Regional land parcel value intelligent evaluation system based on deep learning
By introducing a variational hypergraph autoencoder mechanism and a Mamba multimodal learning model with a hypergraph structure, the problem of insufficient multimodal data fusion was solved, resulting in more accurate land market value prediction and improved prediction accuracy and adaptability.
Patent Information
- Application Number
- CN202511742391.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-02-24
AI Technical Summary
Existing technologies struggle to effectively integrate multimodal data, especially in regional land valuation, as they cannot accurately model the complex nonlinear relationships between multimodal data, resulting in insufficient prediction accuracy and reliability.
We employ a deep learning-based Mamba multimodal learning model, combined with a variational hypergraph autoencoder mechanism, to model multimodal data. We optimize data fusion through hypergraph structure, extract features using convolutional neural networks, long short-term memory networks, and graph convolutional networks, and optimize latent variables through variational inference to improve the accuracy of data fusion.
It significantly improves the accuracy and robustness of land market value prediction, enhances the system's adaptability under different environmental and data change conditions, and breaks through the bottleneck of traditional methods in multimodal data processing.
Smart Images
Figure CN121563607A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent assessment technology based on deep learning, and in particular to a regional land value intelligent assessment system based on deep learning. Background Technology
[0002] In the process of regional land valuation, traditional methods mainly rely on expert experience, statistical models, and some simple geographic information system (GIS) analysis tools. While expert systems, through manually set rules and experience-based data analysis, can provide some assessment results, their large and complex datasets, and their inability to accurately account for the complex nonlinear relationships between multiple factors, result in low prediction accuracy and reliability. Traditional statistical models, such as regression analysis, are typically only applicable to data with linear relationships and cannot effectively capture and model the complex relationships and higher-order interactions between multiple different data modalities.
[0003] With the rapid development and application of remote sensing technology, market transaction data, environmental monitoring data, and other information, the amount of multi-source data required for regional land valuation has increased significantly, posing greater challenges to traditional single models and rule-based expert systems. Existing methods struggle to fully integrate data from different modalities and maintain efficient and accurate assessment results under conditions of massive and dynamically changing data volumes. The bottleneck of existing technologies in multimodal data fusion lies in the lack of effective models to capture the complex dependencies and nonlinear interactions between different modalities.
[0004] In recent years, deep learning technology has made significant progress in applications across multiple fields. In particular, deep learning models such as convolutional neural networks and long short-term memory networks have been widely applied to tasks such as image recognition and time series prediction. While these technologies can effectively process certain types of data, existing deep learning models have not yet fully addressed the challenges of modeling higher-order relationships between data and fusing cross-modal information when dealing with multimodal data.
[0005] Deep learning methods based on graph neural networks and hypergraph structures have begun to attract attention. These methods can effectively model complex relationships between nodes through operations such as graph convolution, and are particularly suitable for tasks involving data with multiple different modalities. However, there is still room for optimization in these methods, especially in how to combine variational inference mechanisms to optimize the hypergraph structure and further improve the model's ability to model complex dependencies between multimodal data and its prediction accuracy. Effective solutions are still lacking in this area.
[0006] Therefore, how to provide a method that can efficiently integrate multimodal data and accurately predict the market value of land parcels is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0007] One objective of this invention is to propose a regional land value intelligent assessment system based on deep learning. Addressing the problems of insufficient multimodal data fusion and difficulty in modeling complex relationships in existing technologies, this invention proposes a deep Mamba multimodal learning model incorporating a variational hypergraph autoencoder mechanism. This invention uses a hypergraph structure to model the complex relationships between multimodal data and optimizes latent variables through variational inference methods, thereby improving the accuracy of data fusion. This technical solution provides more accurate land market value prediction through the efficient fusion of remote sensing images, market transaction data, environmental monitoring data, and geospatial data, exhibiting high prediction accuracy and good adaptability, thus solving the bottleneck problems of existing technologies in multimodal data processing.
