Low-efficiency land use investigation and identification method

Abstract

The invention discloses a low-efficiency land investigation and identification method. The method comprises the following steps: acquiring a low-efficiency land sample from an established comprehensive characteristic index system for low-efficiency land identification; clustering the low-efficiency land samples by using K-Means, analyzing key influence factors of different types of low-efficiency land samples by using XGBoost-SHAP, and quantifying to obtain an environmental variable feature importance weight; and by using a geographic similarity classification space speculation method, carrying out similarity evaluation on all the to-be-identified land parcels according to the obtained environmental variable feature importance weights to obtain potential low-efficiency land use types. According to the method, different low-efficiency land types are divided by utilizing interpretable machine learning, and the feature importance of the low-efficiency land types is extracted, so that the problems of various low-efficiency land influence factors, complex types and unknown influence mechanisms are solved. The problems of few low-efficiency land use samples and single-class samples are solved by utilizing a geographic similarity classification space speculation method.

Description

Technical Field

[0001] The invention relates to land survey technology, and in particular to a method for surveying and identifying low-efficiency land. Background Art

[0002] With the disorderly expansion of urban construction land, the shift of economic development focus, and population migration [1], some old built-up areas in cities have gradually become inefficient in land use. These areas are usually called low-efficiency land. [3-5] , including brownfield sites, such as old factories, warehouses and heavily polluted industrial areas; underutilized commercial land, such as vacant shopping malls, office buildings, etc.; and residential areas abandoned due to economic recession or population migration.

[0003] At the same time, low-efficiency land is a valuable opportunity to solve the shortage of land for urban development. Redevelopment of low-efficiency land can not only maximize the use of limited land resources in cities and reduce the encroachment on ecological land and cultivated land, but also significantly improve urban functions, social welfare and the quality of human settlements. In addition, the revitalization and utilization of low-efficiency land can attract new investment, create employment opportunities, and contribute to the long-term sustainable development of cities. However, there are currently great challenges in accurately identifying low-efficiency land with complex influencing factors, large number and scattered distribution in cities.

[0004] On-site manual survey is the most widely used method. In actual land use management, the competent department sends land surveyors to gain an in-depth understanding of the land use of the target plots and then evaluate the benefits of their land use. However, this method has the disadvantages of high cost, low efficiency, and strong subjectivity, which leads to slow updates, scattered surveys, and errors in update decisions.

[0005] The second method is a multi-factor evaluation method coupled with GIS. GIS has advantages in multi-source data integration and spatial visualization, and is also good at spatial analysis and modeling. It can integrate methods such as hierarchical analysis, expert scoring method and comprehensive index method to comprehensively evaluate the characteristics and distribution of low-efficiency land. For example, Chrysochoou proposed a GIS-based decision support system for brownfield redevelopment, which optimizes the site selection process for brownfield redevelopment by integrating key indicators such as environmental impact, economic potential and social benefits. In practice, management departments in many places also use this method to implement intensive and economical land use evaluation in towns or industrial parks

[32] . Through the multi-factor evaluation method coupled with GIS, the land in the target area is evaluated for utilization benefits and potential low-efficiency land is found. The main goal is to increase the intensity of land development and promote the transformation and upgrading of old urban areas and old industrial areas.

[0006] However, the multi-factor evaluation method coupled with GIS usually uses multi-source data such as remote sensing and urban perception to quantify indicators, which has high requirements for data quality and timeliness. At the same time, it is difficult to explore the complex nonlinear relationship between inefficient land and multiple factors. At the same time, the research scale still uses towns or industrial parks as evaluation units, and the identification results are not precise enough. In addition, the importance evaluation method of multiple factors is subjective and needs to rely on the prior knowledge of technicians and experts, such as a full understanding of the social and economic development of the evaluation area. Whether the evaluation results can objectively reflect the distribution of inefficient land depends largely on the professional level of the modeler.

[0007] The third method is to use machine learning algorithms to identify inefficient land. Machine learning methods perform well in identifying inefficient land. They can effectively process large-scale and complex multi-source geographic data and have high classification accuracy and generalization ability. For example, Liu used a method that integrates crowdsourcing datasets with traditional datasets to measure the dynamic information of plots, and applied the presence and background data (PBL-BPNN) model to conduct large-scale assessment of the redevelopment suitability of plots. The results show that machine learning methods have high accuracy and reliability for residential and commercial plots, and the introduction of crowdsourcing datasets makes it feasible to conduct large-scale assessments at fine spatiotemporal scales. The powerful processing ability of deep learning for graphics can achieve high-precision classification by automatically extracting remote sensing image features. Ru He improved the PBL-BPNN used by Liu, extracted emotional perception and visual perception from street view images, and constructed a new comprehensive indicator set to evaluate the urban renewal potential of multiple land types, identifying inefficient land from the perspective of street view perception.

