Dynamic remote sensing monitoring surveying and mapping system and method

Abstract

The invention discloses a surveying and mapping system and method for dynamic remote sensing monitoring, and relates to the technical field of remote sensing surveying and mapping. The specific implementation process comprises the steps that firstly, a region segmentation model is provided, region coarse segmentation is carried out according to the same-object same-spectrum principle, fine segmentation is assisted through a multispectral image, meanwhile, a self-adaptive loss function supervision model is introduced, and therefore the accuracy of region segmentation is improved; secondly, designing an edge refinement model, obtaining an edge mask in combination with Fourier transform and depth separable convolution, further updating the pixel value of the region edge, and generating an accurate region distribution map; finally, a region change monitoring model is provided, the change region and the change degree of the target region are analyzed through feature extraction, space registration and space-time fusion technologies, and the precision, timeliness and reliability of surveying and mapping are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing mapping, and in particular to a mapping system and method for dynamic remote sensing monitoring. Background Art

[0002] With the continuous improvement of remote sensing platforms such as satellites and drones, remote sensing mapping has gradually developed into an independent discipline, widely used in resource surveys, environmental monitoring, disaster warning, urban planning and other fields. This technology uses various sensors on ground, aviation and aerospace platforms to receive electromagnetic wave signals reflected, scattered or emitted by ground objects to draw topographic maps, land use maps or other thematic maps. Remote sensing mapping realizes dynamic monitoring of surface information and provides high-precision and large-scale geographic information data for various fields.

[0003] With the rapid development of remote sensing mapping technology, multi-temporal satellite images have provided strong data support for the research and management of many tasks such as natural resource management, environmental monitoring, urban planning and disaster assessment. Among them, land change monitoring, as one of the core research of these tasks, is of great significance for promoting sustainable economic and social development. Land change monitoring aims to accurately reflect the dynamic changes of land area, quality and distribution, and provide timely and accurate land use change data for relevant national and regional departments, mainly including changes in the area of ​​various types of production and construction land such as cultivated land, forest land, grassland, water area, transportation and residential land, as well as the damage and impact of natural disasters on land use.

[0004] At present, the existing method of mapping land changes is often to obtain surface change information from single-source multi-temporal satellite images. However, the monitoring range and data accuracy of single-source satellite images are limited by the sensor, making it difficult to fully capture the dynamic changes of the surface in complex scenes, resulting in inaccurate land segmentation results, which in turn causes large mapping errors.

[0005] In order to solve the above problems, the present invention proposes a surveying and mapping system and method for dynamic remote sensing monitoring. Summary of the invention

[0006] A mapping system and method for dynamic remote sensing monitoring is used for land mapping for real-time monitoring of surface changes. In order to solve the problems existing in the prior art, the present invention proposes a regional segmentation model, which performs rough regional segmentation based on the principle of same-object-same-spectrum, and assists in fine segmentation through multispectral images, while introducing an adaptive loss function supervision model, thereby improving the accuracy of regional identification; secondly, an edge refinement model is designed, which further refines the regional edge through an encoder and decoder structure, combined with Fourier transform and deep separable convolution, and generates an accurate regional edge mask; finally, the present invention proposes a regional change monitoring model, which analyzes the change area and degree of change of the target area through feature extraction, spatial registration and spatiotemporal fusion technology, and finally generates a change area map and converts it into vector data for updating mapping information.

[0007] A surveying and mapping system and method for dynamic remote sensing monitoring, comprising:

[0008] According to the composite spectral imager carried by the satellite, multi-temporal multispectral images and multi-temporal hyperspectral images of the target area are obtained respectively, and processed to obtain standard multi-temporal multispectral images and standard multi-temporal hyperspectral images;

[0009] Performing rough regional segmentation on the first hyperspectral image according to the regional segmentation model, and performing fine regional segmentation in combination with the first multispectral image, to generate a regional segmentation map and a regional confidence map; wherein the first hyperspectral image is the first image of the standard multi-temporal hyperspectral image, and the first multispectral image is the first image of the standard multi-temporal multispectral image;

[0010] Input the region segmentation map and the region confidence map, correct the segmentation edges of different regions through the edge refinement model, and generate a region distribution map;

[0011] Inputting the first hyperspectral image, the second hyperspectral image and the first multispectral image into a regional change monitoring model to obtain a change area and a change degree;

[0012] The changed area and degree of change are converted into vector diagrams, and the historical surveying and mapping data are replaced to obtain the current surveying and mapping data.

[0013] Preferably, the multi-temporal multispectral image and the multi-temporal hyperspectral image of the target area are acquired, and preprocessing is performed including denoising, correction, image mosaicking and data standardization; the correction includes geometric correction and radiation correction.

[0014] Preferably, the specific implementation process of performing rough regional segmentation on the first hyperspectral image according to the regional segmentation model and performing fine regional segmentation in combination with the first multispectral image includes:

[0015] The first hyperspectral image is subjected to regional coarse segmentation according to the regional coarse segmentation module to obtain a regional coarse segmentation map and a category confidence; the regional coarse segmentation map is input into the upsampling module to generate a regional coarse segmentation upsampling map; the regional segmentation upsampling map, the category confidence and the first multispectral image are input, and the regional segmentation map and the regional confidence map are obtained through the regional fine segmentation module; the regional segmentation model is optimized according to an adaptive loss function.