[0008] According to an embodiment of the present invention, a regional land parcel value intelligent assessment system based on deep learning includes the following modules: Data acquisition module: Collects multi-source data from multiple sources, including remote sensing image data of the land parcel, market transaction data, environmental monitoring data, and geospatial data, and integrates the collected data into a multi-source dataset; Deep Mamba multimodal learning model construction module: Based on the multi-source dataset, a deep Mamba multimodal learning model is constructed. The deep Mamba multimodal learning model includes a data preprocessing module, a feature extraction module, a data fusion module, and an output prediction module, and the deep Mamba multimodal learning model is obtained. Deep Mamba multimodal learning model optimization module: A variational hypergraph autoencoder mechanism is introduced to optimize the data fusion module in the deep Mamba multimodal learning model. The optimization is based on the hypergraph structure to model the complex relationship between multimodal data, resulting in an optimized deep Mamba multimodal learning model. Land value prediction module: Input the multi-source dataset into the optimized deep Mamba multimodal learning model to generate the market value assessment result of the land parcel, and the market value assessment result is the predicted market value of the land parcel; Feedback optimization module: Based on the comparison between the market value assessment results of the land parcel and the actual market value of the land parcel, the deep Mamba multimodal learning model is adjusted using a reverse optimization algorithm, and the model parameters of the deep Mamba multimodal learning model are dynamically updated.
[0009] Optionally, the data acquisition module is: Remote sensing image data of regional land parcels is collected. The remote sensing image data includes land parcel surface images from multiple time periods. Interference factors are removed through image processing algorithms to obtain clear land parcel surface images. Collect market transaction data for the target land parcel, including transaction price, transaction time, transaction type, and relevant market trend data; Collect environmental monitoring data of the target site, including parameters such as temperature, humidity, wind speed, and precipitation. Collect geospatial data of the target plot, including the plot's location coordinates, boundary information, and spatial relationships with adjacent plots; Remote sensing image data, market transaction data, environmental monitoring data, and geospatial data are integrated to form a multi-source dataset.
[0010] Optionally, the deep Mamba multimodal learning model building module is: A data preprocessing module is constructed to process the data of each modality according to the type and structure of the multi-source dataset, including denoising of image data, standardization of market transaction data, time synchronization of environmental monitoring data, and coordinate transformation of geospatial data, and outputting the preprocessed data. A feature extraction module is constructed to extract features from preprocessed data. For image data, a convolutional neural network is used to extract spatial features. For market transaction data and environmental monitoring data, a long short-term memory network is used to extract temporal features. For geospatial data, a graph convolutional network is used to extract spatial relationship features. A data fusion module is constructed, which combines the features of each modality using data fusion methods. A weighted concatenation strategy is adopted to merge the feature vectors from different modalities according to their weights to form a fused feature vector. An output prediction module is constructed to map the fused feature vectors into predicted land market values, resulting in a deep Mamba multimodal learning model.
[0011] Optionally, the deep Mamba multimodal learning model optimization module is: A variational hypergraph autoencoder mechanism is introduced to optimize the data fusion module in the deep Mamba multimodal learning model. This optimizes the modeling of multimodal data through a hypergraph structure. Each node in the hypergraph structure represents a type of data, and hyperedges represent higher-order relationships between different types of data. Hyperedges connect multiple nodes to form cross-modal data associations. In a hypergraph structure, the weights of edges between nodes are defined. The weights are determined by calculating the similarity between data modalities. Cosine similarity is used as the similarity metric to measure the strength of the relationship between different data modalities. Hypergraph convolution operations are used to update node features. The hypergraph convolution operation includes weighted aggregation of features of neighboring nodes, using weighting coefficients to weight the features of neighboring nodes, forming a new feature representation for each node. In the process of hypergraph convolution, each update operation is based on the features of the current node and its neighboring nodes. After the update, the fused node features are obtained, and the node features reflect the relationship between multimodal data. By using variational inference, latent variables are introduced into the hypergraph structure. Each latent variable represents the potential relationship between the modes of the nodes. The variational inference algorithm optimizes the distribution of latent variables by maximizing the variational lower bound. By iteratively updating the parameters of the latent variables, the feature representation of each node is adjusted. During the variational inference process, the update of latent variables is based on the node features in the hypergraph structure. The data fusion module outputs a fused feature vector. After optimization, an optimized deep Mamba multimodal learning model is obtained, which includes a data preprocessing module, a feature extraction module, a data fusion module, and an output prediction module.