[0008] However, machine learning requires a large amount of labeled data and feature selection optimization, while deep learning training requires a large amount of computing resources and high-quality training data. For example, in the problem of identifying inefficient land, a small number of samples will lead to overfitting or underfitting of machine learning algorithms. Without enough samples to divide the training set and validation set, it is impossible to reliably evaluate the generalization ability and predictive performance of the model. A single sample will result in the model only being exposed to the data of inefficient land, but not to non-inefficient land, which makes it impossible for the model to distinguish the differences between different categories, resulting in the failure of the classification task. Artificially selecting non-inefficient land as classification training samples is not only a large workload, but also difficult to accurately select non-inefficient land due to the regional differences of inefficient land. Once the sample is misselected, it will cause serious pollution to the model. In addition, the interpretability of machine learning is still an important issue. Black boxes can lead to decision-making biases, and managers find it difficult to understand what exactly caused the analysis results. In particular, urban issues involve the interests of multiple subjects. The game and compromise between stakeholders need to be justified, and the government's major decisions need reliable theoretical and data support. Therefore, the analysis results of machine learning are difficult to transform into a reasonable basis for government decision-making, and more often can only be used as a reference.

[0009] Inefficient land is an important research object in the field of urban renewal. Accurate identification of inefficient land is the premise and basis for urban renewal and has important practical significance. However, the current inefficient land identification technology still has problems such as complex influencing factors of inefficient land, small sample size, lack of non-inefficient land samples, and weak interpretability of identification results. Summary of the invention

[0010] In order to solve the problem of how to establish an interpretable inefficient land identification method with a small number of single-class samples, the present invention provides a method for investigating and identifying inefficient land.

[0011] To achieve the above object, the technical solution of the present invention is:

[0012] A method for investigating and identifying inefficient land use, comprising:

[0013] Obtain low-efficiency land samples from the comprehensive characteristic indicator system for identifying low-efficiency land;

[0014] K-Means was used to cluster the low-efficiency land samples and XGBoost-SHAP was used to analyze the key influencing factors of different types of low-efficiency land samples, and the importance weights of environmental variable features were quantified;

[0015] The spatial inference method of geographic similarity classification is used to evaluate the similarity of all the plots to be identified based on the obtained environmental variable feature importance weights to obtain potential inefficient land use types.

[0016] Preferably, the comprehensive characteristic indicator system for identifying low-efficiency land includes four characteristic indicators: land use conditions, economic efficiency, social efficiency and ecological environmental efficiency; the land use conditions take into account the basic conditions of the land itself, including terrain conditions and building conditions; the economic efficiency takes into account the economic benefits of the land, and in addition to the direct economic output data of the land, also selects indicators that can indirectly demonstrate the level of economic development; the social efficiency takes into account the population carrying capacity and public service level of the land; the ecological environmental efficiency takes into account the ecological sensitivity and environmental fragility of the land.

[0017] Preferably, the clustering of the low-efficiency land sample data using K-Means and the analysis of the key influencing factors of different types of low-efficiency land samples using XGBoost-SHAP include:

[0018] K-Means clustering is used to cluster the low-efficiency land samples, and the low-efficiency land samples that meet the K-class characteristic differences are divided to obtain the clustering results;

[0019] XGBoost-SHAP is introduced to perform classification prediction based on clustering results, analyze the key influencing factors of different types of inefficient land samples, and extract the importance of features; the SHAP value explains the model by calculating the marginal contribution of features to the model output under different combinations:

[0020]

[0021] Where F is the set of all features, S is a subset of F that does not contain feature i; f(S) is the model prediction value containing only the features of subset S; is the SHAP value of feature i, indicating its marginal contribution to the model prediction value.

[0022] Preferably, the obtained feature importance weights are used to perform similarity assessment on all the plots to be identified to obtain potential low-efficiency land use types, including:

[0023] Mapping the environmental characteristic data of known low-efficiency land to all the plots to be identified, and obtaining the environmental covariate vector of each plot to be identified;

[0024] According to the obtained environmental variable feature importance weights, the similarity score of each plot to be identified at the environmental covariate level is calculated, and the maximum similarity score is selected for output;

[0025] The maximum similarity score output is compared with the set threshold. If it is higher than the threshold, the to-be-identified land parcel corresponding to the similarity score is output as a potential low-efficiency land. If it is lower than the threshold, no data is output.

[0026] Preferably, uncertainty calculation is performed before outputting the maximum similarity score to calculate the variance of the similarity score as uncertainty.

[0027] Preferably, the similarity score of each plot to be identified at the environmental covariate level is calculated based on the obtained feature importance weights, and the maximum similarity score is selected for output, including:

[0028] The importance weight of each environmental variable feature is multiplied by the similarity value of each environmental covariate of the plot to be identified, and then all are added together to obtain the similarity score of the plot to be identified at the environmental covariate level;

[0029] Repeat the above steps, but keep replacing the known low-efficiency land samples to obtain the similarity score between each plot to be identified and all known low-efficiency land plots, and select the maximum similarity score as the output.

[0030] Preferably, the algorithm formula of the environmental covariate vector of each plot to be identified is:

[0031] e=(e1,e2,…,e m )

[0032] In the formula, e is the environmental vector at a given location, used to describe the environmental conditions, and m is the number of environmental covariates used.

