch screen (resolution ≥7680×4320 pixels, screen size ≥86 inches), multi-view overlay display is achieved, realizing millisecond-level responsive switching between urban status map, negative sentiment map and inefficient land use distribution map, with interaction latency controlled within 0.2 seconds; compared with traditional two-dimensional GIS platform, the three-dimensional interactive visualization solution improves the spatial understanding efficiency of decision-makers by about 40%, and can quickly complete multiple rounds of updated solution comparison and adjustment in a single meeting. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0057] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0058] like Figure 1 As shown, a method for intelligent identification of inefficient urban land use that integrates social media data includes the following steps:

[0059] Step 1: Multi-source data acquisition and fusion processing

[0060] We collected geotagged text data from social media and policy documents on inefficient land use (from the Ministry of Natural Resources website) through the public interface (Weibo Open API); we scanned urban renewal texts using a high-speed document laser scanner with a horizontal and vertical scanning speed greater than 140ppm / 210ipm; we acquired geospatial imagery data with a resolution better than 1m through satellite remote sensing technology, and simultaneously acquired basic urban data, including building outline vector layers, building construction year attribute tables, and land use status maps; we converted the GPS coordinates in the social media data to the local coordinate system using the WGS84 coordinate system and then overlaid them onto the GIS base map to achieve spatial matching.

[0061] Step Two: Thematic Modeling of Urban Renewal and Inefficient Land Use

[0062] Input the urban renewal planning text and inefficient land use policy documents, and use NLP tools (such as Jieba, spaCy) for word segmentation and denoising, removing stop words and non-spatial description words; calculate the word frequency based on TF-IDF weighting, combine with the BERT model to generate word vectors, and identify keywords for urban renewal and inefficient land use; pre-train the LDA / dynamic topic model to establish a theme-feature word mapping table for the four major types of renewal areas: residential, industrial and mining, commercial, and service.

[0063] The word segmentation and denoising using NLP tools in step 2 specifically include the following sub-steps:

[0064] Use NLP tools (such as Jieba, spaCy) to segment the input urban renewal planning text and inefficient land use policy documents, and divide the continuous text sequence T = {t1, t2,..., tn} into a series of word sequences W = {w1, w2,..., wm}, where n is the total number of characters in the text, m is the total number of words after word segmentation, and m ≤ n; perform denoising, including removing stop words and non-spatial description words. The stop word list S_stopwords contains common words that are meaningless for topic analysis, such as "of", "is", "in", etc.; the non-spatial description word list S_nonspace contains words that do not involve geographical location or spatial feature description. The denoised word sequence W_clean is screened through the following logic:

[0065]

[0066] where "∈" means "belongs to", means "does not belong to", and "∧" means the logical "and" operation. This step ensures that subsequent analysis focuses on key information related to urban renewal and inefficient land use;

[0067] The keyword identification in step 2 specifically includes the following steps:

[0068] Use the TF-IDF algorithm to calculate the weight w_t of each word t, and its calculation formula is:

[0069] w_t = tf(t) * log(N / (df(t) + 1))

[0070] Where tf(t) represents the word frequency of word t in the current document, N represents the total number of documents, and df(t) represents the number of documents containing word t. Then, a pre-trained BERT model is used to convert each word t into a semantic vector v_t∈R^d, capturing the semantic information of the word. The cosine similarity sim(v_t,v_c) = (v_t·v_c) / (||v_t||*||v_c||) between v_t and the predefined concept vector v_c∈R^d related to urban renewal and inefficient land use is calculated. Combining the TF-IDF weight w_t of the word and its cosine similarity sim(v_t,v_c) with v_c, words with scores higher than a preset threshold Threshold are selected as the final identified keywords.

[0071] Step two, which involves pre-training the LDA dynamic topic model to establish a topic-feature word mapping table, specifically includes the following steps:

[0072] The keywords and their contextual information identified in step two are analyzed using a pre-trained LDA topic model. Let the model identify K topics, where each topic θk (k 1,2,…,K) is a probability distribution of words in the vocabulary V, denoted as θk p(w|θk)(p(w1|θk),p(w2|θk),…,p(w|θk)). By setting a topic-word probability threshold ε, words with probabilities higher than ε under each topic are selected, forming the feature word set Wk{w|p(w|θk)ε} for that topic. The selected topics and their corresponding feature word sets Wk are mapped to four preset update regions (residential, industrial / mining, commercial, and service), establishing a topic-feature word mapping table T{(θk,Wk,Type)|k 1,2,…,K,Type∈{residential, industrial / mining, commercial, service}}, where Type represents the type of update region. An example is shown below:

