Urban residence land price evaluation method considering data imbalance and spatial heterogeneity

Abstract

The invention discloses an urban residence land price evaluation method considering data imbalance and spatial heterogeneity. The method comprises the steps of collecting historical residence land sample data of a research area and performing spatial position matching; performing hierarchical index evaluation processing on the historical residential land sample data by using a multi-hierarchical index system, and performing fusion arrangement to obtain a residential hierarchical index land price data set; the land price evaluation combination model carries out model training of residence land price prediction by using a residence level index land price data set, and a differential evolution algorithm model DE dynamically searches a hyper-parameter combination of an XGBoost model and carries out optimization processing; and obtaining residential land multi-source data of the research block to obtain indexes of all levels, inputting the indexes into the land price evaluation combination model, and outputting and obtaining a residential land price evaluation prediction result. According to the method, the double problems of unbalanced sample price data and insufficient spatial heterogeneity identification are solved, high-precision urban residential land price evaluation is realized, and the accuracy and automation level of land price evaluation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of territorial space planning and land resource assessment, and particularly relates to a method for evaluating urban residential land prices considering data imbalance and spatial heterogeneity. BACKGROUND

[0002] The scientific and accurate evaluation of urban residential land prices is an important decision basis for the government to implement precise residential market regulation and optimize land supply planning. The existing mainstream evaluation methods (market comparison method, income restoration method, cost approximation method, residual method and benchmark land price coefficient correction method) are suitable for single plot evaluation, and the evaluation results depend largely on the experience and subjective judgment of land valuers, which cannot meet the demand of large-scale and rapid batch evaluation, and cannot effectively capture the complex nonlinear factors affecting residential land prices. With the deepening of digital transformation in the real estate field, residential land price batch evaluation technology based on machine learning models has become a trend. Among them, general machine learning models such as random forest (Random Forest) are widely used in this field due to their strong fitting ability. However, these models still have obvious bottlenecks in practical application: 1. Imbalanced sample price data: Within the city, residential land transaction prices are mostly concentrated in the low and medium range, and high-value samples are relatively scarce, which is prone to form a local skewed distribution. When the evaluation range is expanded to a large scale, the data imbalance problem is further aggravated. When dealing with such imbalanced data, whether it is direct resampling or "peak clipping and valley filling" smoothing strategy, it is difficult to effectively deal with it. Direct resampling will destroy its statistical representativeness due to simple replication or random discard. The smoothing strategy essentially forces the homogenization of data distribution, which weakens the uniqueness of special value segments, leading to insufficient learning of key samples. This makes the model unable to effectively learn from key samples, and further causes global bias in evaluation.

[0003] 2. Insufficient recognition of spatial heterogeneity: General models are limited by initial weight sensitivity and parameter adjustment limitations, and often have difficulty in effectively identifying and quantifying the spatial variation of key attributes such as location and supporting facilities and evaluation prices, leading to the model's inability to adapt to the dynamic evolution of land price driving factors with geographic location. In the context of large-scale evaluation, the heterogeneity challenge brought by complex geographical environment is further intensified, making the model have fitting difficulties, etc., ultimately affecting the effectiveness and accuracy of price evaluation.

[0004] In summary, the existing methods have not fully solved the two key problems of "imbalanced sample price data" and "insufficient recognition of spatial heterogeneity". It is urgent to build a new generation of residential land price evaluation method and system that can fundamentally solve these problems and accurately depict the prices of urban residential land. SUMMARY

[0005] The purpose of this invention is to provide an urban residential land price assessment method that takes into account data imbalance and spatial heterogeneity. By constructing a three-element fusion architecture of data reconstruction, feature learning, and model optimization, this invention innovatively proposes an urban residential land price assessment method that combines the SMOGN data balancing method with a land price assessment model for the land sector. This method accurately identifies the spatial heterogeneity of urban residential land prices on a large scale, thereby achieving high-precision urban residential land price assessment. All performance indicators have been significantly improved, enhancing the accuracy and automation level of land price assessment.

[0006] The objective of this invention is achieved through the following technical solution: A method for assessing urban residential land prices that takes into account data imbalance and spatial heterogeneity includes the following steps: S1. Collect historical residential land sample data of the study area and match its spatial location. The historical residential land sample data includes multi-source data of residential land and corresponding residential land prices. The study area includes several study blocks.

[0007] S2. Construct a multi-level indicator system with a hierarchical structure. Use the multi-level indicator system to evaluate the historical residential land sample data according to spatial location. Then, link the residential land price of the historical residential land sample data with all the hierarchical indicators and integrate them into a residential hierarchical indicator land price dataset.

[0008] S3. Construct a combined land price assessment model comprising an XGBoost model and a Differential Evolutionary Algorithm (DE) model. This combined model uses a residential tiered indicator land price dataset for training to predict residential land prices. The XGBoost model is used for layer-by-layer feature learning and nonlinear transformation fitting of all tiered indicators, achieving deep learning training to predict residential land prices based on all tiered indicators. The Differential Evolutionary Algorithm (DE) model dynamically searches for and optimizes hyperparameter combinations of the XGBoost model.

[0009] S4. Obtain multi-source data on residential land in the study area, use a multi-level indicator system to obtain all level indicators, input them into the trained land price assessment combination model, and output the residential land price assessment prediction results.

[0010] To better implement this invention, the study area includes several study plots. In method S3, multi-source residential land data of the study plots is obtained, and a multi-level indicator system is used to obtain all levels of indicators for the study plots. All levels of indicators for the study plots are then input into the trained land price assessment combination model, and the residential land price assessment prediction results for the study plots are output.