[0033] Preferably, the similarity score of each plot to be identified at the environmental covariate level is calculated based on the obtained feature importance weights, and the algorithm formula for selecting the maximum similarity score output is:

[0034]

[0035] w m =f(SHAP)

[0036] Similarity i =Max(S i1 ,S i2 ,…,S ij )

[0037] S ij represents the environmental similarity between the vth environmental covariate of the unobserved location and the sample location j, w m environmental variable feature importance weights; is the mth weight of the vth environmental covariate.

[0038] Preferably, the uncertainty calculation method is:

[0039]

[0040] In the formula, (Si,1 ,S i,2 ,…,S i,n ) is the similarity between sample i and all low-efficiency land samples, p(S i,1 ,S i,2 ,…,S i,n ) is the value of sample i (S i,1 ,S i,2 ,…,S i,n ) probability.

[0041] Preferably, the indicator value of each characteristic indicator is quantified at the plot scale and a Pearson correlation analysis is performed.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] At present, the academic community has made a lot of useful work on the identification of inefficient land. For example, coupling GIS and machine learning to identify inefficient land is a very advanced method to provide decision support for urban renewal. However, due to the complex influencing factors of inefficient land, the investigation work has not yet been fully popularized, resulting in a small number of samples. Therefore, how to establish an inefficient land identification method based on explainable machine learning in the case of few samples and single-class samples, and realize the automatic classification and identification of inefficient land is an important research gap that currently exists. The method of coupling explainable machine learning and geographic similarity classification spatial inference proposed in the present invention solves this problem. Different types of inefficient land are divided by interpretable machine learning, and the importance of their features is extracted to solve the problem that there are many influencing factors of inefficient land, the types are complex, and the influencing mechanism is unknown. The method of geographic similarity classification spatial inference is used to solve the problem of few inefficient land samples and single-class samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flowchart of a method for identifying inefficient land use based on explainable machine learning and geographic similarity reasoning provided by an embodiment of the present invention;

[0045] Figure 2 Schematic diagram of spatial inference algorithm for geographic similarity classification;

[0046] Figure 3 is the characteristic mean of various types of low-efficiency land;

[0047] Figure 4 The spatial and quantitative distribution of low-efficiency land sample clusters;

[0048] Figure 5 The importance analysis results of various types of inefficient land features based on XGBoost-SHAP;

[0049] Figure 6This is the result of spatial inference of the similarity of low-efficiency land use: where (a1), (b1) and (c1) are similarities, (a2), (b2) and (c2) are uncertainties;

[0050] Figure 7 is the spatial autocorrelation of similarity and uncertainty;

[0051] Figure 8 Identification results of potential inefficient land use based on the IML-GSR-ULIM model;

[0052] Fig. 9 The difference between the number of different types of inefficient land and the number of verification samples

[0053] Fig.10 A complete flow chart of a method for identifying inefficient land use based on explainable machine learning and geographic similarity reasoning provided by an embodiment of the present invention; DETAILED DESCRIPTION

[0054] In the application, the relevant abbreviations and key terms are defined as follows:

[0055] GIS (Geographic information system)

[0056] PBL-BPNN (Presence and background learning-back propagation neuralnetwork)

[0057] IML-GSR-ULIM (Interpretable machine learning and geographic similarity reasoning based underutilized land identification model)

[0058] K-Means

[0059] XGBoost (Extreme gradient boosting, extreme gradient boosting tree)

[0060] SHAP (Shapley additive explanations)

[0061] OSM (Open street map)

[0062] POIs (Points of Interest)

[0063] SVIs (street view images)

[0064] Example:

[0065] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0066] In order to solve the problem of how to establish an interpretable inefficient land use identification method in the case of few samples or single-class samples, the present invention proposes an inefficient land use identification method based on interpretable machine learning and geographic similarity reasoning. Figure 1 As shown, the method mainly includes the following steps:

[0067] 101. Obtain low-efficiency land samples from the comprehensive characteristic indicator system for identifying low-efficiency land;

[0068] 201. Use K-Means to cluster the low-efficiency land samples and use XGBoost-SHAP to analyze the key influencing factors of different types of low-efficiency land samples, and quantify the importance weights of environmental variable features;

[0069] Thus, in this step, interpretable machine learning is used to divide different types of inefficient land use and extract their feature importance, solving the problem that inefficient land use has many influencing factors, complex types, and unknown influencing mechanisms.

[0070] 301. Using the spatial inference method of geographic similarity classification, the importance weights of environmental variable characteristics are obtained to conduct similarity assessment on all the plots to be identified, and obtain potential inefficient land use types.

[0071] Thus, in this step, the method of spatial inference of geographic similarity classification is used to solve the problem of few inefficient land samples and single-class samples.

