Urban land use classification method based on multi-source data and deep learning fusion

Abstract

This invention discloses an urban land use classification method based on the fusion of multi-source data and deep learning, comprising: Step 1: Constructing a spatiotemporal map using Poisson disk sampling and graph neural networks, integrating the spatiotemporal heterogeneity of multi-source data, and dynamically adjusting weights using an attention mechanism; Step 2: Extracting static features from remote sensing and DEM using Swin Transformer, processing dynamic features from social media and mobile phone signaling using LSTM, fusing static and dynamic features through cross-modal attention, and optimizing data matching degree based on Grad-CAM++; Step 3: Combining the person-to-land flow of mobile phone signaling with the spatial correlation of social media activities, predicting trends using LSTM and embedding ecological indicator constraints; Step 4: Constructing a cross-departmental data sharing platform, collaboratively optimizing planning through a unified indicator system, and dynamically adjusting parameters using social media to mine public demands and POI data; Step 5: Introducing ecological footprint and carbon emission indicators to assess sustainability, establishing a monthly update mechanism, and dynamically calibrating model parameters through real-time data streams.

Description

Technical Field

[0001] This invention relates to the field of land use classification technology, and in particular to an urban land use classification method based on the fusion of multi-source data and deep learning. Background Technology

[0002] Against the backdrop of accelerated urbanization and surging demand for ecological protection, urban land use classification faces prominent contradictions such as lagging dynamic monitoring and insufficient coordination in ecological protection. For example, in a coastal new city, commercial land expanded by 30% within five years, resulting in the filling of 200 hectares of surrounding wetlands for real estate development. Traditional static classification methods, which rely on annual satellite imagery and isolated departmental planning, failed to capture the correlation between the thermal migration of people and changes in wetland ecological indicators in real time, and also failed to balance development needs and ecological red line protection through cross-departmental collaboration mechanisms. Therefore, an urban land use classification method based on the fusion of multi-source data and deep learning is proposed. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing an urban land use classification method based on the fusion of multi-source data and deep learning.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: Urban land use classification methods based on the fusion of multi-source data and deep learning include: Step 1 (Spatiotemporal Alignment and Dynamic Preprocessing of Multi-Source Data): A spatiotemporal graph structure is constructed by using Poisson disk sampling and graph neural network (GNN) to integrate the spatiotemporal heterogeneity of multi-source remote sensing, social media, POI, and mobile signaling data. The weights of each data source are dynamically adjusted by combining an attention mechanism (e.g., strengthening the weights of POI and mobile signaling in commercial areas, and highlighting the weights of remote sensing and DEM in ecological areas). This solves the problems of information loss and classification bias dominated by a single data source in traditional gridding methods. Step 2 (Deep Extraction and Enhancement of Multimodal Features): The Swing Transformer is used to extract static spatial features from remote sensing and DEM, and LSTM is used to process the temporal dynamic features of social media and mobile signaling. Static-dynamic features are fused through a cross-modal attention module, and the matching degree of multi-source data information is optimized based on the Grad-CAM++ visualization model of the region of interest (such as balancing the weights of POI and remote sensing images in industrial land classification). Step 3 (Dynamic Land Use Change Prediction and Ecological Constraints): Combining mobile phone signaling-land flow data with social media activity-spatial correlation, LSTM is used to predict land use change trends (such as the correlation between the surge in mobile phone signaling in commercial areas on weekends and activity hotspots). At the same time, ecological indicators are embedded to ensure the protection of ecological land such as arable land / wetlands, and the model is deployed in a lightweight manner and processed in real time through edge computing. Step 4 (Cross-departmental collaborative verification and policy integration): Construct a cross-departmental data sharing platform for natural resources, agriculture, water conservancy and environmental protection. Through a unified indicator system (such as the rate of concentrated contiguous farmland and the coverage rate of ecological ditches), conduct multi-departmental planning collaborative optimization. Utilize social media to mine public demands and POI human-land relationship data, dynamically adjust classification model parameters, and form a "planning-feedback-optimization" closed loop. Step 5 (Dynamic Updates and Sustainability Assessment): Introduce ecological footprint and carbon emission monitoring indicators to conduct sustainability assessments of the classification results (such as constraining the layout of high-pollution industries in the expansion of industrial land), and establish a monthly data update mechanism. Dynamically calibrate model parameters through real-time data streams from mobile phone signaling and social media to solve the problem that traditional static classification methods cannot adapt to rapid urbanization, and achieve a triple improvement in "accuracy, real-time performance, and sustainability".

[0005] The above technical solution further includes: Further, step one includes: Spatially homogenizing multi-source data (high-resolution pixels from remote sensing images, location check-in points on social media, coordinates of POIs, and trajectory points from mobile phone signaling) is performed. Poisson disk sampling avoids the oversampling in dense areas and undersampling in sparse areas caused by traditional gridded sampling by generating random point sets with minimum distance threshold constraints. For example, in commercial areas where mobile phone signaling points are dense, Poisson disk sampling can ensure that the spacing between sampling points is greater than a set threshold, avoiding redundancy. In ecological areas where data is sparse, the threshold can be dynamically adjusted to ensure coverage of key areas. This provides spatially uniform and non-redundant node inputs for subsequent graph neural networks, solving the problem of uneven spatial sampling in data heterogeneity. Mapping entities in multi-source data to graph nodes, for example: Remote sensing image pixels (including spectral and texture features) → spatial nodes; Social media check-in points (including timestamps and event types) → Spatiotemporal nodes; Points of Interest (POIs) (including category and popularity) → Semantic Nodes; Mobile phone signaling trajectory points (including user ID, time, and location) → dynamic nodes; Edges between nodes are defined based on spatiotemporal correlation, for example: The movement trajectory of the same user within 24 hours → "temporal edges" between dynamic nodes; Social media check-in points and nearby POIs → "semantic-spatial" association edges; Remote sensing image pixels and adjacent pixels → spatial adjacency edges.

[0006] Through the message passing mechanism of GNN, node features (such as remote sensing pixel values ​​and POI heat) are aggregated along the edge to perform "local-global" feature fusion. For example, commercial area nodes aggregate surrounding POIs and mobile signaling data to form comprehensive features of pedestrian flow and commerce; ecological area nodes aggregate remote sensing vegetation index and DEM slope to form ecological features. To avoid analytical bias caused by differences in time scale, the time resolution of different data sources is unified (e.g., aggregating "hourly" check-in data from social media and "minute-level" trajectory data from mobile phone signaling into "daily" data through a time window). Geographic coordinate system transformation and spatial resolution unification ensure that multi-source data can be overlaid and analyzed in space. Semantic features of different data sources are mapped to a unified dimension. For example, POI categories (such as "restaurants" and "retail") are mapped to numerical business popularity indicators, which, together with social media check-in frequency and mobile phone signaling flow density, constitute a comprehensive feature of business activity.

