Automatic land valuation method based on multi-source data fusion and machine learning

Abstract

The invention discloses an automatic land valuation method based on multi-source data fusion and machine learning, and the method comprises the following steps: collecting multi-source heterogeneous data of each land block, and carrying out the preprocessing; constructing a training sample set; dividing the training sample set; constructing and training a region sensitive type multi-level residual error enhancement valuation model; generating a residual error correction valuation result through the basic valuation branch and the multi-stage residual error correction branch; executing direct comparison path correction to generate a direct correction value; if the land parcel number of the comparable sample set is not smaller than a preset threshold value, taking the direct correction value as a final estimation result, and if the land parcel number of the comparable sample set is smaller than the preset threshold value, executing indirect comparison path correction to generate an indirect correction value; and fusing the direct correction value and the indirect correction value to generate a final estimation result. According to the method, multi-source data fusion and machine learning are utilized, land automatic valuation is achieved, and the method has the advantages of being high in accuracy, high in adaptability and good in credibility.

Description

Technical Field

[0001] This invention relates to the field of land valuation, and in particular to an automatic land valuation method based on multi-source data fusion and machine learning. Background Technology

[0002] Existing land valuation methods primarily rely on traditional appraisers' manual assessments of land parcels based on market research and experience, or on statistical modeling using single-source data such as historical transaction prices or limited geographic information. These methods often suffer from two problems: firstly, they struggle to simultaneously integrate multi-source, heterogeneous data such as remote sensing imagery, geographic information, historical transaction records, planning policy documents, basic land parcel attributes, and surrounding environmental information, resulting in insufficient comprehensive reflection of land parcel spatial characteristics, market dynamics, and policy constraints; secondly, the model processing has limited adaptability to regional differences and land use types, leading to significant biases in valuation results, particularly in the valuation of cross-regional or multi-use land parcels.

[0003] Meanwhile, most existing machine learning-based land valuation methods adopt a uniform modeling approach, ignoring regional and land category differences, and failing to achieve differentiated valuation adjustments for specific administrative divisions and land use types. Furthermore, traditional comparative methods rely on manual screening of comparable samples, lacking effective indirect correction methods when comparable samples are insufficient, leading to decreased reliability of valuation results under data scarcity conditions. Therefore, existing technologies still have significant shortcomings in multi-source data fusion, regional sensitivity modeling, and the coordination of direct and indirect comparison paths, necessitating the development of a new method to improve upon these advancements. Summary of the Invention

[0004] One objective of this invention is to propose an automatic land valuation method based on multi-source data fusion and machine learning. This invention utilizes multi-source data fusion and machine learning to achieve automatic land valuation, which has the advantages of high accuracy, high adaptability and high reliability.

[0005] An automatic land valuation method based on multi-source data fusion and machine learning according to an embodiment of the present invention includes the following steps: Collect multi-source heterogeneous data from various land parcels, preprocess the data, and then fuse the preprocessed multi-source heterogeneous data to construct land parcel feature vectors. A training sample set is constructed using the land parcel feature vector as input and the corresponding historical transaction price as label; Based on the administrative division code and land use code of each plot, generate regional identifier vectors and land type adaptation factors, and divide the training sample set to generate multiple regional-land type joint sub-sample sets; Based on the regional-land category joint subsample set, a regionally sensitive multi-level residual enhancement valuation model is constructed and trained. The regionally sensitive multi-level residual enhancement valuation model consists of a basic valuation branch, a multi-level residual correction branch, and a market calibration branch. Input the feature vector of the land parcel to be valued into the basic valuation branch to obtain the preliminary valuation result, and input the preliminary valuation result into the multi-level residual correction branch to generate the residual correction valuation result; Based on the residual correction estimation results, a direct comparison path correction is performed to generate a direct correction value; If the number of plots in the comparable sample set is not less than the preset threshold, the direct correction value will be used as the final valuation result. If the number of plots in the comparable sample set is less than the preset threshold, the indirect comparison path correction will be performed based on the residual correction valuation result to generate the indirect correction value. The direct and indirect correction values ​​are combined using a dynamic weighting strategy to generate the final valuation result.

[0006] Optionally, the multi-source heterogeneous data includes remote sensing image data, geographic information attributes, historical transaction records, planning policy texts, basic land parcel attributes, and surrounding environmental information.

[0007] Optionally, the generation of the land parcel feature vector specifically includes: Collect multi-source heterogeneous data from various plots, link and integrate the multi-source heterogeneous data according to the plot number, and establish a set of original plot data records with consistent structure. Feature extraction processing is performed on remote sensing image data to extract image grayscale values, texture directionality, regional boundary gradient values ​​and patch shape parameters. The extraction results are converted into fixed-length vector representations as the remote sensing image feature part in the plot feature vector. The geographic information attributes are processed by structured numerical coding to extract the boundary contours, administrative division levels, topographic relief features, boundary complexity and spatial shape compactness of the land parcels, and then transformed into spatial structure feature vectors, which serve as the geographic information feature part of the land parcel feature vectors. The relevant fields in the historical transaction records are cleaned, transformed, and standardized. The historical transaction unit price, transaction time interval, land use term, transaction method type, and total transaction area are extracted and arranged in the order of the fields into a structured transaction feature representation, which serves as the transaction history feature part in the land parcel feature vector. Semantic expression transformation processing is performed on the planning policy text, and the planning policy information after semantic extraction and word vector generation is mapped into a fixed-length vector. The mapping result is then embedded into the policy feature part of the land parcel feature vector. The basic attributes and surrounding environmental information of the land parcel are structured and processed. The land use code, plot ratio, land level, ownership category, accessibility of surrounding roads, number of educational facilities, green coverage ratio, and density of commercial facilities are normalized and one-hot encoded. The results are organized into attribute environmental feature items, which constitute the attribute environmental feature part in the land parcel feature vector. The remote sensing image features, geographic information features, transaction history features, policy features, and attribute environment features are spliced ​​together in a preset order to generate a land parcel feature vector with consistent dimensions.