[0072] Specifically, this embodiment builds a comprehensive characteristic indicator system for identifying inefficient land through multi-source social perception data, and strives to provide a comprehensive and objective reference for urban renewal decision support. For the characteristic indicators of inefficient land, the efficiency of land use is measured from four dimensions: land use conditions, economic efficiency, social efficiency, and ecological environmental efficiency. The required data are shown in Table 1. Each indicator value is quantified at the plot scale, and Pearson correlation analysis is performed to ensure that there is no collinearity between the characteristic indicators.

[0073] Table 1 Inefficient land identification index system

[0074]

[0075]

[0076] ① The land use conditions take into account the basic conditions of the plot itself, including terrain conditions and building conditions. The terrain conditions can measure the utilization conditions of the plot, including elevation (A1) and slope (A2) data downloaded from the source website. The building height difference (A3), number of buildings (A4) and building difference (A5), plot perimeter area ratio (A6), and building density (A7) data in the building condition indicators are obtained by downloading the building footprint data from the source website and using ArcGIS pro spatial statistics, overlay analysis, and kernel density analysis methods. The building age data (A8) is obtained from the statistical data of the Shenzhen Natural Resources Statistics Bureau through Kriging interpolation and other methods.

[0077] ② Economic efficiency mainly considers the economic benefits of the plot. In addition to the direct economic output data of the plot, indicators that can indirectly show the level of economic development, such as road network density and floor area ratio, are also selected. The economic output data of the plot (B2) is the GDP data of China's kilometer grid downloaded from the source website. The road network density (B3) is obtained by the kernel density analysis of ArcGIS Pro using the open streetmap road network data downloaded from the source website. The floor area ratio (B1) is obtained by the spatial statistics of the building footprint data using ArcGIS Pro.

[0078] ③ The social effectiveness dimension mainly considers the population carrying capacity and public service level of the plot. Population density and night lights can be used to characterize the population activity of the plot; the POI density index of public service facilities can be used to characterize the public service capacity of the plot, and street view perception data can be used to give a quantitative evaluation perspective of the plot from the perspective of street view perception. Population density (C1) is the population spatialization data based on the interpolation of the Seventh National Population Census downloaded from the source website, and night lights (C2) is the Chinese time series DMSP-OLS data downloaded from the source website by integrating DMSP-OLS and SNPP-VIIRS. Infrastructure (C3) is calculated based on POIs. First, urban infrastructure points, including subway stations, parks and green spaces, shopping malls, schools, cultural and leisure facilities, etc., are obtained from the API service of Amap or Baidu Map, and the infrastructure density is obtained through the kernel density analysis of ArcGIS pro. Street view indicators (C4)-(C9) are a method called Human-Machine Adversarial Scoring Framework, which is used to evaluate urban perception using street view images. This approach combines deep learning, street view imagery, and an iterative feedback mechanism to quickly and cost-effectively assess local urban perception in Chinese cities.

[0079] ④ The ecological environmental performance dimension mainly considers the ecological sensitivity and environmental fragility of the plot, such as the normalized vegetation index (D1) representing the greening rate, the sulfur dioxide concentration (D2) and PM2.5 concentration (D3) representing air pollution, and the urban heat island index (D4) and waterlogging frequency (D5) data representing environmental fragility. These data are downloaded from the data source website, and spatial overlay analysis is performed through ArcGIS pro to obtain specific values ​​on each plot.

[0080] Inefficient land use is relative. Different regions have different influencing factors, and the location difference will expand with the distance. In order to find potential inefficient land use in the region, it is necessary to first classify the inefficient land use samples reasonably according to their internal differences and find the significant factors affecting inefficient land use in different regions. The purpose of this step is to analyze the main factors affecting inefficient land use in different locations. K-Means, XGBoost and SHAP methods are used to classify inefficient land use samples and identify key influencing factors, and extract feature importance as the input weight for the next step.

[0081] First, K-Means clustering is used to cluster the inefficient land samples, and K types of inefficient land samples with significant feature differences are divided, which helps to extract and quantify the features of different types of inefficient land samples. K-Means is simple and intuitive, but the number of clusters needs to be specified in advance. This embodiment uses the commonly used methods of the Elbow Method and the Silhouette Score to determine the optimal K value [70,71]. XGBoost's powerful feature extraction and classification prediction capabilities can well learn and extract the feature differences of clustering results. Combined with SHAP, it can well make up for the shortcomings of the "black box" operation of XGBoost prediction results. Therefore, XGBoost-SHAP is introduced to perform classification prediction with the clustering results as a reference, analyze the key influencing factors of different types of inefficient land samples, and extract the importance of features. The SHAP value explains the model by calculating the marginal contribution of features to the model output under different combinations:

[0082]

[0083] Where F is the set of all features, S is a subset of F that does not contain feature i. f(S) is the model prediction value containing only the features of subset S. is the SHAP value of feature i, indicating its marginal contribution to the model prediction value.