[0007] The attention mechanism assigns dynamic weights to different data sources, with the weights determined by both scenario requirements and data quality. For example, in commercial areas, POI data (reflecting the distribution of commercial facilities) and mobile signaling data (reflecting pedestrian density) have higher weights, while remote sensing data (reflecting building outlines) has lower weights. In ecological areas, remote sensing data (reflecting vegetation cover) and DEM data (reflecting terrain slope) have higher weights, while social media data (reflecting human activities) has lower weights. The weights are updated in real time according to changes in the scenario. For example, when mobile phone signaling data surges in commercial areas on weekends, the weights are automatically increased; when the remote sensing vegetation index in ecological areas declines in winter, the weights are automatically adjusted to reflect ecological pressure.

[0008] Furthermore, the method of using Swin Transformer to extract static spatial features from remote sensing and DEM, LSTM to process temporal dynamic features from social media and mobile signaling, and performing static-dynamic feature fusion through a cross-modal attention module includes the following steps: The remote sensing imagery and DEM data are unified to the same spatial resolution and normalized to eliminate dimensional differences. The Swin Transformer extracts features layer by layer through a layered window self-attention mechanism. The imagery is divided into non-overlapping local windows, and self-attention is calculated within each window to capture detailed features such as local textures and edges (e.g., building outlines and vegetation cover). Through a "window translation" mechanism, features from adjacent windows are fused to gradually expand the receptive field and capture the global spatial structure (e.g., the overall layout of urban areas and topographic relief features). A static spatial feature vector is generated, containing spectral features of the remote sensing imagery, topographic features of the DEM (e.g., slope and elevation), and spatial structure information, providing a "stable and high-precision" static base for subsequent fusion. Social media data (such as user check-in locations and activity times) and mobile signaling data (such as user movement trajectories and time-of-day distribution) are aggregated into time-series sequences according to the time dimension; LSTM processes time-series data through gating mechanisms (input gate, forget gate, output gate) to capture time dependencies; for example: Short-term dependence: capturing the differences in pedestrian traffic between weekdays and weekends, and changes in movement patterns during morning and evening rush hours; Long-term dependence: Identify seasonal activity patterns (such as a surge in foot traffic in commercial areas during holidays and a decrease in visitors to eco-regions during winter). A dynamic temporal feature vector is generated, containing dynamic information such as pedestrian flow heat distribution and activity frequency changes, reflecting the real-time status and changing trends of urban land use. The static feature vector extracted by the Swing Transformer and the dynamic feature vector extracted by the LSTM are input into the cross-modal attention module. The module calculates the correlation between the two features through a query-key-value mechanism. Query vector (Q): Represents the feature dimension that needs to be strengthened (such as "business activity"); Key vector (K): Represents candidate information in static / dynamic features (such as "building density" in remote sensing images and "human density" in mobile phone signaling); Value vector (V): Represents the actual weight allocation of the integration (such as the contribution of "building density" to "commercial activity"). Attention scores are used to assign weights to different features. For example, in commercial areas, "pedestrian density" (a dynamic feature) and "building density" (a static feature) are highly correlated, so their weights are increased. In ecological areas, "vegetation cover" (a static feature) may have a higher weight. The static and dynamic features are weighted and fused to generate a comprehensive feature vector. This vector contains both static spatial structure information (such as building layout and terrain features) and dynamic trends (such as pedestrian heat maps and activity patterns), and is ultimately used for land use classification (such as commercial, residential, industrial, and ecological land).

[0009] Furthermore, the optimization of multi-source data information matching degree based on the Grad-CAM++ visualization model's region of interest includes the following steps: After training the classification model that combines Swin Transformer (static features) and LSTM (dynamic features), test samples are selected and input into the model. The feature map of the last convolutional layer of the model records the response intensity of each neuron to the input data (such as the "building edge" feature of remote sensing images and the "crowd gathering" feature of mobile phone signaling). Class activation graph:

[0010] in, For the class activation graph of category c (e.g., "commercial land"), Let be the activation value of the k-th feature map. The weight of category c on feature map k; Grad-CAM++ generates more accurate heatmaps by improving the weight calculation method, thus locating the "areas of interest" of the model in the input data. For example, in the commercial land classification, the highlighted areas on the heatmap may correspond to "POI dense areas" and "high-density mobile signaling areas"; in the ecological land classification, the highlighted areas may correspond to "high vegetation coverage areas in remote sensing images" and "high slope areas in DEMs". Weight The calculation introduces a second-order gradient correction term, and the formula is optimized as follows:

[0011] Wherein, γ is an adjustable parameter used to balance the contributions of the first and second order gradients and improve the positioning accuracy of the heatmap; Z is the spatial size normalization factor of the feature map, which is equal to the total number of pixels of the feature map width × height; i and j are the spatial position indices of the feature map, corresponding to the horizontal (width direction) and vertical (height direction) coordinates of the pixel, respectively, with values ​​ranging from 1≤i≤W and 1≤j≤H (W and H are the width and height of the feature map); The activation value of the k-th feature map at position (i,j) reflects the feature response intensity at that spatial location; The prediction score of the model for category c represents the confidence level of the model in believing that the input belongs to category c; By analyzing the "focus bias" of the model on multi-source data through heatmap analysis, for example, if the heatmap shows that the model overemphasizes POI data in the industrial land classification while ignoring the "building form" features of remote sensing images, it indicates that the weight of POI data is too high; if the heatmap shows that the model does not pay attention to the "slope" features of DEM in the ecological land classification, it indicates that the weight of DEM data is insufficient or the feature extraction is inadequate. Based on the heatmap analysis results, dynamically adjust the fusion weights or feature extraction strategies of multi-source data. For example, in industrial land classification, if the weight of POI data is found to be too high, reduce the weight of POI data and increase the weight of remote sensing imagery. In ecological land classification, if the slope features of DEM are not being considered, optimize the DEM feature extraction module (e.g., increase the resolution of slope calculation) or increase the weight of DEM data. Verify the optimization effect by observing changes in the classification accuracy of the adjusted model (e.g., precision and recall) to ensure improved matching of multi-source data information.

[0012] Furthermore, the method of combining mobile phone signal-to-land flow data with social media activity-spatial correlation, using LSTM to predict land use change trends, and embedding ecological indicator constraints includes the following steps: The system aligns the human-location flow data from mobile phone signaling (such as user movement trajectory, dwell time, and time period distribution) with the activity-spatial correlation data from social media (such as location check-in, activity type, and timestamp). For example, it constructs a human-activity-location spatiotemporal correlation matrix by aggregating geographic coordinate system and time window. Key features such as human flow density, activity frequency, and activity type distribution are extracted to form a dynamic feature vector, which serves as the input sequence for LSTM. LSTM predicts future land use trends by learning time patterns in historical data (such as differences in pedestrian traffic between weekdays and weekends, and seasonal activity changes). For example, a surge in weekend pedestrian traffic in commercial areas, with activity types concentrated in dining / shopping, can predict commercial land expansion trends; high density of mobile phone signaling in industrial areas during fixed weekday periods, with activity types of commuting, can predict industrial land transfer trends. NDVI (Normalized Difference Vegetation Index, reflecting vegetation cover) and DEM slope (reflecting terrain stability) were selected as ecological constraint indicators. A higher NDVI value indicates denser vegetation and a higher priority for ecological protection; a steeper DEM slope indicates steeper terrain and a higher development risk. NDVI and DEM slope were input as additional features into the LSTM, allowing the model to directly consider the impact of ecological indicators during prediction. For example, when predicting commercial area expansion, if the surrounding NDVI value is high (ecologically sensitive area), the model will reduce the probability of expansion. Add an ecological penalty term to the model's loss function:

[0013] in, To predict losses, Let λ represent the ecological constraint loss (such as the decrease in NDVI, the area of ​​the slope exceeding the standard), and λ be the balance coefficient. Gradient descent optimization is used to minimize ecological damage while maintaining classification accuracy in the prediction results. The λ value is dynamically adjusted according to the city's development stage (such as rapid expansion period / ecological protection period). In the stage of prioritizing ecological protection, the λ value is increased to strengthen ecological constraints; in the stage of prioritizing development, the λ value is appropriately reduced to balance development needs.