[0073]

[0074] Step 3: Generating a negative sentiment map of land use based on social media data

[0075] Using the four established topic-feature word mapping tables as keyword tables, all acquired Weibo text data were retrieved. All Weibo texts containing the keyword tables and their posting locations were extracted, and the texts and posting locations were classified according to the topic-feature word correspondence. A RoBERTa-large pre-trained sentiment analysis model, trained on Chinese text, was used, combined with semantic enhancement rules for degree adverbs and negation words, to calculate the sentiment scores of all classified Weibo texts. A sentiment score of 0.5 was set as the negative sentiment threshold, and a land use negative sentiment map was generated using the kernel density method.

[0076] Step three, which involves constructing the RoBERTa-large model to calculate the text sentiment score, specifically includes the following steps:

[0077] Using the NLPCC2014 Weibo sentiment classification corpus as the training set, 80% of the data was randomly selected as the training set and 20% as the test set. The model was fine-tuned using the cross-entropy loss function, with a batch size of 16, a learning rate of 2e-5, and 3 epochs. The fine-tuned RoBERTa-large model was used to perform vector encoding and forward inference on the text to be analyzed, outputting the probability distribution of sentiment classification labels. The probability corresponding to the positive label was used as the original sentiment score of the text. Pre-defined semantic rules were used to post-process the scores, identifying text containing negative words and reversing their sentiment polarity, and identifying text containing degree adverbs and adjusting the score amplitude according to the weight coefficients set in the dictionary, outputting the final sentiment score after fusing the semantic structure. A fixed threshold (0.5) was used to classify the sentiment scores, identifying text with scores below the threshold as negative sentiment text, achieving the output of sentiment recognition results based on deep semantic understanding and semantic rule fusion.

[0078] Step three, which involves generating a negative sentiment map of land use using the kernel density method, specifically includes the following steps:

[0079] In the GIS platform, kernel density difference parameters are set, and Gaussian kernel functions are used to spatially smooth negative emotion points, achieving continuous spatial modeling of emotion intensity. The bandwidth parameter controls the range of thermal influence, set to 300–500 meters according to the city scale, balancing local sensitivity and overall readability, and generating rasterized thermal output.

[0080] Step 4: Calculation of update potential based on AMP algorithm and entropy weight method

[0081] A composite index system was constructed, which includes plot ratio, road network density, rental vacancy rate, resident vitality, and space satisfaction. The AMP algorithm was used to calculate the composite index, and the weights of the indicators were dynamically allocated using the entropy weight method. A composite index of 50 points was set as the threshold for judging inefficient land use, and plots with composite indices below the threshold were marked.

[0082] Step four involves constructing an indicator system that includes factors such as plot ratio, road network density, rental vacancy rate, resident vitality, and space satisfaction. This process specifically includes the following steps:

[0083] Floor area ratio (FAR) refers to the ratio calculated by a computer using a formula. Calculated; road network density refers to the road network density obtained by computer using the formula. The rental vacancy rate is calculated by a computer using a formula. The calculations show that R is the average land rent of the study unit, Sv is the total building footprint of buildings with low average nighttime light intensity, and Sb is the sum of building footprints within the study unit. Resident vitality is calculated by weighting population dynamic density, walking activity intensity, commercial activity, public space utilization, and social interaction intensity using the entropy method. The specific formula is as follows: Population Dynamic Density Walking activity intensity Business activity Se represents the area of ​​the top 20% of nighttime light intensity in the commercial land area of ​​the study unit, and S represents the area of ​​the study unit and the utilization rate of public space. Social interaction intensity Residents' vitality Where wj is the weighted index weight, and Xj is the calculated index; spatial satisfaction is expressed through the formula for the spatial alienation degree index. The calculation shows that Ni is the frequency of positive affirmation keywords on social media within the research unit, and N is the total frequency of social media keywords within the research unit.

