Target area low-efficiency land extraction method and system based on multi-source big data

Abstract

The invention discloses a target area low-efficiency land extraction method and system based on multi-source big data, and belongs to the field of land management.The extraction method comprises the following specific steps that P1, data are collected from multiple sets of data sources in real time, all kinds of collected multi-source data are preprocessed, and then feature information of all kinds of preprocessed multi-source data is extracted; p2, based on the extracted various feature information, establishing a land use characterization matrix, and according to the established land use characterization matrix, establishing a low-efficiency land use discrimination model; according to the method, the land utilization condition can be depicted from multiple angles, the recognition accuracy and the feature extraction accuracy are improved, meanwhile, the low-efficiency land use recognition accuracy is improved, the terminal calculation cost is reduced, real-time updating and optimization are achieved, and the method can meet the land management requirements of different regions and cities.

Description

Technical Field

[0001] The present invention relates to the field of land management, and in particular to a method and system for extracting low-efficiency land in a target area based on multi-source big data. Background Art

[0002] With the continuous advancement of urbanization, the rational allocation of land resources directly affects the urban economy, ecology, and sustainable development. However, due to the complexity of urban land use patterns, some plots of land have long been inefficiently utilized, such as idle industrial land, low-density commercial areas, and residential land with inconvenient transportation. These inefficient land uses not only lead to a waste of land resources but can also cause imbalances in urban spatial layout and uneven distribution of public service resources. Traditional methods for identifying inefficient land use rely primarily on a single data source, making it difficult to fully capture complex land use characteristics. Furthermore, the frequency of updates is low, resulting in delayed monitoring of inefficient land use and hindering refined urban management. Therefore, how to accurately identify inefficient land use and provide optimization recommendations has become an important research direction in current urban land management and planning.

[0003] Existing methods and systems for extracting inefficient land in target areas are unable to characterize land use from multiple perspectives, which reduces recognition accuracy and feature extraction precision, has high terminal computing costs, and cannot achieve real-time updating and optimization. To this end, we propose a method and system for extracting inefficient land in target areas based on multi-source big data. Summary of the Invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose a method and system for extracting low-efficiency land in a target area based on multi-source big data.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A method for extracting inefficient land in a target area based on multi-source big data. The specific steps of the extraction method are as follows: P1. Collect data from multiple data sources in real time, pre-process the collected multi-source data, and then extract feature information from the pre-processed multi-source data. P2. Based on the extracted characteristic information, a land use characterization matrix is established, and based on the constructed land use characterization matrix, an inefficient land use discrimination model is established; P3: Input the real-time collected data into each terminal device, identify and mark the low-efficiency land through the low-efficiency land discrimination model, and update the model parameters in real time based on the model marking results; P4. Analyze the causes of low-efficiency land use, predict future trends of low-efficiency land use, and generate a land use efficiency change forecast report; P5. Dynamically optimize the rules for identifying low-efficiency land, update the rules of the low-efficiency land identification model, and visualize the distribution and historical change trends of low-efficiency land.

[0006] As a further solution of the present invention, the specific steps of pre-processing the collected multi-source data in P1 are as follows: Q1.1: Collect remote sensing images, POI data, sensor data, and text data from multiple data sources and classify the collected data according to the format. Q1.2: Using an atmospheric correction model, convert the digital values of remote sensing images into surface reflectance. Using preset affine transformation parameters, coordinate deviations of different remote sensing images are corrected through affine transformation. The cosine similarities between each POI data set are calculated and recorded. The recorded cosine similarities are traversed. If any cosine similarity exceeds a preset threshold, the corresponding two sets of POI data are considered duplicated. After the traversal is complete, duplicate POI data is removed according to preset requirements, and the different types of POI data are normalized. Q1.3: We used spatiotemporal kriging to interpolate sensor data collected at inconsistent times, aligning the data collection times. We then used a one-dimensional wavelet transform to decompose the low-frequency components of each sensor data to remove noise. We then used the TF-IDF method to extract keywords from social media data and historical planning texts, converting the extracted keywords into numerical vectors. Q1.4: Use resampling methods to adjust remote sensing images of various raster formats to the same resolution. Based on the raster format of the remote sensing images, perform corresponding rasterization processing on the POI data and traffic flow data to spatially align the remote sensing images, POI data, and sensor data. Use time series interpolation to fill missing values in the remote sensing images and POI data to synchronize the sampling time of the remote sensing images and POI data. Then, use the time window aggregation method to time-match the sensor data with the remote sensing images and POI data. Then, perform geo-resolution on the social media data and historical planning texts, and convert the text addresses into the same raster format as the remote sensing images. Use time deduction methods to match the time scales of the social media data and historical planning texts.