[0014] Furthermore, the lightweight deployment of the model and real-time data stream processing via edge computing includes the following steps: We employ techniques such as weight pruning (removing redundant neuron connections), quantization (converting 32-bit floating-point parameters to 8-bit integers), and knowledge distillation (simulating the output of a large model with a small model) to reduce model size and computational load. To address the computing power limitations of edge devices, we simplify the number of model layers (e.g., reducing the number of LSTM units) and use grouped convolutions instead of standard convolutions to reduce memory usage and computational complexity. Edge devices supporting AI inference are selected, featuring low power consumption and high computing power, making them suitable for deployment in scenarios such as urban base stations and monitoring terminals. Lightweight models are converted into an executable format for edge devices using optimization tools such as TensorRT and deployed to edge nodes. For example, edge nodes can be deployed in commercial areas to receive mobile phone signaling and social media data streams in real time for localized inference. An end-edge-cloud collaborative architecture is constructed, with edge nodes processing real-time data streams and the cloud handling model updates and complex analysis, ensuring efficient allocation of computing resources. By utilizing stream processing frameworks such as Kafka and Flink, real-time access is provided to multi-source data streams including mobile signaling, social media, and Points of Interest (POIs), with buffers set up to prevent data loss or congestion. At edge nodes, features are extracted from the real-time data streams (e.g., pedestrian density calculation, activity frequency statistics) to generate dynamic feature vectors (e.g., hourly pedestrian heatmaps, activity type distribution), which serve as model input. The edge model performs real-time inference on these dynamic features, outputting land use classification results (e.g., commercial area expansion warnings, ecological land protection recommendations), and feeds these results back to the city management platform or mobile terminals via API interfaces, supporting rapid decision-making (e.g., traffic control, development approval). Based on the real-time data streams and model outputs, model parameters are periodically calibrated (e.g., weight adjustments, threshold optimization), and edge node models are updated through an edge-cloud collaborative mechanism to ensure the model continuously adapts to urban changes.

[0015] Furthermore, the construction of a cross-departmental data sharing platform encompassing natural resources, agriculture, water conservancy, and environmental protection, and the use of a unified indicator system for multi-departmental planning and collaborative optimization, includes the following steps: The platform adopts a distributed microservice architecture to support the integrated storage and real-time access of data from departments such as natural resources, agriculture, water conservancy, and environmental protection. It achieves standardized data access through API interfaces and sets up an access control module to ensure data security. Data from various departments is cleaned, deduplicated, and spatiotemporally aligned. For example, by overlaying and analyzing the land use status map from the natural resources department with the arable land quality grade map from the agriculture department, contiguous arable land areas can be identified; and by overlaying the river system map from the water conservancy department with the ecological red line map from the environmental protection department, priority areas for ecological ditch construction can be determined. The indicators selected are quantifiable, comparable, and implementable, covering multiple dimensions such as ecological protection, farmland protection, and water resource management; for example: Farmland consolidation rate: defined as "contiguous farmland area / total farmland area" × 100%, reflecting the continuity and scale of farmland space (e.g., ≥85% is considered highly concentrated and contiguous, supporting large-scale agricultural development). Ecological ditch coverage rate: defined as "ecological ditch length / total regional ditch length" × 100%, reflecting the level of ecological transformation of farmland drainage system (e.g., ≥30% is considered compliant, reducing non-point source pollution). Ecological red line compliance rate: defined as "development area that does not infringe on the ecological red line / total development area" × 100%, ensuring that development activities do not exceed the bottom line of ecological protection; Quantify indicator values ​​using GIS spatial analysis tools (such as ArcGIS overlay analysis). For example, the contiguous farmland concentration rate is automatically calculated using a "contiguous area identification algorithm"; the ecological ditch coverage rate is calculated using "ditch network extraction + ecological attribute labeling" to determine the ecological proportion. The platform integrates planning data from various departments and uses spatial overlay analysis to detect conflict areas (such as the overlap between development boundaries and farmland protection areas, and conflicts between ecological ditches and flood control and drainage facilities). Based on a unified indicator system, optimization plans for conflict areas are formulated; if the development boundary overlaps with the farmland protection zone, the development needs and farmland protection are balanced by adjusting the development boundary or delineating a farmland protection buffer zone; if ecological ditches conflict with flood control and drainage facilities, functional synergy is achieved by optimizing the ditch route or adding ecological drainage facilities; if the compliance rate of the ecological red line does not meet the standard, the compliance rate is improved by adjusting the site selection of development projects or strengthening ecological restoration measures. The platform supports real-time uploading of planning adjustment data by various departments, and achieves dynamic collaboration through a closed loop of planning-detection-optimization-feedback. For example, after the agricultural department uploads a farmland quality improvement plan, the platform automatically detects its impact on the ecological red line and provides optimization suggestions to the environmental protection department.

[0016] Furthermore, step five includes: Ecological footprint is defined as the ratio of human activity's demand for natural resources to the carrying capacity of the ecosystem.

[0017] in, For the consumption of the i-th type of resource or service, Let i be the global average production of resource i. EF is the equivalence factor of the i-th type of resource (reflecting the relative scarcity of different resource types, such as the conversion coefficient of arable land, forest land, and water area), and EF is the total ecological footprint. Carbon emissions are calculated based on activity data and emission factors:

[0018] in, This refers to data for activity type j (such as industrial electricity consumption and motor vehicle mileage). CE represents the total carbon emissions, where CE is the emission factor for activity type j (e.g., CO2 emission factor per kilowatt-hour, fuel consumption emission factor per kilometer driven). The sustainability of the land use classification results is assessed by comparing the ecological footprint and carbon emissions with preset thresholds or ecological red lines. For example, if a region's ecological footprint exceeds the global sustainable threshold of "hectares per person," or its carbon emissions exceed the dual-carbon target constraints, the classification result is deemed unsustainable. If the ecological footprint is less than the threshold and carbon emissions meet the standards, the classification result is deemed sustainable and can be used as a policy basis (such as for ecological land protection and the designation of low-emission development zones). Based on the assessment results, the land use classification model can be optimized in reverse. For example, if the ecological footprint of an industrial zone is too high, the model weights can be adjusted to reduce the probability of industrial land expansion; if carbon emissions in an ecological zone exceed the standards, ecological protection measures can be strengthened (such as increasing vegetation cover and restricting development intensity). Monthly data streams from mobile phone signaling, social media, and public posts (POs) are collected. Through spatiotemporal alignment and feature extraction, monthly dynamic feature vectors (such as changes in pedestrian flow heatmaps and activity type distribution) are generated. Based on the monthly data, model parameters (such as weights and thresholds) are adjusted using gradient descent or Bayesian optimization algorithms. If mobile phone signaling shows a continuous increase in pedestrian density in a certain area, the weight of commercial land use in that area can be increased. If social media shows an increase in public demand for ecological protection, ecological indicator constraints can be strengthened (such as increasing the weight of NDVI and DEM slope). The calibrated model parameters are fed back to the classification model, forming a closed loop of evaluation-adjustment-optimization to ensure that the classification results continuously adapt to the dynamic changes and sustainability requirements of the city.