[0084] Step four, which involves calculating the composite index using the AMP algorithm and dynamically allocating index weights using the entropy weight method, specifically includes the following steps:

[0085] The Adjusted Mazziotta–Pareto (AMP) algorithm is used to normalize each indicator before non-compensatory weighted integration. Normalization of the original indicators is performed. Given the original indicator matrix X = {x...} ij}, which needs to be converted into a normalized matrix R = {r ij The formula is as follows:

[0086]

[0087] Among them, Min xj and Max xj These are the minimum and maximum values ​​of index j, respectively. Ref xj Use this as a reference value. Calculate the weighted average for each unit. If there exists an indicator weight w... j Satisfy 0 <w j <1 and Then the weighted average of unit i The calculation formula is:

[0088]

[0089] Calculate the weighted standard deviation of unit i The weighted coefficient of variation can be obtained by combining the weighted mean. The formula used to measure the relative imbalance in the distribution of indicators is as follows:

[0090]

[0091] The formula for calculating the synthetic AMPI index is as follows:

[0092]

[0093] Finally, the weights of the indicators are calculated based on the differences reflected by information entropy, using the following formula:

[0094]

[0095] Where m is the total number of evaluation indicators, and all weights satisfy ∑w j =1.

[0096] Step 5: Whitelist Verification and Parameter Iteration

[0097] Whitelisted areas are delineated based on the building's construction year and government purple line documents; the results of inefficient land use assessment are compared with the spatial distribution of the whitelisted areas, and if the deviation rate exceeds 10%, the AMPA algorithm weight parameters are readjusted until the deviation rate is ≤10%.

[0098] Step five, which involves filtering the whitelist area based on the construction year and function, specifically includes the following steps:

[0099] Based on building age, areas with buildings ≤ 5 years old and areas with buildings 5 ​​years < 15 years old and renovation / renewal rates ≥ 30% are selected. Based on function, areas with commercial area > 50% are first identified, and then further filtered using the High-End Commercial Index (HCI). The formula is as follows:

[0100] Luxury goods density =

[0101] Luxury store count / area, international brand percentage = international brand store count / total commercial store count, nighttime economy intensity = 22:00-02:00 foot traffic / average daily foot traffic, filter out areas with HCI > 0.7; overlay building age and functional dimensions to find spatial intersection, filter out whitelist areas;

[0102] If the deviation from the actual data exceeds 10% in step five, the specific steps include:

[0103] Spatial distribution deviation rate using formula The calculation shows that A = inefficient output area of ​​the model, and B = whitelist area;

[0104] Step Six: Visualization and Decision Support

[0105] The generated negative sentiment map and inefficient land use distribution heatmap are overlaid and calculated on a WebGIS platform to identify the final urban inefficient land use. Vector spatial information such as urban road network and urban functional zoning is then combined with the urban inefficient land use identification results to generate an urban inefficient land use distribution map. The urban spatial status map, residents' negative sentiment map, and urban inefficient land use distribution map are then integrated into an AR interactive touchscreen with a resolution of at least 7680x4320 pixels and a screen size of at least 86 inches, enabling interactive display of the three views.