[0008] Optionally, the generation of the multiple region-land type joint subsample sets specifically includes: The administrative division code of each training sample is parsed hierarchically. The four levels of administrative division information of province, city, district / county and street are transformed into discrete feature vectors by using one-hot coding method. The feature vectors of different levels are combined to generate the corresponding regional identifier vector, which is used to reflect the regional feature differences in the sample set. For each training sample, the land use code is converted to reflect the land use. The seven land use categories—arable land, residential land, commercial land, industrial land, transportation land, water area land, and public management service land—are mapped to land use adaptation factors using a one-hot coding method. These factors are used to reflect the differences in land use in the sample set. The training sample set is divided according to the generated regional identifier vector and land type adaptation factor, and the samples are grouped into multiple regional-land type joint sub-sample sets.

[0009] Optionally, the construction and training of the region-sensitive multi-level residual enhancement estimation model specifically includes: Based on the regional-land category joint subsample set, a regionally sensitive multi-level residual enhancement valuation model is constructed. The regionally sensitive multi-level residual enhancement valuation model consists of a basic valuation branch, a multi-level residual correction branch, and a market calibration branch. The basic valuation branch uses a ResMLP network, which takes the land parcel feature vector as input and generates a preliminary valuation result for the target land parcel. The ResMLP network is a multilayer perceptron with residual connections, including multiple fully connected layers and nonlinear activation layers, with residual connections added between each layer; The multi-level residual correction branch includes a residual calculation unit, a step-by-step residual fitting unit, and a residual superposition unit. The residual calculation unit calculates the residual based on the difference between the preliminary valuation result and the historical transaction price. The step-by-step residual fitting unit fits the residuals of each layer in sequence. The residual superposition unit accumulates the residual results of each layer and outputs the residual correction value. During the training process, the land use adaptation factor is used to adjust the residual correction weight of each layer. The market calibration branch includes a direct comparison path and an indirect comparison path. The direct comparison path constructs a comparable sample set based on the similarity of land parcel feature vectors and generates a direct correction value by combining the transaction time. The indirect comparison path constructs an indirect reference set based on the mapping of administrative division levels, similarity of land use types, and similarity of the surrounding environment and generates an indirect correction value. Using historical transaction prices as a supervisory signal, the model is iteratively trained to minimize the error between the predicted results and the actual transaction prices. The model parameters are then updated until convergence, resulting in a trained region-sensitive multi-level residual enhancement valuation model.

[0010] Optionally, the generation of the residual correction estimation result specifically includes: The feature vector of the land parcel to be valued is input into the basic valuation branch, and the residual multilayer perceptron structure is used for calculation. During the layer-by-layer calculation, the input feature is multiplied by the corresponding weight matrix and a bias vector is added. After processing by a nonlinear activation function, it is added to the input to form the output feature vector of the next layer. When the basic estimation branch reaches the last layer, the weight matrix and bias vector of the last layer are used to perform a linear mapping on the output feature vector of the previous layer to obtain the preliminary estimation result. Input the preliminary valuation results into the multi-level residual correction branch to calculate the initial residuals; Based on the preliminary valuation results and the initial residuals, the residuals are fitted step by step. In the residual correction branch of the k-th layer, the land feature vector and the residual value of the previous layer are used as inputs and the k-th layer residual correction value is generated by mapping through the multilayer perceptron. All residual correction values ​​are summed sequentially and then added to the preliminary valuation result to generate the residual correction valuation result.

[0011] Optionally, the generation of the direct correction value specifically includes: Use the residual corrected estimation results as input; For the feature vector of the land parcel to be estimated, cosine similarity is calculated between it and the feature vector of each land parcel in the training sample set to generate a similarity score. All plots in the training sample set are sorted according to their similarity scores, and the top N plots with the highest similarity scores are selected to form a comparable sample set. For each plot of land in the comparable sample set, calculate the transaction time difference factor; Calculate the weighting coefficients; Weighted fitting is performed on the historical transaction prices in the comparable sample set based on weighted coefficients. The historical transaction price of each comparable land parcel is multiplied by the corresponding weighted coefficient, and then all multiplications are summed to obtain the weighted correction value. The direct correction value is obtained by adding the residual correction estimate to the weighted correction value.

[0012] Optionally, the generation of the indirect correction value specifically includes: If the number of plots in the comparable sample set is greater than or equal to the preset threshold, the direct correction value obtained from the direct comparison path will be used as the final valuation result. If the number of plots in the comparable sample set is less than a preset threshold, the indirect comparison path correction is initiated, and the residual correction valuation result is used as input. Based on the administrative division code of the land parcel to be valued, an administrative division hierarchy mapping is performed to obtain an administrative division reference set; Based on the land use type code of the land parcel to be valued, the land use type similarity is calculated, and a land use type reference set is generated. The land use type reference set consists of reference land parcels with the same land use type as the land parcel to be valued. Based on the surrounding environment information of the land parcel to be valued, the accessibility of surrounding roads is extracted to form a surrounding environment reference set, which consists of reference land parcels with similar environmental conditions to the land parcel to be valued. The environmental conditions being similar to those of the land parcel to be valued means that when the difference between the accessibility of the surrounding roads of the reference land parcel and that of the land parcel to be valued is less than 10%, the environmental conditions are considered similar. The administrative division reference set, the land use type reference set, and the surrounding environment reference set are combined to form an indirect reference set, which consists of reference plots that simultaneously meet the administrative division conditions, land use conditions, and surrounding environment conditions. Within the indirect reference set, calculate the mean deviation of the region; Based on the regional mean deviation, an indirect correction is performed on the residual correction estimation result. The indirect correction value is obtained by adding the regional mean deviation to the residual correction estimation result.

