Terrain similarity analysis method and system fusing multi-source and multi-scale terrain information
By fusing DEM and remote sensing image data using an improved DeepLab v3+ network, joint feature representations are generated and multi-scale weighted fusion is performed. This solves the misjudgment and integration problems in traditional terrain similarity analysis and achieves high-accuracy and stable terrain similarity analysis results.
Patent Information
- Application Number
- CN202610740392.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-27
- Publication Date
- 2026-08-25
AI Technical Summary
Traditional terrain similarity analysis methods are difficult to effectively integrate multi-source data and lack explicit modeling of terrain structure, leading to misjudgment and scale sensitivity issues in complex landform areas. Furthermore, deep learning models are difficult to integrate with GIS platforms.
An improved DeepLab v3+ network is used to fuse DEM data and remote sensing image data through a terrain structure coding module to generate a joint feature representation. Similarity is calculated at different spatial scales, weights are dynamically assigned for weighted fusion, and finally a raster weight map is output.
It significantly improves the accuracy and stability of terrain similarity analysis, solves the problem of misjudgment under complex landforms, and enables the results to be directly integrated into the GIS platform, thereby improving the efficiency of practical engineering applications.
Smart Images
Figure CN122637142A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of terrain analysis technology, and more specifically to a terrain similarity analysis method and system that integrates multi-source, multi-scale terrain information. Background Technology
[0002] Terrain similarity analysis is a fundamental technology in Geographic Information Systems (GIS), widely used in scenarios such as route planning, regional site selection, and geomorphological comparison analysis. Traditional terrain similarity analysis methods typically rely on terrain parameters such as slope, aspect, and curvature extracted from Digital Elevation Models (DEMs), employing statistical analysis or rule-based matching for judgment. However, these methods have the following significant limitations:
[0003] 1. This type of method relies on manually designed parameters, which are mostly based on local differential geometry calculations. While these parameters can reflect basic attributes such as surface symmetry and slope, they are difficult to effectively characterize the comprehensive features of the terrain in terms of spatial structure, morphological distribution, and multi-scale levels. Especially in complex landforms (regions where mountains, hills, and valleys intertwine), where terrain features are highly nonlinear and spatially heterogeneous, these methods are prone to losing key structural information, leading to misjudgments of similarity.
[0004] 2. This type of method performs analysis at a single scale or simply superimposes parameters at multiple scales. It lacks an explicit modeling mechanism for the consistency of terrain features across different scales, which can easily lead to scale sensitivity issues. For example, terrains that are similar at a macro scale may be misjudged as dissimilar at a micro scale due to differences in local undulations.
[0005] 3. Most of these methods rely solely on DEM data and fail to effectively integrate semantic information such as land cover type, land use, and vegetation cover provided by multi-source data such as remote sensing images. This can lead to misjudgments in practice, such as "semantically similar but with different terrain" or "with similar terrain but with different semantics," thus reducing the accuracy of the analysis results.
[0006] In recent years, deep learning technology has been widely used in the field of remote sensing image interpretation, especially semantic segmentation models represented by DeepLab v3+, which have made significant progress in remote sensing image interpretation. However, the existing DeepLab v3+ network design is mainly geared towards pixel-level land cover classification tasks, lacking the ability to explicitly model terrain structure features, and does not consider the cross-regional feature matching and multi-scale consistency constraints required for terrain similarity analysis. Therefore, it cannot be directly applied to terrain similarity analysis tasks.
[0007] Furthermore, existing deep learning-based methods mostly output probability maps or feature vectors, which are difficult to integrate effectively with traditional GIS platforms, thus limiting their application value in practical engineering.
[0008] Therefore, how to provide an intelligent solution that can deeply integrate multi-source data, explicitly model terrain structure, achieve cross-scale consistency analysis, and effectively integrate with existing GIS platforms is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0009] In view of the above problems, the present invention proposes a terrain similarity analysis method and system that integrates multi-source and multi-scale terrain information, so as to overcome the above problems or at least partially solve the above problems.
[0010] To achieve the above objectives, the present invention adopts the following technical solution:
[0011] In a first aspect, the present invention provides a terrain similarity analysis method that integrates multi-source, multi-scale terrain information, comprising the following steps:
[0012] Acquire remote sensing image data and corresponding DEM data of the terrain area to be analyzed and the target terrain area, and perform preprocessing;
[0013] The terrain structure parameters are calculated based on the preprocessed DEM data, and the terrain structure parameters are stacked according to the channel dimension to generate a terrain structure feature tensor.