[0106] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for intelligent identification of urban underutilized land by fusing social media data, characterized in that, Comprising the following steps: Step one: Multi-source data collection and fusion processing Collect social media geotagged text data and low utility land policy files through public interfaces; Scan urban renewal texts using high-speed document laser scanners with a horizontal and vertical scanning speed greater than 140ppm / 210ipm; Obtain geospatial image data with a resolution better than 1m through satellite remote sensing technology, and synchronously obtain urban basic data, including building outline vector layers, building construction year attribute tables, and land use status maps; Convert the GPS coordinates in the social media data to local coordinates through the WGS84 coordinate system, and then superimpose them on the GIS base map to achieve spatial matching; Step two: Urban renewal and low utility land keyword modeling Input urban renewal planning texts and low utility land policy files, and use NLP tools for word segmentation and noise reduction processing to remove stop words and non-spatial description words; Calculate the word frequency based on TF-IDF weighting, generate word vectors using the BERT model, and identify urban renewal and low utility land keywords; Pre-train LDA / dynamic topic models to establish a theme-feature word mapping table for four types of renewal areas: residential, industrial, commercial, and service; Step three: Generation of land use negative sentiment map based on social media data Use the constructed four-type theme-feature word mapping table as a keyword table to search all obtained microblog text data, extract all microblog texts and posting locations containing the keyword table, and classify the microblog texts and posting locations according to the theme-feature word correspondence; Use the RoBERTa-large pre-trained sentiment analysis model with Chinese text as the training set, combined with the semantic enhancement rules of degree adverbs and negative words, to calculate the sentiment scores of all classified microblog texts; Set the negative sentiment threshold at a sentiment score of 0.5, and generate a land use negative sentiment map using kernel density method; Step four: Update potential calculation based on AMP algorithm and entropy weight method Construct an index system including plot ratio, road network density, rent vacancy rate, resident vitality, and spatial satisfaction, calculate the composite index using the AMP algorithm, and dynamically assign index weights using the entropy weight method; Set the composite index threshold at 50 points for low utility land, and mark land parcels with a composite index below the threshold; Step five: White list verification and parameter iteration Define the white list area through building construction year attributes and government purple line files; Compare the low utility land determination results with the spatial distribution of the white list area, and if the deviation rate exceeds 10%, adjust the AMPA algorithm weight parameters until the deviation rate is ≤10%; Step six: Visualization and decision support Overlay and calculate the generated negative sentiment map and low utility land distribution heat map in the WebGIS platform to identify the final urban low utility land, combine the urban road network, urban functional area division, and other vector spatial information with the urban low utility land recognition results to generate a city low utility land distribution map; Place the city spatial status map, resident negative sentiment map, and city low utility land distribution map in an AR interactive touch screen with a resolution not less than 7680x4320 pixels and a screen size not less than 86 inches to achieve interactive display of the three views. 2.The method of claim 1, wherein, Step two, which involves using NLP tools for word segmentation and noise reduction, specifically includes the following sub-steps: NLP tools were used to segment the input urban renewal planning text and inefficient land use policy documents into words, dividing the continuous text sequence T = {t1, t2, ..., tn} into a series of word sequences W = {w1, w2, ..., wm}, where n is the total number of characters in the text, m is the total number of words after segmentation, and m ≤ n. Noise removal was then performed, including the removal of stop words and non-spatial descriptive words. The stop word list S_stopwords contains common words that are meaningless for topic analysis; the non-spatial descriptive word list S_nonspace contains words that do not involve geographical location or spatial feature descriptions. The denoised word sequence W_clean was then filtered using the following logic: Where "∈" means "belongs to". The symbol "not belonging" indicates "not belonging to", and "∧" indicates a logical AND operation; this step ensures that subsequent analysis focuses on key information related to urban renewal and inefficient land use. 3.The method of claim 2, wherein, The keyword identification in step two specifically includes the following steps: The weight w_t of each word t is calculated using the TF-IDF algorithm, and the formula is as follows: w_t=tf(t)*log(N / (df(t)+1)) Where tf(t) represents the word frequency of word t in the current document, N represents the total number of documents, and df(t) represents the number of documents containing word t. Then, the pre-trained BERT model is used to convert each word t into a semantic vector v_t∈R^d to capture the semantic information of the word. The cosine similarity sim(v_t,v_c) = (v_t·v_c) / (||v_t||*||v_c||) between v_t and the predefined concept vector v_c∈R^d related to urban renewal and inefficient land use is calculated. Combining the TF-IDF weight w_t of the word and its cosine similarity sim(v_t,v_c) with v_c, words with scores higher than the preset threshold Threshold are selected as the final identified keywords. 4.The method of claim 3, wherein, Step two, which involves pre-training the LDA dynamic topic model to establish a topic-feature word mapping table, specifically includes the following steps: The keywords and their contextual information identified in step two are analyzed using a pre-trained LDA topic model. The model identifies K topics, each θk (k 1,2,…,K) being a probability distribution of words in vocabulary V, denoted as θk p(w|θk)(p(w1|θk),p(w2|θk),…,p(w|θk)). By setting a topic-word probability threshold ε, words with probabilities higher than ε under each topic are selected, forming the feature word set Wk{w|p(w|θk)ε} for that topic. The selected topics and their corresponding feature word sets Wk are mapped to four preset update regions: residential, industrial / mining, commercial, and service. A topic-feature word mapping table T{(θk,Wk,Type)|k 1,2,…,K,Type∈{residential, industrial / mining, commercial, service}} is established, where Type represents the type of update region.