[0016] Preferably, the implementation process of the region coarse segmentation module includes:

[0017] The first hyperspectral image is input into the backbone network to generate a hyperspectral feature map. According to the same spectrum principle for the same object, the spectral weight map of the hyperspectral feature map in the n×n area around the pixel point is obtained to generate a hyperspectral feature enhancement map. The specific formula is:

[0018]

[0019] F en =LM(Concat(Map1V1,…,Map i V i ,…,Map h V h ));

[0020] Among them, Map i represents the spectral weight map of the ith n×n pixel region; Q i ,K i ,V i respectively represent the query, key and value matrices of the ith n×n pixel region of the hyperspectral feature map; Respectively represent Q i and K i Spectral vectors of different pixel points; · represents vector dot product operation; || || represents vector modulus operation; F en represents the hyperspectral feature enhancement map; LM() and Concat() represent linear mapping and dimension concatenation operations respectively; h represents the number of n×n local regions in the hyperspectral feature map;

[0021] The hyperspectral feature enhancement map is input into the fully connected layer and the Softmax classifier to generate a rough regional segmentation map and category confidence.

[0022] Preferably, the implementation process of the regional segmentation module includes:

[0023] According to the regional coarse segmentation upsampling map and the category confidence, a regional coarse segmentation probability map is generated; the first multispectral image and the regional coarse segmentation probability map are input into the self-attention mechanism to obtain a regional refined feature map; based on the fully connected layer and the Softmax classifier, a regional segmentation map and a regional confidence map are generated.

[0024] Preferably, the implementation process of optimizing the region segmentation model according to the adaptive loss function includes:

[0025] The region segmentation map is supervised according to the intra-class difference loss function and the inter-class difference loss function, and the specific calculation formula is:

[0026]

[0027] L=L intra -λL iner ;

[0028] Among them, L intra represents the intra-class difference loss function; Liner represents the inter-class difference loss function; L represents the adaptive loss function; They respectively represent the features of the region coarse segmentation upsampling map and the region segmentation map in k categories; C represents the number of categories of region segmentation; || ||1 represents a norm function; λ represents an adjustment parameter and 0<λ≤1.

[0029] Preferably, the implementation process of correcting the segmentation edges of different regions by the edge refinement model includes:

[0030] The region segmentation map and the first multispectral image are input, and edge features are generated according to the encoder, Fourier transform and depth-separable convolution. The specific calculation formula of the edge features is:

[0031]

[0032] Among them, F edge represents the edge feature; DW() represents the depth-separable convolution; IFFT() and FFT() represent inverse fast Fourier transform and fast Fourier transform respectively; HPF() represents high-pass filter; F encoder represents the output features of the encoder; Represents element-by-element subtraction operation;

[0033] Inputting the edge feature into a decoder and a classifier to generate a region edge mask; wherein the edge pixel value of the region edge mask is 1, and the non-edge pixel value is 0;

[0034] According to the region confidence map and the region edge mask, a region edge confidence map is generated; according to the region segmentation map and the region edge confidence map, the edge pixels of the region segmentation map are updated to generate a region distribution map, and the specific formula is expressed as:

[0035]

[0036] in, Represents the pixel value of the regional distribution map at the position (h, w); represents the pixel value of the region segmentation map at the position (h, w); Represents the pixel value of the edge mask of the region at the position (h,w); represents the pixel value of the edge confidence map of the region at the position (h, w); α represents the edge confidence threshold.

[0037] Preferably, the regional change monitoring model includes:

[0038] According to the first hyperspectral image, the second hyperspectral image and the first multispectral image, the first hyperspectral upsampling feature, the second hyperspectral upsampling feature and the first multispectral feature are obtained; according to the regional change monitoring model, a spatial alignment operation is performed to generate a change mask, and the specific calculation formula is:

[0039]

[0040] Where, ΔF spa represents the motion feature between the first hyperspectral upsampling feature and the second hyperspectral upsampling feature; ΔF res represents the spatial difference feature between the first hyperspectral upsampling feature and the first multispectral feature; F diff represents the change mask; FLOW() represents the optical flow network; Concat() represents the dimension concatenation operation; Respectively represent the first hyperspectral upsampling feature and the second hyperspectral upsampling feature; represents the first multi-spectral feature; CNN() represents a convolutional neural network; Conv() represents a convolution operation; Represents a linear interpolation operation;

[0041] The first hyperspectral image, the second hyperspectral image and the first multispectral image are input, and the second multispectral image is obtained according to the spatiotemporal fusion technology; the change area and the change degree are generated according to the change mask and the second multispectral image; the specific formula of the change degree is:

[0042]

[0043] in, represents the degree of change; ΔLA represents the change in regional category; ΔA represents the changed area; A represents the area corresponding to the first multispectral image; SL() regional damage function; I p ,I c Respectively represent the multispectral images of the corresponding areas before and after the change.

[0044] Preferably, the process of converting the change area and the change degree into a vector diagram includes:

[0045] The geometric features of the changed area are obtained; according to the geometric features and the degree of change, the changed area is converted into a vector data format, a vector layer is generated and optimized to obtain a vector map; wherein the geometric features include: contour boundaries, vertex coordinates, area and shape parameters; the vector map includes: vertex coordinates, line segments and face area information; the optimization process includes removing redundant vertices, smoothing boundaries and adjusting coordinate accuracy.