[0011] Preferably, in method S1, the SMOGN algorithm is used to calculate the coefficient of variation of historical residential land sample data according to price range and / or spatial location, and to identify whether the distribution is skewed. Data sets with good coefficient of variation and no distribution skewness are evaluated as balanced samples, and data sets with poor coefficient of variation and / or distribution skewness are evaluated as imbalanced samples. Based on The probability distribution function is inversely proportional to its probability density function. Set a correlation threshold The imbalanced sample was divided into zones for scarce, high-value residential land samples. and the normal distribution of residential land sample zoning The expression is as follows: ,in For the sample The actual value of residential land price For the sample Features This is an imbalanced sample.

[0012] Zoning of residential land samples with normal distribution A random undersampling process is implemented by setting an undersampling ratio. The sample of scarce, high-value residential land is then zoned. Set an oversampling ratio for oversampling processing.

[0013] Preferably, the sample of scarce, high-value residential land is divided into zones. Set the oversampling ratio, find and calculate the sum of the oversampling ratios of the data points that need to be oversampled. The distance to each neighboring sample is used, while a safe distance threshold is set. To determine the interpolation strategy, if the data points that need to be oversampled are greater than the safe distance threshold between them and their neighboring samples... Then, a Gaussian noise injection method is used to interpolate and generate new residential land sample data points. If the data points requiring oversampling are less than or equal to a safe distance threshold from neighboring samples... Then, the SMOTER method is used to interpolate new residential land sample data points within a local area.

[0014] Preferably, in method S1, the multi-source data sources for residential land include POI data of the study area, satellite remote sensing image data, statistical yearbook data, residential land transfer data, and population density datasets. Historical residential land sample data and residential land prices are correlated and matched according to spatial location. The multi-level indicator system, from top to bottom, includes first-level dimensions, second-level indicators, and third-level indicators. The first-level dimensions include spatial service field strength dimension, convenience topology network dimension, development potential dimension, land parcel attribute dimension, and economic level radiation field dimension. The spatial service field strength dimension includes second-level indicators for educational facilities, commercial facilities, and other supporting facilities. The convenience topology network dimension includes second-level indicators for the accessibility potential field of living service facilities. The development potential dimension includes second-level indicators for policy restrictions and housing demand. The land parcel attribute dimension includes second-level indicators for land parcel information. The economic level radiation field dimension includes second-level indicators for regional economic background. Each second-level indicator includes several third-level indicators.

[0015] Preferably, in method S2, a multi-level indicator system is used to evaluate historical residential land sample data according to spatial location. The hierarchical indicators are the sum of first-level, second-level, and third-level indicators obtained from the evaluation and statistical processing of historical residential land sample data, arranged hierarchically from bottom to top. Spatial service field strength dimension-related indicator data is obtained from the historical residential land sample data, and location calibration coding processing is performed based on the classification coding system to correspond to the location of the plots in the residential land price area. Convenience topology network dimension-related indicator data is obtained from the historical residential land sample data. Based on the road network topology, the Dijkstra algorithm (single-source shortest path algorithm) is used to calculate the shortest path distance between the location of the plot corresponding to the residential land price and various public service facilities. Development potential dimension-related indicator data is obtained from the historical residential land sample data. Bilinear interpolation and AOI masking technology are used to construct a plot-level standardized parameter system containing development intensity and demand potential, and the plot-level density index is evaluated. Data on local attributes of land parcels were obtained from historical residential land sample data. Kriging interpolation was used to construct a spatially continuous surface representing the benchmark land price. Topological overlay was then employed to achieve a precise mapping to land parcel units, obtaining standardized location value parameters. Based on the acquired imagery, the Normalized Difference Vegetation Index (NDVI) was calculated to quantify the ecological quality level of the land parcels. Data on economic radiation field dimensions were also obtained from historical residential land sample data and mapped to the corresponding location areas of residential land parcels using a spatial allocation mechanism.

[0016] Preferably, in method S3, the hyperparameter combination optimized by the differential evolution algorithm model DE includes maximum tree depth, learning rate, feature sampling ratio, training sample ratio, number of learners, and regularization coefficient. The method for optimizing the hyperparameter combination by the differential evolution algorithm model DE is as follows: S31. Define the parameters in the hyperparameter combination as candidate solutions.

[0017] S32. Perturb the current individual vector, introduce a difference vector to enhance the search capability, perform mutation operation, explore unknown areas through the difference vectors between individuals, enhance the model's ability to perceive low sample density areas, and improve the generalization of data-sparse high-value plots.

[0018] S33. By controlling the crossover probability (CR), information exchange is achieved between the current individual and the mutated individual, and the crossover operation is performed: ,in For the test vector dimension, The first of the variant vectors dimension, For the original individual's first dimension, For the first dimensional random numbers, For crossover probability, For randomly selected dimension indexes, via To ensure traversal of the parameter space, maintain population diversity to adapt to the multimodal spatial distribution of housing prices.

[0019] S34, Each Individual A corresponding land price assessment portfolio model is constructed, and the residential land price assessment error (MSE) of the validation set is used as the fitness function, as shown in the following expression: ,in For individuals fitness function, The mean squared error of the validation set. M To determine the number of samples in the validation set, For the first j The true value of the validation sample. For the model to the first j The predicted value of each validation sample.

[0020] ,in For the first in the next generation population individual, These are the experimental vectors generated through mutation and crossover.