[0014] Based on the improved DeepLab v3+ network, feature extraction and fusion are performed on the terrain structure feature tensor and the preprocessed remote sensing image data to generate a joint feature representation.
[0015] Simultaneously extract joint feature representations of feature maps at different spatial scales, and calculate the similarity between each sub-region of the terrain region to be analyzed and the target terrain region at each spatial scale;
[0016] The confidence weights are dynamically assigned based on the degree of difference between similarity results at each scale, and the similarity results at each scale are weighted and fused to obtain the final similarity result.
[0017] The final similarity results are converted into a raster weighted map and output.
[0018] Furthermore, the preprocessing of remote sensing image data and corresponding DEM data includes:
[0019] The remote sensing image data and the corresponding DEM data are subjected to unified spatial reference coordinate transformation, spatial registration and cropping. Resolution resampling is used to unify the resolution of the remote sensing image data and DEM data, so that each pixel is aligned in spatial position.
[0020] After aligning each pixel in spatial position, radiometric and atmospheric corrections are performed on the remote sensing image data, and hole filling and smoothing are performed on the DEM data.
[0021] Furthermore, the topographic structure parameters calculated based on the preprocessed DEM data include at least: slope, aspect, curvature, and topographic potential index, wherein the aspect parameter is decomposed into sine and cosine components.
[0022] Furthermore, after stacking the terrain structure parameters of each region according to the channel dimension, the terrain structure parameters of each channel are normalized and the numerical range is mapped to the [0,1] interval to generate the final terrain structure feature tensor.
[0023] Furthermore, the steps for generating joint feature representations based on the improved DeepLab v3+ network include:
[0024] A terrain structure coding module is introduced between the backbone network and the hollow spatial pyramid pooling module of the original DeepLab v3+ network; the backbone network is used to extract features from remote sensing image data and generate semantic features.
[0025] The terrain structure encoding module is used to encode the terrain structure feature tensor to obtain terrain structure features that can express the spatial morphological relationships of the terrain.
[0026] Semantic features and terrain structure features are concatenated along the channel dimension and then fused and dimensionality reduced using a 1×1 convolutional layer to generate a joint feature representation.
[0027] Furthermore, the terrain structure encoding module contains a three-layer convolutional structure. The first layer of the convolutional structure extracts local spatial features from the terrain structure feature tensor using a 3×3 convolutional kernel, learns the spatial relationship between each individual terrain structure parameter, and obtains a local spatial structure representation.
[0028] The second convolutional structure extracts the nonlinear coupling relationship between different terrain structure parameters in the local spatial structure feature representation through a 3×3 convolutional kernel, thus obtaining multi-parameter coupling features.
[0029] The third convolutional layer uses a 1×1 convolutional kernel to keep the multi-parameter coupled features and semantic features consistent in the channel dimension.
[0030] Furthermore, the similarity calculation process between any sub-region within the topographic region to be analyzed and the target topographic region at each spatial scale includes:
[0031] The joint feature representations of any sub-region A and the target region B within the terrain region to be analyzed are input into the hollow spatial pyramid pooling module. Parallel convolutional layers with different void ratios are used to simultaneously extract feature maps at micro, meso, and macro scales, resulting in a multi-scale feature set. and ,in, These are the microscale features of region A. This represents the mesoscale characteristics of region A. For the macro-scale features of region A, These are the microscale features of region B. This represents the mesoscale characteristics of region B. For the macro-scale features of region B;
[0032] Calculate each scale Cosine similarity between region A and region B :
[0033]
[0034] in, .
[0035] Furthermore, the final similarity result is calculated as follows:
[0036] Calculate the absolute difference between the similarity results at each pairwise scale to form a difference vector. : ;
[0037] Based on the degree of consistency between the similarity calculation results of one scale and those of other scales, an initial confidence weight is assigned to the similarity calculation result of that scale. The initial confidence weights for the similarity calculation results of the three scales are... , , The calculation formulas are as follows:
[0038]
[0039]
[0040]
[0041] in, It is a very small constant;
[0042] The initial confidence weights are normalized to obtain the final fusion weights: ;
[0043] The similarity calculation results at each scale are weighted and fused based on the final fusion weight to obtain the final similarity result. :
[0044]
[0045] Set threshold ,like If the multi-scale results are considered to be highly consistent, then the output will be directly determined. Otherwise, there is scale uncertainty in the similarity judgment of the two terrain regions being compared.