[0013] Optionally, generating the final valuation result specifically includes: After obtaining the direct and indirect correction values, a dynamic weighting strategy is constructed. In the dynamic weighting strategy, the number of plots in the comparable sample set is determined; Determine the feature similarity score; Determine the confidence level of the regional valuation bias; The direct correction weight is calculated by multiplying the number of plots in the comparable sample set by the feature similarity score, and the regional estimation deviation confidence level is used as the indirect correction weight. The direct and indirect correction values ​​are weighted using direct and indirect correction weights respectively, and the two weighted results are added together to generate the final valuation result.

[0014] The beneficial effects of this invention are: This invention, by introducing multi-source data fusion and machine learning technologies, can effectively integrate heterogeneous data from multiple sources, such as remote sensing imagery, geographic information attributes, historical transaction records, planning policy texts, basic land parcel attributes, and surrounding environmental information, during the automatic land valuation process. By constructing a unified land parcel feature vector, this invention achieves standardization and unified expression of data from different sources and with different structures, enabling the valuation model to comprehensively reflect the differences in land parcels in terms of spatial pattern, market behavior, and policy conditions. This data fusion method avoids the limitations of existing technologies that rely on only a single data source or limited features, making the valuation input base richer and more comprehensive, thereby improving the scientificity and credibility of the valuation results.

[0015] In terms of modeling, the region-sensitive multi-level residual enhancement valuation model proposed in this invention possesses differentiated modeling capabilities. By introducing regional identifier vectors and land use adaptation factors, and dividing the training sample set into multiple region-land use joint sub-sample sets, the model can achieve differentiated parameter updates and residual corrections for land parcels of different regions and uses during iterative training. This mechanism ensures the model's adaptability in cross-regional and cross-use scenarios, solving the shortcomings of traditional methods in taking into account both regional and use differences. In particular, the basic valuation branch generates preliminary valuation results, the multi-level residual correction branch progressively fits and accumulates residual correction values, and the market calibration branch achieves secondary correction through direct and indirect comparison paths. This enables the model to both fit historical transaction patterns and maintain stability when samples are insufficient, thereby significantly improving the model's valuation accuracy and robustness.

[0016] Furthermore, this invention balances the complementarity of direct and indirect comparisons in its comparison method design. When there are sufficient comparable samples, direct comparison is used for correction, ensuring that the valuation results closely reflect market transaction behavior. When comparable samples are insufficient, an indirect comparison path is used to construct an indirect reference set based on administrative division level, usage type, and similarity of the surrounding environment. This, combined with regional mean deviation, achieves valuation correction, solving the problem of insufficient valuation reliability under conditions of scarce samples in traditional methods. Simultaneously, the dynamic weighting strategy proposed in this invention determines the fusion weights based on the number of comparable samples, feature similarity scores, and regional valuation deviation confidence levels, achieving a reasonable balance between direct and indirect corrections in the final valuation results, further enhancing the robustness and generalizability of the model. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1This is a flowchart of an automatic land valuation method based on multi-source data fusion and machine learning proposed in this invention; Figure 2 This is a structural framework diagram of a regionally sensitive multi-level residual enhancement valuation model based on multi-source data fusion and machine learning, which is proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 An automatic land valuation method based on multi-source data fusion and machine learning includes the following steps: Collect multi-source heterogeneous data from various land parcels, preprocess the data, and then fuse the preprocessed multi-source heterogeneous data to construct land parcel feature vectors. A training sample set is constructed using the land parcel feature vector as input and the corresponding historical transaction price as label; Based on the administrative division code and land use code of each plot, generate regional identifier vectors and land type adaptation factors, and divide the training sample set to generate multiple regional-land type joint sub-sample sets; Based on the regional-land category joint subsample set, a regionally sensitive multi-level residual enhancement valuation model is constructed and trained. The regionally sensitive multi-level residual enhancement valuation model consists of a basic valuation branch, a multi-level residual correction branch, and a market calibration branch. Input the feature vector of the land parcel to be valued into the basic valuation branch to obtain the preliminary valuation result, and input the preliminary valuation result into the multi-level residual correction branch to generate the residual correction valuation result; Based on the residual correction estimation results, a direct comparison path correction is performed to generate a direct correction value; If the number of plots in the comparable sample set is not less than the preset threshold, the direct correction value will be used as the final valuation result. If the number of plots in the comparable sample set is less than the preset threshold, the indirect comparison path correction will be performed based on the residual correction valuation result to generate the indirect correction value. The direct and indirect correction values ​​are combined using a dynamic weighting strategy to generate the final valuation result.

[0021] In this embodiment, the multi-source heterogeneous data includes remote sensing image data, geographic information attributes, historical transaction records, planning policy texts, basic land parcel attributes, and surrounding environmental information.