[0021] S35. Finally, iterate until convergence to obtain the optimal hyperparameter combination of the XGBoost model.

[0022] Preferably, in method S2, the residential land price dataset consists of several residential land price data, and each residential land price data is associated with and integrated with residential land prices and all corresponding level indicators.

[0023] Preferably, the multi-source residential land data in methods S1 and S4 are preprocessed. The data preprocessing method is as follows: the Isolation Forest algorithm is used to identify and remove abnormal samples in the data in an unsupervised learning mode.

[0024] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) In the data processing stage, the present invention uses the SMOGN algorithm as the core processing module to effectively solve the problem of imbalance in price sample distribution. In the feature learning stage, the deep nonlinear transformation structure of the XGBoost model is used to explore the complex correlation mechanism between various influencing factors and urban residential land prices. In the model optimization stage, the coupled differential evolution algorithm model DE is used as the core of adaptive optimization to realize intelligent optimization of network structure and hyperparameters. The present invention constructs a collaborative optimization architecture of "data reconstruction-feature learning-model optimization" and forms a complete land price assessment technical solution. The present invention not only maintains the prediction accuracy, but also significantly improves the adaptability and stability of the model in the case of imbalanced samples, which can provide scientific and reliable technical support for land asset management, market monitoring and urban planning decision-making.

[0025] (2) This invention constructs a three-element fusion architecture of data reconstruction-feature learning-model optimization, and innovatively proposes an urban residential land price assessment method based on the SMOGN data balancing method and land price assessment combined model for the land field. It accurately identifies the spatial heterogeneity of urban residential land prices on a large scale, thereby achieving high-precision urban residential land price assessment. All performance indicators have been significantly improved, enhancing the accuracy and automation level of land price assessment.

[0026] (3) This invention significantly enhances the adaptability to unbalanced sample distribution and effectively captures the complex nonlinear relationship between multi-source features and land prices; it also demonstrates higher reliability and automation in large-scale cross-regional assessments and can play an important supporting role in land asset management, market dynamic monitoring and urban planning.

[0027] (4) This invention overcomes the problem of imbalanced urban residential land price data samples. From a data perspective, this invention effectively alleviates the long-standing problem of price assessment blind spots caused by sample imbalance in the field of land appraisal by optimizing the distribution structure of training samples. Based on the adaptive mechanism of the SMOGN method, this invention can automatically execute data reconstruction strategies according to sample distribution characteristics, realizing full automation of the data processing process. Through intelligent sample balancing technology, the human resource input and time consumption in the data processing process are significantly reduced, effectively improving the efficiency and economic benefits of the data preprocessing stage. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the method flow of the present invention; Figure 3 This is a schematic diagram illustrating the principle of sample partitioning using the SMOGN algorithm in the embodiment. Figure 4 This is a schematic diagram illustrating the principle of hyperparameter combination optimization using the differential evolution algorithm (DE) model in the land price assessment combination model in the embodiment. Figure 5 This is a schematic diagram comparing the statistical results before and after undersampling or oversampling of the imbalanced samples in the example. Figure 6 This is a graph showing the statistical results of data correction for the balanced sample in the example. Figure 7 This is a schematic diagram of the fitting results for residential land prices in the example study area; Figure 8 This is a schematic diagram illustrating the comparison between actual values ​​and predicted values ​​for various cities in the example study area. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to embodiments: Example like Figure 1 As shown, a method for assessing urban residential land prices that takes into account data imbalance and spatial heterogeneity includes the following steps: S1. Collect historical residential land sample data of the study area and match its spatial location. The historical residential land sample data includes multi-source data of residential land and corresponding residential land prices. The study area includes several study blocks, and each study block includes several study plots.

[0030] The multi-source data sources for residential land in this invention include POI data of the study area, satellite remote sensing image data, statistical yearbook data, residential land transfer data, and population density datasets. Historical residential land sample data is correlated and matched with residential land prices based on spatial location. In this embodiment, an example is given: the historical residential land sample data is collected from 2020 to 2023 from 11 prefecture-level cities in a certain province (example: Zhejiang Province), yielding a total of 4299 original samples (i.e., sample data for 4299 land parcels); spatial location matching is assisted by using the unified coordinate system CGCS2000 to complete the spatial registration of all land parcel data. The study area covers 11 prefecture-level cities in a certain province. For example, POI data corresponding to the indicator features were obtained from the official API of Gaode Maps. For example, Landsat 8 images of the study area were obtained from the Earth Explorer platform (https: / / earthexplorer.usgs.gov) (satellite remote sensing image data was extracted from Landsat 8 images). For example, data on per capita living consumption expenditure, annual residential housing sales area, residential housing sales volume, and per capita GDP of urban and rural residents in various cities and counties were obtained from the "Statistical Yearbook of a Certain Province". Population density data were also collected from the LandScan dataset published by Oak Ridge National Laboratory (https: / / landscan.ornl.gov).

[0031] S2. Construct a multi-level indicator system with a hierarchical structure. Use the multi-level indicator system to evaluate the historical residential land sample data according to spatial location. Preferably, the multi-level indicator system includes first-level dimensions, second-level indicators and third-level indicators from top to bottom. The first-level dimensions include spatial service field strength dimension, convenience topology network dimension, development potential dimension, land parcel attribute dimension and economic level radiation field dimension.