[0019] The present invention has the following beneficial effects: This invention employs Poisson disk sampling and graph neural networks to construct a spatiotemporal graph structure that integrates the spatiotemporal heterogeneity of multi-source data, dynamically adjusting the weights of each data source using an attention mechanism. Static spatial features are extracted using a Swing Transformer, and temporal dynamic features are processed by LSTM. Information matching is optimized through cross-modal attention module fusion and Grad-CAM++ visualization. Based on mobile phone signaling-land flow data and social media activity-spatial correlation, LSTM is used to predict land use change trends and embed ecological indicator constraints, coupled with lightweight edge computing deployment and real-time data stream processing. A cross-departmental data sharing platform integrating natural resources, agriculture, water conservancy, and environmental protection is constructed. A unified indicator system enables collaborative optimization of multi-departmental planning, and social media is used to mine public demands and POI human-land relationship data to dynamically adjust model parameters. Ecological footprint and carbon emission monitoring indicators are introduced to conduct sustainability assessments of the classification results. A monthly data update mechanism is established, and model parameters are dynamically calibrated through mobile phone signaling and real-time social media data streams to ensure that the classification results continuously adapt to urban dynamic changes and sustainability requirements, effectively achieving effective synergy between dynamic land use change monitoring and ecological protection. Attached Figure Description

[0020] Figure 1This is a flowchart illustrating the steps of the urban land use classification method based on the fusion of multi-source data and deep learning proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, this invention is an urban land use classification method based on the fusion of multi-source data and deep learning, comprising: Step 1 (Spatiotemporal Alignment and Dynamic Preprocessing of Multi-Source Data): A spatiotemporal graph structure is constructed by using Poisson disk sampling and graph neural network (GNN) to integrate the spatiotemporal heterogeneity of multi-source remote sensing, social media, POI, and mobile signaling data. The weights of each data source are dynamically adjusted by combining an attention mechanism (e.g., strengthening the weights of POI and mobile signaling in commercial areas, and highlighting the weights of remote sensing and DEM in ecological areas). This solves the problems of information loss and classification bias dominated by a single data source in traditional gridding methods. Step 2 (Deep Extraction and Enhancement of Multimodal Features): The Swing Transformer is used to extract static spatial features from remote sensing and DEM, and LSTM is used to process the temporal dynamic features of social media and mobile signaling. Static-dynamic features are fused through a cross-modal attention module, and the matching degree of multi-source data information is optimized based on the Grad-CAM++ visualization model of the region of interest (such as balancing the weights of POI and remote sensing images in industrial land classification). Step 3 (Dynamic Land Use Change Prediction and Ecological Constraints): Combining mobile phone signaling-land flow data with social media activity-spatial correlation, LSTM is used to predict land use change trends (such as the correlation between the surge in mobile phone signaling in commercial areas on weekends and activity hotspots). At the same time, ecological indicators are embedded to ensure the protection of ecological land such as arable land / wetlands, and the model is deployed in a lightweight manner and processed in real time through edge computing. Step 4 (Cross-departmental collaborative verification and policy integration): Construct a cross-departmental data sharing platform for natural resources, agriculture, water conservancy and environmental protection. Through a unified indicator system (such as the rate of concentrated contiguous farmland and the coverage rate of ecological ditches), conduct multi-departmental planning collaborative optimization. Utilize social media to mine public demands and POI human-land relationship data, dynamically adjust classification model parameters, and form a "planning-feedback-optimization" closed loop. Step 5 (Dynamic Updates and Sustainability Assessment): Introduce ecological footprint and carbon emission monitoring indicators to conduct sustainability assessments of the classification results (such as constraining the layout of high-pollution industries in the expansion of industrial land), and establish a monthly data update mechanism. Dynamically calibrate model parameters through real-time data streams from mobile phone signaling and social media to solve the problem that traditional static classification methods cannot adapt to rapid urbanization, and achieve a triple improvement in "accuracy, real-time performance, and sustainability".

[0023] In one embodiment, step one includes: Spatially homogenizing multi-source data (high-resolution pixels from remote sensing images, location check-in points on social media, coordinates of POIs, and trajectory points from mobile phone signaling) is performed. Poisson disk sampling avoids the oversampling in dense areas and undersampling in sparse areas caused by traditional gridded sampling by generating random point sets with minimum distance threshold constraints. For example, in commercial areas where mobile phone signaling points are dense, Poisson disk sampling can ensure that the spacing between sampling points is greater than a set threshold, avoiding redundancy. In ecological areas where data is sparse, the threshold can be dynamically adjusted to ensure coverage of key areas. This provides spatially uniform and non-redundant node inputs for subsequent graph neural networks, solving the problem of uneven spatial sampling in data heterogeneity. Mapping entities in multi-source data to graph nodes, for example: Remote sensing image pixels (including spectral and texture features) → spatial nodes; Social media check-in points (including timestamps and event types) → Spatiotemporal nodes; Points of Interest (POIs) (including category and popularity) → Semantic Nodes; Mobile phone signaling trajectory points (including user ID, time, and location) → dynamic nodes; Edges between nodes are defined based on spatiotemporal correlation, for example: The movement trajectory of the same user within 24 hours → "temporal edges" between dynamic nodes; Social media check-in points and nearby POIs → "semantic-spatial" association edges; Remote sensing image pixels and adjacent pixels → spatial adjacency edges.

[0024] Through the message passing mechanism of GNN, node features (such as remote sensing pixel values ​​and POI heat) are aggregated along the edge to perform "local-global" feature fusion. For example, commercial area nodes aggregate surrounding POIs and mobile signaling data to form comprehensive features of pedestrian flow and commerce; ecological area nodes aggregate remote sensing vegetation index and DEM slope to form ecological features. To avoid analytical bias caused by differences in time scale, the time resolution of different data sources is unified (e.g., aggregating "hourly" check-in data from social media and "minute-level" trajectory data from mobile phone signaling into "daily" data through a time window). Geographic coordinate system transformation and spatial resolution unification ensure that multi-source data can be overlaid and analyzed in space. Semantic features of different data sources are mapped to a unified dimension. For example, POI categories (such as "restaurants" and "retail") are mapped to numerical business popularity indicators, which, together with social media check-in frequency and mobile phone signaling flow density, constitute a comprehensive feature of business activity.