[0012] Optionally, the data preprocessing module specifically includes: A denoising autoencoder in a convolutional neural network is used to denoise remote sensing image data. The denoising autoencoder learns the latent features of the image to remove noise introduced by environmental factors, thus obtaining preprocessed remote sensing image data. The market transaction data of the target land parcel is standardized using the min-max normalization method, which maps the range of transaction price, transaction time and transaction type to the interval [0, 1] to obtain standardized market transaction data. Missing values are filled in using linear interpolation, data from different time periods are aligned based on timestamps, and environmental monitoring data is time-synchronized to obtain synchronized environmental monitoring data. By employing coordinate transformation methods, geographic data from different sources are uniformly converted into a standard geographic coordinate system. Geospatial data is transformed into the geospatial data by performing coordinate system transformation. The preprocessed remote sensing image data, standardized market transaction data, synchronously processed environmental monitoring data, and transformed geospatial data are integrated to form a unified multimodal dataset.
[0013] Optionally, the feature extraction module specifically includes: The original remote sensing image is input into the input layer of the convolutional neural network. The image is filtered by the convolution kernel to extract low-level features and obtain a feature map. The feature map is sampled by the pooling layer to extract high-level features. The extracted high-level features are flattened by the fully connected layer to generate a feature vector containing spatial information of land parcels. Market transaction data and environmental monitoring data are input into the input layer of the LSTM network in chronological order. The LSTM network extracts the patterns and trends of data changes over time through a gating mechanism and outputs a vector representing the time series characteristics, which includes the time dependence and dynamic change characteristics of the data. Geospatial data is constructed as a graph structure, with each land parcel as a node in the graph and the spatial relationships between adjacent land parcels as edges between nodes. Through graph convolution operations, the neighboring nodes of a node are weighted and aggregated to obtain the feature representation of each node.
[0014] Optionally, the output prediction module specifically includes: The fused feature vector is input into a fully connected neural network. In the fully connected layer, an activation function is used to perform a non-linear transformation on the output of each layer, and the transformed feature vector is output. The transformed feature vector is then processed by a regression algorithm. The regression algorithm adopts a linear regression model. Based on the feature vector output by the network, the network output is mapped to the predicted value of the land parcel market value through the regression weight parameters obtained during training.
[0015] Optionally, the feedback optimization module is: The market value assessment results of the land parcel are compared with the actual market value of the land parcel, and the error between the assessment results and the actual market value is calculated. Based on the error, a backpropagation optimization algorithm is adopted. The gradient of the error with respect to the model parameters is calculated through the backpropagation algorithm. During the backpropagation process, the partial derivatives of each model parameter are updated through the chain rule. The calculated gradients are passed to the various parameters in the model, and the model parameters are adjusted using optimization algorithms. Dynamically update the model parameters of the deep Mamba multimodal learning model.
[0016] The beneficial effects of this invention are This invention achieves effective multimodal data fusion by introducing a variational hypergraph autoencoder mechanism and combining it with a hypergraph structure to model multimodal data. This technical solution overcomes the bottleneck of traditional methods in fully modeling complex relationships between data sources, significantly improving the accuracy of land market value prediction. Through an optimized data fusion module and a deep Mamba multimodal learning model, this invention effectively solves the problems of incomplete multimodal data fusion and difficulty in modeling nonlinear relationships in existing technologies, improving the accuracy and robustness of land market value prediction and enhancing the system's adaptability under different environmental and data change conditions. 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: Figure 1 This is a system framework diagram of a regional land value intelligent assessment system based on deep learning proposed in this invention.