[0032] The spatial service field strength dimension includes secondary indicators such as educational facilities, commercial facilities, and other supporting facilities. Below these secondary indicators are several levels of tertiary and quaternary indicators. The spatial service field strength dimension is a crucial explanatory variable for urban residential land value assessment, its core function being to accurately quantify the gradient impact of the spatial distribution of public service facilities on urban residential land prices. To comprehensively characterize the potential impact of residents' activity range on land prices, this embodiment constructs an adaptive bandwidth radiation circle based on the centroid of the land parcel, achieving a refined characterization and quantitative expression of the spatial layout of facilities. After spatial calibration, the acquired data is used to calculate the supporting capacity of educational, commercial, and other public facilities within the range, characterizing the spatial agglomeration gradient and radiation intensity of different public service resources.

[0033] The educational facilities (also known as the educational density field) address the issue of the vague consideration of educational facilities in existing evaluation systems. This invention proposes using the "number of kindergartens, primary schools, and secondary schools" as the core quantitative indicator to comprehensively measure the educational convenience of land. This indicator covers all stages of basic education, objectively reflects the degree of concentration of regional educational resources, and directly reflects the concentration and richness of educational resources within the region. Commercial facilities (also known as the commercial density field): To accurately quantify the convenience of commercial services, this invention constructs a multi-level indicator system covering commercial service institutions, shopping centers, hotels, bars, and convenience stores. This system achieves a comprehensive quantitative representation of everything from regional commercial radiation to daily life convenience through three scales: macro-level commercial capacity (number of shopping centers and commercial service institutions), business and leisure vitality (number of hotels and bars), and community service convenience (number of convenience stores). This multi-level indicator structure collectively achieves a comprehensive quantitative assessment of commercial facilities from macro to micro, from daily life to consumption, and from daytime to nighttime. In the land price evaluation system, in addition to core elements such as education and commerce, supporting factors such as transportation accessibility, medical resources, and living environment have a significant impact on land price formation. Therefore, this invention selects "number of bus stops" as a direct quantitative indicator of public transportation convenience, reflecting the commuting efficiency and transportation network density of a region; uses "number of medical points" to characterize the spatial distribution of public health service resources, reflecting the region's medical security level and emergency service capabilities; and introduces "number of villa areas" as an indirect indicator of living environment quality, reflecting the region's cultural environment, ecological quality, and potential socio-economic characteristics through the concentration of high-end residences. By incorporating these three indicators into the evaluation system, a deep understanding of the implicit value of land and its comprehensive locational potential is achieved, providing important support for establishing a scientific and comprehensive land price evaluation system.

[0034] The convenience topology network dimension includes secondary indicators such as the accessibility potential field of living service facilities. Establishing this dimension further quantifies the accessibility efficiency of residents' actual travel and constructs the accessibility potential field of living service facilities. Based on urban road network data, this invention uses Dijkstra's algorithm to calculate the shortest path distance from residential land parcels to eight types of facilities, constructing the accessibility potential field. This dimension overcomes the limitations of Euclidean distance, integrates road network topology, and can accurately reflect residents' travel costs and spatial interaction efficiency.

[0035] The development potential dimension (also known as the development potential dimension field) includes secondary indicators such as policy restrictions and housing demand. This invention constructs a development potential dimension field to quantify the future appreciation potential of land. This dimension field operates through a "population-planning" coupling mechanism. Its core lies in using "population density" as the core proxy variable on the demand side to assess the current market's housing demand pressure. It can comprehensively analyze the dynamic spatial relationship between regional development intensity and housing demand, thereby scientifically predicting the long-term value of land.

[0036] The land parcel ontological attribute dimension (also known as the land parcel ontological attribute dimension field) includes secondary indicators such as land parcel information. In the land price formation process, the benchmark land price set by the government serves as a basic price benchmark, influencing the price benchmark for land transfers. Simultaneously, ecological environment quality, as an important value influencing factor, objectively characterizes the ecological livability of a region through quantitative indicators such as the Normalized Difference Vegetation Index (NDVI). This invention uses the Kriging Interpolation method to construct a continuous surface model for the benchmark land price. Through topological overlay, it completes the spatial discretization processing at the parcel level to obtain the benchmark land price coverage for the entire province. The NDVI is used to measure the vegetation environment of the parcels, and a multi-dimensional feature space is constructed by integrating a cadastral ontological semantic network to systematically quantify the ecological quality status of the parcels. The calculation process of the NDVI is as follows: .

[0037] The economic level radiation field dimension (also known as the economic level radiation dimension field) includes secondary indicators such as regional economic background, and each secondary indicator includes several tertiary indicators. The regional economic situation is an important macroeconomic factor affecting urban residential land prices. Factors such as the level of economic development within a region affect the market demand for land and land transfer prices. By quantitatively analyzing various economic indicators, the radiation effect of economic development on land value can be characterized.

[0038] In method S1, historical residential land sample data is obtained. This invention adheres to the principles of scientific rigor, comprehensiveness, systematic approach, and feasibility, systematically considering the differentiated land market transaction levels and price drivers in different prefecture-level cities within the region. It accurately and comprehensively depicts the impact characteristics of urban residential land prices under different land market conditions. Therefore, based on thorough screening and verification, a multi-dimensional characteristic system (i.e., a multi-level indicator system) for analyzing urban residential land prices is constructed. This system integrates geospatial perception and urban big data technologies, covering the price composition of prefecture-level cities under different land market conditions with five primary dimensions (i.e., spatial service field strength, convenience topology network, development potential, land parcel attributes, and economic radiation field). This multi-level indicator system achieves a comprehensive depiction of the factors influencing urban residential land prices. An example of some contents of the multi-level indicator system in this embodiment is shown in the table below: Table 1. Examples of some hierarchical indicators in a multi-level indicator system.