[0025] The attention mechanism assigns dynamic weights to different data sources, with the weights determined by both scenario requirements and data quality. For example, in commercial areas, POI data (reflecting the distribution of commercial facilities) and mobile signaling data (reflecting pedestrian density) have higher weights, while remote sensing data (reflecting building outlines) has lower weights. In ecological areas, remote sensing data (reflecting vegetation cover) and DEM data (reflecting terrain slope) have higher weights, while social media data (reflecting human activities) has lower weights. The weights are updated in real time according to changes in the scenario. For example, when mobile phone signaling data surges in commercial areas on weekends, the weights are automatically increased; when the remote sensing vegetation index in ecological areas declines in winter, the weights are automatically adjusted to reflect ecological pressure.

[0026] In one embodiment, the step of using Swing Transformer to extract static spatial features from remote sensing and DEM, LSTM to process temporal dynamic features from social media and mobile signaling, and a cross-modal attention module to perform static-dynamic feature fusion includes the following steps: The remote sensing imagery and DEM data are unified to the same spatial resolution and normalized to eliminate dimensional differences. The Swin Transformer extracts features layer by layer through a layered window self-attention mechanism. The imagery is divided into non-overlapping local windows, and self-attention is calculated within each window to capture detailed features such as local textures and edges (e.g., building outlines and vegetation cover). Through a "window translation" mechanism, features from adjacent windows are fused to gradually expand the receptive field and capture the global spatial structure (e.g., the overall layout of urban areas and topographic relief features). A static spatial feature vector is generated, containing spectral features of the remote sensing imagery, topographic features of the DEM (e.g., slope and elevation), and spatial structure information, providing a "stable and high-precision" static base for subsequent fusion. Social media data (such as user check-in locations and activity times) and mobile signaling data (such as user movement trajectories and time-of-day distribution) are aggregated into time-series sequences according to the time dimension; LSTM processes time-series data through gating mechanisms (input gate, forget gate, output gate) to capture time dependencies; for example: Short-term dependence: capturing the differences in pedestrian traffic between weekdays and weekends, and changes in movement patterns during morning and evening rush hours; Long-term dependence: Identify seasonal activity patterns (such as a surge in foot traffic in commercial areas during holidays and a decrease in visitors to eco-regions during winter). A dynamic temporal feature vector is generated, containing dynamic information such as pedestrian flow heat distribution and activity frequency changes, reflecting the real-time status and changing trends of urban land use. The static feature vector extracted by the Swing Transformer and the dynamic feature vector extracted by the LSTM are input into the cross-modal attention module. The module calculates the correlation between the two features through a query-key-value mechanism. Query vector (Q): Represents the feature dimension that needs to be strengthened (such as "business activity"); Key vector (K): Represents candidate information in static / dynamic features (such as "building density" in remote sensing images and "human density" in mobile phone signaling); Value vector (V): Represents the actual weight allocation of the integration (such as the contribution of "building density" to "commercial activity"). Attention scores are used to assign weights to different features. For example, in commercial areas, "pedestrian density" (a dynamic feature) and "building density" (a static feature) are highly correlated, so their weights are increased. In ecological areas, "vegetation cover" (a static feature) may have a higher weight. The static and dynamic features are weighted and fused to generate a comprehensive feature vector. This vector contains both static spatial structure information (such as building layout and terrain features) and dynamic trends (such as pedestrian heat maps and activity patterns), and is ultimately used for land use classification (such as commercial, residential, industrial, and ecological land).

[0027] In one embodiment, optimizing the matching degree of multi-source data information based on the region of interest of the Grad-CAM++ visualization model includes the following steps: After training the classification model that combines Swin Transformer (static features) and LSTM (dynamic features), test samples are selected and input into the model. The feature map of the last convolutional layer of the model records the response intensity of each neuron to the input data (such as the "building edge" feature of remote sensing images and the "crowd gathering" feature of mobile phone signaling). Class activation graph:

[0028] in, For the class activation graph of category c (e.g., "commercial land"), Let be the activation value of the k-th feature map. The weight of category c on feature map k; Grad-CAM++ generates more accurate heatmaps by improving the weight calculation method, thus locating the "areas of interest" of the model in the input data. For example, in the commercial land classification, the highlighted areas on the heatmap may correspond to "POI dense areas" and "high-density mobile signaling areas"; in the ecological land classification, the highlighted areas may correspond to "high vegetation coverage areas in remote sensing images" and "high slope areas in DEMs". Weight The calculation introduces a second-order gradient correction term, and the formula is optimized as follows:

[0029] Wherein, γ is an adjustable parameter used to balance the contributions of the first and second order gradients and improve the positioning accuracy of the heatmap; Z is the spatial size normalization factor of the feature map, which is equal to the total number of pixels of the feature map width × height; i and j are the spatial position indices of the feature map, corresponding to the horizontal (width direction) and vertical (height direction) coordinates of the pixel, respectively, with values ​​ranging from 1≤i≤W and 1≤j≤H (W and H are the width and height of the feature map); The activation value of the k-th feature map at position (i,j) reflects the feature response intensity at that spatial location; The prediction score of the model for category c represents the confidence level of the model in believing that the input belongs to category c; By analyzing the "focus bias" of the model on multi-source data through heatmap analysis, for example, if the heatmap shows that the model overemphasizes POI data in the industrial land classification while ignoring the "building form" features of remote sensing images, it indicates that the weight of POI data is too high; if the heatmap shows that the model does not pay attention to the "slope" features of DEM in the ecological land classification, it indicates that the weight of DEM data is insufficient or the feature extraction is inadequate. Based on the heatmap analysis results, dynamically adjust the fusion weights or feature extraction strategies of multi-source data. For example, in industrial land classification, if the weight of POI data is found to be too high, reduce the weight of POI data and increase the weight of remote sensing imagery. In ecological land classification, if the slope features of DEM are not being considered, optimize the DEM feature extraction module (e.g., increase the resolution of slope calculation) or increase the weight of DEM data. Verify the optimization effect by observing changes in the classification accuracy of the adjusted model (e.g., precision and recall) to ensure improved matching of multi-source data information.

[0030] In one embodiment, the method of combining mobile phone signal-to-land flow data with social media activity-spatial correlations to predict land use change trends using LSTM, while embedding ecological indicator constraints, includes the following steps: The system aligns the human-location flow data from mobile phone signaling (such as user movement trajectory, dwell time, and time period distribution) with the activity-spatial correlation data from social media (such as location check-in, activity type, and timestamp). For example, it constructs a human-activity-location spatiotemporal correlation matrix by aggregating geographic coordinate system and time window. Key features such as human flow density, activity frequency, and activity type distribution are extracted to form a dynamic feature vector, which serves as the input sequence for LSTM. LSTM predicts future land use trends by learning time patterns in historical data (such as differences in pedestrian traffic between weekdays and weekends, and seasonal activity changes). For example, a surge in weekend pedestrian traffic in commercial areas, with activity types concentrated in dining / shopping, can predict commercial land expansion trends; high density of mobile phone signaling in industrial areas during fixed weekday periods, with activity types of commuting, can predict industrial land transfer trends. NDVI (Normalized Difference Vegetation Index, reflecting vegetation cover) and DEM slope (reflecting terrain stability) were selected as ecological constraint indicators. A higher NDVI value indicates denser vegetation and a higher priority for ecological protection; a steeper DEM slope indicates steeper terrain and a higher development risk. NDVI and DEM slope were input as additional features into the LSTM, allowing the model to directly consider the impact of ecological indicators during prediction. For example, when predicting commercial area expansion, if the surrounding NDVI value is high (ecologically sensitive area), the model will reduce the probability of expansion. Add an ecological penalty term to the model's loss function:

[0031] in, To predict losses, Let λ represent the ecological constraint loss (such as the decrease in NDVI, the area of ​​the slope exceeding the standard), and λ be the balance coefficient. Gradient descent optimization is used to minimize ecological damage while maintaining classification accuracy in the prediction results. The λ value is dynamically adjusted according to the city's development stage (such as rapid expansion period / ecological protection period). In the stage of prioritizing ecological protection, the λ value is increased to strengthen ecological constraints; in the stage of prioritizing development, the λ value is appropriately reduced to balance development needs.