[0046] A dynamic remote sensing monitoring surveying and mapping system, comprising:

[0047] The data acquisition and preprocessing module obtains multi-temporal multispectral images and multi-temporal hyperspectral images of the target area according to the composite spectral imager carried by the satellite, and performs preprocessing to obtain standard multi-temporal multispectral images and standard multi-temporal hyperspectral images;

[0048] A region segmentation module, inputting the first hyperspectral image and the first multispectral image into the region segmentation model to generate a region segmentation map and a region confidence map; wherein the first hyperspectral image is the first image of the standard multi-temporal hyperspectral image, and the first multispectral image is the first image of the standard multi-temporal multispectral image;

[0049] An edge refinement module, which corrects the segmentation edge of the region segmentation map according to the edge refinement model to generate a region distribution map;

[0050] A regional change monitoring module, which obtains a second multispectral image and a second hyperspectral image, combines the first hyperspectral image and the first multispectral image, and generates a change area and a change degree;

[0051] The surveying and mapping information updating module converts the changed area and the change degree into a vector map, replaces the vector map of the corresponding area in the historical surveying and mapping data, and generates updated surveying and mapping data.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The present invention proposes a regional segmentation model, which not only performs coarse regional segmentation based on hyperspectral images, but also combines multispectral images for fine segmentation, making regional classification more detailed and accurate, improving the accuracy of regional classification, and facilitating subsequent change monitoring and updating of surveying and mapping information. In addition, the model also constructs an adaptive loss function, using the regional coarse segmentation upsampling map as a pseudo-label to supervise the regional fine segmentation results, so that the model can adapt to the complex environment of different regions, effectively improving the training efficiency and segmentation accuracy of the model.

[0054] 2. The present invention designs an edge refinement model, uses Fourier transform and depth-separable convolution to obtain edge information, and updates the edge pixels of the region segmentation map through the edge confidence map. The model can capture the edge features of the region more finely, further improve the integrity and accuracy of the region edge, make the region boundary clearer, improve the accuracy of the overall surveying and mapping data, and provide more reliable basic data for subsequent change monitoring.

[0055] 3. The present invention proposes a regional change monitoring model to determine the change area based on time, space and spectral data; the model makes change monitoring more comprehensive and accurate through feature extraction and spatial alignment operations, reduces false alarms and missed alarms, and further improves the accuracy and reliability of change monitoring; the second multispectral image is obtained through time-space fusion technology, and the change area and degree of change are determined through change masks, thereby realizing real-time monitoring and quantitative evaluation of surface changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart of a surveying and mapping method for dynamic remote sensing monitoring provided by an embodiment of the present invention;

[0057] Figure 2 A structural diagram of a region segmentation model provided in an embodiment of the present invention;

[0058] Figure 3 A structural diagram of a regional change monitoring model provided by an embodiment of the present invention;

[0059] Figure 4 A flow chart of a dynamic remote sensing monitoring surveying and mapping system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Land change refers to the changes in land use, land cover type and ownership. It is the result of multiple factors such as social and economic development, natural environment changes and human activities. This change may manifest as the conversion of farmland to urban land, deforestation, disappearance of wetlands, land degradation and other phenomena, which have a profound impact on the ecological environment, resource utilization and social and economic development. In order to monitor land change in a timely and accurate manner, dynamic remote sensing mapping has become an important technical means. This technology can accurately identify and quantify the changes in surface types, spatial distribution and changes by comparing images at different times or conditions, providing scientific support for regional economic and social development.

[0062] The present invention proposes a dynamic remote sensing monitoring mapping system and method to monitor land changes to achieve real-time mapping of land changes. In order to illustrate the effectiveness of the method and system of the present invention, a specific description will be given in conjunction with the accompanying drawings of this embodiment and the following two embodiments.

[0063] Embodiment 1

[0064] The present application embodiment discloses a dynamic remote sensing monitoring mapping method to realize monitoring and mapping of land changes in area A. Figure 1 The specific steps of the method proposed in the present invention include: S1. Acquire multi-temporal multispectral images and multi-temporal hyperspectral images, and perform preprocessing; S2. Perform regional segmentation according to the regional segmentation model; S3. Perform edge detail division on the segmentation map through the edge refinement model; S4. Acquire the change area and the degree of change according to the regional change monitoring model; S5. Convert the change area and the degree of change into a vector map, and generate current surveying and mapping data in combination with historical surveying and mapping data.

[0065] Furthermore, according to the composite spectral imager carried by the satellite, multi-temporal multispectral images and multi-temporal hyperspectral images of the target area are respectively obtained, and pre-processed to obtain standard multi-temporal multispectral images and standard multi-temporal hyperspectral images, corresponding to the above-mentioned step S1; the implementation process of the pre-processing includes:

[0066] De-noising, removing the equipment noise and environmental noise interference of the multi-temporal multispectral image and the multi-temporal hyperspectral image according to a filtering method; wherein the filtering method includes: smoothing filtering, median filtering and wavelet transform;

[0067] Correction processing includes geometric correction and radiation correction; the geometric correction converts the remote sensing image into a unified geographic coordinate system, and corrects it through an affine transformation method, while performing a registration operation on the remote sensing image; the radiation correction adjusts the image data according to a specific model and auxiliary information, and removes image distortion caused by factors such as sensors, atmosphere and the sun in the remote sensing image;

[0068] Image mosaicking, through preprocessing, registration, tone adjustment and de-overlapping steps, multiple remote sensing images of different areas are stitched into a remote sensing image of a larger area;

[0069] Data standardization: Perform channel normalization on remote sensing images, limit the value of each pixel to the range of [0,1], and obtain standard multi-temporal multispectral images and standard multi-temporal hyperspectral images.