[0039] The residential land prices from historical residential land sample data are correlated and integrated with all hierarchical indicators to form a residential hierarchical indicator land price dataset. Preferably, the residential hierarchical indicator land price dataset consists of several residential hierarchical indicator land price data, each of which is correlated and integrated with the residential land price and all corresponding hierarchical indicators.

[0040] In some embodiments, a multi-level indicator system is used to evaluate historical residential land sample data according to spatial location. The hierarchical indicators are the sum of first-level, second-level, and third-level indicators, which are associated in a hierarchical structure from bottom to top, obtained from the evaluation and statistical processing of historical residential land sample data. Spatial service field strength dimension-related indicator data are obtained from historical residential land sample data, and location calibration coding processing is performed on the corresponding plot location area based on a classification coding system. Convenience topology network dimension-related indicator data are obtained from historical residential land sample data. Based on the road network topology, the Dijkstra algorithm (single-source shortest path algorithm) is used to calculate the shortest path distance between the corresponding plot location and various public service facilities. Development potential dimension-related indicator data are obtained from historical residential land sample data. Bilinear interpolation and AOI masking technology are used to construct a plot-level standardized parameter system containing development intensity and demand potential, and the plot-level density index is evaluated. Data on local attributes of land parcels were obtained from historical residential land sample data. Kriging interpolation was used to construct a spatially continuous surface representing the benchmark land price. Topological overlay was then employed to achieve a precise mapping to land parcel units, obtaining standardized location value parameters. Based on the acquired imagery, the Normalized Difference Vegetation Index (NDVI) was calculated to quantify the ecological quality level of the land parcels. Data on economic radiation field dimensions were also obtained from historical residential land sample data and mapped to the corresponding location areas of residential land parcels using a spatial allocation mechanism.

[0041] In some embodiments, before using a multi-level indicator system for evaluation, the present invention further performs data evaluation processing on historical residential land sample data and, based on the evaluation results, performs the following strategy processing: Utilizing the SMOGN algorithm (which achieves balance correction of urban residential land sample data; the SMOGN algorithm discretizes the continuous land price spectrum into urban residential land price intervals with economic semantics through a price binning strategy, calculates the coefficient of variation of the number of plots in each price interval, and uses this to accurately measure the degree of balance in price distribution). When a significant distribution skewness is identified in the sample set, a dual-mode sampling mechanism consistent with urban residential land price data is initiated. If a city's... High-value land samples are relatively continuous and highly similar within the feature space. Therefore, the SMOTER strategy is used to synthesize representative new samples through reasonable interpolation in their nearest neighbor space. Conversely, when the characteristics of these scarce samples are relatively isolated or dissimilar, a strategy of injecting Gaussian noise is adopted to simulate the impact of potential factors such as policy regulation and market expectations, generating diversified new samples. This dual-mechanism synergistic operation aims to enhance the model's ability to learn and capture the inherent formation patterns of various price ranges for urban residential land, especially scarce price ranges. Historical residential land sample data is calculated by calculating the coefficient of variation of sample size according to price range and / or spatial location, and identifying whether the distribution is skewed (e.g., ...). Figure 3 As shown), datasets with good coefficients of variation and no skewed distribution are evaluated as balanced samples (e.g., ...). Figure 6 As shown, after correcting the balanced sample, the statistical results of the balanced sample after processing are as follows: Figure 6 As shown in the figure, datasets with poor coefficients of variation and / or skewed distributions are evaluated as imbalanced samples. Based on The probability distribution function is inversely proportional to its probability density function. Set a correlation threshold The imbalanced sample was divided into zones for scarce, high-value residential land samples. and the normal distribution of residential land sample zoning The expression is as follows: ,in For the sample The actual value of residential land price For the sample Features This is an imbalanced sample.

[0042] Zoning of residential land samples with normal distribution Setting an undersampling ratio for random undersampling aims to reduce the model's overfitting to concentrated low- and medium-priced land, thus mitigating the risk of overfitting to the main price range. This involves zoning the sample of scarce, high-value residential land. Oversampling is performed by setting an oversampling ratio. This applies to imbalanced samples (including residential land samples with normal distribution). Sample zones for scarce, high-value residential land ) Statistical results before and after undersampling or oversampling processing, for example Figure 5 As shown.

[0043] In some embodiments, such as Figure 3 As shown, the sample zoning of scarce, high-value residential land. Two strategies are used for oversampling, where the distance between the sampled data point and its neighboring samples is greater than... (In a case study of historical residential land sample data from a certain province between 2020 and 2023, the safe distance threshold) Using a value of 0.8, Gaussian noise injection was employed for interpolation to generate new samples of urban residential land transaction prices. This simulates unobserved random disturbances in the land market, generating diverse new samples that help improve the model's ability to learn about potential price fluctuations. When the distance between a sampled data point and its neighboring samples is less than a safety threshold... In this case, the SMOTER method is used to interpolate within this local area to form new samples, in order to maintain the local continuity and geographical feature structure of urban residential land prices in the feature space. The method includes: partitioning the scarce, high-value residential land samples. Set the oversampling ratio, find and calculate the sum of the oversampling ratios of the data points that need to be oversampled. (In a case study of historical residential land sample data from a certain province between 2020 and 2023,) The distance to 5 neighboring samples is selected, and a safe distance threshold is set. To determine the interpolation strategy, if the data points that need to be oversampled are greater than the safe distance threshold between them and their neighboring samples... Then, a Gaussian noise injection method is used to interpolate and generate new residential land sample data points. If the data points requiring oversampling are less than or equal to a safe distance threshold from neighboring samples... Then, the SMOTER method is used to interpolate new residential land sample data points within a local area.