[0018] Figure 2 This is a schematic diagram illustrating the optimization process of the deep Mamba multimodal learning model proposed in this invention. Detailed Implementation
[0019] Combination Figures 1-2 The present invention will be described in further detail below. These accompanying drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and showing the main components relevant to the invention. Figure 1 and Figure 2 The present invention provides a regional land value intelligent assessment system based on deep learning, comprising the following modules: Data acquisition module: Collects multi-source data from multiple sources, including remote sensing image data of the land parcel, market transaction data, environmental monitoring data, and geospatial data, and integrates the collected data into a multi-source dataset; Deep Mamba multimodal learning model construction module: Based on the multi-source dataset, a deep Mamba multimodal learning model is constructed. The deep Mamba multimodal learning model includes a data preprocessing module, a feature extraction module, a data fusion module, and an output prediction module, and the deep Mamba multimodal learning model is obtained. Deep Mamba multimodal learning model optimization module: A variational hypergraph autoencoder mechanism is introduced to optimize the data fusion module in the deep Mamba multimodal learning model. The optimization is based on the hypergraph structure to model the complex relationship between multimodal data, resulting in an optimized deep Mamba multimodal learning model. Land value prediction module: Input the multi-source dataset into the optimized deep Mamba multimodal learning model to generate the market value assessment result of the land parcel, and the market value assessment result is the predicted market value of the land parcel; Feedback optimization module: Based on the comparison between the market value assessment results of the land parcel and the actual market value of the land parcel, the deep Mamba multimodal learning model is adjusted using a reverse optimization algorithm, and the model parameters of the deep Mamba multimodal learning model are dynamically updated.
[0020] In this embodiment, the data acquisition module is: Remote sensing image data of regional land parcels is collected. The remote sensing image data includes land parcel surface images from multiple time periods. Interference factors are removed through image processing algorithms to obtain clear land parcel surface images. Collect market transaction data for the target land parcel, including transaction price, transaction time, transaction type, and relevant market trend data; Collect environmental monitoring data of the target site, including parameters such as temperature, humidity, wind speed, and precipitation. Collect geospatial data of the target plot, including the plot's location coordinates, boundary information, and spatial relationships with adjacent plots; The remote sensing image data, market transaction data, environmental monitoring data, and geospatial data are integrated to form a multi-source dataset.
[0021] In this embodiment, the deep Mamba multimodal learning model construction module is: A data preprocessing module is constructed to process the data of each modality according to the type and structure of the multi-source dataset, including denoising of image data, standardization of market transaction data, time synchronization of environmental monitoring data, and coordinate transformation of geospatial data, and outputting the preprocessed data. A feature extraction module is constructed to extract features from preprocessed data. For image data, a convolutional neural network is used to extract spatial features. For market transaction data and environmental monitoring data, a long short-term memory network is used to extract temporal features. For geospatial data, a graph convolutional network is used to extract spatial relationship features. A data fusion module is constructed, which combines the features of each modality using data fusion methods. A weighted concatenation strategy is adopted to merge the feature vectors from different modalities according to their weights to form a fused feature vector. An output prediction module is constructed to map the fused feature vectors into predicted land market values, resulting in a deep Mamba multimodal learning model.
[0022] In this embodiment, the deep Mamba multimodal learning model optimization module is: A variational hypergraph autoencoder mechanism is introduced to optimize the data fusion module in the deep Mamba multimodal learning model. This optimizes the modeling of multimodal data through a hypergraph structure. Each node in the hypergraph structure represents a type of data, and hyperedges represent higher-order relationships between different types of data. Hyperedges connect multiple nodes to form cross-modal data associations. In a hypergraph structure, the weights of edges between nodes are defined. The weights are determined by calculating the similarity between data modalities. Cosine similarity is used as the similarity metric to measure the strength of the relationship between different data modalities. Hypergraph convolution operations are used to update node features. The hypergraph convolution operation includes weighted aggregation of features of neighboring nodes, using weighting coefficients to weight the features of neighboring nodes, forming a new feature representation for each node. In the process of hypergraph convolution, each update operation is based on the features of the current node and its neighboring nodes. After the update, the fused node features are obtained, and the node features reflect the relationship between multimodal data. By introducing latent variables into the hypergraph structure through variational inference, each latent variable represents the potential relationship between the modes of the nodes. The variational inference algorithm optimizes the distribution of latent variables by maximizing the variational lower bound and adjusts the feature representation of each node by iteratively updating the parameters of the latent variables. During the variational inference process, the update of latent variables is based on the node features in the hypergraph structure, which enables data fusion to more accurately reflect the interrelationship between the various data modes. The data fusion module outputs a fused feature vector. After optimization, an optimized deep Mamba multimodal learning model is obtained, which includes a data preprocessing module, a feature extraction module, a data fusion module, and an output prediction module.