[0070] Due to the inherent limitations of remote sensing sensors, it is difficult to obtain remote sensing images with high spatial resolution, high spectral resolution, and high temporal resolution. Therefore, the combination of multi-temporal hyperspectral images and multi-temporal multispectral images can give full play to their advantages and achieve comprehensive, accurate, and real-time acquisition of surface information. The high-precision spectral information provided by multi-temporal hyperspectral images can be used for land analysis and classification, while the high spatial resolution information of multi-temporal multispectral images is helpful for the fine division of different land types. At the same time, the acquisition cycle of multi-temporal hyperspectral images is shorter than that of multi-temporal multispectral images, which realizes dynamic change monitoring. The two complement each other and jointly improve the accuracy and efficiency of remote sensing mapping, providing more reliable data support for environmental protection, resource management, and scientific research.

[0071] The embodiment of the present application performs preprocessing operations on the multi-phase multispectral image and the multi-phase hyperspectral image, including denoising, correction, image mosaicking and data standardization. Among them, denoising significantly reduces the interference of equipment and environmental noise, and improves the quality of remote sensing images; geometric correction and radiation correction ensure the geographic positioning accuracy of the image and the authenticity of the radiation information, and remove image distortion caused by various factors such as sensors, atmosphere and sun; image mosaicking seamlessly splices multiple regional images, expanding the scope of dynamic remote sensing monitoring; data standardization further unifies the image data format, providing convenience for subsequent analysis. This series of processing measures significantly improves the availability, reliability and analysis accuracy of the image, and provides strong data support for the subsequent segmentation and change detection of remote sensing images.

[0072] Further, the first hyperspectral image is subjected to a rough regional segmentation according to the regional segmentation model, and the first multispectral image is combined to perform a fine regional segmentation to generate a regional segmentation map and a regional confidence map, corresponding to the above-mentioned step S2; wherein the first hyperspectral image is the first image of the standard multi-temporal hyperspectral image, and the first multispectral image is the first image of the standard multi-temporal multispectral image; the regional segmentation categories include: grassland, cultivated land, garden, woodland, residential land and water area, etc.; refer to Figure 2 , the implementation process of the region segmentation model includes:

[0073] The regional coarse segmentation module performs regional coarse segmentation on the first hyperspectral image according to the regional coarse segmentation module to obtain a regional coarse segmentation map and a category confidence; the implementation process of the regional coarse segmentation module includes:

[0074] The first hyperspectral image is input into the backbone network to generate a hyperspectral feature map; based on the principle of the same object with the same spectrum, the similarity between spectra in different regions is calculated to obtain the spectral weight map of the hyperspectral feature map in the n×n area around the pixel point, and generate a hyperspectral feature enhancement map; wherein the backbone network is a convolutional neural network or Resnet; the specific calculation formula of the hyperspectral feature enhancement map is:

[0075]

[0076] F en =LM(Concat(Map1V1,…,Map i V i ,…,Map h V h ));

[0077] Among them, Map i represents the spectral weight map of the ith n×n pixel region; Q i ,K i ,V i respectively represent the query, key and value matrices of the ith n×n pixel region of the hyperspectral feature map; Respectively represent Q i and K i Spectral vectors of different pixel points; · represents vector dot product operation; || || represents vector modulus operation; F en represents the hyperspectral feature enhancement map; LM() and Concat() represent linear mapping and dimension concatenation operations respectively; h represents the number of n×n local regions in the hyperspectral feature map;

[0078] The hyperspectral feature enhancement map is input into the fully connected layer and the Softmax classifier to generate a rough regional segmentation map and category confidence.

[0079] The embodiment of the present application uses a regional coarse segmentation module to process the first hyperspectral image, extracts the hyperspectral feature map through the backbone network, and calculates the similarity between spectral vectors at different pixel points based on the principle of the same object and the same spectrum, which effectively enhances the spectral characteristics of key areas in the image and improves the discrimination and robustness of the feature map. Subsequently, the enhanced feature map is parsed through a fully connected layer and a Softmax classifier, which not only achieves the coarse segmentation of different regions in the hyperspectral image, but also outputs the confidence of each category, significantly improving the efficiency and accuracy of hyperspectral image segmentation, and also provides a reliable basis for subsequent fine classification or decision-making.

[0080] Furthermore, the region coarse segmentation map is input into an upsampling module to generate a region coarse segmentation upsampling map; the region segmentation upsampling map, the category confidence and the first multispectral image are input, and the region segmentation map and the region confidence map are obtained through a region fine segmentation module; the implementation process of the region fine segmentation module includes:

[0081] According to the regional coarse segmentation upsampling map and the category confidence, a regional coarse segmentation probability map is obtained; the spatial refinement relationship between the first multispectral image and the regional coarse segmentation probability map is obtained through the self-attention mechanism, so as to obtain the regional refinement feature map, and further according to the fully connected layer and the Softmax classifier, the regional segmentation map and the regional confidence map are obtained.

[0082] The embodiment of the present application performs upsampling processing on the region coarse segmentation map, which not only retains the basic outline of the original segmentation, but also enhances the ability to capture detailed information. Subsequently, the category confidence and the first multispectral image are combined to deeply explore the spatial refinement relationship between images through the self-attention mechanism, effectively integrating spatial information and semantic information, making the region fine segmentation process more accurate and detailed. The fully connected layer and the Softmax classifier generate a high-confidence region confidence map, which can significantly improve the accuracy and efficiency of region segmentation.