[0084] The similarity between known inefficient land samples and unknown plots is calculated using the geographic similarity spatial inference algorithm. The core of the geographic similarity spatial inference algorithm is to quantify the similarity of geographical phenomena, and on this basis, evaluate the similarity between spatial units. Based on the connotation and characteristics of inefficient land and the management thinking of urban land management that it is better to misidentify than to miss, this embodiment assumes that as long as an unknown sample has a high degree of similarity with a low-efficiency land sample, it may be a potential inefficient land. Here, the characteristic indicators related to the target sample are called environmental variables (EnvironmentalVariable). By using environmental vectors and similarity functions, the environmental similarities between unobserved locations and sample locations can be systematically compared and evaluated, thereby helping to understand and predict the spatial distribution of target samples. The algorithm principle is as follows Figure 2 As shown, the specific steps include:

[0085] Step 1: Characteristics of environmental conditions: Map the environmental characteristic data of known low-efficiency land to all the plots to be identified, and obtain the environmental covariate vector of each plot to be identified;

[0086] Step 2: Similarity evaluation: Based on the obtained environmental variable feature importance weights, calculate the similarity score of each plot to be identified at the environmental covariate level, and select the maximum similarity score for output;

[0087] Specifically, in this step, taking a known low-efficiency land use patch and an unknown plot as an example, the calculation method is to multiply the importance weight of each environmental covariate feature by the similarity value of each environmental covariate of the pair of unknown plots, and then add them all up to obtain the similarity score of the pair of unknown plots at the environmental covariate level. Repeat the above steps in a loop, but constantly replace the known low-efficiency land sample patches, so that in the end, a similarity score between an unknown plot and all known plots will be obtained, and the maximum similarity score is selected as Si (the similarity score between the unknown plot and the low-efficiency land) output.

[0088] Step 3: Uncertainty quantification: According to step 2, we can get the similarity score between an unknown plot and all known low-utility plots, which is a two-dimensional matrix. Calculate the variance of this two-dimensional matrix, that is, the distribution characteristics of the similarity score data, as the "uncertainty"

[0089] Step 4: Identification of inefficient land: Compare the maximum similarity score output with the set threshold. If it is higher than the threshold, the land parcel to be identified corresponding to the similarity score is output as a potential inefficient land. If it is lower than the threshold, it is impossible to reliably infer whether the land parcel is an inefficient land based on the current inefficient land parcel map, so the output is 'no data'.

[0090] Specifically, the core algorithm of geographic similarity spatial inference is as follows:

[0091] e=(e1,e2,…,e m ) (2)

[0092]

[0093] In the formula, e is the environmental vector at a given location, used to describe the environmental conditions, and m is the number of environmental covariates used. v is the value of the vth environmental covariate at the unobserved location. vj is the value of the vth environmental covariate at sample location j. ij represents the environmental similarity between the vth environmental covariate at the unobserved location and the sample location j. E(e v ,e vj ) describes the environmental similarity between the unobserved location and the sample location. v is the index of the environmental covariate, which represents the specific environmental factors (such as road network density, NDVI, heat island index, etc.). ev is the standard deviation of the vth environmental covariate in the study area. is the root mean square difference of the vth environmental covariate at all unobserved locations and the sample location j.

[0094] In formula (3), it is indicated that the similarity between the calculated single unknown sample and all known environmental covariates is considered in general. The weights can be determined by using weighted average method, hierarchical analysis method, minimum method, entropy weight method and other methods. In order to improve the objectivity and scientificity of the research results, this embodiment uses an interpretable machine learning method to determine the weights of environmental covariates, and calculates the similarity between unknown samples and known samples by weighted summation. Here, according to the characteristics of the research object, this embodiment modifies the similarity model.

[0095] (① Based on the scenario assumed in this embodiment, the maximum value is selected as the similarity of the unknown sample between the unknown sample and all known samples, so as to better identify potential inefficient land. In practical applications, the algorithm can be adjusted according to management needs. For example, if a high accuracy rate of recognition result is required at the expense of missed recognition rate, the average value of the upper quartile of the similarity of all known samples can be taken as the similarity of the unknown sample.

[0096] Similarity i =Max(S i1 ,S i2 ,…,S ij ) (6)

[0097] (②Uncertainty measures the stability of the similarity score and can be used to determine the reliability of the low-efficiency land identification results. This embodiment designs two optional ways to calculate uncertainty to enhance its expression of the connotation of "uncertainty", namely standard deviation and Shannon entropy.

[0098]

[0099] In the formula, (S i,1 ,S i,2 ,…,S i,n ) is the similarity between sample i and all low-efficiency land samples, p(S i,1 ,S i,2 ,…,S i,n ) is the value of sample i (S i,1 ,S i,2 ,…,S i,n ) probability.

[0100] This embodiment uses five-fold cross validation to test the accuracy of the results. In each iteration, 20% of the low-efficiency land samples are added to all the plot samples to calculate the similarity together. At the same time, n background samples equal to the low-efficiency land verification samples are randomly selected as the control group. Therefore, each iteration calculates the similarity between 20% of the low-efficiency land samples and 100 non-low-efficiency land samples. According to Pontius Jr, the cross-tab matrix provides a more accurate classification evaluation and comparison than the Kappa coefficient. Therefore, this embodiment combines the two methods to evaluate the accuracy of the results. The verification process includes random sampling from the identified blocks and using a confusion matrix to evaluate several key indicators, including producer accuracy, user accuracy, quantity inconsistency, allocation inconsistency, overall accuracy, and Kappa coefficient.