[0023] In this embodiment, the data preprocessing module specifically includes: A denoising autoencoder in a convolutional neural network is used to denoise remote sensing image data. The denoising autoencoder learns the latent features of the image to remove noise introduced by environmental factors, thus obtaining preprocessed remote sensing image data. The market transaction data of the target land parcel is standardized using the min-max normalization method, which maps the range of transaction price, transaction time, and transaction type to the interval [0, 1] to obtain standardized market transaction data, thereby eliminating the impact of differences in different units and data distributions. The processing steps ensure that different types of market transaction data can be compared and analyzed on a uniform scale. Missing values are filled in using linear interpolation, and data from different time periods are aligned according to timestamps. Environmental monitoring data is then time-synchronized to obtain synchronized environmental monitoring data. This time synchronization ensures the consistency of various environmental parameter data in the time dimension, enabling subsequent analysis to accurately reflect the real-time environmental changes of the site. By employing coordinate transformation methods, geographic data from different sources are uniformly converted into a standard geographic coordinate system. The geospatial data is transformed into a coordinate system to obtain the transformed geospatial data, ensuring that the spatial relationships of different plots of data can be accurately matched and that spatial consistency is maintained. Coordinate transformation is the process of converting geographic data from different sources from one coordinate system to another standard coordinate system. Typically, the source and target coordinate systems are first determined, and then an appropriate transformation algorithm is selected, such as affine transformation or projection transformation. Affine transformation is suitable for simple transformations between planar coordinate systems, while projection transformation is suitable for converting geographic coordinates to planar coordinates. By applying transformation parameters between the coordinate systems, the coordinates in the source coordinate system are calculated into the target coordinate system, ensuring spatial consistency between different data sources.
[0024] Preprocessed remote sensing image data, standardized market transaction data, synchronously processed environmental monitoring data, and transformed geospatial data are integrated to form a unified multimodal dataset. This dataset undergoes standardization and cleaning to ensure seamless integration across all data sources.
[0025] In this embodiment, the feature extraction module specifically includes: The original remote sensing image is input into the input layer of the convolutional neural network. After passing through multiple convolutional layers, feature extraction is performed. The convolutional layers filter the image through convolutional kernels to extract low-level features such as edges, textures, and shapes to obtain feature maps. Pooling layers sample the feature maps to extract high-level features, reducing computation and preserving key spatial information. Fully connected layers flatten the extracted high-level features to generate feature vectors containing spatial information of land parcels. Market transaction data and environmental monitoring data are input into the input layer of the LSTM network in chronological order. The LSTM network effectively remembers and forgets historical information through a gating mechanism, extracting the patterns and trends of data changes over time. Specifically, the LSTM processes the time-series data through forget gates, input gates, and output gates, enabling the model to capture the price fluctuation patterns in market transactions and the time-series patterns of parameters such as temperature and humidity in environmental monitoring. Finally, it outputs a vector representing the time-series characteristics, which includes the time dependence and dynamic change characteristics of the data. Geospatial data is constructed as a graph structure, with each land parcel as a node in the graph and the spatial relationships between adjacent land parcels as edges between nodes. Through graph convolution operations, the neighboring nodes of a node are weighted and aggregated to capture the spatial relationships and interdependencies between land parcels. Specifically, GCN iteratively propagates the feature information of nodes, aggregating the features of adjacent nodes layer by layer, and finally obtains the feature representation of each node, reflecting the geographical location, adjacency relationship and spatial correlation of the land parcel.
[0026] In this embodiment, the output prediction module specifically includes: The fused feature vector is input into a fully connected neural network. In the fully connected layer, an activation function is used to perform a non-linear transformation on the output of each layer, and the transformed feature vector is output. The transformed feature vector is then processed by a regression algorithm. The regression algorithm adopts a linear regression model. Based on the feature vector output by the network, the network output is mapped to the predicted value of the land parcel market value through the regression weight parameters obtained during training.
[0027] The weights are updated by subtracting a value proportional to the gradient from the current weights. If the gradient is positive, it means the error has increased and the weights need to be decreased; if the gradient is negative, it means the error has decreased and the weights need to be increased.