[0083] Furthermore, according to the region coarse segmentation up-sampling map and the region segmentation map, an adaptive loss function is constructed to optimize the region segmentation model; the adaptive loss function includes:

[0084] The region rough segmentation upsampled map is regarded as a pseudo label, and the intra-class difference loss function and the inter-class difference loss function are constructed to supervise the region segmentation map. The specific calculation formula is:

[0085]

[0086] L=L intra -λL iner ;

[0087] Among them, L intra represents the intra-class difference loss function; L iner represents the inter-class difference loss function; L represents the adaptive loss function; They respectively represent the features of the region coarse segmentation upsampling map and the region segmentation map in k categories; C represents the number of categories of region segmentation; || ||1 represents a norm function; λ represents an adjustment parameter and 0<λ≤1.

[0088] The embodiment of the present application uses an adaptive loss function to optimize the region segmentation model, which can significantly improve the segmentation performance and robustness of the model. By treating the upsampled map of the coarse segmentation of the region as a pseudo-label, and constructing the intra-class difference loss function and the inter-class difference loss function accordingly, this method effectively utilizes the useful information in the coarse segmentation results and performs fine supervision on the fine segmentation process. The intra-class difference loss function aims to reduce the inconsistency of the prediction results within the same category, while the inter-class difference loss function enhances the distinction between different categories. This dual supervision mechanism not only promotes the model's accurate recognition of regional boundaries, but also improves the model's accuracy in regional segmentation in complex remote sensing scenarios.

[0089] The embodiment of the present application segments different land types in area A through regional coarse segmentation, upsampling, regional fine segmentation and adaptive loss function. Among them, the regional coarse segmentation module quickly identifies the approximate area in the image, laying the foundation for subsequent processing; the upsampling module effectively retains the segmentation contour and enhances the detail information, providing richer data support for fine segmentation; the regional fine segmentation module makes full use of the spatial information of the multispectral image and the semantic information of the coarse segmentation result, deeply mines the spatial refinement relationship through the self-attention mechanism, and realizes high-precision regional segmentation; in view of the characteristics of the regional segmentation task, the regional segmentation model introduces an adaptive loss function, accurately supervises the model training process, effectively improves the generalization ability and stability of the model, and significantly improves the accuracy and efficiency of regional segmentation.

[0090] Furthermore, the region segmentation map and the region confidence map are input, and the segmentation edges of different regions are corrected by an edge refinement model to generate a region distribution map, corresponding to the above step S3; the edge refinement model includes:

[0091] The region segmentation map and the first multispectral image are spliced ​​and input into an encoder to extract features, and edge features are obtained by Fourier transform and depthwise separable convolution; the specific calculation formula of the edge features is:

[0092]

[0093] Among them, F edge represents the edge feature; DW() represents the depth-separable convolution; IFFT() and FFT() represent inverse fast Fourier transform and fast Fourier transform respectively; HPF() represents high-pass filter; F encoder represents the output features of the encoder; Represents element-by-element subtraction operation;

[0094] Inputting the edge features into a decoder and a classifier to generate a region edge mask; wherein the encoder and the decoder are both convolutional neural networks; the edge pixel value of the region edge mask is 1, and the non-edge pixel value is 0;

[0095] According to the region confidence map and the region edge mask, a region edge confidence map is generated; according to the region segmentation map and the region edge confidence map, the edge pixels of the region segmentation map are updated to generate a region distribution map, and the specific formula is expressed as:

[0096]

[0097] in, Represents the pixel value of the regional distribution map at the position (h, w); represents the pixel value of the region segmentation map at the position (h, w); Represents the pixel value of the edge mask of the region at the position (h,w); represents the pixel value of the edge confidence map of the region at the position (h, w); α=0.8 represents the edge confidence threshold.

[0098] The embodiment of the present application uses an edge refinement model to post-process the region segmentation map and the region confidence map. Through Fourier transform and depth-separable convolution, the subtle edge features between regions are effectively captured and enhanced, which not only improves the accuracy of edge detection, but also enhances the model's ability to recognize different types of boundaries. Subsequently, the decoder and classifier collaborate to generate an accurate region edge mask, providing a reliable basis for the correction of region boundaries. Combined with the region confidence map, the edge refinement model further generates a region edge confidence map, and through the edge pixel value update strategy, ensures the accuracy and continuity of the region distribution map in the edge area. This process not only improves the accuracy of region segmentation, but also enhances the spatial consistency and semantic coherence of the region distribution map.

[0099] Further, the first hyperspectral image, the second hyperspectral image and the first multispectral image are input into the regional change monitoring model to obtain the change area and the change degree, corresponding to the above step S4; wherein the second hyperspectral image is the hyperspectral image data obtained according to the revisit period; refer to Figure 3 , the regional change monitoring model includes:

[0100] The first hyperspectral image, the second hyperspectral image and the first multispectral image are input into the feature extraction module to obtain the first hyperspectral feature, the second hyperspectral feature and the first multispectral feature respectively; the first hyperspectral upsampling feature and the second hyperspectral upsampling feature are obtained through the upsampling module, and a spatial alignment operation is performed to obtain a change mask. The specific calculation formula is:

[0101]

[0102] Where, ΔF spa represents the motion feature between the first hyperspectral upsampling feature and the second hyperspectral upsampling feature; ΔF res represents the spatial difference feature between the first hyperspectral upsampling feature and the first multispectral feature; F diff Represents a change mask; FLOW() represents an optical flow network; Concat() represents a dimension concatenation operation; Represent the first hyperspectral upsampling feature and the second hyperspectral upsampling feature respectively; represents the first multi-spectral feature; CNN() represents a convolutional neural network; Conv() represents a convolution operation; Represents a linear interpolation operation;