[0007] As a further solution of the present invention, the specific steps of establishing a land use characterization matrix based on the extracted various feature information in P2 are as follows: S1.1: Collect pre-processed remote sensing images, POI data, sensor data, and text data from multiple data sources. Then, check the spatiotemporal alignment of these data to see if there are any deviations. If so, correct them. Then, separate and extract traffic flow data from the sensor data. S1.2: Use the DeepLabV3+ model to convert each remote sensing image into a pixel-level feature classification map. Then use the gray-level co-occurrence matrix to extract the texture features of each remote sensing image. Use the normalized vegetation index in each remote sensing data to obtain the corresponding spectral features. Then use KDE to calculate the spatial distribution characteristics of the POI data and calculate the POI diversity entropy to obtain the diversity characteristics of the POI data type in each region. S1.3: Extract the road network structure of traffic flow data based on the GNN network and construct an adjacency matrix to extract road connectivity features. Calculate the average traffic flow within a preset unit time and obtain traffic flow aggregation features within different time intervals. Statistical analysis is then performed on the sensor data to extract mean and variance features. Fourier transform is used to extract the spectral features of the sensor data. S1.4: Use the BERT model and the LDA model to extract the feature texts from each text data. Then, based on the extracted texture features, spectral features, distribution features, diversity features, road connectivity features, traffic flow aggregation features, mean, variance, spectral features, and feature texts, establish the land use representation matrix Z for the corresponding area. The specific form of the land use representation matrix Z is as follows: ;in, Represents remote sensing image features, including texture features and spectral features; Represents POI characteristics, including distribution characteristics and diversity characteristics; Represents traffic flow characteristics, including road connectivity characteristics and traffic flow aggregation characteristics, Represents sensor characteristics, including mean, variance, and spectral characteristics, Represents feature text.

[0008] As a further solution of the present invention, the specific steps of establishing the low-efficiency land discrimination model described in P2 are as follows: S2.1: After standardizing the various feature data in the constructed land use representation matrix Z, the various feature data are used as input data to establish a low-efficiency land discrimination model. The low-efficiency land discrimination model includes an input layer, a self-attention layer, a GNN layer, an LSTM layer, a correction layer, a classification layer, and an output layer. S2.2: Each set of input data is divided into a training set, a validation set, and a test set. The training set is input into the low-efficiency land discrimination model. The input layer receives the training set and performs forward propagation. The self-attention layer receives remote sensing image features and splits the remote sensing image features into multiple sub-blocks. Each sub-block is then mapped to a multidimensional vector space through a linear projection unit to form a patch embedding. S2.3: Calculate the positional encoding of each patch embedding using the sine function. Then, calculate the query matrix Q, key matrix K, and value matrix V of each patch embedding using the independent projection matrix of each attention head in the self-attention layer. Calculate the attention score of each attention head based on the query matrix Q and key matrix K. Normalize each attention score using the softmax function to obtain the attention weight of each attention head. S2.4: Based on the attention weights and the value matrix V, the global features of each attention head are obtained. The global features output by each head are then concatenated and projected through the output projection matrix to obtain the global features of the remote sensing image. S2.5: The GNN layer receives POI features and traffic flow features, and uses the geographic location, type, and density information in the POI features, as well as the traffic facilities in the traffic flow features, as nodes, and the connections between the nodes as edges to build a road network topology map for the corresponding area. Then, through repeated graph convolution operations, the information of each node in the road network topology map and its corresponding neighboring nodes is aggregated and the information of each node is updated. After multiple rounds of updates, the change value of each node information converges to the preset threshold. Then, the update stops and the enhanced POI features and traffic flow features are output. S2.6: The LSTM layer receives sensor features and processes them sequentially through the LSTM layer's forget gate, input gate, and output gate to obtain a predicted value for the current land use efficiency trend. The correction layer then receives feature text and corrects it based on the feature text of the current urban planning rules to obtain the corresponding corrected features. S2.7: Concatenate the global features, enhanced POI features, traffic flow features, predicted values of land efficiency trends, and corrected features output by each layer, and input the concatenated feature information into the classification layer. Then, based on the classification layer weights and biases, the concatenated feature information is processed using a sigmoid activation function, and the predicted probability of the current area being an inefficient land use is output. If the predicted probability is higher than a preset threshold, the current area is determined to be an inefficient land use. S2.8: Calculate the loss between the predicted result and the actual result using the Focal loss function, and backpropagate the loss value from the output layer of the low-efficiency land discrimination model to calculate the gradient value of each layer corresponding to the loss value. Use the AdamW optimizer to optimize the parameters of each network layer. After each round of training, input the validation set into the low-efficiency land discrimination model and calculate the loss value of the validation set for the low-efficiency land discrimination model to evaluate the model performance index. If the performance index does not reach the preset threshold, retrain; S2.9: Repeatedly train and verify the low-efficiency land discrimination model. If the model loss value of the validation set converges to the preset threshold after multiple rounds of iterations, stop training, and then input the test set data into the low-efficiency land discrimination model. The low-efficiency land discrimination model processes the test set layer by layer through forward propagation, and outputs the low-efficiency land judgment results of each region. At the same time, it calculates the SHAP value of each feature, and analyzes the impact of each feature on the judgment result based on each SHAP value.