[0044] In this embodiment, taking historical residential land sample data from a certain province from 2020 to 2023 as an example, the SMOGN algorithm is used for balance correction. First, the Isolation Forest algorithm is used to identify and remove outlier samples in an unsupervised learning mode. This process has the technical characteristics of not relying on sample labeling and being robust to high-dimensional data. Then, the identified outliers are effectively removed, and finally a clean model training set is obtained. A city-specific differentiated processing strategy is adopted to effectively address the heterogeneity of land markets in different regions. The specific implementation process is as follows: First, the training data is grouped by city label according to administrative divisions. Then, a standardized cleaning process is performed on the data subset of each city, including numerical feature screening and missing value detection, to form a standardized local dataset. On this basis, the system automatically diagnoses the distribution balance of each city dataset. When the sample distribution skewness is identified to reach a preset threshold, the SMOGN data reconstruction mechanism is intelligently triggered to ensure that the subsequent modeling process obtains the optimal training samples.

[0045] S3, such as Figures 2-4 As shown, a combined land price assessment model is constructed, comprising an XGBoost model and a differential evolution algorithm (DE) model. This combined model uses a residential land price dataset to train a residential land price prediction model. The XGBoost model is used to perform layer-by-layer feature learning and nonlinear transformation fitting on all level indicators, achieving deep learning training for predicting residential land prices based on all level indicators. Preferably, multi-source residential land data for the study plot is obtained, and all level indicators for the study plot are obtained using a multi-level indicator system. These indicators are then input into the trained combined land price assessment model, outputting the predicted residential land price for the study plot.

[0046] The Differential Evolutionary Algorithm (DE) model dynamically searches for and optimizes hyperparameter combinations of the XGBoost model. By integrating the high-dimensional nonlinear mapping capability of the XGBoost model with the adaptive hyperparameter optimization of the Differential Evolutionary Algorithm (DE), a two-layer implicit adaptive land price assessment combination model is formed, achieving accurate simulation of the complex interaction between regional characteristics and price drivers. Through the multi-layer network structure of the XGBoost model, the multi-dimensional land attributes and regional characteristics of the input are nonlinearly transformed and abstracted layer by layer, gradually uncovering the deep-seated patterns affecting urban residential land prices and achieving a fit to the complex price formation mechanism under different regional market environments. Then, DE, based on mutation, crossover, and selection mechanisms, is embedded to dynamically search for and optimize hyperparameter configurations in the XGBoost model, such as maximum tree depth, learning rate, feature sampling ratio, training sample ratio, number of learners, and regularization coefficient, to further improve the assessment accuracy of urban residential land at a large scale, enabling the land price assessment combination model to adapt to different regional market environments and price structures. In some embodiments, such as... Figure 4As shown, the hyperparameter combination for optimization of the Differential Evolutionary Algorithm (DE) model includes maximum tree depth, learning rate, feature sampling ratio, training sample ratio, number of learners, and regularization coefficient. The optimization method for the hyperparameter combination of the Differential Evolutionary Algorithm (DE) model is as follows: S31. Vector Encoding: Using parameters in a hyperparameter combination as candidate solution individuals, define each candidate solution individual. It includes the following content: ; The maximum tree depth; The learning rate; The coefficient of the regularization term; The number of learners; The proportion of training samples; This represents the feature sampling ratio.

[0047] S32, Mutation Operation: Perturb the current individual vector, introduce a difference vector to enhance the search capability, perform the mutation operation, explore unknown areas through the difference vectors between individuals, enhance the model's ability to perceive low sample density areas, and improve the generalization ability in sparse and high-value plots.

[0048] ,in This is the scaling factor.

[0049] S33. Crossover Operation: Information exchange between the current individual and the mutated individual is achieved through crossover probability (CR) control (its core mechanism lies in both introducing new genes to explore a broader solution space and retaining the effective components of the current individual to inherit optimization results, thus ensuring robust convergence of the algorithm). The crossover operation is executed as follows: ,in For the test vector dimension, The first of the variant vectors dimension, For the original individual's first dimension, For the first Dimensional random numbers (random numbers selected in this embodiment) ), Crossover probability (selected in this example) ), For randomly selected dimension indexes, via To ensure traversal of the parameter space, maintain population diversity to adapt to the multimodal spatial distribution of housing prices.

[0050] S34. Fitness Function and Selection: Each Individual A corresponding land price assessment portfolio model is constructed, and the residential land price assessment error (MSE) of the validation set is used as the fitness function, as shown in the following expression: ,in For individuals fitness function, The mean squared error of the validation set. M To determine the number of samples in the validation set, For the first j The true value of the validation sample. For the model to the first j The predicted value of each validation sample.

[0051] ,in For the first in the next generation population individual, These are the experimental vectors generated through mutation and crossover.