[0103] According to the first hyperspectral image, the second hyperspectral image and the first multispectral image, a second multispectral image is obtained by using a spatiotemporal fusion technology; according to the change mask and the second multispectral image, a change area and a change degree are obtained; the specific formulas of the change area and the change degree are:

[0104]

[0105]

[0106] Among them, D diff Representing the image of the changed area; Indicates the degree of said change; represents the second multispectral image; ⊙ represents the pixel-by-pixel multiplication operation; ΔLA represents the change of regional category, which is 1 when the regional category changes, otherwise it is 0; ΔA represents the changed area; A represents the area corresponding to the first multispectral image; SL() regional damage function; I p ,I c Respectively represent the multispectral images of the corresponding areas before and after the changes; They represent the pixel values ​​at the (i, j) position of the multispectral image of the corresponding area before and after the change, respectively; m and n represent the width and height of the multispectral image of the corresponding area, respectively.

[0107] Specifically, in order to further illustrate the role of the degree of land change proposed by the present invention, the changes in land use types in area A within two years are given as an example for comparison, and the changes in various land types are listed in Table 1. According to the table data, it can be seen that the degree of change is closely related to the size of the transformation area, that is, the land type with a larger transformation area has a relatively higher degree of change; among them, the degree of change of residential land reached 0.74, which is because it received a large amount of transferred area from cultivated land, grassland and water area, indicating that the area has experienced large-scale urbanization or land development activities; in addition, although the forest land has undergone land changes, its land type has not changed, which means that the area may have experienced large-scale logging or tree planting activities, resulting in changes in surface cover or landscape; therefore, the degree of land change is not only related to the change area, but also highly correlated with whether the land type has changed.

[0108] Table 1. Changes in land use types in region A within two years

[0109]

[0110] The embodiment of the present application proposes a regional change monitoring model, which can accurately extract key features from multi-source remote sensing data, including hyperspectral features and multispectral features, and effectively utilize the high spectral resolution advantages of hyperspectral images and the high spatial resolution of multispectral images. The combination of the upsampling module and the spatial alignment operation not only realizes the spatial alignment of features, but also deeply integrates the motion features and spatial difference features between different images through the optical flow network and the convolutional neural network, significantly improving the accuracy of change detection. Furthermore, the second multispectral image generated by the spatiotemporal fusion technology, combined with the change mask, can carefully depict the specific location and degree of change of the changed area, including the transformation of regional categories and the assessment of damage, providing scientific and intuitive decision support for fields such as land monitoring, disaster assessment and resource management.

[0111] Further, the change area and degree of change are converted into vector diagrams, and the historical surveying and mapping data are replaced to obtain the current surveying and mapping data. Figure 1 The specific implementation process of step S5 includes:

[0112] The geometric features of the changed area are obtained; according to the geometric features and the degree of change, the changed area is converted into a vector data format, a vector layer is generated and optimized to obtain a vector map; wherein the geometric features include: contour boundaries, vertex coordinates, area and shape parameters; the vector map contains: vertex coordinates, line segments and surface information; the optimization process includes removing redundant vertices, smoothing boundaries and adjusting coordinate accuracy; it is worth noting that when conducting initial land surveying, the geometric features are obtained and initial surveying data are generated based on the regional distribution map generated by the edge refinement model.

[0113] The embodiment of the present application realizes efficient updating of surveying and mapping data by converting the change area and degree of change into a vector diagram and replacing the historical surveying and mapping data. This method not only captures the latest dynamics of surface changes, but also ensures the accurate expression of change information by extracting the geometric features of the change area (such as contour boundaries, vertex coordinates, area and shape parameters). After conversion to vector data format, the generated vector layer is optimized, such as removing redundant vertices, smoothing boundaries and adjusting coordinate accuracy, which improves the simplicity and accuracy of the data, and further significantly improves the timeliness, reliability and application value of surveying and mapping data.

[0114] The embodiment of the present application realizes dynamic monitoring and mapping of land changes through multi-phase multispectral images and multi-phase hyperspectral images taken by satellites. The specific implementation process mainly includes the following steps: (a) land type segmentation of the target area; (b) refinement and updating of the edges of the segmentation results; (c) change monitoring of the land use situation in the target area. The present invention proposes a regional segmentation model, an edge refinement model and a regional change monitoring model for the above three processes respectively; the regional segmentation model combines hyperspectral images and multispectral images to realize progressive regional segmentation from coarse to fine, making regional classification more detailed and accurate, and improving the accuracy of regional classification; at the same time, the segmentation results are supervised according to the adaptive loss function, which improves the segmentation accuracy of the model in complex environments in different regions; the edge refinement model updates the edge pixels of the regional segmentation map according to the edge mask and the edge confidence map, making the regional boundary clearer, and further improving the integrity and accuracy of the regional edge; the regional change monitoring model makes the change monitoring more comprehensive and accurate through feature extraction and spatial alignment operations, reduces false alarms and omissions, and further improves the accuracy and reliability of change monitoring.