[0009] As a further solution of the present invention, the specific steps of updating the model parameters in real time according to the model marking results described in P3 are as follows: S3.1: The trained low-efficiency land discrimination model is used as a global model. The weights, activation values, and inference accuracy of the global model are quantized to generate a lightweight low-efficiency land discrimination model with the same architecture. The generated lightweight low-efficiency land discrimination model is then distributed to each terminal device as a local model. S3.2: When the loss value of a local model exceeds a preset threshold, the parameter update mechanism is triggered. The parameter information of the current low-efficiency land use discrimination model is used as the initial global parameter and sent to each terminal device to replace the parameter information of each local model. After that, each terminal device inputs the local data set into the local model. S3.3: Each local model performs forward propagation on the local dataset, processes each group of data in the local dataset layer by layer, and outputs the corresponding low-efficiency land discrimination results. It then calculates the loss between each discrimination result and the true value, and then updates the parameters of each local model through backpropagation and gradient descent algorithm. S3.4: Each terminal device uploads the updated local model parameters to the global model. The global model then aggregates the local model parameters through weighted averaging to update the original model parameters, and then calculates the global model loss value. S3.5: If the loss value is higher than the preset threshold, the updated global model parameters will be re-sent to each terminal device, and local training, parameter upload and global parameter update will be performed again until the global model loss value is lower than the preset threshold. The parameter update will be stopped, the final global model parameters will be sent to each terminal device, and the parameters of each local device will be updated.

[0010] A system for extracting inefficient land in target areas based on multi-source big data, including a data acquisition module, a processing and alignment module, an extraction and fusion module, a land use identification module, an update monitoring module, a simulation and analysis module, a rule optimization module, and a data display module; The data acquisition module collects various types of data from different data sources through API interfaces, satellite remote sensing and web crawling; The processing alignment module is used to preprocess various types of data from different data sources and uniformly map the various types of data to the same analysis framework; The extraction and fusion module extracts various data features after preprocessing, fuses them, and constructs a corresponding representation matrix; The land use identification module identifies inefficient land use in the target area based on the constructed land use representation matrix; The update monitoring module is used to monitor changes in various types of data from different data sources and update the land use identification module parameter information in real time; The simulation analysis module analyzes the causes of low-efficiency land use based on the identification results of low-efficiency land use and simulates land use changes under different decision-making interventions; The rule optimization module crawls decision-making documents, academic papers, and public opinion data in real time, extracts the latest planning guidance information, and dynamically adjusts the inefficient land use identification rules; The data display module is used to build an interactive data panel to visually display the distribution of low-efficiency land and historical change trends, and automatically generate an analysis report.

[0011] As a further solution of the present invention, the specific steps of the simulation analysis module to simulate land use changes under different decision interventions are as follows: S4.1: The simulation analysis module receives the low-efficiency land identification results of each region and collects variables affecting low-efficiency land use, including economic factors, transportation and infrastructure, decision-making and regulation, and market behavior in each region. Based on the preset initial settings, it defines the causal relationship between these variables. S4.2: Calculate the conditional independence between variables through the structure learning algorithm and eliminate redundant causal relationships. Then, adjust the direction of the causal relationship between variables based on domain expert knowledge to construct a variable causal graph. , where V is a node in the variable causal graph G, representing each variable, and E is a directed edge in the variable causal graph G, representing the causal relationship between each variable; S4.3: Based on the variable causal graph G, a structural equation model is established through linear equations. Then, a set of instrumental variables is randomly set and the structural equation model is solved using the two-stage least squares method to calculate the causal impact of each variable on low-efficiency land use; S4.4: Assign a probability distribution to each variable. Then, based on historical data and causal inference results, construct a relationship between the proportion of inefficient land and each variable. Randomly extract variable values from the set probability distribution and input the extracted variable values into the land use identification module. Calculate the inefficient land change rate based on the inefficient land identification results of the land use identification module. S4.5: Repeat the random sampling and low-efficiency land change rate calculation multiple times until a preset number of iterations is reached, collect the calculated low-efficiency land change rates to generate a distribution of low-efficiency land changes, and then calculate the mean and variance of the low-efficiency land change rates; S4.6: If the mean of the low-efficiency land change rate is less than 0, it means that the decision is to reduce low-efficiency land. Otherwise, the feedback requires adjustment of the decision. Then, based on the mean and variance of the low-efficiency land change rate, calculate the confidence value of each decision, and select the decision with the highest confidence value as the optimal decision and feedback it to the staff for reference.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention collects remote sensing imagery, POIs, traffic flow, sensor, and text data, performs data alignment and correction, extracts feature categories, textures, and spectral features from remote sensing images, calculates POI distribution and diversity, simultaneously acquires sensor features based on statistical analysis and Fourier transform, and extracts text features using BERT and LDA. The extracted feature data is then fed into a low-efficiency land discrimination model. The model's self-attention layer extracts global remote sensing features, the GNN layer enhances POI and traffic features, the LSTM layer predicts land efficiency trends, and the text correction layer optimizes the discrimination rules. The classification layer ultimately outputs the probability of low-efficiency land use, and the model is optimized using a Focal loss function. After training, the model is quantized to generate a lightweight version. Based on federated learning, model parameters are updated locally on the terminal device. This model can characterize land use from multiple perspectives, improving recognition accuracy and feature extraction precision, while also improving low-efficiency land identification accuracy, reducing terminal computing costs, and enabling real-time updates and optimization. It is applicable to land management needs in different regions and cities. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0014] Figure 1 This is a flowchart of a method for extracting low-efficiency land in a target area based on multi-source big data proposed by the present invention; Figure 2 This is a system block diagram of a target area low-efficiency land extraction system based on multi-source big data proposed by the present invention. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0016] Example 1, reference Figure 1 This embodiment discloses a method for extracting low-efficiency land in a target area based on multi-source big data. The specific steps of the extraction method are as follows: Collect data from multiple data sources in real time, pre-process the collected multi-source data, and then extract the feature information of the pre-processed multi-source data.