[0101]

[0102] Where K represents the Kappa coefficient, X ii represents the elements of the matrix, represents the sum of the i-th row, Represents the sum of the i-th column, M represents the total number of samples used for detection, and N represents the number of rows and columns of the confusion matrix.

[0103] In addition, we used Google high-resolution remote sensing images combined with manual on-site visits to conduct case verification on some typical plots of land identified. like Fig.10 A complete flowchart of the inefficient land identification method based on explainable machine learning and geographic similarity reasoning provided for this embodiment.

[0104] The method of the present invention is further verified and explained below in conjunction with an application scenario example:

[0105] This method was applied to Shenzhen City, Guangdong Province. A total of 9,668 potential inefficient plots were identified among all 38,008 plots, accounting for 25.44% of all unknown plots. Among them, there were 790 Class I inefficient plots, accounting for 7.44%; 6,322 Class II inefficient plots, accounting for 59.57%; and 3,500 Class III inefficient plots, accounting for 32.98%. If more reliable results with uncertainty less than 0.6 are considered, there are 7,256 plots, accounting for 19.09% of all plots. The following are the detailed results of the model application:

[0106] (1) Clustering of low-efficiency land samples and analysis of key influencing factors

[0107] The application case uses Elbow Method and Silhouette Score cross-validation to determine the optimal cluster, and finally chooses to divide the low-efficiency land samples into three categories. Figure 3 The characteristic means of each type of inefficient land are displayed, fully demonstrating the differences between inefficient land.

[0108] According to the clustering results, three types of inefficient land use are visualized, such as Figure 4 As shown, the number of samples of different types of inefficient land is counted based on the districts of Shenzhen. It can be seen that the distribution of the three types of inefficient land has their own characteristics.

[0109] Type 1 is a typical city central underutilized land. They have a high GDP level, and the corresponding road network density, infrastructure construction, and population activity are all very high. At the same time, the living environment here is the most uncomfortable, with high volume ratio, low greening rate, high building density, and old buildings, which make people feel depressed. Type 1 inefficient land prefers to be distributed in Nanshan District, Luohu District, Futian District, and Bao'an District. These are the core areas of Shenzhen City, with a long construction period, some old buildings, or the decline of commercial centers has become inefficient land.

[0110] Type 2 is marginal underutilized land, which has a low economic level, low population activity, and the lowest level of transportation and infrastructure. However, it has a comfortable living environment, a high greening rate, a high sense of beauty, and is relatively boring. Type 2 inefficient land is distributed in Longgang District, Yantian District, Pingshan District, etc. These are the peripheral areas of Shenzhen City, mainly industrial land and idle land, and there are also some old urban village houses.

[0111] Type 3 is an intermediate underutilized land between the first two categories. It has a high building density and number of buildings. Its population activity is second only to the first type of inefficient land. Its GDP level is not high, and its living atmosphere is strong, but also relatively boring. Type 3 inefficient land is concentrated in Longhua District, Longgang District, Guangming District and other areas, and is mainly residential land. In addition, the air pollution of this type of inefficient land is high, indicating that there are still some industrial inefficient plots.

[0112] (2) Using XGBoost-SHAP to analyze the locational factors affecting inefficient land use

[0113] Using XGBoost, we take cluster labels as dependent variables and feature indicators as independent variables to learn different types of inefficient land use, and use SHAP to analyze the location influencing factors of inefficient land use and quantify the importance of features. The results are as follows: Figure 5 As shown. It can be seen that the dominant locational influence characteristics of the three types of inefficient land are different. Specifically, for Type 1 inefficient land, characteristics such as GDP (C3), Infrastructure (B3), and Roadnet density (B2) have a strong locational influence; while Type 2 inefficient land is more affected by characteristics such as NDVI (D1), Flood (D5), and Infrastructure (C3); for Type 3 inefficient land, characteristics such as Heat Island (D4), Building age (A7), and NDVI (D1) have a higher influence. We use the locational influence calculated based on the SHAP value as the weight of the characteristic indicators in subsequent analysis.

[0114] (3) Using geographic similarity to calculate similarity and uncertainty of plots

[0115] Geographic similarity can be used to infer the similarity of unknown samples based on a single sample, but a single sample must be input. In this example, the low-efficiency land in the study area is subdivided into three types, so the similarity and uncertainty of the three types of low-efficiency land and all urban plots are calculated in turn. The results are as follows: Figure 6 As shown in the figure, the redder the color, the higher the similarity and uncertainty, and the greener the color, the lower the similarity and uncertainty.

[0116] It can be seen from the figure that different types of low-efficiency land samples have different similarity scores, and the uncertainty distribution also varies greatly. In order to further analyze the quantitative characteristics and spatial patterns of similarity and uncertainty, we draw a local spatial autocorrelation Lisa diagram, such as Figure 7 shown.