[0028] In this embodiment, the feedback optimization module specifically includes: The market value assessment results of the land parcel are compared with the actual market value of the land parcel, and the error between the assessment results and the actual market value is calculated. Based on the error, a backpropagation optimization algorithm is used to calculate the gradient of the error with respect to the model parameters through the backpropagation algorithm. During the backpropagation process, the partial derivatives of each model parameter are updated through the chain rule. The calculated gradient is passed to the various parameters in the model, and the model parameters are adjusted using the gradient descent optimization algorithm or its variants to reduce errors and optimize the model performance. Dynamically update the model parameters of the deep Mamba multimodal learning model, including the weights and biases in each module such as the data fusion module, feature extraction module, and relation modeling module; Through multiple rounds of feedback optimization, the model parameters are gradually adjusted, and finally an updated deep Mamba multimodal learning model is obtained. This model can provide more accurate prediction results in subsequent land value assessments.
[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a real-world scenario of regional land market value assessment. This scenario involves the collection of multi-source data, including remote sensing image data, market transaction data, environmental monitoring data, and geospatial data. The remote sensing image data in the dataset includes multi-time-segment images of each land parcel; the market transaction data includes the transaction price, transaction time, and transaction type of each land parcel; the environmental monitoring data includes temperature, humidity, wind speed, etc.; and the geospatial data includes the location coordinates of the land parcels, boundary information, and spatial relationships between adjacent land parcels.
[0030] In the data preprocessing stage, remote sensing image data is denoised using convolutional neural networks. A ResNet50 architecture is used for convolutional layer extraction, and pooling layers are used to reduce feature map size. Market transaction data undergoes min-max normalization to standardize the range of price, transaction time, and transaction type values. In environmental monitoring data processing, linear interpolation is used to fill in missing values, and monitoring data from different parcels are aligned through time synchronization. Geospatial data is transformed from geographical data from different sources into a unified standard coordinate system using coordinate transformation algorithms.
[0031] In the feature extraction stage, convolutional neural networks are used to extract spatial features from remote sensing images, LSTM is used to extract temporal features from market transaction data and environmental monitoring data, and graph convolutional networks are used to extract spatial relationship features from geospatial data.
[0032] In the data fusion stage, a weighted concatenation strategy was adopted to fuse data features from different modalities. A variational hypergraph autoencoder mechanism was used to optimize the data fusion process, employing cosine similarity as the similarity metric to calculate the similarity between data modalities and weighted aggregate node features. Hypergraph convolution operations were then performed based on the calculated similarity weights to achieve weighted aggregation, resulting in the optimized fused feature vector.
[0033] The distribution of latent variables was adjusted using a variational inference algorithm. Specifically, the variational inference employed a strategy of maximizing the variational lower bound to optimize the parameters of the latent variables. During training, 1000 iterations were performed, with batch gradient descent used for parameter updates in each iteration. The initial learning rate was set to 0.001, and the Adam optimization algorithm was employed. The error in each iteration gradually decreased on the training set, eventually reaching 0.028.
[0034] In the prediction phase, a regression algorithm is used to map the fused feature vectors to predicted land market values. In the regression model, regression weight parameters are trained, and the loss function is optimized using the least squares method. During the optimization process, the mean squared error on the training set was 0.032. After 1000 rounds of iterative optimization, the mean squared error was reduced to 0.028, demonstrating high prediction accuracy.
[0035] The feedback optimization module dynamically updates model parameters based on the error between the predicted results and the actual market value, using the backpropagation algorithm and gradient descent method. Specifically, the backpropagation optimization process uses batch gradient descent with a learning rate set to 0.0001. The optimization process undergoes 500 iterations, and the feedback error continuously decreases during the update process, eventually reducing the error to 0.027.