[0017] Specifically, remote sensing images, POI data, sensor data and text data are collected from multiple data sources, and the collected data are classified according to the data format. The atmospheric correction model is used to convert the digital values of the remote sensing images into surface reflectance. The preset affine transformation parameters are then used to correct the coordinate deviations of different remote sensing images through affine transformation. The cosine similarities between the POI data are calculated and recorded. The recorded cosine similarities are traversed. If the cosine similarity is higher than the preset threshold, it is determined that the corresponding two groups of POI data are repeated. After the traversal is completed, the repeated POI data are removed according to the preset requirements, and the different types of POI data are normalized. The different sensor data with inconsistent collection times are interpolated by the spatiotemporal Kriging interpolation method to align the collection times of the sensor data. The low-frequency components of the sensor data are then decomposed by one-dimensional wavelet transform to remove the low-frequency components of the sensor data. Noise interference is eliminated, and then the TF-IDF method is used to extract keywords from social media data and historical planning texts, and the extracted keywords are converted into numerical vectors. The remote sensing images of various raster formats are adjusted to the same resolution through the resampling method. Based on the raster format of the remote sensing images, the POI data and traffic flow data are rasterized accordingly to spatially align the remote sensing images, POI data and sensor data. Time series interpolation is used to fill the missing values of remote sensing images and POI data to synchronize the sampling time of remote sensing images and POI data. The time window aggregation method is then used to time-match the sensor data with the remote sensing images and POI data. After that, the social media data and historical planning texts are geographically parsed, and the text addresses are converted into the same raster format as the remote sensing images. The time deduction method is used to match the time scales of social media data and historical planning texts.

[0018] Based on the extracted characteristic information, a land use characterization matrix is established, and based on the constructed land use characterization matrix, a low-efficiency land use discrimination model is established.

[0019] Specifically, we collect pre-processed remote sensing images, POI data, sensor data, and text data from multiple data sources, and then detect whether there is any deviation in the various types of data after spatiotemporal alignment. If so, we correct it again, and then separate and extract the traffic flow data from the sensor data. Through the DeepLabV3+ model, we convert each remote sensing image into a pixel-level feature category map, and then use the gray-level co-occurrence matrix to extract the texture features of each remote sensing image. The vegetation index in each remote sensing data is normalized to obtain the corresponding spectral features. Then, we calculate the spatial distribution characteristics of the POI data through KDE, and calculate the POI diversity entropy to obtain the diversity characteristics of the POI data type in each region. The GNN network extracts the road network structure of traffic flow data and constructs an adjacency matrix to extract road connectivity features, calculate the average traffic flow in a preset unit time, and obtain traffic flow aggregation features in different time intervals. It then performs statistical analysis on the sensor data, extracts the mean and variance feature information, and uses Fourier transform to extract the spectral features of the sensor data. The BERT model and LDA model are used to extract the feature texts in each text data. Then, based on the extracted texture features, spectral features, distribution features, diversity features, road connectivity features, traffic flow aggregation features, mean, variance, spectral features and feature texts, a land use representation matrix Z for the corresponding area is established.