[0046] Furthermore, the process of converting the final similarity results into a raster weighted map includes:
[0047] The terrain region to be analyzed is divided into multiple sub-regions in advance. Each sub-region is compared with the target terrain region one by one to obtain multiple similarity results. The multiple similarity results are summarized into a two-dimensional matrix. The element values in the two-dimensional matrix represent the final similarity results between the corresponding region and the target terrain region.
[0048] The final similarity results are linearly normalized to transform their value range to the [0,1] interval;
[0049] The normalized similarity results are mapped to the grayscale range of [0, 255] to generate a raster weight map. In the raster weight map, the higher the grayscale value, the more similar the terrain.
[0050] Secondly, the present invention provides a terrain similarity analysis system that integrates multi-source, multi-scale terrain information, which employs the method described above, including:
[0051] The data acquisition and preprocessing module is used to acquire remote sensing image data and corresponding DEM data of the terrain area to be analyzed and the target terrain area, and to perform preprocessing.
[0052] The terrain structure analysis module is used to calculate terrain structure parameters based on preprocessed DEM data and stack the terrain structure parameters according to the channel dimension to generate a terrain structure feature tensor.
[0053] The fusion module is used to extract and fuse features from the terrain structure feature tensor and preprocessed remote sensing image data based on the improved DeepLab v3+ network to generate a joint feature representation.
[0054] The similarity calculation module is used to simultaneously extract feature maps of joint feature representation at different spatial scales, and calculate the similarity between the terrain region to be analyzed and the target terrain region at each spatial scale.
[0055] The similarity consistency constraint module dynamically allocates confidence weights based on the degree of difference between similarity results at each scale, and performs weighted fusion of the similarity results at each scale to obtain the final similarity result.
[0056] The output module is used to convert the final similarity results into a raster weighted map and output it.
[0057] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects:
[0058] 1. This invention is the first to construct a parallel multi-source feature extraction path through an improved DeepLab v3+ network, realizing the explicit fusion of remote sensing image semantic information and DEM terrain structure parameters at the depth feature level. This enables the model to simultaneously perceive the surface cover type and the terrain morphology that supports that type, solving the misjudgment problem faced by traditional single data source methods in complex scenarios where semantics are similar but terrains are different or terrains are similar but semantics are different. This significantly improves the accuracy of terrain similarity analysis in real complex environments.
[0059] 2. This invention explicitly models the scale dependence of terrain features through a multi-scale feature extraction and consistency fusion mechanism, thereby enhancing the stability and interpretability of the model in cross-scale analysis.
[0060] 3. This invention is based on dynamic weight allocation with inter-scale consistency, which enables the model to effectively suppress local noise and single-scale misjudgment, thereby improving the robustness of terrain similarity analysis.
[0061] 4. The terrain similarity analysis results output by this invention can be converted into GIS standard data formats such as rasterized weighted maps, which are not only more intuitive, but also easier to integrate with existing GIS platforms, thereby improving the application efficiency of practical engineering. Attached Figure Description
[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0063] Figure 1 This is a flowchart of the terrain similarity analysis method that integrates multi-source and multi-scale terrain information provided in this embodiment of the invention;
[0064] Figure 2 This is an architecture diagram of the improved DeepLab v3+ network provided in an embodiment of the present invention;
[0065] Figure 3 This is a schematic diagram of the terrain structure encoding module provided in an embodiment of the present invention;
[0066] Figure 4 This is a flowchart of multi-scale feature extraction and similarity calculation provided in an embodiment of the present invention;
[0067] Figure 5 This is a schematic diagram of the standardized output and integrated analysis process of the final similarity results provided in the embodiments of the present invention. Detailed Implementation
[0068] 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.
[0069] This invention discloses a terrain similarity analysis method that integrates multi-source, multi-scale terrain information, comprising the following steps:
[0070] S1. Acquire remote sensing image data and corresponding DEM data of the terrain area to be analyzed and the target terrain area, and perform preprocessing;
[0071] S2. Calculate terrain structure parameters based on the preprocessed DEM data, and stack the terrain structure parameters according to the channel dimension to generate a terrain structure feature tensor.
[0072] S3. Based on the improved DeepLab v3+ network, feature extraction and fusion are performed on the terrain structure feature tensor and the preprocessed remote sensing image data to generate a joint feature representation.