[0032] In one embodiment, the lightweight deployment of the model and real-time data stream processing via edge computing includes the following steps: We employ techniques such as weight pruning (removing redundant neuron connections), quantization (converting 32-bit floating-point parameters to 8-bit integers), and knowledge distillation (simulating the output of a large model with a small model) to reduce model size and computational load. To address the computing power limitations of edge devices, we simplify the number of model layers (e.g., reducing the number of LSTM units) and use grouped convolutions instead of standard convolutions to reduce memory usage and computational complexity. Edge devices supporting AI inference are selected, featuring low power consumption and high computing power, making them suitable for deployment in scenarios such as urban base stations and monitoring terminals. Lightweight models are converted into an executable format for edge devices using optimization tools such as TensorRT and deployed to edge nodes. For example, edge nodes can be deployed in commercial areas to receive mobile phone signaling and social media data streams in real time for localized inference. An end-edge-cloud collaborative architecture is constructed, with edge nodes processing real-time data streams and the cloud handling model updates and complex analysis, ensuring efficient allocation of computing resources. By utilizing stream processing frameworks such as Kafka and Flink, real-time access is provided to multi-source data streams including mobile signaling, social media, and Points of Interest (POIs), with buffers set up to prevent data loss or congestion. At edge nodes, features are extracted from the real-time data streams (e.g., pedestrian density calculation, activity frequency statistics) to generate dynamic feature vectors (e.g., hourly pedestrian heatmaps, activity type distribution), which serve as model input. The edge model performs real-time inference on these dynamic features, outputting land use classification results (e.g., commercial area expansion warnings, ecological land protection recommendations), and feeds these results back to the city management platform or mobile terminals via API interfaces, supporting rapid decision-making (e.g., traffic control, development approval). Based on the real-time data streams and model outputs, model parameters are periodically calibrated (e.g., weight adjustments, threshold optimization), and edge node models are updated through an edge-cloud collaborative mechanism to ensure the model continuously adapts to urban changes.

[0033] In one embodiment, the construction of a cross-departmental data sharing platform for natural resources, agriculture, water conservancy, and environmental protection, and the use of a unified indicator system for multi-departmental collaborative optimization of planning, includes the following steps: The platform adopts a distributed microservice architecture to support the integrated storage and real-time access of data from departments such as natural resources, agriculture, water conservancy, and environmental protection. It achieves standardized data access through API interfaces and sets up an access control module to ensure data security. Data from various departments is cleaned, deduplicated, and spatiotemporally aligned. For example, by overlaying and analyzing the land use status map from the natural resources department with the arable land quality grade map from the agriculture department, contiguous arable land areas can be identified; and by overlaying the river system map from the water conservancy department with the ecological red line map from the environmental protection department, priority areas for ecological ditch construction can be determined. The indicators selected are quantifiable, comparable, and implementable, covering multiple dimensions such as ecological protection, farmland protection, and water resource management; for example: Farmland consolidation rate: defined as "contiguous farmland area / total farmland area" × 100%, reflecting the continuity and scale of farmland space (e.g., ≥85% is considered highly concentrated and contiguous, supporting large-scale agricultural development). Ecological ditch coverage rate: defined as "ecological ditch length / total regional ditch length" × 100%, reflecting the level of ecological transformation of farmland drainage system (e.g., ≥30% is considered compliant, reducing non-point source pollution). Ecological red line compliance rate: defined as "development area that does not infringe on the ecological red line / total development area" × 100%, ensuring that development activities do not exceed the bottom line of ecological protection; Quantify indicator values ​​using GIS spatial analysis tools (such as ArcGIS overlay analysis). For example, the contiguous farmland concentration rate is automatically calculated using a "contiguous area identification algorithm"; the ecological ditch coverage rate is calculated using "ditch network extraction + ecological attribute labeling" to determine the ecological proportion. The platform integrates planning data from various departments and uses spatial overlay analysis to detect conflict areas (such as the overlap between development boundaries and farmland protection areas, and conflicts between ecological ditches and flood control and drainage facilities). Based on a unified indicator system, optimization plans for conflict areas are formulated; if the development boundary overlaps with the farmland protection zone, the development needs and farmland protection are balanced by adjusting the development boundary or delineating a farmland protection buffer zone; if ecological ditches conflict with flood control and drainage facilities, functional synergy is achieved by optimizing the ditch route or adding ecological drainage facilities; if the compliance rate of the ecological red line does not meet the standard, the compliance rate is improved by adjusting the site selection of development projects or strengthening ecological restoration measures. The platform supports real-time uploading of planning adjustment data by various departments, and achieves dynamic collaboration through a closed loop of planning-detection-optimization-feedback. For example, after the agricultural department uploads a farmland quality improvement plan, the platform automatically detects its impact on the ecological red line and provides optimization suggestions to the environmental protection department.

[0034] In one embodiment, step five includes: Ecological footprint is defined as the ratio of human activity's demand for natural resources to the carrying capacity of the ecosystem.

[0035] in, For the consumption of the i-th type of resource or service, Let i be the global average production of resource i. EF is the equivalence factor of the i-th type of resource (reflecting the relative scarcity of different resource types, such as the conversion coefficient of arable land, forest land, and water area), and EF is the total ecological footprint. Carbon emissions are calculated based on activity data and emission factors:

[0036] in, This refers to data for activity type j (such as industrial electricity consumption and motor vehicle mileage). CE represents the total carbon emissions, where CE is the emission factor for activity type j (e.g., CO2 emission factor per kilowatt-hour, fuel consumption emission factor per kilometer driven). The sustainability of the land use classification results is assessed by comparing the ecological footprint and carbon emissions with preset thresholds or ecological red lines. For example, if a region's ecological footprint exceeds the global sustainable threshold of "hectares per person," or its carbon emissions exceed the dual-carbon target constraints, the classification result is deemed unsustainable. If the ecological footprint is less than the threshold and carbon emissions meet the standards, the classification result is deemed sustainable and can be used as a policy basis (such as for ecological land protection and the designation of low-emission development zones). Based on the assessment results, the land use classification model can be optimized in reverse. For example, if the ecological footprint of an industrial zone is too high, the model weights can be adjusted to reduce the probability of industrial land expansion; if carbon emissions in an ecological zone exceed the standards, ecological protection measures can be strengthened (such as increasing vegetation cover and restricting development intensity). Monthly data streams from mobile phone signaling, social media, and public posts (POs) are collected. Through spatiotemporal alignment and feature extraction, monthly dynamic feature vectors (such as changes in pedestrian flow heatmaps and activity type distribution) are generated. Based on the monthly data, model parameters (such as weights and thresholds) are adjusted using gradient descent or Bayesian optimization algorithms. If mobile phone signaling shows a continuous increase in pedestrian density in a certain area, the weight of commercial land use in that area can be increased. If social media shows an increase in public demand for ecological protection, ecological indicator constraints can be strengthened (such as increasing the weight of NDVI and DEM slope). The calibrated model parameters are fed back to the classification model, forming a closed loop of evaluation-adjustment-optimization to ensure that the classification results continuously adapt to the dynamic changes and sustainability requirements of the city.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for urban land use classification based on the fusion of multi-source data and deep learning, characterized in that, include: Step 1: Construct a spatiotemporal graph structure by sampling from a Poisson disk and using a graph neural network (GNN) to integrate the spatiotemporal heterogeneity of multi-source remote sensing, social media, POI, and mobile signaling data, and dynamically adjust the weights of each data source using an attention mechanism. Step 2: Use Swing Transformer to extract static spatial features from remote sensing and DEM, LSTM to process temporal dynamic features from social media and mobile signaling, perform static-dynamic feature fusion through cross-modal attention module, and optimize the matching degree of multi-source data information based on Grad-CAM++ visualization model of the region of interest. Step 3: Combining mobile phone signal-to-land flow data with social media activity-spatial correlation, LSTM is used to predict land use change trends, while ecological indicators are embedded as constraints, and edge computing is used for lightweight model deployment and real-time data stream processing; Step 4: Construct a cross-departmental data sharing platform for natural resources, agriculture, water conservancy, and environmental protection. Through a unified indicator system, conduct collaborative optimization of multi-departmental planning and utilize social media to mine public demands and POI human-land relationship data, and dynamically adjust classification model parameters. Step 5: Introduce ecological footprint and carbon emission monitoring indicators to conduct sustainability assessments of the classification results, and establish a monthly data update mechanism to dynamically calibrate model parameters through real-time data streams from mobile phone signaling and social media.

2. The urban land use classification method based on the fusion of multi-source data and deep learning according to claim 1, characterized in that, Step one includes: Spatial homogenization is performed on multi-source data using Poisson disk sampling, which generates a random point set constrained by a minimum distance threshold. In commercial areas with dense mobile signaling points, Poisson disk sampling ensures the distance between sampling points exceeds a set threshold. In ecological areas with sparse data, the threshold is dynamically adjusted to cover key areas. This provides node input for subsequent graph neural networks and maps entities in the multi-source data to graph nodes. Remote sensing image pixels → spatial nodes; Social media check-in points → temporal and spatial nodes; POI point → semantic node; Mobile phone signaling trajectory points → dynamic nodes; Edges between nodes are defined based on spatiotemporal correlation: The movement trajectory of the same user within 24 hours → temporal edges between dynamic nodes; Social media check-in points and nearby POIs → semantic-spatial association edges; Remote sensing image pixels and adjacent pixels → spatial adjacency edges; Through the message passing mechanism of GNN, node features are aggregated along the edge to perform local-global feature fusion; commercial area nodes aggregate surrounding POI and mobile signaling data to form comprehensive pedestrian-commercial features; ecological area nodes aggregate remote sensing vegetation index and DEM slope to form ecological features. Unify the temporal resolution of different data sources to avoid analytical bias caused by differences in time scales; enable spatial overlay analysis of multi-source data through geographic coordinate system transformation and spatial resolution unification; map the semantic features of different data sources to a unified dimension; map POI categories to numerical business popularity indicators, which together with social media check-in frequency and mobile phone message flow density constitute a comprehensive feature of business activity. An attention mechanism is used to assign dynamic weights to different data sources, with the weights determined by both scenario requirements and data quality. In commercial areas, POI data and mobile signaling data have higher weights, while remote sensing data has lower weights. In ecological areas, remote sensing data and DEM data have higher weights, while social media data has lower weights. The weights are updated in real time as the scenario changes. When mobile signaling data in commercial areas surges on weekends, the weights are automatically increased. When the remote sensing vegetation index in ecological areas declines in winter, the weights are automatically adjusted.

3. The urban land use classification method based on the fusion of multi-source data and deep learning according to claim 1, characterized in that, The method employs Swing Transformer to extract static spatial features from remote sensing and DEM, LSTM to process temporal dynamic features from social media and mobile signaling, and a cross-modal attention module to fuse static and dynamic features, including the following steps: The remote sensing imagery and DEM data are unified to the same spatial resolution and normalized. The Swin Transformer extracts features layer by layer through a hierarchical window self-attention mechanism, dividing the imagery into non-overlapping local windows, calculating self-attention within each window, and capturing segment features. The features of adjacent windows are fused by window translation, gradually expanding the receptive field and capturing the global spatial structure. A static spatial feature vector is generated, which includes the spectral features of the remote sensing imagery, the topographic features of the DEM, and spatial structure information. Social media data and mobile signaling data are aggregated into time-series sequences based on the time dimension; LSTM processes the time-series data through a gating mechanism to capture time dependencies; dynamic time-series feature vectors are generated, containing dynamic information that reflects the real-time status and changing trends of urban land use; the static feature vectors extracted by the Swin Transformer and the dynamic feature vectors extracted by the LSTM are input into a cross-modal attention module; the module calculates the correlation between the two features through a query-key-value mechanism. Query vector Q: represents the feature dimension that needs to be enhanced; Key vector K: Represents candidate information in static / dynamic features; Value vector V: represents the actual weight allocation in the fusion process; Attention scores are used to assign weights to different features. In commercial areas, pedestrian density and building density are highly correlated, so their weights are increased. In ecological areas, vegetation cover has a higher weight. Based on the weights, static and dynamic features are weighted and fused to generate a comprehensive feature vector that includes both static spatial structure information and dynamic change trends.

4. The urban land use classification method based on the fusion of multi-source data and deep learning according to claim 1, characterized in that, The optimization of multi-source data information matching based on the Grad-CAM++ visualization model's region of interest includes the following steps: After training the classification model that combines Swing Transformer and LSTM, test samples are selected and input into the model; the feature map of the last convolutional layer of the model records the response intensity of each neuron to the input data. Class activation graph: in, For the class activation graph of category c, Let be the activation value of the k-th feature map. The weight of category c on feature map k; Grad-CAM++ improves the weight calculation method, generates heatmaps, and locates the areas of interest of the model in the input data; Weight The calculation introduces a second-order gradient correction term, and the formula is optimized as follows: Where γ is an adjustable parameter; Z is the spatial size normalization factor of the feature map; i and j are the spatial position indices of the feature map, corresponding to the horizontal and vertical coordinates of the pixel, respectively, with values ​​ranging from 1≤i≤W and 1≤j≤H. Let be the activation value of the k-th feature map at position (i,j); The prediction score of the model for category c represents the confidence level of the model in believing that the input belongs to category c; By analyzing the model's focus on multi-source data through heatmap analysis, if the heatmap shows that the model overemphasizes POI data in industrial land classification while ignoring the "building form" features of remote sensing images, then the weight of POI data is too high; if the heatmap shows that the model does not pay attention to the slope features of DEM in ecological land classification, then the weight of DEM data is insufficient or the feature extraction is inadequate. Based on the heatmap analysis results, dynamically adjust the fusion weights or feature extraction strategies of multi-source data; in the industrial land classification, if the weight of POI data is found to be too high, reduce the weight of POI and increase the weight of remote sensing imagery; in the ecological land classification, if the slope features of DEM are not considered, optimize the DEM feature extraction module or increase the weight of DEM data, and verify the optimization effect by the change in the classification accuracy of the adjusted model.