[0020] Specifically, after standardizing various feature data in the constructed land representation matrix Z, various feature data are used as input data, and then a low-efficiency land discrimination model is established, wherein the low-efficiency land discrimination model includes an input layer, a self-attention layer, a GNN layer, a LSTM layer, a correction layer, a classification layer and an output layer. Each group of input data is divided into a training set, a validation set and a test set. The training set is input into the low-efficiency land discrimination model. The input layer receives the training set and performs forward propagation. The self-attention layer receives remote sensing image features and splits the remote sensing image features into multiple sub-blocks. Each sub-block is then mapped to a multidimensional vector space through a linear projection unit to form a patch embedding. The position of each patch embedding is calculated by a sine function. The self-attention layer is used to encode the image, and then the query matrix Q, key matrix K and value matrix V of each patch are calculated through the independent projection matrix of each attention head in the self-attention layer. The attention score of each attention head is calculated based on the query matrix Q and the key matrix K. The softmax function is used to normalize the attention scores to obtain the attention weights of each attention head. According to the attention weights and the value matrix V, the global features of each attention head are obtained, and the global features of the output of each head are spliced. After projection through the output projection matrix, the global features of the remote sensing image are obtained. The GNN layer receives POI features and traffic flow features, and combines the geographical location, type and density information in the POI features, and the traffic flow features in the traffic flow features. The communication facilities are used as nodes, and the connections between the nodes are used as edges to establish the road network topology of the corresponding area. Then, through repeated graph convolution operations, the information of each node in the road network topology and its corresponding neighboring nodes are aggregated, and the information of each node is updated. After multiple rounds of updates, the change value of each node information converges to the preset threshold, the update is stopped, and the enhanced POI features and traffic flow features are output. The LSTM layer receives the sensor features and processes them in sequence through the LSTM layer forget gate, input gate, and output gate to obtain the predicted value of the current land use efficiency trend. After that, the correction layer receives the feature text and corrects the feature text based on the current urban planning field rules to obtain the corresponding correction features, splicing the global features output by each layer, Enhanced POI features and traffic flow features, predicted values of land use efficiency trends and corrected features, and the spliced feature information is input into the classification layer. Then, according to the classification layer weights and biases, the spliced feature information is processed through the sigmoid activation function, and the predicted probability that the current area is low-efficiency land is output. If the predicted probability is higher than the preset threshold, the current area is judged to be low-efficiency land. The loss value between the predicted result and the actual result is calculated through the Focal loss function, and the loss value is output from the low-efficiency land discrimination model output layer, and the gradient value of each layer corresponding to the loss value is calculated through back propagation, and the parameters of each network layer are optimized using the AdamW optimizer. After each round of training, the verification set is input into the low-efficiency land discrimination model.Calculate the loss value of the validation set for the low-efficiency land discrimination model to evaluate the model performance index. If the performance index does not reach the preset threshold, retrain and repeatedly train the low-efficiency land discrimination model. If the model loss value of the validation set converges to the preset threshold after multiple rounds of iterations, stop training, and then input the test set data into the low-efficiency land discrimination model. The low-efficiency land discrimination model processes the test set layer by layer through forward propagation and outputs the low-efficiency land judgment results of each region. At the same time, calculate the SHAP value of each feature, and analyze the impact of each feature on the discrimination result based on each SHAP value.

[0021] It should be further explained that the specific form of the land characterization matrix Z is as follows: ;in, Represents remote sensing image features, including texture features and spectral features; Represents POI characteristics, including distribution characteristics and diversity characteristics; Represents traffic flow characteristics, including road connectivity characteristics and traffic flow aggregation characteristics, Represents sensor characteristics, including mean, variance, and spectral characteristics, Represents feature text.

[0022] The real-time collected data is input into each terminal device, and the low-efficiency land is identified and marked through the low-efficiency land discrimination model. At the same time, the model parameters are updated in real time according to the model marking results.

[0023] Specifically, the trained low-efficiency land discrimination model is used as the global model, and the weights, activation values and inference accuracy of the global model are quantified to generate a lightweight low-efficiency land discrimination model with the same architecture, and the generated lightweight low-efficiency land discrimination model is sent to each terminal device as the local model of each terminal device. When the loss value of the local model is higher than the preset threshold, the parameter update mechanism is triggered, and the parameter information of the current low-efficiency land discrimination model is used as the initial global parameter and sent to each terminal device to replace the parameter information of each local model. After that, each terminal device inputs the local data set into the local model, and each local model performs forward propagation on the local data set respectively, and processes each group in the local data set layer by layer. Data is collected and the corresponding low-efficiency land discrimination result is output. The loss value between each discrimination result and the true value is then calculated. The local model parameters are then updated through the back-propagation method and the gradient descent algorithm. Each terminal device uploads the updated local model parameters to the global model. The global model then aggregates the local model parameters through weighted averaging to update the original model parameters. The global model loss value is then calculated. If the loss value is higher than the preset threshold, the updated global model parameters are re-sent to each terminal device, and local training, parameter upload, and global parameter update are performed again until the global model loss value is lower than the preset threshold. The parameter update is then stopped, and the final global model parameters are sent to each terminal device, and the parameters of each local device are updated.