[0073] S4. Simultaneously extract joint feature representations of feature maps at different spatial scales, and calculate the similarity between the terrain region to be analyzed and the target terrain region at each spatial scale.
[0074] S5. Based on the degree of difference between similarity results at each scale, dynamically allocate confidence weights and perform weighted fusion on the similarity results at each scale to obtain the final similarity result.
[0075] S6. Convert the final similarity results into a raster weighted map and output it.
[0076] The following provides a further explanation of the specific implementation process for each of the above steps. The entire process is as follows: Figure 1 As shown.
[0077] S1. Data Acquisition and Preprocessing:
[0078] This embodiment selects a typical mountainous area as the analysis object. The terrain of this area is undulating and includes various terrain forms such as ridges, valleys, slopes and gentle hills, which is suitable for verifying the similarity analysis capability under complex terrain conditions.
[0079] Acquire remote sensing image data and corresponding DEM data of the terrain area to be analyzed. The remote sensing image is a high-resolution multispectral image, and the DEM data is in regular raster format. Its spatial resolution is consistent with that of the remote sensing image after resampling.
[0080] The remote sensing image data and the corresponding DEM data undergo unified spatial reference coordinate transformation, spatial registration, and cropping. Resolution resampling is then used to unify the resolution of the remote sensing image data and DEM data, aligning each pixel in spatial location to form a spatially sized dataset. A pair of dual-source inputs;
[0081] After aligning each pixel in spatial position, radiometric and atmospheric corrections are performed on the remote sensing image data to eliminate the influence of differences in illumination conditions on image feature extraction. Hole filling and smoothing are performed on the DEM data to ensure the stability of terrain parameter calculation.
[0082] S2. Calculation of terrain structure parameters and characteristic tensor:
[0083] 1) Calculate a set of topographic structure parameters based on the preprocessed DEM data, including at least: slope (S), aspect (A), curvature (P) and topographic potential index (TPI).
[0084] The slope (S) can be expressed as:
[0085]
[0086] The aspect (A) can be expressed as:
[0087]
[0088] To eliminate discontinuities around the slope aspect, the slope aspect parameter is decomposed into its sinusoidal component sin(A) and cosine component cos(A).
[0089] Curvature (P): A weighted combination of profile curvature and planar curvature is used to reflect the concave and convex shape of the terrain surface;
[0090] The Topographic Potential Index (TPI) can be expressed as:
[0091]
[0092] in The elevation value of the center pixel. Here, n represents the elevation value of a neighboring cell, and n is the total number of neighboring cells. In this embodiment, the size of the neighboring window is set to... .
[0093] 2) Stack the above terrain structure parameters according to the channel dimension: [S, sin(A), cos(A), P, TPI] to form a 5-channel initial terrain structure tensor.
[0094] 3) After stacking the terrain structure parameters of each region according to the channel dimension, the terrain structure parameters of each channel are normalized and the numerical range is mapped to the [0,1] interval, and finally a terrain structure feature tensor T with size [H, W, 5] is generated.
[0095] S3, the improved DeepLab v3+ network architecture, such as Figure 2 As shown, its input data processing flow is as follows:
[0096] 1) The backbone network of DeepLab v3+ is used to extract features from remote sensing images and output semantic features. Its dimensions are [H, W, C].
[0097] 2) A terrain structure encoding module is introduced between the backbone network and the hollow spatial pyramid pooling module of the original DeepLab v3+ network. This module encodes the terrain structure feature tensor T to obtain terrain structure features that express the spatial morphological relationships of the terrain. The composition of the terrain structure encoding module is shown in Table 1. Figure 3 As shown:
[0098] Table 1. Composition of the Terrain Structure Coding Module
[0099]
[0100] The terrain structure encoding module comprises a three-layer convolutional structure. The first layer uses a 3×3 convolutional kernel to extract local spatial features from the terrain structure feature tensor, learning the spatial relationships between individual terrain structure parameters to obtain a local spatial structure representation, expressed as:
[0101]
[0102] This convolutional operation can learn the spatial relationships between terrain parameters within a 3×3 neighborhood, such as slope continuity, ridge structure, and valley morphology, and its output feature map... The dimensions are [H, W, 16];
[0103] The second convolutional layer extracts the nonlinear coupling relationship between different terrain structure parameters in the local spatial structure feature representation using a 3×3 convolutional kernel, resulting in a multi-parameter coupling feature, the expression of which is:
[0104]
[0105] This process can learn the nonlinear coupling relationships between parameters such as slope, aspect, and curvature, thereby obtaining more stable terrain structure features, including the output feature map. The dimensions are [H, W, 32];
[0106] The third convolutional layer uses a 1×1 convolutional kernel to maintain consistency between multi-parameter coupled features and semantic features in the channel dimension. Its expression is as follows:
[0107]
[0108] in, The dimensions are [H, W, C].