[0115] Embodiment 2

[0116] In the first embodiment, the method of the present invention realizes the dynamic remote sensing mapping of the land change situation in area A. In the embodiment of the present application, a dynamic remote sensing monitoring mapping system proposed by the present invention will be described. Figure 4 , the system comprising:

[0117] The data acquisition and preprocessing module obtains multi-temporal multispectral images and multi-temporal hyperspectral images of the target area according to the composite spectral imager carried by the satellite, and performs preprocessing to obtain standard multi-temporal multispectral images and standard multi-temporal hyperspectral images;

[0118] A region segmentation module, inputting the first hyperspectral image and the first multispectral image into the region segmentation model to generate a region segmentation map and a region confidence map; wherein the first hyperspectral image is the first image of the standard multi-temporal hyperspectral image, and the first multispectral image is the first image of the standard multi-temporal multispectral image;

[0119] An edge refinement module, which corrects the segmentation edge of the region segmentation map according to the edge refinement model to generate a region distribution map and region types;

[0120] A regional change monitoring module obtains a second multispectral image and a second hyperspectral image, combines the first hyperspectral image and the first multispectral image, obtains a change mask through feature extraction and spatial alignment operations, and obtains a second multispectral image according to a spatiotemporal fusion technology to generate a change area and a change degree of the target area;

[0121] The surveying and mapping information updating module converts the changed area and the change degree into a vector map, replaces the vector map of the corresponding area in the historical surveying and mapping data, and generates updated surveying and mapping data.

[0122] The embodiment of the present application can efficiently and accurately realize the multi-phase multi-spectral and multi-phase hyperspectral image acquisition, processing and analysis of the target area. The system not only uses the high-quality image data obtained by the composite spectral imager for standardized preprocessing to ensure data accuracy, but also accurately divides different geographical areas and evaluates confidence through the regional segmentation model, laying a solid foundation for subsequent analysis; the introduction of the edge refinement module further optimizes the definition of regional boundaries, improves the accuracy of regional distribution maps and the ability to identify regional types; the regional change monitoring module can effectively identify and quantify the change areas and degrees of the target area at different time points, and realize dynamic monitoring of the land use of the target area; the surveying and mapping information update module converts the change information into a vector diagram, updates the surveying and mapping database in real time, and ensures the timeliness of the surveying and mapping information; the system effectively solves the shortcomings of traditional remote sensing surveying and mapping technology in long-term large-scale monitoring, severe weather impact, data processing complexity and identification of complex ground feature areas, and improves the accuracy, timeliness and reliability of surveying and mapping.

[0123] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A surveying and mapping method for dynamic remote sensing monitoring, characterized in that: include: According to the composite spectral imager carried by the satellite, multi-temporal multispectral images and multi-temporal hyperspectral images of the target area are obtained respectively, and processed to obtain standard multi-temporal multispectral images and standard multi-temporal hyperspectral images; Performing rough regional segmentation on the first hyperspectral image according to the regional segmentation model, and performing fine regional segmentation in combination with the first multispectral image, to generate a regional segmentation map and a regional confidence map; wherein the first hyperspectral image is the first image of the standard multi-temporal hyperspectral image, and the first multispectral image is the first image of the standard multi-temporal multispectral image; Input the region segmentation map and the region confidence map, correct the segmentation edges of different regions through the edge refinement model, and generate a region distribution map; The second hyperspectral image, the first hyperspectral image and the first multispectral image are input into the regional change monitoring model to obtain the change area and the change degree; the specific formula of the change degree is: in, represents the degree of change; ΔLA represents the change in regional category; ΔA represents the changed area; A represents the area corresponding to the first multispectral image; SL() regional damage function; I p ,I c Respectively represent the multispectral images of the corresponding areas before and after the changes; The changed area and degree of change are converted into vector diagrams, and the historical surveying and mapping data are replaced to obtain the current surveying and mapping data.

2. A dynamic remote sensing monitoring surveying and mapping method according to claim 1, characterized in that: The multi-temporal multispectral image and the multi-temporal hyperspectral image of the target area are acquired, and preprocessed including denoising, correction, image mosaicking and data standardization; the correction includes geometric correction and radiation correction.

3. A dynamic remote sensing monitoring surveying and mapping method according to claim 1, characterized in that: The specific implementation process of performing rough regional segmentation on the first hyperspectral image according to the regional segmentation model and performing fine regional segmentation in combination with the first multispectral image includes: The first hyperspectral image is subjected to regional coarse segmentation according to the regional coarse segmentation module to obtain a regional coarse segmentation map and a category confidence; the regional coarse segmentation map is input into the upsampling module to generate a regional coarse segmentation upsampling map; the regional segmentation upsampling map, the category confidence and the first multispectral image are input, and the regional segmentation map and the regional confidence map are obtained through the regional fine segmentation module; the regional segmentation model is optimized according to an adaptive loss function.

4. A dynamic remote sensing monitoring surveying and mapping method according to claim 3, characterized in that: The implementation process of the regional coarse segmentation module includes: The first hyperspectral image is input into the backbone network to generate a hyperspectral feature map. According to the same spectrum principle for the same object, the spectral weight map of the hyperspectral feature map in the n×n area around the pixel point is obtained to generate a hyperspectral feature enhancement map. The specific formula is: F en =LM(Concat(Map1V1,…,Map i V i ,…,Map h V h )); Among them, Map i represents the spectral weight map of the ith n×n pixel region; Q i ,K i ,V i respectively represent the query, key and value matrices of the ith n×n pixel region of the hyperspectral feature map; Respectively represent Q i and K i Spectral vectors of different pixel points; · represents vector dot product operation; || || represents vector modulus operation; F en represents the hyperspectral feature enhancement map; LM() and Concat() represent linear mapping and dimension concatenation operations respectively; h represents the number of n×n local regions in the hyperspectral feature map; The hyperspectral feature enhancement map is input into the fully connected layer and the Softmax classifier to generate a rough regional segmentation map and category confidence.

