Urban functional area surveying and mapping method based on multi-feature ensemble learning
By employing a multi-feature ensemble learning method and utilizing the MDSA-Former and AdaBoost models to fuse high-resolution remote sensing images and point-of-interest data, the problem of data heterogeneity and sparsity in urban functional area mapping was solved, achieving high-precision and efficient urban functional area mapping.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-13
AI Technical Summary
Existing urban functional area mapping methods are difficult to effectively integrate high-resolution remote sensing images and point of interest (POI) data. They are particularly robust in sparse POI data areas and are difficult to fit the complex nonlinear relationships between multi-source features, resulting in insufficient mapping accuracy and stability.
A multi-feature ensemble learning approach is adopted. By constructing a multimodal dual-stream adaptive fusion Transformer model (MDSA-Former) and an ensemble learning classifier based on the adaptive boosting algorithm (AdaBoost), and combining the gray-level co-occurrence matrix, visible light atmospheric impedance index and word frequency-inverse document frequency model, spectral, texture and socio-semantic features are extracted from high-resolution remote sensing images and point of interest data, and dynamic weighted deep fusion and classification are performed.
It achieves high robustness and high accuracy data fusion in sparse POIs areas, significantly improving the reliability and efficiency of large-scale functional area mapping, enhancing classification accuracy and model generalization ability, and enabling unified processing of complex mixed functional areas with urban core areas and peripheral land cover.
Smart Images
Figure CN121661188A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of urban planning, remote sensing image processing and geographic information systems, and in particular to a method for urban functional zone mapping based on multi-feature integrated learning. Background Technology
[0002] With the acceleration of global urbanization, urban spatial structures are becoming increasingly complex. Efficient and accurate mapping of urban functional areas is fundamental to achieving refined urban management, optimizing resource allocation, and supporting sustainable planning. Traditional urban functional area mapping methods, such as field surveys and manual interpretation, suffer from high subjectivity, high costs, and time-consuming processes, making them unsuitable for large-scale, high-time-sensitivity mapping needs. In recent years, high-resolution (VHR) remote sensing imagery and socially-sensory data such as points of interest (POIs) have become major data sources. However, VHR imagery lacks socio-semantic information, leading to prominent "dissimilar objects" phenomena (e.g., difficulty in distinguishing between schools and office buildings); while POI data suffers from uneven spatial distribution, being particularly sparse in suburban areas, resulting in low mapping accuracy. Therefore, integrating VHR remote sensing imagery and POI data has become the current mainstream trend.
[0003] Existing research can be broadly categorized into several types: one type is based on traditional feature engineering methods. For example, Zhang et al. attempted to extract physical features using spectral and gray-level co-occurrence matrices (GLCM), Liu et al. combined remote sensing imagery and social media data, and Zhong et al. proposed a framework integrating multi-source semantic objects of points, lines, and surfaces. These methods typically rely on traditional machine learning classifiers (such as random forests and support vector machines) for simple feature concatenation and classification. Another type turns to more complex deep learning models to seek better fusion results. For instance, Tao et al. proposed a graph-based fusion framework (G2MF) to capture spatial topological relationships, and Qiao et al. designed a multi-source dynamic fusion network (MSDFN) to incorporate spatial attention mechanisms. Furthermore, some research has begun to explore the applications of ensemble learning, such as Wu et al.'s SATL multi-feature ensemble learning framework.
[0004] Despite previous research efforts in data fusion, several challenges remain: First, existing fusion methods (whether simple stitching or complex networks) struggle to effectively handle the high heterogeneity between VHR imagery and POIs data, especially in sparsely populated suburban areas where robust mechanisms for dynamically adjusting weights are lacking. Second, large-scale, multi-category (e.g., 11 categories) detailed mapping, particularly the unified processing of mixed-use urban cores and peripheral land cover to monitor dynamic changes, places extremely high demands on the model. Finally, complex nonlinear relationships exist among multi-source features, making it difficult for traditional classifiers to fit effectively.
[0005] In summary, existing technologies either employ simple feature stitching, which is insufficient to handle the high heterogeneity between VHR imagery and POIs data, or rely on complex networks, which fail to effectively address the robustness issue of POIs data in sparse suburban areas. Furthermore, they generally struggle to fit the complex nonlinear relationships between multi-source features, resulting in insufficient overall accuracy and stability.