[0117] Figure 7 The bivariate spatial autocorrelation clustering of three types of similarity and uncertainty is shown. Ideally, the clustering attribute of low-efficiency land should be high-low, while the clustering attribute of ordinary plots is low-high, which means that low-efficiency land is an area with high similarity and low uncertainty, and ordinary plots are areas with low similarity and low uncertainty. As can be seen from Figures (a1) and (a2), there are large areas of high-high clustering in the northeast and central parts of the study area, and the significant P value is less than 0.001, indicating that the Type 1 sample has a false high similarity to this area, and the credibility of the results needs further analysis. There are some high-low clustering plots in the core area, which are relatively credible results. Similarly, for the peripheral areas of Shenzhen City, since Type 1 is rarely distributed, it presents a low-low clustering situation, indicating that those areas have low similarity and low uncertainty, and they do not belong to Type 1 low-efficiency land. On the contrary, (b1) and (b2) have many high-low clustering plots in the peripheral areas, and the closer to the city center, the more the clustering type tends to be low-low. In stark contrast to Type 1, there are large areas of low-high concentration in the middle of the study area, which shows that there is a strong credibility that these areas do not belong to Type 2 inefficient land. It is worth noting that Type 3 inefficient land is concentrated in Guangming District, Longhua District, Longgang District, etc., due to its wide distribution range, so it can be seen that there are more high-high concentration and high-low concentration plots in these districts.

[0118] Recognition results and accuracy verification

[0119] A total of 9,668 plots were identified as potentially inefficiently utilized plots, accounting for 25.44% of all plots. Among them, Type 1 accounted for 6.99%, Type 2 accounted for 56.17%, Type 3 accounted for 27.18%, and mixed types accounted for 9.67%. If only the results with uncertainty less than the threshold of 0.6 are considered, 7,256 plots are identified, accounting for 19.09% of all plots. In terms of quantity, the number of Type 1 is the least, indicating that the land use efficiency in the city center is relatively high; on the contrary, Type 2 accounts for the majority, indicating that the land use efficiency in the suburbs of Shenzhen can be improved. From a spatial perspective, the potentially underutilized land is mainly distributed in the east and north of Shenzhen, which is the main distribution area of ​​Type 2 land and also the area with low urban population density and economic activity.

[0120] In order to verify the reliability of the results, 20% of the low-efficiency land samples and the similarity score results of 100 known non-low-efficiency lands are extracted in each round of cross-validation to evaluate the model's false recognition rate and misrecognition rate. In addition, because the similarity score calculated by the geographic similarity model is more like the classification probability (continuous variable) output in the machine learning model, rather than the category (nominal variable), it is necessary to manually define the threshold to classify the results into two categories. In order to scientifically determine the threshold, considering the tolerable error, the average lower quartile point of the similarity score of 20% sampled low-efficiency land in the five-fold cross-validation is selected as the threshold for low-efficiency land and non-low-efficiency land. The similarity threshold of Type 1 is 0.832, the similarity threshold of Type 2 is 0.752, and the similarity threshold of Type 3 is 0.866. The confusion matrix is ​​calculated based on the threshold to obtain the evaluation of the model's performance. The recognition accuracy of low-efficiency land of different categories is shown in Table 3. The accuracy, recall rate, Kappa coefficient, etc. of each type of low-efficiency land are calculated in the table. Figure 10 shows the inconsistency ratio between low-efficiency land use and actual results. The quantitative disagreement shows the difference between the proportion of identified land categories and actual land categories, while the distribution disagreement shows the misidentification rate and omission rate in the identification results. From these data, we can analyze the reasons for the differences in identification accuracy between different low-efficiency land types.

[0121] Table 3. Accuracy verification results

[0122]

[0123] As can be seen from Table 3, Type 1 has the highest recognition accuracy, reaching 87%, and the Kappa coefficient is 0.739. This is related to the fact that the distribution of Type 1 inefficient land samples is concentrated in the central area of ​​Shenzhen City, and the impact of spatial heterogeneity on the model is relatively small. The accuracy of Type 2 is 81.3%, and the Kappa coefficient is 0.609; the accuracy of Type 3 is slightly lower, at 80.4%, and the Kappa coefficient is 0.804. The reason is that these two types of inefficient land are distributed over a wide range, which increases spatial heterogeneity and affects the recognition accuracy of the model. The average accuracy of the three categories reached 82.9%, and the Kappa coefficient reached 0.652, which shows that the model can well identify the disorderly distributed inefficient land in the city. Figure 8It also describes in more detail the way the model makes mistakes when identifying different types of inefficient land. For example, Type 1 and Type 3 are mainly based on inconsistent quantity, while Type 2 is mainly based on inconsistent allocation (including missed identification and misidentification). Among them, Type 3 has the highest inconsistency ratio, while Type 1 and Type 2 have similar inconsistencies. This shows that the results of Type 1 and Type 2 are more reliable, while the identification results of Type 3 still have room for improvement. The red and blue lines show the proportion of inconsistent types in the overall verification sample. It can be seen that the inconsistent quantity accounts for about 0.15, while the inconsistent allocation accounts for 0.1, indicating that this model performs well in both the quantity and allocation of identification results, and can cope with the complex spatial heterogeneity of inefficient land.