5. The urban land use classification method based on the fusion of multi-source data and deep learning according to claim 1, characterized in that, The method of combining mobile phone signal-to-land flow data with social media activity-spatial correlation, using LSTM to predict land use change trends, and embedding ecological indicator constraints includes the following steps: The system aligns the person-location flow data from mobile phone signaling with the activity-spatial correlation data from social media in a spatiotemporal manner; it then constructs a spatiotemporal correlation matrix of people-activity-location by aggregating geographic coordinate system and time window; and extracts key features to form a dynamic feature vector, which serves as the input sequence for LSTM. LSTM predicts future land use change trends by learning temporal patterns from historical data. NDVI and DEM slope are selected as ecological constraint indicators. Higher NDVI values ​​indicate denser vegetation and a higher priority for ecological protection; steeper DEM slopes indicate steeper terrain and higher development risks. NDVI and DEM slope are input as additional features into the LSTM, allowing the model to directly consider the impact of ecological indicators during prediction. An ecological penalty term is added to the model's loss function. in, To predict losses, The ecological constraint loss is represented by λ, which is the balance coefficient. Through gradient descent optimization, the prediction results are optimized to minimize ecological damage while meeting classification accuracy requirements. The λ value is dynamically adjusted according to the stage of urban development. In the stage of prioritizing ecological protection, the λ value is increased to strengthen ecological constraints; in the stage of prioritizing development, the λ value is appropriately reduced to balance development needs.

6. The urban land use classification method based on the fusion of multi-source data and deep learning according to claim 1, characterized in that, The lightweight deployment of models and real-time data stream processing via edge computing includes the following steps: We employ weight pruning, quantization, and knowledge distillation techniques to reduce model size and computational cost. To address the computing power limitations of edge devices, we simplify the number of model layers and use grouped convolutions instead of standard convolutions to reduce memory usage and computational complexity. Select edge devices that support AI inference, convert lightweight models into an edge-device-executable format using optimization tools, and deploy them to edge nodes; deploy edge nodes in commercial areas to receive mobile phone signaling and social media data streams in real time for localized inference; build an end-edge-cloud collaborative architecture, with edge nodes processing real-time data streams and the cloud responsible for model updates and complex analysis, as well as allocating computing resources. By utilizing streaming processing frameworks such as Kafka and Flink, multi-source data streams are accessed in real time, and buffers are set up to prevent data loss or congestion. Features are extracted from the real-time data streams at edge nodes to generate dynamic feature vectors, which serve as model inputs. The edge model performs real-time inference on the dynamic features, outputs land use classification results, and feeds them back to the city management platform or mobile terminal via API interfaces to support rapid decision-making. Based on the real-time data streams and model outputs, model parameters are calibrated periodically, and edge node models are updated through an edge-cloud collaborative mechanism to ensure that the model continuously adapts to urban changes.

7. The urban land use classification method based on the fusion of multi-source data and deep learning according to claim 1, characterized in that, The construction of a cross-departmental data sharing platform encompassing natural resources, agriculture, water resources, and environmental protection, and the use of a unified indicator system for multi-departmental planning and collaborative optimization, includes the following steps: The platform adopts a distributed microservice architecture to support the integrated storage and real-time access of data from natural resources, agriculture, water resources, and environmental protection departments. It standardizes data access through API interfaces and implements a permission management module to protect data security. Data from each department is cleaned, deduplicated, and spatiotemporally aligned. The platform overlays and analyzes land use maps from the natural resources department with farmland quality maps from the agriculture department to identify contiguous farmland areas. It also overlays river system maps from the water resources department with ecological red line maps from the environmental protection department to determine priority areas for ecological ditch construction. Quantifiable, comparable, and actionable indicators are selected, covering multiple dimensions of ecological protection, farmland protection, and water resource management, and the indicator values ​​are quantified using GIS spatial analysis tools. The platform integrates planning data from various departments and detects conflict areas through spatial overlay analysis. Based on a unified indicator system, optimization plans for conflict areas are formulated; if the development boundary overlaps with the farmland protection zone, the development needs and farmland protection are balanced by adjusting the development boundary or delineating a farmland protection buffer zone; if ecological ditches conflict with flood control and drainage facilities, functional coordination is achieved by optimizing the ditch route or adding ecological drainage facilities; if the compliance rate of the ecological red line does not meet the standards, the compliance rate is improved by adjusting the site selection of development projects or strengthening ecological restoration measures; the platform supports real-time uploading of planning adjustment data by various departments, and achieves dynamic coordination through a closed loop of planning-detection-optimization-feedback.

8. The urban land use classification method based on the fusion of multi-source data and deep learning according to claim 1, characterized in that, Step five includes: Ecological footprint is defined as the ratio of human activity's demand for natural resources to the carrying capacity of the ecosystem. in, For the consumption of the i-th type of resource or service, Let i be the global average production of resource i. Let EF be the equivalence factor for the i-th type of resource, and EF be the total ecological footprint. Carbon emissions are calculated based on activity data and emission factors: in, For the data of the j-th type of activity, The emission factor CE for activity category j is the total carbon emissions; The sustainability of the classification results is assessed by comparing the ecological footprint and carbon emissions with preset thresholds or ecological red lines. If the ecological footprint of a region exceeds the global sustainable threshold per hectare per person, or the carbon emissions exceed the dual carbon target constraints, the classification result is deemed unsustainable. If the ecological footprint is less than the threshold and the carbon emissions meet the standards, the classification result is deemed sustainable and can be used as a policy basis. Based on the assessment results, the land use classification model is optimized in reverse. If the ecological footprint of an industrial zone is too high, the model weights are adjusted to reduce the probability of industrial land expansion. If the carbon emissions of an ecological zone exceed the standards, ecological protection measures are strengthened. Monthly data streams from mobile phone signaling, social media, and public posts are collected. Through spatiotemporal alignment and feature extraction, monthly dynamic feature vectors are generated. Based on the monthly data, model parameters are adjusted using gradient descent or Bayesian optimization algorithms. If mobile phone signaling shows a continuous increase in population density in a certain area, the weight of the area's commercial land classification is increased. If social media shows an increase in public demand for ecological protection, ecological indicator constraints are strengthened. The calibrated model parameters are fed back to the classification model, forming a closed loop of evaluation-adjustment-optimization.

Images