[0006] Therefore, there is an urgent need for an automated method that can efficiently integrate heterogeneous data, overcome data sparsity and nonlinearity problems, and unify the mapping of urban and rural areas, so as to overcome the limitations of existing methods and improve mapping efficiency and accuracy. Summary of the Invention
[0007] The purpose of this invention is to provide a method for urban functional area mapping based on multi-feature ensemble learning, which solves the problems of existing methods mentioned in the background art in terms of integrating heterogeneous data, handling sparse data areas, and classification accuracy.
[0008] To achieve the above objectives, this invention provides a method for urban functional area mapping based on multi-feature ensemble learning, comprising the following steps: S1. Construct the multimodal dual-stream adaptive fusion Transformer model MDSA-Former; S2. Based on the Open Street Map (OSM) road network data and combined with building footprint data, delineate basic spatial units; S3. Using the visible light atmospheric impedance index VARI, the gray-level co-occurrence matrix GLCM model and the term frequency-inverse document frequency TF-IDF model, spectral features, texture features and socio-semantic features are extracted from VHR remote sensing images and point of interest (POI) data, respectively, and TSE feature sets are constructed. S4. Input the TSE feature set into MDSA-Former to generate the fused TSE' feature set; S5. The AdaBoost ensemble learning classifier based on the adaptive boosting algorithm is used to classify urban functional area units according to the TSE' feature set and obtain the classification results. S6. Based on the classification results, generate a map of urban functional areas and complete the surveying and mapping of urban functional areas.
[0009] Preferably, the MDSA-Former in S1 adopts a dual-branch Transformer encoder architecture.
[0010] Preferably, the specific steps of S2 are as follows: S21. Perform topology correction on OSM road network data and connect suspended road sections; S22. Generate road buffer zones according to road network levels; S23. Remove the buffer road area from the entire study area and extract the urban functional area plots; S24. Use building footprint data as a supplement to enhance the spatial integrity of sparse road network areas in the suburbs.
[0011] Preferably, the specific steps of S3 are as follows: S31. Spatial texture features are extracted from VHR remote sensing images using the Gray-Level Co-occurrence Matrix (GLCM) model. S32. Use the visible light atmospheric impedance index (VARI) to extract spectral features from VHR remote sensing images; S33. Extract social semantic features from POIs data.
[0012] Preferably, the specific steps of S33 are as follows: S331. Preprocess the Points of Interest (POIs) data, including reclassification and cleaning. S332. Apply the term frequency-inverse document frequency (TF-IDF) method to extract socio-semantic features from the attribute information of the preprocessed POIs data.
[0013] Preferably, the specific steps of S4 are as follows: S41. Use MDSA-Former to process the features of the two modalities, which include remote sensing features and socio-semantic features. S42. Perform operations on all plot samples in a batch through a multi-head self-attention mechanism to capture cross-plot context information; S43. Introduce an attention-based dynamic weighted fusion strategy to predict fusion weights for features from the two modalities; S44. Based on the learned weights, the features of the two modalities are weighted and connected to form a unified TSE' feature representation.
[0014] Preferably, the remote sensing features in S41 include texture and spectral features.
[0015] Preferably, the specific steps of S5 are as follows: S51. Input the fused TSE' feature set and the corresponding functional area labels into the ensemble learning classifier AdaBoost for model training. S52. The trained model is used to classify all spatial units within the study area to generate the final urban functional area map.
[0016] Therefore, the above-mentioned urban functional area mapping method based on multi-feature ensemble learning, adopted in this invention, has the following beneficial effects: (1) This invention constructs a TSE feature set, uses a multimodal dual-stream adaptive fusion Transformer model (MDSA-Former fusion network) for dynamic weighted deep fusion, and combines an ensemble learning classifier based on an adaptive boosting algorithm, which can achieve high robustness and high accuracy in heterogeneous data fusion and classification in challenging scenarios where POIs are sparse. Compared with traditional fusion methods, it effectively overcomes the limitations of existing methods and significantly improves the reliability, accuracy and efficiency of large-scale functional area mapping.
[0017] (2) This invention comprehensively utilizes the gray-level co-occurrence matrix (GLCM), visible light atmospheric impedance index (VARI), and term frequency-inverse document frequency (TF-IDF) model to extract texture, spectral, and socio-semantic features from VHR remote sensing images and points of interest (POIs) data, respectively, and constructs a multi-dimensional TSE feature set of "texture-spectrum-semantic".