5. The method of claim 4, wherein the method further comprises: Step three, which involves constructing the RoBERTa-large model to calculate the text sentiment score, specifically includes the following steps: The NLPCC2014 micro-blog sentiment classification corpus is used as the training set, 80% of the data is randomly extracted as the training set, and 20% is used as the test set. The cross-entropy loss function is used for model fine-tuning training. During the training process, the batch size is set to 16, the learning rate is 2e-5, and the training rounds are 3 rounds. The fine-tuned RoBERTa-large model is used for vector coding and forward inference of the text to be analyzed. The probability distribution of the sentiment classification label is output, and the probability corresponding to the positive label is used as the original sentiment score of the text. The preset semantic rules are used for post-processing of the score. The text containing negative words is identified and the sentiment polarity is reversed. The text containing degree adverbs is identified and the score amplitude is added or subtracted according to the weight coefficient set in the dictionary. The final sentiment score after fusing the semantic structure is output. The fixed threshold value 0.5 is used for classification of the sentiment score. The text with a score lower than the threshold value is identified as a negative sentiment text. The sentiment recognition result output based on deep semantic understanding and semantic rule fusion is realized. 6.The method of claim 5, wherein, The step three of generating the negative sentiment map by kernel density method specifically comprises the following steps: In the gis platform, set the kernel density difference parameter, use the gaussian kernel function to perform spatial smoothing processing on the negative sentiment points, realize continuous spatial modeling of the sentiment intensity, control the heat influence range through the bandwidth parameter, and set 300-500 meters according to the city scale, balance the local sensitivity and the overall readability, and generate the rasterized heat output.

7. The method of claim 6, wherein the method further comprises: The step four of constructing the index system of volume rate, road network density, rent vacancy rate, resident vitality and space satisfaction specifically comprises the following steps: Volume rate refers to the computer through the formula volume rate Calculated; road network density refers to the computer through the formula road network density Calculated; rent vacancy rate refers to the computer through the formula rent vacancy rate Calculated, wherein R is the average land rent of the study unit, Sv is the integrated average intensity of night light of building floor area, Sb is the sum of building floor area in the study unit; The resident vitality is calculated by the population dynamic density, walking activity intensity, commercial activity, public space utilization and social interaction intensity by entropy method, the specific formula is as follows: Population dynamic density Walking activity intensity Commercial activity Se is the area of the top 20% of the commercial land of the study unit in the night light intensity, S is the area of the study unit, public space utilization Social interaction intensity Resident vitality Wherein wj refers to the index weighted weight, Xj is the calculation index; The spatial satisfaction is calculated by the spatial alienation degree index formula Calculated, wherein Ni is the frequency of positive identification type keywords in the study unit of social media, N is the total frequency of keywords in the study unit of social media. 8.The method of claim 7, wherein, The step four of calculating the composite index by using the AMP algorithm and dynamically allocating the index weight by using the entropy weight method specifically comprises the following steps: The Adjusted Mazziotta-Pareto algorithm is used to normalize each index and then perform non-compensatory weighted integration; the original index is normalized; the original index matrix X = {x ij} is given, which needs to be converted into a normalized matrix R = {r ij}, the formula is as follows: where Min xj and Max xj are the minimum and maximum values of the index j, respectively, and Ref xj is the reference value; the weighted average of each cell is calculated; if there exists an index weight w j satisfying 0 < w j < 1 and then the weighted average of cell i is The calculation formula is: Weighted standard deviation of the calculation unit i Weighted coefficient of variation in combination with the weighted mean For measuring the relative imbalance of the index distribution, the formula is: The synthetic AMPI index is calculated by the following formula: Finally, the index weight is calculated according to the difference reflected by the information entropy, and the formula is: where m is the total number of evaluation indexes, all weights satisfy∑w j = 1. 9.The method of claim 8, wherein, The step five of screening the white list area according to the construction period and function specifically comprises the following steps: The building age dimension screens the areas with building age ≤ 5 years, and the areas with 5 years < building age ≤ 15 years and renovation rate update rate ≥ 30%; the function dimension first extracts the areas with commercial area proportion > 50%, and then screens through the high-end commercial index HCI calculation, the formula is Wherein the luxury goods density = luxury goods store number / area of the region, the proportion of international brands = international brand store number / total commercial store number, the night economy intensity = 22.00-02.00 passenger flow / daily passenger flow, and the regions with HCI>0.7 are screened out; the construction period dimension and the function dimension are superimposed to obtain the spatial intersection, and the white list area is screened out. The step five of deviation exceeding 10% from the actual data specifically comprises the following steps: 10.The method of claim 9, wherein, ​ By formula spatial distribution deviation rate Computed where A = model output low efficiency area, B = white list area.

Images