[0036] The following is a comparison of the prediction results in this embodiment. The table shows the error comparison between the deep Mamba multimodal learning model of this invention and traditional methods in terms of real market value and predicted value: By comparing the market value predictions of this invention with those of traditional methods in the table, it can be seen that the deep Mamba multimodal learning model of this invention exhibits higher accuracy in predicting land market value. The error for most land parcels is controlled within 100,000 yuan, especially for parcels 1, 4, and 5, where the error is 50,000 yuan each, demonstrating the effectiveness of this invention in multimodal data fusion and prediction. In contrast, the prediction errors of traditional methods are generally larger; for example, the errors for parcels 1 and 3 using traditional methods are 200,000 yuan and 400,000 yuan respectively, far exceeding the error values of this invention. The mean squared error is better in the model of this invention, indicating that the model has a significant advantage in data fitting and can provide more accurate prediction results in modeling complex relationships in multimodal data.
[0037] These data further validate the significant advantages of this invention in handling complex data relationships and improving prediction accuracy. Compared with traditional methods, this invention has higher accuracy and robustness in practical applications.
Claims
1. A regional land value intelligent assessment system based on deep learning, characterized in that, Includes the following modules: Data acquisition module: Collects multi-source data from multiple sources and integrates the collected data into a multi-source dataset; Deep Mamba multimodal learning model construction module: Based on the multi-source dataset, a deep Mamba multimodal learning model is constructed. The deep Mamba multimodal learning model includes a data preprocessing module, a feature extraction module, a data fusion module, and an output prediction module, and the deep Mamba multimodal learning model is obtained. Deep Mamba multimodal learning model optimization module: A variational hypergraph autoencoder mechanism is introduced to optimize the data fusion module in the deep Mamba multimodal learning model. The optimization is based on the hypergraph structure to model the complex relationship between multimodal data, resulting in an optimized deep Mamba multimodal learning model. Land value prediction module: Input the multi-source dataset into the optimized deep Mamba multimodal learning model to generate the market value assessment result of the land parcel, and the market value assessment result is the predicted market value of the land parcel; Feedback optimization module: Based on the comparison between the market value assessment results of the land parcel and the actual market value of the land parcel, the deep Mamba multimodal learning model is adjusted using a reverse optimization algorithm, and the model parameters of the deep Mamba multimodal learning model are dynamically updated.
2. The regional land value intelligent assessment system based on deep learning according to claim 1, characterized in that, The data acquisition module is: Remote sensing image data of regional land parcels is collected. The remote sensing image data includes land parcel surface images from multiple time periods. Interference factors are removed through image processing algorithms to obtain clear land parcel surface images. Collect market transaction data for the target land parcel, including transaction price, transaction time, transaction type, and relevant market trend data; Collect environmental monitoring data of the target site, including parameters such as temperature, humidity, wind speed, and precipitation. Collect geospatial data of the target plot, including the plot's location coordinates, boundary information, and spatial relationships with adjacent plots; Remote sensing image data, market transaction data, environmental monitoring data, and geospatial data are integrated to form a multi-source dataset.
3. The regional land value intelligent assessment system based on deep learning according to claim 1, characterized in that, The deep Mamba multimodal learning model construction module is as follows: A data preprocessing module is constructed to process the data of each modality according to the type and structure of the multi-source dataset, including denoising of image data, standardization of market transaction data, time synchronization of environmental monitoring data, and coordinate transformation of geospatial data, and outputting the preprocessed data. A feature extraction module is constructed to extract features from preprocessed data. For image data, a convolutional neural network is used to extract spatial features. For market transaction data and environmental monitoring data, a long short-term memory network is used to extract temporal features. For geospatial data, a graph convolutional network is used to extract spatial relationship features. A data fusion module is constructed, which combines the features of each modality using data fusion methods. A weighted concatenation strategy is adopted to merge the feature vectors from different modalities according to their weights to form a fused feature vector. An output prediction module is constructed to map the fused feature vectors into predicted land market values, resulting in a deep Mamba multimodal learning model.