[0109] Through the above process, the original terrain structure feature tensor T is encoded into structural features that can express the spatial morphological relationships of the terrain. It can also be fused with image semantic features.
[0110] The improvements of this invention in terrain structure feature encoding are as follows:
[0111] The single DEM parameter input is expanded into a multi-channel terrain structure tensor that includes slope, aspect, curvature, and topographic potential index, thereby improving the ability to express terrain features. The spatial combination relationship between different terrain parameters is automatically learned through convolutional networks, avoiding the limitations of traditional manually designed terrain features. Efficient encoding is achieved through lightweight convolutional structures, which improves the expressive power of terrain structure information with less computational overhead. Spatial alignment of structural features and semantic features is achieved through 1×1 convolution, enabling deep fusion of the two types of features in a unified feature space.
[0112] 3) Integrating semantic features with and topographical features The data is concatenated along the channel dimension and then fused and reduced in dimensionality using a 1×1 convolutional layer to generate a joint feature representation. :
[0113]
[0114] S4. Multi-scale feature extraction and similarity calculation, the specific process is as follows: Figure 4 As shown:
[0115] The entire terrain region to be analyzed is divided into multiple sub-regions, and a similarity analysis is performed between each sub-region and the target terrain region. For any sub-region A and the target region B, their joint feature representations are input into the hollow spatial pyramid pooling module, and parallel convolutional layers with different void ratios (rates=6, 12, 18) are used to simultaneously extract feature maps at microscale (s), mesoscale (m), and macroscale (l), resulting in a multi-scale feature set. and ,in, These are the microscale features of region A. This represents the mesoscale characteristics of region A. For the macro-scale features of region A, These are the microscale features of region B. This represents the mesoscale characteristics of region B. For the macro-scale features of region B;
[0116] Calculate each scale Cosine similarity between region A and region B :
[0117]
[0118] in, .
[0119] S5. Multi-scale consistency constraints and weighted fusion:
[0120] 1) Calculate the difference vector between scales: Calculate the absolute difference between the similarity results at each pair of scales to form the difference vector. : ;
[0121] 2) Dynamic Confidence Weight Calculation: Based on the consistency between the similarity calculation results of one scale and other scales, an initial confidence weight is assigned to the similarity calculation results of that scale. The higher the consistency (the smaller the difference), the larger the weight. The initial confidence weights of the similarity calculation results of the three scales are calculated as follows. , , The calculation formulas are as follows:
[0122]
[0123]
[0124]
[0125] in, It is a minimal constant, and its value is This is used to prevent division by zero errors.
[0126] 3) Weight Normalization: The initial confidence weights are normalized to obtain the final fusion weights. .
[0127] 4) Weighted Fusion: Based on the final fusion weights, the similarity calculation results at each scale are weighted and fused to obtain the final similarity result. :
[0128]
[0129] 5) Cross-scale consistency judgment and output: setting a threshold Its value is set to 0.15, if If the multi-scale results are considered to be highly consistent, then the output will be directly determined. Otherwise, the similarity judgment of the two terrain regions being compared will have scale uncertainty, which will be left for subsequent analysts to assess.
[0130] S6. GIS-standardized output and integrated analysis of analysis results. Specific workflow as follows: Figure 5 As shown, the process includes:
[0131] 1) Compare each sub-region of the terrain region to be analyzed with the target terrain region one by one to obtain multiple similarity results. Summarize the multiple similarity results into a two-dimensional matrix. The element values in the two-dimensional matrix represent the final similarity results between the corresponding region and the target terrain region.
[0132] 2) Linearly normalize each final similarity result, transforming its value range linearly to the [0,1] interval:
[0133]
[0134] Here, min and max represent the minimum and maximum similarity values for the entire analysis region, respectively.
[0135] 3) Map the normalized similarity results to the grayscale range of [0, 255] to obtain the grayscale value of each pixel, forming an 8-bit grayscale raster image. This refers to a raster weighted map, where higher grayscale values indicate more similar terrain.