[0124] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for investigating and identifying inefficient land use, characterized in that: include: Obtain low-efficiency land samples from the comprehensive characteristic indicator system for identifying low-efficiency land; K-Means was used to cluster the low-efficiency land samples and XGBoost-SHAP was used to analyze the key influencing factors of different types of low-efficiency land samples, and the importance weights of environmental variable features were quantified; Using the spatial inference method of geographic similarity classification, the obtained environmental variable feature importance weights are used to conduct similarity assessment on all the plots to be identified, and the potential inefficient land use types are obtained.

2. The method for investigating and identifying low-efficiency land use according to claim 1, characterized in that: The comprehensive characteristic indicator system for identifying low-efficiency land includes four characteristic indicators: land use conditions, economic efficiency, social efficiency and ecological environmental efficiency; the land use conditions take into account the basic conditions of the land itself, including terrain conditions and building conditions; the economic efficiency takes into account the economic benefits of the land, and in addition to the direct economic output data of the land, it also selects indicators that can indirectly demonstrate the level of economic development; the social efficiency takes into account the population carrying capacity and public service level of the land; the ecological environmental efficiency takes into account the ecological sensitivity and environmental fragility of the land.

3. The method for investigating and identifying low-efficiency land use according to claim 1, characterized in that: The key factors affecting the clustering of inefficient land sample data using K-Means and the analysis of different types of inefficient land samples using XGBoost-SHAP include: K-Means clustering is used to cluster the low-efficiency land samples, and the low-efficiency land samples that meet the K-class characteristic differences are divided to obtain the clustering results; XGBoost-SHAP was introduced to perform classification prediction based on clustering results, analyze the key influencing factors of different types of inefficient land samples, and extract the importance weights of environmental variable features; the SHAP value explains the model by calculating the marginal contribution of features to the model output under different combinations: Where F is the set of all features, S is a subset of F that does not contain feature i; f(S) is the model prediction value containing only the features of subset S; is the SHAP value of feature i, indicating its marginal contribution to the model prediction value.

4. The method for investigating and identifying low-efficiency land use according to claim 1, characterized in that: The obtained environmental variable feature importance weights are used to perform similarity assessment on all the plots to be identified, and potential low-efficiency land use types are obtained, including: Mapping the environmental characteristic data of known low-efficiency land to all the plots to be identified, and obtaining the environmental covariate vector of each plot to be identified; According to the obtained environmental variable feature importance weights, the similarity score of each plot to be identified at the environmental covariate level is calculated, and the maximum similarity score is selected for output; The maximum similarity score output is compared with the set threshold. If it is higher than the threshold, the to-be-identified land parcel corresponding to the similarity score is output as a potential low-efficiency land. If it is lower than the threshold, no data is output.

5. The method for investigating and identifying low-efficiency land use according to claim 4, characterized in that: Before outputting the maximum similarity score, uncertainty calculation is also performed to calculate the variance of the similarity score as uncertainty.

6. The method for investigating and identifying low-efficiency land use according to claim 4, characterized in that: The method of calculating the similarity score of each plot to be identified at the environmental covariate level according to the obtained feature importance weights and selecting the maximum similarity score for output includes: The importance weight of each environmental variable feature is multiplied by the similarity value of each environmental covariate of the plot to be identified, and then all are added together to obtain the similarity score of the plot to be identified at the environmental covariate level; Repeat the above steps, continuously replace the known low-efficiency land samples, obtain the similarity score between each plot to be identified and all known low-efficiency land plots, and select the maximum similarity score as the output.

7. The method for investigating and identifying low-efficiency land use according to claim 4, characterized in that: The algorithm formula of the environmental covariate vector of each plot to be identified is: e=(e1,e2,…,e m ) In the formula, e is the environmental vector at a given location, used to describe the environmental conditions, and m is the number of environmental covariates used.

8. The method for investigating and identifying low-efficiency land use according to claim 7, characterized in that: The algorithm formula for calculating the similarity score of each plot to be identified at the environmental covariate level based on the obtained environmental variable feature importance weights and selecting the maximum similarity score output is: w m =f(SHAP) Similarity i =Max(S i1 ,S i2 ,…,S ij ) S ij represents the environmental similarity between the vth environmental covariate of the unobserved location and the sample location j, w m environmental variable feature importance weights; is the mth weight of the vth environmental covariate.

9. The method for investigating and identifying low-efficiency land use according to claim 5, characterized in that: The uncertainty calculation method is: In the formula, (S i,1 ,S i,2 ,…,S i,n ) is the similarity between sample i and all low-efficiency land samples, p(S i,1 ,S i,2 ,…,S i,n ) is the value of sample i (S i,1 ,S i,2 ,…,S i,n ) probability.

10. The method for investigating and identifying low-efficiency land use according to claim 1, characterized in that: The indicator values ​​of each characteristic indicator were quantified at the plot scale and Pearson correlation analysis was performed.

Images