[0024] Analyze the causes of low-efficiency land use, predict future trends of low-efficiency land use, and generate a land use efficiency change forecast report.

[0025] Dynamically optimize the low-efficiency land identification rules, update the identification rules of the low-efficiency land identification model, and visualize the distribution and historical change trends of low-efficiency land.

[0026] Example 2, reference Figure 2 This embodiment discloses a system for extracting inefficient land in a target area based on multi-source big data, including a data acquisition module, a processing and alignment module, an extraction and fusion module, a land use identification module, an update monitoring module, a simulation analysis module, a rule optimization module, and a data display module; The data acquisition module collects various types of data from different data sources through API interfaces, satellite remote sensing and web crawling; the processing and alignment module is used to preprocess various types of data from different data sources and uniformly map various types of data to the same analysis framework; the extraction and fusion module extracts the features of various types of preprocessed data, fuses them, and constructs a land use representation matrix; the land use identification module identifies inefficient land use in the target area based on the constructed land use representation matrix; the update monitoring module is used to monitor changes in various types of data from different data sources and update the land use identification module parameter information in real time.

[0027] The simulation analysis module analyzes the causes of low-efficiency land use based on the identification results of low-efficiency land use and simulates land use changes under different decision-making interventions.

[0028] Specifically, the simulation analysis module receives the low-efficiency land identification results of each region, and collects the variables of economic factors, transportation and infrastructure, decision-making and regulation, and market behavior that affect low-efficiency land in each region. Based on the preset initial settings, it defines the causal relationship between each variable, calculates the conditional independence between each variable through the structure learning algorithm, and eliminates redundant causal relationships. Then, based on the knowledge of domain experts, it adjusts the direction of the causal relationship between each variable to construct a variable causal diagram. , where V is a node in the variable causal graph G, representing each variable, and E is a directed edge in the variable causal graph G, representing the causal relationship between each variable. Based on the variable causal graph G, a structural equation model is established through a linear equation. Then, a set of instrumental variables are randomly set, and the structural equation model is solved by the two-stage least squares method to calculate the causal impact of each variable on low-efficiency land. A probability distribution is assigned to each variable, and then based on historical data and causal reasoning results, the relationship between the proportion of low-efficiency land and each variable is constructed. Variable values are randomly extracted from the set probability distribution and the extracted variable values are input into the land use identification module. According to The low-efficiency land identification results of the land identification module are used to calculate the low-efficiency land change rate. Random sampling and low-efficiency land change rate calculations are repeated multiple times until the preset number of iterations is reached. The calculated low-efficiency land change rates are collected to generate the distribution of low-efficiency land changes, and then the mean and variance of the low-efficiency land change rate are calculated. If the mean of the low-efficiency land change rate is less than 0, it means that the decision is to reduce low-efficiency land. Otherwise, the feedback requires adjustment of the decision. Then, based on the mean and variance of the low-efficiency land change rate, the confidence value of each decision is calculated, and the decision with the highest confidence value is selected as the optimal decision and fed back to the staff for reference.

[0029] The rule optimization module crawls decision-making documents, academic papers and public opinion data in real time, extracts the latest planning guidance information, and dynamically adjusts the low-efficiency land identification rules; the data display module is used to build an interactive data panel to visually display the distribution of low-efficiency land and historical change trends, and automatically generate analysis reports.

Claims

1. A method for extracting inefficient land in target areas based on multi-source big data, characterized by: The specific steps of the extraction method are as follows: P1. Collect data from multiple data sources in real time, pre-process the collected multi-source data, and then extract feature information from the pre-processed multi-source data. P2. Based on the extracted characteristic information, a land use characterization matrix is established, and based on the constructed land use characterization matrix, an inefficient land use discrimination model is established; P3: Input the real-time collected data into each terminal device, identify and mark the low-efficiency land through the low-efficiency land discrimination model, and update the model parameters in real time based on the model marking results; P4. Analyze the causes of low-efficiency land use, predict future trends of low-efficiency land use, and generate a land use efficiency change forecast report; P5. Dynamically optimize the rules for identifying low-efficiency land, update the rules of the low-efficiency land identification model, and visualize the distribution and historical change trends of low-efficiency land.