[0022] In this embodiment, the generation of the land parcel feature vector specifically includes: Collect multi-source heterogeneous data from various plots, link and integrate the multi-source heterogeneous data according to the plot number, establish a plot original data record set with consistent structure, ensure that each plot has a complete information source, and maintain the consistency of field order and dimension in the data structure; Feature extraction processing is performed on remote sensing image data to extract image grayscale values, texture directionality, regional boundary gradient values ​​and patch shape parameters. The extraction results are converted into fixed-length vector representations as the remote sensing image feature part in the plot feature vector. The geographic information attributes are processed by structured numerical coding to extract the boundary contours, administrative division levels, topographic relief features, boundary complexity and spatial shape compactness of the land parcels, and then transformed into spatial structure feature vectors, which serve as the geographic information feature part of the land parcel feature vectors. The terrain undulation features include the maximum elevation difference, average slope, and terrain variance. The boundary complexity is the square of the perimeter divided by the area. The spatial shape compactness is four times pi multiplied by the area divided by the square of the perimeter. The relevant fields in the historical transaction records are cleaned, transformed, and standardized. The historical transaction unit price, transaction time interval, land use term, transaction method type, and total transaction area are extracted and arranged in the order of the fields into a structured transaction feature representation, which serves as the transaction history feature part in the land parcel feature vector. Semantic expression transformation processing is performed on the planning policy text, and the planning policy information after semantic extraction and word vector generation is mapped into a fixed-length vector to reflect the applicable use constraints, development conditions, functional positioning and planning guidelines of the target plot. The mapping result is embedded into the policy feature part of the plot feature vector. The basic attributes and surrounding environmental information of the land parcel are structured and processed. The land use code, plot ratio, land level, ownership category, accessibility of surrounding roads, number of educational facilities, green coverage ratio, and density of commercial facilities are normalized and one-hot encoded. The results are organized into attribute environmental feature items, which constitute the attribute environmental feature part of the land parcel feature vector. The land use code refers to the code used to identify the use of a land parcel in planning or management. The accessibility of surrounding roads refers to the distribution density of road intersections within a unit area around the target land parcel, reflecting the degree of structural embedding of the land parcel in the urban road network and the diversity of travel paths, and is used to measure the convenience of land parcel transportation access and network accessibility. Remote sensing image features, geographic information features, transaction history features, policy features, and attribute environment features are spliced ​​together in a preset order to generate land parcel feature vectors with consistent dimensions. Each land parcel feature vector is associated with its corresponding land parcel number to ensure that the land parcel feature dataset has a unified representation format and organizational structure.

[0023] In this embodiment, the generation of the multiple region-land type joint sub-sample sets specifically includes: The administrative division code of each training sample is parsed hierarchically. The four levels of administrative division information of province, city, district / county and street are transformed into discrete feature vectors by using one-hot coding method. The feature vectors of different levels are combined to generate the corresponding regional identifier vector, which is used to reflect the regional feature differences in the sample set. For each training sample, the land use code is converted to reflect the land use. The seven land use categories—arable land, residential land, commercial land, industrial land, transportation land, water area land, and public management service land—are mapped to land use adaptation factors using a one-hot coding method. These factors are used to reflect the differences in land use in the sample set. The training sample set is divided according to the generated regional identifier vector and land type adaptation factor. The samples are grouped into multiple regional-land type joint sub-sample sets. The number of regional-land type joint sub-sample sets is determined by the actual regional and land type combination relationships that appear in the training sample set.

[0024] In this embodiment, the construction and training of the region-sensitive multi-level residual enhancement estimation model specifically includes: Based on the regional-land category joint subsample set, a regionally sensitive multi-level residual enhancement valuation model is constructed. The regionally sensitive multi-level residual enhancement valuation model consists of a basic valuation branch, a multi-level residual correction branch, and a market calibration branch. The basic valuation branch uses a ResMLP network, which takes the land parcel feature vector as input and generates a preliminary valuation result for the target land parcel. The ResMLP network is a multilayer perceptron with residual connections, including multiple fully connected layers and nonlinear activation layers, with residual connections added between each layer; The multi-level residual correction branch includes a residual calculation unit, a step-by-step residual fitting unit, and a residual superposition unit. The residual calculation unit calculates the residual based on the difference between the preliminary valuation result and the historical transaction price. The step-by-step residual fitting unit fits the residuals of each layer in sequence. The residual superposition unit accumulates the residual results of each layer and outputs the residual correction value. During the training process, the land use adaptation factor is used to adjust the residual correction weight of each layer. The market calibration branch includes a direct comparison path and an indirect comparison path. The direct comparison path constructs a comparable sample set based on the similarity of land parcel feature vectors and generates a direct correction value by combining the transaction time. The indirect comparison path constructs an indirect reference set based on the mapping of administrative division levels, similarity of land use types, and similarity of the surrounding environment and generates an indirect correction value. During model training, the region identification vector acts on the parameter update process of the basic estimation branch and the multi-level residual correction branch to control the differentiated modeling of different regions. The land use adaptation factor acts on the multi-level residual correction branch to adjust the residual correction weight under different land use conditions. During model training, the region identifier vector is not directly input into the forward propagation calculation of the enhanced estimation model, but only participates in parameter updates during the back propagation stage, specifically including: In the parameter update phase, the regional identifier vector is linearly mapped to the regional adjustment coefficient. After calculating the error between the residual correction valuation result and the actual transaction price and obtaining the gradient of the loss function, the parameters of the basic valuation branch and the multi-level residual correction branch are updated. The update amount is equal to the product of the learning rate and the gradient of the loss function, and then multiplied by the regional adjustment coefficient. In this way, the update magnitude of samples from different regions is different during backpropagation, so that the model exhibits differentiated learning effects for different regions during training. Finally, after iterative iteration on multiple regional-land category joint subsample sets, the model can converge to a parameter distribution that is sensitive to different regional conditions, realizing the effective training of the regionally sensitive multi-level residual enhancement valuation model. During model training, the specific roles of the land use adaptation factor include: During the backpropagation phase, the land use adaptation factor does not participate in the forward calculation. Instead, it generates adjustment coefficients through mapping and weights the gradient of the loss function of the multi-level residual correction branch. This is used to adjust the parameter update magnitude under different land use code conditions, thereby forming differentiated residual correction weights. Finally, through joint training under all land use code conditions, the multi-level residual correction branch can dynamically update the residual correction weights according to the land use adaptation factor under different use types, thereby improving the model's adaptability and accuracy in various land use valuation scenarios.