[0136]
[0137] 4) Weight the raster map Exported as a standard GeoTIFF format, it can be directly imported into mainstream GIS platforms (such as ArcGIS).
[0138] 5) In the GIS platform, this weighted map can be used as a cost surface or suitability layer, and can be comprehensively analyzed in conjunction with spatial constraint layers such as roads, water systems, ecological red lines, and basic farmland. The specific process is as follows:
[0139] (a) Unify all layers involved in the analysis to the same spatial reference frame and cell size to ensure the accuracy of the overlay analysis;
[0140] (b) Normalize each layer according to the analysis objective to make their numerical range consistent;
[0141] (c) A multi-criteria decision analysis method is used to overlay the layers and calculate the comprehensive score pixel by pixel:
[0142]
[0143] in This is the normalized terrain similarity weight map. , Other spatial constraint layers are normalized. , , Equal to the corresponding weight coefficients (the sum of the weights is 1);
[0144] (d) Based on the overall score The threshold is used to determine suitable areas. Areas with a score higher than 0.8 are classified as highly suitable areas, and areas with a score lower than 0.3 are classified as unsuitable areas, which can provide a basis for subsequent planning decisions.
[0145] The above process can be automated by building workflows in ArcGIS Model Builder, or executed in batches via Python scripts, thereby effectively integrating deep learning analysis results into the geographic decision support process.
[0146] In another embodiment, the present invention provides a terrain similarity analysis system that integrates multi-source, multi-scale terrain information, which employs the method described above, including:
[0147] The data acquisition and preprocessing module is used to acquire remote sensing image data and corresponding DEM data of the terrain area to be analyzed and the target terrain area, and to perform preprocessing.
[0148] The terrain structure analysis module is used to calculate terrain structure parameters based on preprocessed DEM data and stack the terrain structure parameters according to the channel dimension to generate a terrain structure feature tensor.
[0149] The fusion module is used to extract and fuse features from the terrain structure feature tensor and preprocessed remote sensing image data based on the improved DeepLab v3+ network to generate a joint feature representation.
[0150] The similarity calculation module is used to simultaneously extract feature maps of joint feature representation at different spatial scales, and calculate the similarity between the terrain region to be analyzed and the target terrain region at each spatial scale. The confidence weights are dynamically allocated according to the degree of difference between the similarity results at each scale, and the similarity results at each scale are weighted and fused to obtain the final similarity result.
[0151] The output module is used to convert the final similarity results into a raster weighted map and output it.
[0152] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0153] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A terrain similarity analysis method that integrates multi-source, multi-scale terrain information, characterized in that, Includes the following steps: Acquire remote sensing image data and corresponding DEM data of the terrain area to be analyzed and the target terrain area, and perform preprocessing; The terrain structure parameters are calculated based on the preprocessed DEM data, and the terrain structure parameters are stacked according to the channel dimension to generate a terrain structure feature tensor. Based on the improved DeepLab v3+ network, feature extraction and fusion are performed on the terrain structure feature tensor and the preprocessed remote sensing image data to generate a joint feature representation. Simultaneously extract joint feature representations of feature maps at different spatial scales, and calculate the similarity between each sub-region of the terrain region to be analyzed and the target terrain region at each spatial scale; The confidence weights are dynamically assigned based on the degree of difference between similarity results at each scale, and the similarity results at each scale are weighted and fused to obtain the final similarity result. The final similarity results are converted into a raster weighted map and output.
2. The method as described in claim 1, characterized in that, The preprocessing of remote sensing image data and corresponding DEM data includes: The remote sensing image data and the corresponding DEM data are subjected to unified spatial reference coordinate transformation, spatial registration and cropping. Resolution resampling is used to unify the resolution of the remote sensing image data and DEM data, so that each pixel is aligned in spatial position. After aligning each pixel in spatial position, radiometric and atmospheric corrections are performed on the remote sensing image data, and hole filling and smoothing are performed on the DEM data.
3. The method as described in claim 1, characterized in that, The topographic structure parameters calculated based on the preprocessed DEM data include at least: slope, aspect, curvature, and topographic potential index, wherein the aspect parameter is decomposed into sine and cosine components.
4. The method as described in claim 1, characterized in that, After stacking the terrain structure parameters of each region according to the channel dimension, the terrain structure parameters of each channel are normalized and the numerical range is mapped to the [0,1] interval to generate the final terrain structure feature tensor.