2. The method for extracting low-efficiency land in target areas based on multi-source big data according to claim 1 is characterized in that: The specific steps for establishing the land use representation matrix based on the extracted various feature information in P2 are as follows: S1.1: Collect pre-processed remote sensing images, POI data, sensor data, and text data from multiple data sources. Then, check the spatiotemporal alignment of these data to see if there are any deviations. If so, correct them. Then, separate and extract traffic flow data from the sensor data. S1.2: Use the DeepLabV3+ model to convert each remote sensing image into a pixel-level feature classification map. Then use the gray-level co-occurrence matrix to extract the texture features of each remote sensing image. Use the normalized vegetation index in each remote sensing data to obtain the corresponding spectral features. Then use KDE to calculate the spatial distribution characteristics of the POI data and calculate the POI diversity entropy to obtain the diversity characteristics of the POI data type in each region. S1.3: Extract the road network structure of traffic flow data based on the GNN network and construct an adjacency matrix to extract road connectivity features. Calculate the average traffic flow within a preset unit time and obtain traffic flow aggregation features within different time intervals. Statistical analysis is then performed on the sensor data to extract mean and variance features. Fourier transform is used to extract the spectral features of the sensor data. S1.4: Use the BERT model and the LDA model to extract the feature texts from each text data. Then, based on the extracted texture features, spectral features, distribution features, diversity features, road connectivity features, traffic flow aggregation features, mean, variance, spectral features, and feature texts, establish the land use representation matrix Z for the corresponding area. The specific form of the land use representation matrix Z is as follows: ;in, Represents remote sensing image features, including texture features and spectral features; Represents POI characteristics, including distribution characteristics and diversity characteristics; Represents traffic flow characteristics, including road connectivity characteristics and traffic flow aggregation characteristics, Represents sensor characteristics, including mean, variance, and spectral characteristics, Represents feature text.

3. The method for extracting low-efficiency land in target areas based on multi-source big data according to claim 2 is characterized in that: The specific steps for establishing the low-efficiency land use discrimination model described in P2 are as follows: S2.1: After standardizing the various feature data in the constructed land use representation matrix Z, the various feature data are used as input data to establish a low-efficiency land discrimination model. The low-efficiency land discrimination model includes an input layer, a self-attention layer, a GNN layer, an LSTM layer, a correction layer, a classification layer, and an output layer. S2.2: Each set of input data is divided into a training set, a validation set, and a test set. The training set is input into the low-efficiency land discrimination model. The input layer receives the training set and performs forward propagation. The self-attention layer receives remote sensing image features and splits the remote sensing image features into multiple sub-blocks. Each sub-block is then mapped to a multidimensional vector space through a linear projection unit to form a patch embedding. S2.3: Calculate the positional encoding of each patch embedding using the sine function. Then, calculate the query matrix Q, key matrix K, and value matrix V of each patch embedding using the independent projection matrix of each attention head in the self-attention layer. Calculate the attention score of each attention head based on the query matrix Q and key matrix K. Normalize each attention score using the softmax function to obtain the attention weight of each attention head. S2.4: Based on the attention weights and the value matrix V, the global features of each attention head are obtained. The global features output by each head are then concatenated and projected through the output projection matrix to obtain the global features of the remote sensing image. S2.5: The GNN layer receives POI features and traffic flow features, and uses the geographic location, type, and density information in the POI features, as well as the traffic facilities in the traffic flow features, as nodes, and the connections between the nodes as edges to build a road network topology map for the corresponding area. Then, through repeated graph convolution operations, the information of each node in the road network topology map and its corresponding neighboring nodes is aggregated and the information of each node is updated. After multiple rounds of updates, the change value of each node information converges to the preset threshold. Then, the update stops and the enhanced POI features and traffic flow features are output. S2.6: The LSTM layer receives sensor features and processes them sequentially through the LSTM layer's forget gate, input gate, and output gate to obtain a predicted value for the current land use efficiency trend. The correction layer then receives feature text and corrects it based on the feature text of the current urban planning rules to obtain the corresponding corrected features. S2.7: Concatenate the global features, enhanced POI features, traffic flow features, predicted values of land efficiency trends, and corrected features output by each layer, and input the concatenated feature information into the classification layer. Then, based on the classification layer weights and biases, the concatenated feature information is processed using a sigmoid activation function, and the predicted probability of the current area being an inefficient land use is output. If the predicted probability is higher than a preset threshold, the current area is determined to be an inefficient land use. S2.8: Calculate the loss between the predicted result and the actual result using the Focal loss function, and backpropagate the loss value from the output layer of the low-efficiency land discrimination model to calculate the gradient value of each layer corresponding to the loss value. Use the AdamW optimizer to optimize the parameters of each network layer. After each round of training, input the validation set into the low-efficiency land discrimination model and calculate the loss value of the validation set for the low-efficiency land discrimination model to evaluate the model performance index. If the performance index does not reach the preset threshold, retrain; S2.9: Repeatedly train and verify the low-efficiency land discrimination model. If the model loss value of the validation set converges to the preset threshold after multiple rounds of iterations, stop training, and then input the test set data into the low-efficiency land discrimination model. The low-efficiency land discrimination model processes the test set layer by layer through forward propagation, and outputs the low-efficiency land judgment results of each region. At the same time, it calculates the SHAP value of each feature, and analyzes the impact of each feature on the judgment result based on each SHAP value.