[0025] Using historical transaction prices as a monitoring signal, the model is iteratively trained to minimize the error between the predicted results and the actual transaction prices, and the model parameters are updated until convergence, resulting in a trained region-sensitive multi-level residual enhancement estimation model. Specifically, let the training set... Divided into multiple region-land type joint subsample sets ,in Indicates a region. Representing land categories, in each iteration, from a region-land category joint subset. Extract data and input it into the model to obtain the predicted estimate for that subset of samples, and calculate the error based on the loss function: ; in, Indicates the first The model parameters corresponding to the region-land type joint subsample set in the next iteration. This represents the model parameters corresponding to the region-land type joint subset at the t-th iteration. Indicates the learning rate. Indicates the region-land category joint subsample set The gradient of the loss function on the mean squared error loss; The parameter updates are performed on a specific region-land category joint subsample set, thereby enabling the region-sensitive multilevel residual enhancement estimation model to achieve region sensitivity and application adaptability during the training phase.

[0026] In this embodiment, the generation of the residual correction estimation result specifically includes: The feature vector of the land parcel to be valued is input into the basic valuation branch, and the residual multilayer perceptron structure is used for calculation. During the layer-by-layer calculation, the input feature is multiplied by the corresponding weight matrix and a bias vector is added. After processing by a nonlinear activation function, it is added to the input to form the output feature vector of the next layer. When the basic estimation branch reaches the last layer, the weight matrix and bias vector of the last layer are used to perform a linear mapping on the output feature vector of the previous layer to obtain the preliminary estimation result. Input the preliminary valuation results into the multi-level residual correction branch to calculate the initial residuals; Based on the preliminary valuation results and the initial residuals, the residuals are fitted step by step. In the residual correction branch of the k-th layer, the land feature vector and the residual value of the previous layer are used as inputs and the k-th layer residual correction value is generated by mapping through the multilayer perceptron. All residual correction values ​​are summed sequentially and then added to the preliminary valuation result to generate the residual correction valuation result.

[0027] In this embodiment, the generation of the direct correction value specifically includes: Use the residual corrected estimation results as input; For the feature vector of the land parcel to be estimated, cosine similarity is calculated between it and the feature vector of each land parcel in the training sample set to generate a similarity score. All land parcels in the training sample set are sorted according to their similarity scores. The top N land parcels in terms of similarity are selected to form a comparable sample set. Each land parcel in the comparable sample set includes a land parcel feature vector and a historical transaction price. For each plot in the comparable sample set, a transaction time difference factor is calculated. The transaction time difference factor is the reference time of the plot to be estimated minus the transaction time of the plot. The difference is then normalized. The weighting coefficient is calculated, which consists of two parts. The first part is the proportion of the similarity score of the land parcel to the sum of all similarity scores in the comparable sample set. The second part is the result of applying an exponential decay function to the transaction time difference factor with the time decay parameter as the base. The weighting coefficient is equal to the product of the two parts. Weighted fitting is performed on the historical transaction prices in the comparable sample set based on weighted coefficients. The historical transaction price of each comparable land parcel is multiplied by the corresponding weighted coefficient, and then all multiplications are summed to obtain the weighted correction value. The direct correction value is obtained by adding the residual correction estimate to the weighted correction value.

[0028] In this embodiment, the generation of the indirect correction value specifically includes: If the number of plots in the comparable sample set is greater than or equal to the preset threshold, the direct correction value obtained from the direct comparison path will be used as the final valuation result. If the number of plots in the comparable sample set is less than a preset threshold, the indirect comparison path correction is initiated, and the residual correction valuation result is used as input. Based on the administrative division code of the land parcel to be valued, an administrative division level mapping is performed to obtain an administrative division reference set, which consists of reference land parcels at the same administrative division level as the land parcel to be valued. Based on the land use type code of the land parcel to be valued, the land use type similarity is calculated, and a land use type reference set is generated. The land use type reference set consists of reference land parcels with the same land use type as the land parcel to be valued. Based on the surrounding environment information of the land parcel to be valued, the accessibility of surrounding roads is extracted to form a surrounding environment reference set, which consists of reference land parcels with similar environmental conditions to the land parcel to be valued. The environmental conditions being similar to those of the land parcel to be valued means that when the difference between the accessibility of the surrounding roads of the reference land parcel and that of the land parcel to be valued is less than 10%, the environmental conditions are considered similar. The administrative division reference set, the land use type reference set, and the surrounding environment reference set are combined to form an indirect reference set, which consists of reference plots that simultaneously meet the administrative division conditions, land use conditions, and surrounding environment conditions. Within the indirect reference set, the regional mean deviation is calculated, which is equal to the average of the historical transaction prices of all reference land parcels within the indirect reference set minus the residual correction valuation result. Based on the regional mean deviation, an indirect correction is performed on the residual correction estimation result. The indirect correction value is obtained by adding the regional mean deviation to the residual correction estimation result.