[0018] (3) The present invention adopts an ensemble learning classifier AdaBoost based on an adaptive boosting algorithm. By iteratively adjusting the sample weights, the model training automatically focuses on difficult samples with high feature confusion and difficulty in differentiation (such as mixed functional areas). Compared with a single classifier, it significantly improves the classification accuracy, effectively reduces the risk of overfitting, and enhances the generalization ability of the model.
[0019] (4) The present invention can uniformly handle the complex mixed functions of the urban core area and the land cover (such as cultivated land, forest and grassland) on the outskirts of the city, providing reliable support for large-scale, multi-category (such as 11 categories) fine mapping and dynamic monitoring.
[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating an embodiment of an urban functional area mapping method based on multi-feature ensemble learning according to the present invention. Figure 2 This is a schematic diagram illustrating the urban functional area unit construction process of an embodiment of the urban functional area mapping method based on multi-feature ensemble learning according to the present invention. Figure 3 This is a schematic diagram illustrating the TSE feature set extraction process in an embodiment of the urban functional area mapping method based on multi-feature ensemble learning according to the present invention. Figure 4 This is a schematic diagram of the overall architecture of the MDSA-Former feature fusion network, which is an embodiment of the urban functional area mapping method based on multi-feature ensemble learning according to the present invention. Figure 5This is a schematic diagram of the ensemble learning model of an embodiment of the urban functional area mapping method based on multi-feature ensemble learning according to the present invention. Detailed Implementation
[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0024] Example Please see Figures 1-5 This invention provides a method for urban functional area mapping based on multi-feature ensemble learning, comprising the following steps: like Figure 1 The diagram shown is a flowchart illustrating an embodiment of the urban functional area mapping method based on multi-feature ensemble learning according to the present invention. The specific steps are as follows: Step S1: Delineate basic spatial units based on Open Street Map (OSM) road network data, supplemented by building footprint data.
[0025] Land parcels are the basic units of urban planning; this step aims to generate fundamental spatial units for subsequent analysis. The specific steps are as follows: Step S11: Based on the Open Street Map (OSM) road network data, perform topological corrections on it, such as connecting “suspended” road segments through trimming and extension operations to ensure the continuity of the road network.
[0026] Step S12: Based on the road category in the OSM data, divide it into three levels of road buffer zones and generate buffer zones of corresponding widths to establish the road space range; wherein, the buffer distance of the first-level road buffer zone is 20m, the buffer distance of the second-level road buffer zone is 10m, and the buffer distance of the third-level road buffer zone is 5m.
[0027] Step S13: By removing the generated road buffer zones from the entire study area, the non-road areas segmented by the road network are extracted, i.e., the preliminary urban functional area plots.
[0028] Step S14: Considering that OSM road network data in urban suburbs or fringe areas may be sparse, building footprint data is introduced as a supplement to improve the integrity of spatial units. Building edge information is integrated into the road network to further subdivide large plots, especially distinguishing large natural plots from small-scale human activity areas.
[0029] Step S2: Extract spectral features, texture features, and socio-semantic features from high-resolution (VHR) remote sensing images and point of interest (POI) data to construct a TSE feature set.
[0030] This step extracts three types of features from VHR remote sensing imagery and POIs data: spectral features, texture features, and socio-semantic features. The specific steps are as follows: Step S21: For VHR imagery (such as Google Earth imagery containing only red, green, and blue bands), the Visible Atmospheric Impedance Index (VARI) is used as a vegetation index to enhance the differentiation of functions such as green space, forest, and grassland. The formula for calculating VARI is: ; Green, Red, and Blue represent the reflectance values of the green, red, and blue bands, respectively.
[0031] Step S22: VHR imagery possesses rich surface texture and spatial structure information. This embodiment employs the Gray-Level Co-occurrence Matrix (GLCM) model to calculate typical texture features, including mean, contrast, entropy, correlation, and dissimilarity, to characterize the physical texture properties of the land parcels.
[0032] Step S23: Extract the socio-economic functions of urban space reflected in the POIs data, i.e., social semantic features.
[0033] Step S231 involves preprocessing the acquired POIs data, including reclassifying POIs according to application requirements (e.g., merging detailed catering categories into "commercial services"), and cleaning and removing POIs such as ATMs and parking lots that cannot independently reflect the function of the land parcel.