[0052] S35. Finally, iterate until convergence to obtain the optimal hyperparameter combination of the XGBoost model; in the case study of historical residential land sample data from a certain province from 2020 to 2023, the optimized hyperparameter combination of the XGBoost model includes... (Maximum tree depth) (Learning rate) (Regularization coefficient) (Number of learners) (Proportion of training samples) (Feature sampling ratio), the population size of the differential evolution algorithm model DE is 50, the maximum number of iterations is 100, the scaling factor F=0.5, and the crossover probability CR=0.7. During the model training phase, 20% of the data in the training set is independently divided as the validation set, and the mean square error on the validation set is used as the fitness evaluation function to guide the global search process in the hyperparameter space. In this embodiment, the land price assessment combination model is evaluated and compared with the traditional XGBoost model and RF model applied to residential land prices, as shown in Table 2.

[0053] Table 2 Comparison of this invention with other models

[0054] The SMOGN-DE-XGBoost method and system for accurate urban residential land price assessment constructed in this invention demonstrates excellent performance and robust generalization ability in assessments. The prediction model constructed using this invention's technical solution, validated on an independent test set, exhibits a mean standard error of 0.38 and a goodness of fit of 0.84. These technical indicators demonstrate the model's excellent generalization performance, maintaining stable prediction accuracy in new sample assessments. Compared to other commonly used models in current land price assessments, SMOGN-DE-XGBoost demonstrates higher fitting accuracy and stronger generalization stability on the test set, verifying its excellent ability to capture land price formation mechanisms in large-scale, complex, spatially heterogeneous environments. Furthermore, the XGBoost model is re-initialized for final training to ensure that the final model possesses good generalization ability while minimizing training error. The optimized model uses a two-layer adaptive mechanism to characterize the spatial heterogeneity of large-scale residential land prices, accurately capturing its attribute and process differentiation features, thereby outputting land transfer price assessment results with high accuracy and high stability. Preferably, the mean absolute percentage error (MAPE) and R² are used for model evaluation and generalization performance verification.

[0055] S4. Obtain multi-source data on residential land in the study area, utilize a multi-level indicator system to obtain all level indicators, and input them into the trained land price assessment combination model to output the residential land price assessment prediction results. This embodiment provides an example of residential land price fitting results for several plots in the study area, as shown below. Figure 7 As shown.

[0056] In some embodiments, the multi-source residential land data in methods S1 and S4 are preprocessed. The data preprocessing method is as follows: the Isolation Forest algorithm is used to identify and remove abnormal samples in the data in an unsupervised learning mode.

[0057] In a model training case study for predicting residential land prices using a combined land price assessment model based on historical residential land sample data from a certain province from 2020 to 2023, residential land prices in various cities within the province were assessed and predicted. The comparison results between the actual values ​​and the assessed and predicted values ​​for each city in the province are as follows: Figure 8 As shown. This invention improves the ability to identify spatial heterogeneity. The land price assessment combination model of this invention constructs a two-layer implicit adaptive structure, which effectively identifies and handles the technical bottleneck of spatial heterogeneity faced in the large-scale urban residential land price assessment. In the model verification stage, the MAPE (0.38) of the method proposed in this invention on the test set is significantly lower than that of the comparative models such as XGBoost (0.52) and Random Forest (0.5), while the R² (0.84) is also significantly higher.

[0058] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for urban residential land price evaluation considering data imbalance and spatial heterogeneity, characterized in that: The method comprises: S1, collecting historical residential land sample data of a research area and spatial position matching, the historical residential land sample data comprising residential land multi-source data and corresponding residential land price, the research area comprising a plurality of research blocks; S2, constructing a multi-level index system according to the upper and lower hierarchical architecture, and using the multi-level index system to perform hierarchical index evaluation processing on the historical residential land sample data according to the spatial position, and fusing and arranging the residential land price of the historical residential land sample data and all the hierarchical indexes into a residential hierarchical index land price data set in position; S3, constructing a land price evaluation combination model comprising an XGBoost model and a differential evolution algorithm model DE, the land price evaluation combination model being used for model training of residential land price prediction using the residential hierarchical index land price data set, the XGBoost model being used for layer-by-layer feature learning and nonlinear transformation fitting processing of all the hierarchical indexes and realizing deep learning training based on all the hierarchical indexes to predict the residential land price; the differential evolution algorithm model DE dynamically searches for the hyperparameter combination of the XGBoost model and optimizes the processing; S4, obtaining the residential land multi-source data of the research block, using the multi-level index system to obtain all the hierarchical indexes, and inputting the trained land price evaluation combination model to output the residential land price evaluation prediction result.

2. The urban residential land price evaluation method according to claim 1, characterized in that: The research block comprises a plurality of research plots; in method S3, the residential land multi-source data of the research plot is obtained, all the hierarchical indexes of the research plot are obtained using the multi-level index system, all the hierarchical indexes of the research plot are input into the trained land price evaluation combination model, and the residential land price evaluation prediction result of the research plot is output.

3. The urban residential land price evaluation method considering data imbalance and spatial heterogeneity according to claim 1, characterized in that: In the method S1, the SMOGN algorithm is used to calculate the sample number variation coefficient and identify the distribution skewness of the historical residential land sample data according to the price interval or / and spatial location. The data set with good variation coefficient and no distribution skewness is evaluated as balanced sample, and the data set with poor variation coefficient or / and distribution skewness is evaluated as unbalanced sample. Based on the correlation function of the probability distribution function and the probability density in inverse proportion , the correlation threshold is set , the unbalanced sample is divided into the scarce high-value residential land sample partition and the normal distribution residential land sample partition , and the expression is as follows: wherein is the actual value of residential land price for the sample , is the characteristic for the sample , is the unbalanced sample; Dividing a normal distribution residential land sample into zones Setting an undersampling ratio to perform random undersampling processing; and dividing a scarce high-value residential land sample into zones Setting an oversampling ratio to perform oversampling processing.