5. A dynamic remote sensing monitoring surveying and mapping method according to claim 3, characterized in that: The implementation process of the regional fine segmentation module includes: According to the regional coarse segmentation upsampling map and the category confidence, a regional coarse segmentation probability map is generated; the first multispectral image and the regional coarse segmentation probability map are input into the self-attention mechanism to obtain a regional refined feature map; based on the fully connected layer and the Softmax classifier, a regional segmentation map and a regional confidence map are generated.

6. A dynamic remote sensing monitoring surveying and mapping method according to claim 3, characterized in that: The implementation process of optimizing the region segmentation model according to the adaptive loss function includes: The region segmentation map is supervised according to the intra-class difference loss function and the inter-class difference loss function, and the specific calculation formula is: L=L intra -λL iner ; Among them, L intra represents the intra-class difference loss function; L iner represents the inter-class difference loss function; L represents the adaptive loss function; They respectively represent the features of the region coarse segmentation upsampling map and the region segmentation map in k categories; C represents the number of categories of region segmentation; || ||1 represents a norm function; λ represents an adjustment parameter and 0<λ≤1.

7. A dynamic remote sensing monitoring surveying and mapping method according to claim 1, characterized in that: The implementation process of correcting the segmentation edges of different regions by the edge refinement model includes: The region segmentation map and the first multispectral image are input, and edge features are generated according to the encoder, Fourier transform and depth-separable convolution. The specific calculation formula of the edge features is: Among them, F edge represents the edge feature; DW() represents the depth-separable convolution; IFFT() and FFT() represent inverse fast Fourier transform and fast Fourier transform respectively; HPF() represents high-pass filter; F encoder represents the output features of the encoder; Represents element-by-element subtraction operation; Inputting the edge feature into a decoder and a classifier to generate a region edge mask; wherein the edge pixel value of the region edge mask is 1, and the non-edge pixel value is 0; According to the region confidence map and the region edge mask, a region edge confidence map is generated; according to the region segmentation map and the region edge confidence map, the edge pixels of the region segmentation map are updated to generate the region distribution map, and the specific formula is expressed as: in, Represents the pixel value of the regional distribution map at the position (h, w); represents the pixel value of the region segmentation map at the position (h, w); Represents the pixel value of the edge mask of the region at the position (h,w); represents the pixel value of the edge confidence map of the region at the position (h, w); α represents the edge confidence threshold.

8. A dynamic remote sensing monitoring surveying and mapping method according to claim 1, characterized in that: The regional change monitoring model includes: According to the first hyperspectral image, the second hyperspectral image and the first multispectral image, the first hyperspectral upsampling feature, the second hyperspectral upsampling feature and the first multispectral feature are obtained; according to the regional change monitoring model, a spatial alignment operation is performed to generate a change mask, and the specific calculation formula is: Where, ΔF spa represents the motion feature between the first hyperspectral upsampling feature and the second hyperspectral upsampling feature; ΔF res represents the spatial difference feature between the first hyperspectral upsampling feature and the first multispectral feature; F diff represents the change mask; FLOW() represents the optical flow network; Concat() represents the dimension concatenation operation; Respectively represent the first hyperspectral upsampling feature and the second hyperspectral upsampling feature; represents the first multi-spectral feature; CNN() represents a convolutional neural network; Conv() represents a convolution operation; Represents a linear interpolation operation; The first hyperspectral image, the second hyperspectral image and the first multispectral image are input, and a second multispectral image is obtained according to the spatiotemporal fusion technology; and a change area and a change degree are generated according to the change mask and the second multispectral image.

9. A dynamic remote sensing monitoring surveying and mapping method according to claim 1, characterized in that: The implementation process of converting the change area and the change degree into the vector diagram includes: The geometric features of the changed area are obtained; according to the geometric features and the degree of change, the changed area is converted into a vector data format, a vector layer is generated and optimized to obtain a vector map; wherein the geometric features include: contour boundaries, vertex coordinates, area and shape parameters; the vector map includes: vertex coordinates, line segments and face area information; the optimization process includes removing redundant vertices, smoothing boundaries and adjusting coordinate accuracy.

10. A dynamic remote sensing monitoring mapping system, characterized in that: The surveying and mapping system comprises: The data acquisition and preprocessing module obtains multi-temporal multispectral images and multi-temporal hyperspectral images of the target area according to the composite spectral imager carried by the satellite, and performs preprocessing to obtain standard multi-temporal multispectral images and standard multi-temporal hyperspectral images; A region segmentation module, inputting the first hyperspectral image and the first multispectral image into the region segmentation model to generate a region segmentation map and a region confidence map; wherein the first hyperspectral image is the first image of the standard multi-temporal hyperspectral image, and the first multispectral image is the first image of the standard multi-temporal multispectral image; An edge refinement module, which corrects the segmentation edge of the region segmentation map according to the edge refinement model to generate a region distribution map; A regional change monitoring module, which obtains a second multispectral image and a second hyperspectral image, combines the first hyperspectral image and the first multispectral image, and generates a change area and a change degree; The surveying and mapping information updating module converts the changed area and the change degree into a vector map, replaces the vector map of the corresponding area in the historical surveying and mapping data, and generates updated surveying and mapping data.

Images