4. The method for extracting low-efficiency land in target areas based on multi-source big data according to claim 3 is characterized in that: The specific steps for updating model parameters in real time based on model labeling results described in P3 are as follows: S3.1: The trained low-efficiency land discrimination model is used as a global model. The weights, activation values, and inference accuracy of the global model are quantized to generate a lightweight low-efficiency land discrimination model with the same architecture. The generated lightweight low-efficiency land discrimination model is then distributed to each terminal device as a local model. S3.2: When the loss value of a local model exceeds a preset threshold, the parameter update mechanism is triggered. The parameter information of the current low-efficiency land use discrimination model is used as the initial global parameter and sent to each terminal device to replace the parameter information of each local model. After that, each terminal device inputs the local data set into the local model. S3.3: Each local model performs forward propagation on the local dataset, processes each group of data in the local dataset layer by layer, and outputs the corresponding low-efficiency land discrimination results. It then calculates the loss between each discrimination result and the true value, and then updates the parameters of each local model through backpropagation and gradient descent algorithm. S3.4: Each terminal device uploads the updated local model parameters to the global model. The global model then aggregates the local model parameters through weighted averaging to update the original model parameters, and then calculates the global model loss value. S3.5: If the loss value is higher than the preset threshold, the updated global model parameters will be re-sent to each terminal device, and local training, parameter upload and global parameter update will be performed again until the global model loss value is lower than the preset threshold. The parameter update will be stopped, the final global model parameters will be sent to each terminal device, and the parameters of each local device will be updated.

5. A system for extracting low-efficiency land in target areas based on multi-source big data, used to implement the method for extracting low-efficiency land in target areas based on multi-source big data according to any one of claims 1 to 4, characterized in that: It includes data acquisition module, processing and alignment module, extraction and fusion module, land use identification module, update monitoring module, simulation analysis module, rule optimization module and data display module; The data acquisition module collects various types of data from different data sources through API interfaces, satellite remote sensing and web crawling; The processing alignment module is used to preprocess various types of data from different data sources and uniformly map the various types of data to the same analysis framework; The extraction and fusion module extracts various data features after preprocessing, fuses them, and constructs a corresponding representation matrix; The land use identification module identifies inefficient land use in the target area based on the constructed land use representation matrix; The update monitoring module is used to monitor changes in various types of data from different data sources and update the land use identification module parameter information in real time; The simulation analysis module analyzes the causes of low-efficiency land use based on the identification results of low-efficiency land use and simulates land use changes under different decision-making interventions; The rule optimization module crawls decision-making documents, academic papers, and public opinion data in real time, extracts the latest planning guidance information, and dynamically adjusts the inefficient land use identification rules; The data display module is used to build an interactive data panel to visually display the distribution of low-efficiency land and historical change trends, and automatically generate an analysis report.

6. The target area low-efficiency land extraction system based on multi-source big data according to claim 5 is characterized in that: The specific steps of the simulation analysis module to simulate land use changes under different decision interventions are as follows: S4.1: The simulation analysis module receives the low-efficiency land identification results of each region and collects variables affecting low-efficiency land use, including economic factors, transportation and infrastructure, decision-making and regulation, and market behavior in each region. Based on the preset initial settings, it defines the causal relationship between these variables. S4.2: Calculate the conditional independence between variables through the structure learning algorithm and eliminate redundant causal relationships. Then, adjust the direction of the causal relationship between variables based on domain expert knowledge to construct a variable causal graph. , where V is a node in the variable causal graph G, representing each variable, and E is a directed edge in the variable causal graph G, representing the causal relationship between each variable; S4.3: Based on the variable causal graph G, a structural equation model is established through linear equations. Then, a set of instrumental variables is randomly set and the structural equation model is solved using the two-stage least squares method to calculate the causal impact of each variable on low-efficiency land use; S4.4: Assign a probability distribution to each variable. Then, based on historical data and causal inference results, construct a relationship between the proportion of inefficient land and each variable. Randomly extract variable values from the set probability distribution and input the extracted variable values into the land use identification module. Calculate the inefficient land change rate based on the inefficient land identification results of the land use identification module. S4.5: Repeat the random sampling and low-efficiency land change rate calculation multiple times until a preset number of iterations is reached, collect the calculated low-efficiency land change rates to generate a distribution of low-efficiency land changes, and then calculate the mean and variance of the low-efficiency land change rates; S4.6: If the mean of the low-efficiency land change rate is less than 0, it means that the decision is to reduce low-efficiency land. Otherwise, the feedback requires adjustment of the decision. Then, based on the mean and variance of the low-efficiency land change rate, calculate the confidence value of each decision, and select the decision with the highest confidence value as the optimal decision and feedback it to the staff for reference.

Images