5. The method as described in claim 1, characterized in that, The steps for generating joint feature representations based on the improved DeepLab v3+ network include: A terrain structure coding module is introduced between the backbone network and the hollow spatial pyramid pooling module of the original DeepLab v3+ network; the backbone network is used to extract features from remote sensing image data and generate semantic features. The terrain structure encoding module is used to encode the terrain structure feature tensor to obtain terrain structure features that can express the spatial morphological relationships of the terrain. Semantic features and terrain structure features are concatenated along the channel dimension and then fused and dimensionality reduced using a 1×1 convolutional layer to generate a joint feature representation.
6. The method as described in claim 5, characterized in that, The terrain structure encoding module contains a three-layer convolutional structure. The first layer of convolutional structure extracts local spatial features from the terrain structure feature tensor using a 3×3 convolutional kernel, learns the spatial relationship of each individual terrain structure parameter, and obtains a local spatial structure representation. The second convolutional structure extracts the nonlinear coupling relationship between different terrain structure parameters in the local spatial structure feature representation through a 3×3 convolutional kernel, thus obtaining multi-parameter coupling features. The third convolutional layer uses a 1×1 convolutional kernel to keep the multi-parameter coupled features and semantic features consistent in the channel dimension.
7. The method as described in claim 1, characterized in that, The similarity calculation process between any sub-region and the target terrain region within the terrain region to be analyzed at each spatial scale includes: The joint feature representations of any sub-region A and the target region B within the terrain region to be analyzed are input into the hollow spatial pyramid pooling module. Parallel convolutional layers with different void ratios are used to simultaneously extract feature maps at micro, meso, and macro scales, resulting in a multi-scale feature set. and ,in, These are the microscale features of region A. This represents the mesoscale characteristics of region A. For the macro-scale features of region A, These are the microscale features of region B. This represents the mesoscale characteristics of region B. For the macro-scale features of region B; Calculate each scale Cosine similarity between region A and region B : ； in, .
8. The method as described in claim 7, characterized in that, The final similarity result is calculated as follows: Calculate the absolute difference between the similarity results at each pairwise scale to form a difference vector. : ; Based on the degree of consistency between the similarity calculation results of one scale and those of other scales, an initial confidence weight is assigned to the similarity calculation result of that scale. The initial confidence weights for the similarity calculation results of the three scales are... , , The calculation formulas are as follows: ； ； ； in, It is a very small constant; The initial confidence weights are normalized to obtain the final fusion weights: ; The similarity calculation results at each scale are weighted and fused based on the final fusion weight to obtain the final similarity result. : ； Set threshold ,like If the multi-scale results are considered to be highly consistent, then the output will be directly determined. Otherwise, there is scale uncertainty in the similarity judgment of the two terrain regions being compared.
9. The method as described in claim 1, characterized in that, The process of converting the final similarity results into a raster weighted map includes: The terrain region to be analyzed is divided into multiple sub-regions in advance. Each sub-region is compared with the target terrain region one by one to obtain multiple similarity results. The multiple similarity results are summarized into a two-dimensional matrix. The element values in the two-dimensional matrix represent the final similarity results between the corresponding region and the target terrain region. The final similarity results are linearly normalized to transform their value range to the [0,1] interval; The normalized similarity results are mapped to the grayscale range of [0, 255] to generate a raster weight map. In the raster weight map, the higher the grayscale value, the more similar the terrain.
10. A terrain similarity analysis system integrating multi-source, multi-scale terrain information, characterized in that, It employs the method described in any one of claims 1-9, comprising: The data acquisition and preprocessing module is used to acquire remote sensing image data and corresponding DEM data of the terrain area to be analyzed and the target terrain area, and to perform preprocessing. The terrain structure analysis module is used to calculate terrain structure parameters based on preprocessed DEM data and stack the terrain structure parameters according to the channel dimension to generate a terrain structure feature tensor. The fusion module is used to extract and fuse features from the terrain structure feature tensor and preprocessed remote sensing image data based on the improved DeepLab v3+ network to generate a joint feature representation. The similarity calculation module is used to simultaneously extract feature maps of joint feature representation at different spatial scales, and calculate the similarity between the terrain region to be analyzed and the target terrain region at each spatial scale. The similarity consistency constraint module dynamically allocates confidence weights based on the degree of difference between similarity results at each scale, and performs weighted fusion of the similarity results at each scale to obtain the final similarity result. The output module is used to convert the final similarity results into a raster weighted map and output it.