4. The urban residential land price evaluation method according to claim 3, characterized in that: Partitioning a sample of scarce high-value residential land Setting an oversampling ratio to find and calculate the distance between the data point needing oversampling and its nearest neighbor, while setting a safety distance threshold to divide the interpolation strategy, if the data point needing oversampling is greater than the safety distance threshold from its nearest neighbor, then a Gaussian noise injection method is used to interpolate a new residential land sample data point; if the data point needing oversampling is less than or equal to the safety distance threshold from its nearest neighbor, then a SMOTER method is used to interpolate a new residential land sample data point in a local range.

5. The urban residential land price evaluation method according to claim 1, which takes into account data imbalance and spatial heterogeneity, characterized in that: In method S1, the sources of the residential land multi-source data include research area POI data, satellite remote sensing image data, statistical yearbook data, residential land transfer data and population density data set, and the residential land multi-source data and the residential land price of the historical residential land sample data are associated and matched according to the spatial position; the multi-level index system comprises, from top to bottom, a first dimension, a second index and a third index, the first dimension comprises a spatial service field strength dimension, a convenience degree topological network dimension, a development potential dimension, a plot ontology attribute dimension and an economic energy level radiation field dimension, the spatial service field strength dimension comprises second indexes of education supporting, commercial supporting and other supporting, the convenience degree topological network dimension comprises second indexes of life service facility accessibility potential field, the development potential dimension comprises second indexes of policy restriction and housing demand, the plot ontology attribute dimension comprises second indexes of plot information, and the economic energy level radiation field dimension comprises second indexes of regional economic background, and each second index comprises a plurality of third indexes.

6. The urban residential land price evaluation method according to claim 5, which takes into account data imbalance and spatial heterogeneity, characterized in that: In method S2, the historical residential land sample data is evaluated by a multi-level index system according to the spatial position, and the hierarchical index is obtained by evaluating and statistically processing the historical residential land sample data, which is associated with the first-level dimension, the second-level index and the third-level index from bottom to top according to the hierarchical architecture; the spatial service field strength dimension related index data is obtained from the historical residential land sample data, and the position calibration coding processing is performed based on the classification coding system corresponding to the residential land price of the block position area; the convenience topology network dimension related index data is obtained from the historical residential land sample data, and the shortest path distance between the residential land price corresponding block position and each public service facility is calculated based on the road network topology structure by using the single-source shortest path algorithm Dijkstra; the development potential dimension related index data is obtained from the historical residential land sample data, and the block-level standardized parameter system including development intensity and demand potential is constructed by using bilinear interpolation and AOI mask technology, and the block-level density index is obtained by evaluation; the block local attribute dimension related index data is obtained from the historical residential land sample data, and the benchmark land price is constructed as a spatial continuous surface by Kriging interpolation, and the standardized location value parameter is obtained by accurately mapping the surface to the block unit by means of topological superposition; the normalized difference vegetation index NDVI is calculated based on the obtained image to quantify the ecological quality level of the block; the economic energy level radiation field dimension related index data is obtained from the historical residential land sample data, and it is mapped to the residential land price corresponding block position area by the spatial distribution mechanism.

7. The urban residential land price evaluation method according to claim 1, which takes into account data imbalance and spatial heterogeneity, characterized in that: In method S3, the hyperparameter combination optimized by the differential evolution algorithm model DE includes the maximum tree depth, the learning rate, the feature sampling ratio, the training sample ratio, the number of learners and the regularization coefficient. The differential evolution algorithm model DE optimizes the hyperparameter combination as follows: S31, the parameter definition in the hyperparameter combination is used as the candidate solution individual; S32, the current individual vector is disturbed, a difference vector is introduced to enhance the search ability, a mutation operation is performed, the unknown area is explored through the difference vector between individuals, the perception ability of the model to the low sample density area is enhanced, and the generalization of the data sparse high value block is improved; S33, the information exchange between the current individual and the mutated individual is realized by the crossover probability CR control, and the crossover operation is performed: ,in For the test vector dimension, The first of the variant vectors dimension, For the original individual's first dimension, For the first dimensional random numbers, For crossover probability, For randomly selected dimension indexes, via To ensure traversal of the parameter space and maintain population diversity to adapt to the multimodal spatial distribution of housing prices; S34, each individual A combination model of land price assessment is constructed, and the error MSE of residential land price assessment of the validation set is taken as the fitness function, expressed as follows: ,in For individuals fitness function, The mean squared error of the validation set. M To determine the number of samples in the validation set, For the first j The true value of each validation sample; For the model to the first j Predicted values ​​for each validation sample; wherein is the th individual in the next generation population, is the th individual, is a trial vector generated by mutation and crossover; S35, finally iterate to convergence, and obtain the optimal hyperparameter combination of the XGBoost model.

8. The urban residential land price evaluation method according to claim 1, which takes into account data imbalance and spatial heterogeneity, characterized in that: In method S2, the residential hierarchical index land price data set is composed of a plurality of residential hierarchical index land price data, and each residential hierarchical index land price data is associated with the residential land price and all corresponding hierarchical indexes.

9. The urban residential land price evaluation method according to claim 1, which takes into account data imbalance and spatial heterogeneity, characterized in that: The residential land multi-source data in method S1 and method S4 are all preprocessed, and the data preprocessing method is as follows: the isolated forest algorithm is used to identify and eliminate abnormal samples in the data in an unsupervised learning mode.

Images