4. The regional land value intelligent assessment system based on deep learning according to claim 1, characterized in that, The deep Mamba multimodal learning model optimization module is as follows: A variational hypergraph autoencoder mechanism is introduced to optimize the data fusion module in the deep Mamba multimodal learning model. This optimizes the modeling of multimodal data through a hypergraph structure. Each node in the hypergraph structure represents a type of data, and hyperedges represent higher-order relationships between different types of data. Hyperedges connect multiple nodes to form cross-modal data associations. In a hypergraph structure, the weights of edges between nodes are defined. The weights are determined by calculating the similarity between data modalities. Cosine similarity is used as the similarity metric to measure the strength of the relationship between different data modalities. Hypergraph convolution operations are used to update node features. The hypergraph convolution operation includes weighted aggregation of features of neighboring nodes, using weighting coefficients to weight the features of neighboring nodes, forming a new feature representation for each node. In the process of hypergraph convolution, each update operation is based on the features of the current node and its neighboring nodes. After the update, the fused node features are obtained. By using variational inference, latent variables are introduced into the hypergraph structure. Each latent variable represents the potential relationship between the modes of the nodes. The variational inference algorithm optimizes the distribution of latent variables by maximizing the variational lower bound. By iteratively updating the parameters of the latent variables, the feature representation of each node is adjusted. During the variational inference process, the update of latent variables is based on the node features in the hypergraph structure. The data fusion module outputs a fused feature vector. After optimization, an optimized deep Mamba multimodal learning model is obtained, which includes a data preprocessing module, a feature extraction module, a data fusion module, and an output prediction module.
5. A regional land value intelligent assessment system based on deep learning according to claim 4, characterized in that, The data preprocessing module specifically includes: A denoising autoencoder in a convolutional neural network is used to denoise remote sensing image data. The denoising autoencoder learns the latent features of the image to remove noise introduced by environmental factors, thus obtaining preprocessed remote sensing image data. The market transaction data of the target land parcel is standardized using the min-max normalization method, which maps the range of transaction price, transaction time and transaction type to the interval [0, 1] to obtain standardized market transaction data. Missing values are filled in using linear interpolation, data from different time periods are aligned based on timestamps, and environmental monitoring data is time-synchronized to obtain synchronized environmental monitoring data. By employing coordinate transformation methods, geographic data from different sources are uniformly converted into a standard geographic coordinate system. Geospatial data is transformed into the geospatial data by performing coordinate system transformation. The preprocessed remote sensing image data, standardized market transaction data, synchronously processed environmental monitoring data, and transformed geospatial data are integrated to form a unified multimodal dataset.
6. A regional land value intelligent assessment system based on deep learning according to claim 4, characterized in that, The feature extraction module specifically includes: The original remote sensing image is input into the input layer of the convolutional neural network. The image is filtered by the convolution kernel to extract low-level features and obtain a feature map. The feature map is sampled by the pooling layer to extract high-level features. The extracted high-level features are flattened by the fully connected layer to generate a feature vector containing spatial information of land parcels. Market transaction data and environmental monitoring data are input into the input layer of the LSTM network in chronological order. The LSTM network extracts the patterns and trends of data changes over time through a gating mechanism and outputs a vector representing the time series characteristics, which includes the time dependence and dynamic change characteristics of the data. Geospatial data is constructed as a graph structure, with each land parcel as a node in the graph and the spatial relationships between adjacent land parcels as edges between nodes. Through graph convolution operations, the neighboring nodes of a node are weighted and aggregated to obtain the feature representation of each node.
7. A regional land value intelligent assessment system based on deep learning according to claim 4, characterized in that, The output prediction module specifically includes: The fused feature vector is input into a fully connected neural network. In the fully connected layer, an activation function is used to perform a non-linear transformation on the output of each layer, and the transformed feature vector is output. The transformed feature vector is then processed by a regression algorithm. The regression algorithm adopts a linear regression model. Based on the feature vector output by the network, the network output is mapped to the predicted value of the land parcel market value through the regression weight parameters obtained during training.
8. A regional land value intelligent assessment system based on deep learning according to claim 1, characterized in that, The feedback optimization module is: The market value assessment results of the land parcel are compared with the actual market value of the land parcel, and the error between the assessment results and the actual market value is calculated. Based on the error, a backpropagation optimization algorithm is adopted. The gradient of the error with respect to the model parameters is calculated through the backpropagation algorithm. During the backpropagation process, the partial derivatives of each model parameter are updated through the chain rule. The calculated gradients are passed to the various parameters in the model, and the model parameters are adjusted using optimization algorithms. Dynamically update the model parameters of the deep Mamba multimodal learning model.