[0029] In this embodiment, generating the final valuation result specifically includes: After obtaining the direct and indirect correction values, a dynamic weighting strategy is constructed. In the dynamic weighting strategy, the number of plots in the comparable sample set is determined; A feature similarity score is determined, which is the average cosine similarity between the feature vector of the land parcel to be estimated and the feature vector of the comparable sample set. Determine the confidence level of the regional valuation deviation, which is the number of plots in the indirect reference set divided by a preset upper limit value, which is set to 50 or 100; The direct correction weight is calculated by multiplying the number of plots in the comparable sample set by the feature similarity score, and the regional estimation deviation confidence level is used as the indirect correction weight, ensuring that the sum of the direct correction weight and the indirect correction weight is 1. The direct and indirect correction values ​​are weighted using direct and indirect correction weights respectively, and the two weighted results are added together to generate the final valuation result.

[0030] Example 1: To verify the feasibility and effectiveness of this invention in practical applications, it was applied to the core urban area and suburban areas of a provincial capital city. This region experienced frequent land transactions between 2020 and 2023, encompassing various types of land including residential, commercial, industrial, and public management service land. Common problems in land valuation included inconsistent valuations, fragmented data sources, and a lack of persuasiveness in the process. Particularly in financial mortgages and land transfers, valuation results often exhibited excessively high or low deviations, leading to uncertainty. Therefore, this invention was introduced to explore whether an automatic land valuation method based on multi-source data fusion and machine learning could improve the accuracy and credibility of the assessment in this scenario.

[0031] In this embodiment, the spatial pattern characteristics of the target plot are first acquired through remote sensing imagery, and information such as boundary contours, topographic relief, and road intersection density are collected in conjunction with geographic information system data. Simultaneously, historical transaction records from the past three years are collected, including fields such as transaction price, transaction method, usage period, and transaction time. Furthermore, semantic information such as planned use, development conditions, and functional positioning is extracted using planning text data. Through the fusion of these data from different sources, a plot feature vector with a unified structure is constructed. Each plot feature vector corresponds to a historical transaction price tag, and administrative division codes and plot use codes are appended to the dataset to ensure that regional sensitivity and use sensitivity can be formed during subsequent training.

[0032] During the training phase, regional identifier vectors and land use adaptation factors are first generated based on administrative division codes and land use codes. The overall sample set is then divided into multiple region-land use joint sub-sample sets; for example, residential land in the central urban area and industrial land in the suburbs are assigned to different sub-sample sets. Subsequently, a region-sensitive multi-level residual enhancement estimation model is constructed and trained based on these sub-sample sets. The basic estimation branch adopts a multi-layer perceptron structure with residual connections to ensure that the initial estimation results have strong nonlinear fitting capabilities. The multi-level residual correction branch fits the residual error step by step based on the initial estimation results and adjusts the weights of different land use parcels in the correction process using land use adaptation factors. The market calibration branch performs secondary correction on the results after residual correction through a combination of direct and indirect comparison paths. During training, the regional identifier vector does not directly participate in forward propagation but acts as an adjustment coefficient in the gradient update during backpropagation, resulting in different update magnitudes of model parameters under different regional conditions, ensuring that the model can learn regional differentiation patterns.

[0033] In the inference phase, taking a newly added land parcel to be valued in the third quarter of 2023 as an example, its remote sensing imagery, geographic information attributes, planning text, and surrounding environmental information are first input into the model to generate a feature vector for the land parcel. After the feature vector enters the basic valuation branch, a preliminary valuation result is obtained. Subsequently, the residuals are corrected level by level in the multi-level residual correction branch to generate a residual-corrected valuation result. If sufficient comparable samples are found in the training sample set, the residual correction result is corrected through a direct comparison path, calculating the similarity and combining it with the transaction time for weighted fitting to obtain a direct correction value. If there are insufficient comparable samples, an indirect comparison path is used, utilizing administrative division reference sets, usage type reference sets, and surrounding environment reference sets to form an indirect reference set, which is then corrected by combining regional mean deviation to obtain an indirect correction value. Finally, the model uses a dynamic weighting strategy to fuse the direct and indirect correction values ​​to generate the final valuation result.

[0034] To verify the performance of the present invention in practice, it was compared with traditional methods, and the results are shown in Table 1.

[0035] Table 1. Comparison of the Invention Method and Traditional Methods (Experimental Table) ; As shown in Table 1, the method of this invention outperforms traditional valuation methods in many core indicators. Traditional single-data regression models rely on historical transaction records, making it difficult to capture land spatial characteristics and policy constraints. Therefore, their mean absolute error and root mean square error are relatively high, their valuation stability index is less than 0.65, and their cross-regional generalization ability is extremely weak. In practical applications, this method is often only barely applicable in regions with sufficient sample size and simple transaction structures.

[0036] While traditional comparative methods combine some GIS data and manual selection of comparable samples, their accuracy improvement is limited due to the uncertainty of human judgment and the lack of comparable samples. In particular, there is still a significant bias when estimating across regions. Although the improved machine learning model incorporates remote sensing imagery and geographic information features, which can capture spatial patterns better, it does not incorporate planning policy texts and surrounding environmental information. This results in insufficient sensitivity of the valuation results to policy constraints and functional positioning, and therefore its performance remains limited in complex scenarios.