[0034] Step S232: The Term Frequency-Inverse Document Frequency (TF-IDF) method is used to quantify the importance of POI categories in each functional area unit and extract socio-semantic features. The formula for calculating TF-IDF is: ; in, For the first plot of land The number of POIs This represents the total number of POIs within the land parcel. This represents the total number of functional area units. For including the first The number of units in the POI class.
[0035] By constructing a TSE feature set that includes spectral, textural, and socio-semantic information, the feature extraction process is not only convenient and fast, but it can also take into account both physical features and socio-semantic information, thus effectively solving the problems of insufficient information from a single data source and difficulty in distinguishing between "different objects with the same spectrum".
[0036] Step S3: Input the TSE feature set into the multimodal dual-stream adaptive fusion Transformer model (MDSA-Former) to perform feature fusion and generate a fused TSE' feature set.
[0037] To address the high heterogeneity between VHR image features and POIs features, and to prepare fused features for subsequent classification, the specific steps are as follows: Step S31: To ensure no data leakage during the training process, the TSE feature set constructed in step S2 and its corresponding manually labeled functional area labels (e.g., public administration, commercial services, residential areas, industry, green space, farmland, etc.) are pre-divided into TSE training set, TSE validation set and TSE test set.
[0038] Step S32 involves using a multimodal dual-stream adaptive fusion Transformer (MDSA-Former) network for feature fusion. MDSA-Former employs a dual-branch Transformer encoder architecture, with one branch processing remote sensing features (a combination of spectral and texture features) and the other branch processing socio-semantic features.
[0039] Step S33: The core of this network lies in its multi-head self-attention mechanism. During training, each branch treats all plot samples in a batch as a sequence. The self-attention mechanism calculates the relationships between all samples (plots) in this sequence, thereby generating a feature representation for each plot that incorporates "cross-plot" contextual information, making it more discriminative.
[0040] In step S34, after context aggregation, the strategy uses a lightweight multilayer perceptron (MLP) and a sigmoid activation function to predict a fusion weight for each parcel's remote sensing features and socio-semantic features.
[0041] Step S35: Train and validate the MDSA-Former network (as a supervised model) on the TSE training set and TSE validation set partitioned in step S31 to determine the optimal model weights.
[0042] Step S36: Using the MDSA-Former model trained on the TSE training set, forward propagation (i.e., feature fusion) is performed on the TSE training set, TSE validation set, and TSE test set respectively to generate the corresponding TSE' training set, TSE' validation set, and TSE' test set. This mechanism enables the model to adaptively balance the contributions of the two modalities: in regions rich in POIs, the model may rely more on socio-semantic features; while in regions sparse in POIs, the model automatically increases the weight of remote sensing features.
[0043] By employing the MDSA-Former network for feature fusion, deep interaction and adaptive weighting of heterogeneous features are achieved, effectively solving the problem of uneven spatial distribution of POIs data (such as sparse distribution in suburban areas), and significantly enhancing the discriminative power of features and the robustness of the model.
[0044] Step S4: Use an ensemble learning classifier to classify urban functional area units based on the TSE' feature set to obtain classification results.
[0045] This step uses an ensemble learning classifier to classify the TSE' feature set generated in step S3. The specific steps are as follows: Step S41: In this embodiment, the AdaBoost ensemble learning model is used as the classifier. It is an ensemble learning algorithm based on Adaptive Boosting. AdaBoost iteratively trains a series of weak learners (usually decision trees), increasing the weight of samples misclassified in the previous round in each round, making subsequent learners focus more on these "difficult cases," and finally combining all weak learners in a weighted manner, resulting in robust performance when processing complex data.
[0046] Step S42: Train the AdaBoost classifier using the TSE' training set generated in step S36. In this embodiment, the hyperparameters of AdaBoost can be set as follows: the base learner (base_estimator) can be a decision stump, the number of iterations (n_estimators) can be set to 100, and the learning rate (learning_rate) can be set to 1.0 or tuned based on the validation set.
[0047] Step S43: Apply the trained AdaBoost model to predict the TSE' test set (or the TSE' features of all test spatial units within the study area) to obtain the functional area category of each unit as the classification result.
[0048] By employing the AdaBoost ensemble learning strategy, the model can adaptively increase its attention to complex functional area samples that are difficult to distinguish, thereby significantly reducing the misclassification rate and improving the model's generalization ability and final mapping accuracy in multi-class classification tasks.