[0037] This invention proposes an automatic land valuation method based on multi-source data fusion and machine learning. By constructing a unified land parcel feature vector, it comprehensively integrates remote sensing imagery, geographic information, historical transactions, planning policies, and surrounding environmental features. Simultaneously, it introduces regional identifier vectors and land use adaptation factors, enabling the valuation model to possess regional and land use sensitivity. Experimental results show that the method's mean absolute error is controlled within 7%, and the root mean square error is significantly reduced, indicating a breakthrough in valuation accuracy. Furthermore, the valuation stability index reaches 0.86, and the cross-regional generalization ability improves to 0.84, demonstrating the model's robustness across different regions and land use types.

[0038] The performance improvement is due to the multi-level residual enhancement mechanism, which can fit the valuation error step by step and add corrections, so that the model can maintain accuracy under complex market fluctuations. The direct comparison path and indirect comparison path of the market calibration branch form a complementary mechanism, which allows the model to closely follow market transactions when the sample is sufficient, and can still make reasonable corrections by using regional mean deviation when the sample is insufficient. The dynamic weighting strategy further optimizes the balance between direct and indirect corrections, thus ensuring a double improvement in the accuracy and stability of the final valuation result.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic land valuation based on multi-source data fusion and machine learning, characterized in that, Includes the following steps: Collect multi-source heterogeneous data from various land parcels, preprocess the data, and then fuse the preprocessed multi-source heterogeneous data to construct land parcel feature vectors. A training sample set is constructed using the land parcel feature vector as input and the corresponding historical transaction price as label; Based on the administrative division code and land use code of each plot, generate regional identifier vectors and land type adaptation factors, and divide the training sample set to generate multiple regional-land type joint sub-sample sets; Based on the regional-land category joint subsample set, a regionally sensitive multi-level residual enhancement valuation model is constructed and trained. The regionally sensitive multi-level residual enhancement valuation model consists of a basic valuation branch, a multi-level residual correction branch, and a market calibration branch. Input the feature vector of the land parcel to be valued into the basic valuation branch to obtain the preliminary valuation result, and input the preliminary valuation result into the multi-level residual correction branch to generate the residual correction valuation result; Based on the residual correction estimation results, a direct comparison path correction is performed to generate a direct correction value; If the number of plots in the comparable sample set is not less than the preset threshold, the direct correction value will be used as the final valuation result. If the number of plots in the comparable sample set is less than the preset threshold, the indirect comparison path correction will be performed based on the residual correction valuation result to generate the indirect correction value. The direct and indirect correction values ​​are combined using a dynamic weighting strategy to generate the final valuation result.

2. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The multi-source heterogeneous data includes remote sensing image data, geographic information attributes, historical transaction records, planning policy texts, basic land parcel attributes, and surrounding environmental information.

3. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The generation of the land parcel feature vector specifically includes: Collect multi-source heterogeneous data from various plots, link and integrate the multi-source heterogeneous data according to the plot number, and establish a set of original plot data records with consistent structure. Feature extraction processing is performed on remote sensing image data to extract image grayscale values, texture directionality, regional boundary gradient values ​​and patch shape parameters. The extraction results are converted into fixed-length vector representations as the remote sensing image feature part in the plot feature vector. The geographic information attributes are processed by structured numerical coding to extract the boundary contours, administrative division levels, topographic relief features, boundary complexity and spatial shape compactness of the land parcels, and then transformed into spatial structure feature vectors, which serve as the geographic information feature part of the land parcel feature vectors. The relevant fields in the historical transaction records are cleaned, transformed, and standardized. The historical transaction unit price, transaction time interval, land use term, transaction method type, and total transaction area are extracted and arranged in the order of the fields into a structured transaction feature representation, which serves as the transaction history feature part in the land parcel feature vector. Semantic expression transformation processing is performed on the planning policy text, and the planning policy information after semantic extraction and word vector generation is mapped into a fixed-length vector. The mapping result is then embedded into the policy feature part of the land parcel feature vector. The basic attributes and surrounding environmental information of the land parcel are structured and processed. The land use code, plot ratio, land level, ownership category, accessibility of surrounding roads, number of educational facilities, green coverage ratio, and density of commercial facilities are normalized and one-hot encoded. The results are organized into attribute environmental feature items, which constitute the attribute environmental feature part in the land parcel feature vector. The remote sensing image features, geographic information features, transaction history features, policy features, and attribute environment features are spliced ​​together in a preset order to generate a land parcel feature vector with consistent dimensions.

4. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The generation of the multiple region-land type joint sub-sample sets specifically includes: The administrative division code of each training sample is parsed hierarchically. The four levels of administrative division information of province, city, district / county and street are transformed into discrete feature vectors by using one-hot coding method. The feature vectors of different levels are combined to generate the corresponding regional identifier vector, which is used to reflect the regional feature differences in the sample set. For each training sample, the land use code is converted to reflect the land use. The seven land use categories—arable land, residential land, commercial land, industrial land, transportation land, water area land, and public management service land—are mapped to land use adaptation factors using a one-hot coding method. These factors are used to reflect the differences in land use in the sample set. The training sample set is divided according to the generated regional identifier vector and land type adaptation factor, and the samples are grouped into multiple regional-land type joint sub-sample sets.

5. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The construction and training of the region-sensitive multi-level residual enhancement estimation model specifically includes: Based on the regional-land category joint subsample set, a regionally sensitive multi-level residual enhancement valuation model is constructed. The regionally sensitive multi-level residual enhancement valuation model consists of a basic valuation branch, a multi-level residual correction branch, and a market calibration branch. The basic valuation branch uses a ResMLP network, which takes the land parcel feature vector as input and generates a preliminary valuation result for the target land parcel. The ResMLP network is a multilayer perceptron with residual connections, including multiple fully connected layers and nonlinear activation layers, with residual connections added between each layer; The multi-level residual correction branch includes a residual calculation unit, a step-by-step residual fitting unit, and a residual superposition unit. The residual calculation unit calculates the residual based on the difference between the preliminary valuation result and the historical transaction price. The step-by-step residual fitting unit fits the residuals of each layer in sequence. The residual superposition unit accumulates the residual results of each layer and outputs the residual correction value. During the training process, the land use adaptation factor is used to adjust the residual correction weight of each layer. The market calibration branch includes a direct comparison path and an indirect comparison path. The direct comparison path constructs a comparable sample set based on the similarity of land parcel feature vectors and generates a direct correction value by combining the transaction time. The indirect comparison path constructs an indirect reference set based on the mapping of administrative division levels, similarity of land use types, and similarity of the surrounding environment and generates an indirect correction value. Using historical transaction prices as a supervisory signal, the model is iteratively trained to minimize the error between the predicted results and the actual transaction prices. The model parameters are then updated until convergence, resulting in a trained region-sensitive multi-level residual enhancement valuation model.

6. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The generation of the residual correction estimation result specifically includes: The feature vector of the land parcel to be valued is input into the basic valuation branch, and the residual multilayer perceptron structure is used for calculation. During the layer-by-layer calculation, the input feature is multiplied by the corresponding weight matrix and a bias vector is added. After processing by a nonlinear activation function, it is added to the input to form the output feature vector of the next layer. When the basic estimation branch reaches the last layer, the weight matrix and bias vector of the last layer are used to perform a linear mapping on the output feature vector of the previous layer to obtain the preliminary estimation result. Input the preliminary valuation results into the multi-level residual correction branch to calculate the initial residuals; Based on the preliminary valuation results and the initial residuals, the residuals are fitted step by step. In the residual correction branch of the k-th layer, the land feature vector and the residual value of the previous layer are used as inputs and the k-th layer residual correction value is generated by mapping through the multilayer perceptron. All residual correction values ​​are summed sequentially and then added to the preliminary valuation result to generate the residual correction valuation result.

7. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The generation of the direct correction value specifically includes: Use the residual corrected estimation results as input; For the feature vector of the land parcel to be estimated, cosine similarity is calculated between it and the feature vector of each land parcel in the training sample set to generate a similarity score. All plots in the training sample set are sorted according to their similarity scores, and the top N plots with the highest similarity scores are selected to form a comparable sample set. For each plot of land in the comparable sample set, calculate the transaction time difference factor; Calculate the weighting coefficients; Weighted fitting is performed on the historical transaction prices in the comparable sample set based on weighted coefficients. The historical transaction price of each comparable land parcel is multiplied by the corresponding weighted coefficient, and then all multiplications are summed to obtain the weighted correction value. The direct correction value is obtained by adding the residual correction estimate to the weighted correction value.

8. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The generation of the indirect correction value specifically includes: If the number of plots in the comparable sample set is greater than or equal to the preset threshold, the direct correction value obtained from the direct comparison path will be used as the final valuation result. If the number of plots in the comparable sample set is less than a preset threshold, the indirect comparison path correction is initiated, and the residual correction valuation result is used as input. Based on the administrative division code of the land parcel to be valued, an administrative division hierarchy mapping is performed to obtain an administrative division reference set; Based on the land use type code of the land parcel to be valued, the land use type similarity is calculated, and a land use type reference set is generated. The land use type reference set consists of reference land parcels with the same land use type as the land parcel to be valued. Based on the surrounding environment information of the land parcel to be valued, the accessibility of surrounding roads is extracted to form a surrounding environment reference set, which consists of reference land parcels with similar environmental conditions to the land parcel to be valued. The environmental conditions being similar to those of the land parcel to be valued means that when the difference between the accessibility of the surrounding roads of the reference land parcel and that of the land parcel to be valued is less than 10%, the environmental conditions are considered similar. The administrative division reference set, the land use type reference set, and the surrounding environment reference set are combined to form an indirect reference set, which consists of reference plots that simultaneously meet the administrative division conditions, land use conditions, and surrounding environment conditions. Within the indirect reference set, calculate the mean deviation of the region; Based on the regional mean deviation, an indirect correction is performed on the residual correction estimation result. The indirect correction value is obtained by adding the regional mean deviation to the residual correction estimation result.

9. The automatic land valuation method based on multi-source data fusion and machine learning according to claim 1, characterized in that, The generation of the final valuation result specifically includes: After obtaining the direct and indirect correction values, a dynamic weighting strategy is constructed. In the dynamic weighting strategy, the number of plots in the comparable sample set is determined; Determine the feature similarity score; Determine the confidence level of the regional valuation bias; The direct correction weight is calculated by multiplying the number of plots in the comparable sample set by the feature similarity score, and the regional estimation deviation confidence level is used as the indirect correction weight. The direct and indirect correction values ​​are weighted using direct and indirect correction weights respectively, and the two weighted results are added together to generate the final valuation result.

Images