[0049] Step S5: Based on the classification results, generate a map of urban functional areas to complete the urban functional area surveying.
[0050] This step generates the final urban functional area map based on the classification results obtained in step S4. The specific steps are as follows: Step S51: Associate the functional zone categories (i.e., classification results) of each spatial unit obtained in step S43 with the geospatial information (such as plot boundaries) of the urban functional zone units constructed in step S1.
[0051] Step S52: Using Geographic Information System (GIS) software or visualization tools, the associated data is colored and rendered (e.g., red is given to "commercial area" and green is given to "green space") to generate the final urban functional area map and realize the visualization of the surveying results.
[0052] Therefore, the present invention adopts the above-mentioned urban functional area mapping method based on multi-feature ensemble learning, which solves the problems of strong heterogeneity and difficulty in deep fusion between VHR imagery and POIs data in existing urban functional area mapping methods, as well as the problems of poor robustness and low classification accuracy in sparse areas of POIs data. It efficiently realizes adaptive fusion and nonlinear classification of multi-source heterogeneous features, and provides a high-precision and high-robust technical solution for automated mapping of large-scale and multi-category urban functional areas.
[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for urban functional zone mapping based on multi-feature ensemble learning, characterized in that, Includes the following steps: S1. Construct the multimodal dual-stream adaptive fusion Transformer model MDSA-Former; S2. Based on the Open Street Map (OSM) road network data and combined with building footprint data, delineate basic spatial units; S3. Using the visible light atmospheric impedance index VARI, the gray-level co-occurrence matrix GLCM model and the term frequency-inverse document frequency TF-IDF model, spectral features, texture features and socio-semantic features are extracted from VHR remote sensing images and point of interest (POI) data, respectively, and TSE feature sets are constructed. S4. Input the TSE feature set into MDSA-Former to generate the fused TSE' feature set; S5. The AdaBoost ensemble learning classifier based on the adaptive boosting algorithm is used to classify urban functional area units according to the TSE' feature set and obtain the classification results. S6. Based on the classification results, generate a map of urban functional areas and complete the surveying and mapping of urban functional areas.
2. The urban functional area mapping method based on multi-feature ensemble learning according to claim 1, characterized in that: The MDSA-Former in S1 adopts a dual-branch Transformer encoder architecture.
3. The urban functional area mapping method based on multi-feature ensemble learning according to claim 1, characterized in that, The specific steps of S2 are as follows: S21. Perform topology correction on OSM road network data and connect suspended road sections; S22. Generate road buffer zones according to road network levels; S23. Remove the buffer road area from the entire study area and extract the urban functional area plots; S24. Use building footprint data as a supplement to enhance the spatial integrity of sparse road network areas in the suburbs.
4. The urban functional area mapping method based on multi-feature ensemble learning according to claim 1, characterized in that, The specific steps of S3 are as follows: S31. Spatial texture features are extracted from VHR remote sensing images using the Gray-Level Co-occurrence Matrix (GLCM) model. S32. Use the visible light atmospheric impedance index (VARI) to extract spectral features from VHR remote sensing images; S33. Extract social semantic features from POIs data.
5. The urban functional area mapping method based on multi-feature ensemble learning according to claim 4, characterized in that, The specific steps of S33 are as follows: S331. Preprocess the Points of Interest (POIs) data, including reclassification and cleaning. S332. Apply the term frequency-inverse document frequency (TF-IDF) method to extract socio-semantic features from the attribute information of the preprocessed POIs data.
6. The urban functional area mapping method based on multi-feature ensemble learning according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Use MDSA-Former to process the features of the two modalities, which include remote sensing features and socio-semantic features. S42. Perform operations on all plot samples in a batch through a multi-head self-attention mechanism to capture cross-plot context information; S43. Introduce an attention-based dynamic weighted fusion strategy to predict fusion weights for features from the two modalities; S44. Based on the learned weights, the features of the two modalities are weighted and connected to form a unified TSE' feature representation.
7. The urban functional area mapping method based on multi-feature ensemble learning according to claim 6, characterized in that: The remote sensing features in S41 include texture and spectral features.
8. The urban functional area mapping method based on multi-feature ensemble learning according to claim 1, characterized in that, The specific steps of S5 are as follows: S51. Input the fused TSE' feature set and the corresponding functional area labels into the ensemble learning classifier AdaBoost for model training. S52. The trained model is used to classify all spatial units within the study area to generate the final urban functional area map.