Remote sensing image change detection method and related equipment

Abstract

The embodiment of the invention discloses a remote sensing image change detection method and related equipment, and relates to the technical field of remote sensing image change detection, shallow layer features can accurately capture fine geometric changes such as building boundaries, the boundary blur problem is effectively relieved, middle layer features enhance the continuity and definition of contours, and the detection accuracy is improved. According to the method, the shallow-layer features, the middle-layer features and the deep-layer features are used for identifying semantic features, so that the phenomenon of leak detection of complex scenes such as farmland fragmentation areas is remarkably reduced, and therefore, a binary change graph of the remote sensing image is obtained according to the shallow-layer features, the middle-layer features and the deep-layer features, and the comprehensiveness and accuracy of change detection of the remote sensing image are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image change detection, and in particular to a remote sensing image change detection method and related equipment. BACKGROUND

[0002] The existing remote sensing image change detection method directly performs feature splicing or simple difference operation on the dual-time-phase images when processing dual-time-phase data. Although this method can capture part of the difference information, it lacks multi-scale analysis capability for change dynamics, resulting in problems such as blurred building boundaries and missed detection of fragmented farmland areas, which affects the accuracy of the remote sensing image change detection result. SUMMARY

[0003] Therefore, the present application provides a remote sensing image change detection method and related equipment.

[0004] The specific technical scheme of the first embodiment of the present application is as follows: a remote sensing image change detection method, the method comprising: acquiring dual-time-phase remote sensing images of a region to be detected; the dual-time-phase remote sensing image data comprising a first remote sensing image at a first time and a second remote sensing image at a second time; extracting a first feature set of the first remote sensing image; the first feature set comprising a first shallow feature, a first middle feature and a first deep feature; the shallow feature comprising an edge feature or a texture feature, the middle feature comprising a structural contour feature, and the deep feature comprising a semantic abstract feature, the semantic abstract feature comprising a land cover type, a land cover type change pattern, a city expansion direction and a farmland rotation rule; extracting a second feature set of the second remote sensing image; the second feature set comprising a second shallow feature, a second middle feature and a second deep feature; and obtaining a remote sensing image binary change map of the dual-time-phase remote sensing images according to the first feature set and the second feature set.

[0005] Preferably, the obtaining of the remote sensing image binary change map of the dual-time-phase remote sensing images according to the first feature set and the second feature set further comprises: fusing a preset shallow prompt vector with the first shallow feature and the second shallow feature respectively to obtain a first shallow fusion feature and a second shallow fusion feature; fusing a preset middle prompt vector with the first middle feature and the second middle feature respectively to obtain a first middle fusion feature and a second middle fusion feature; fusing a preset deep prompt vector with the first deep feature and the second deep feature respectively to obtain a first deep fusion feature and a second deep fusion feature; and obtaining the remote sensing image binary change map according to the first shallow fusion feature, the second shallow fusion feature, the first middle fusion feature, the second middle fusion feature, the first deep fusion feature and the second deep fusion feature.

[0006] Preferably, the obtaining of the binary change map of the remote sensing image according to the first shallow fusion feature, the second shallow fusion feature, the first middle fusion feature, the second middle fusion feature, the first deep fusion feature and the second deep fusion feature comprises: obtaining a shallow difference feature according to the first shallow fusion feature and the second shallow fusion feature; obtaining a middle difference feature according to the first middle fusion feature and the second middle fusion feature; obtaining a deep difference feature according to the first deep fusion feature and the second deep fusion feature; and obtaining the binary change map of the remote sensing image according to the shallow difference feature, the middle difference feature and the deep difference feature.

[0007] Preferably, the obtaining of the binary change map of the remote sensing image according to the shallow difference feature, the middle difference feature and the deep difference feature comprises: splicing and fusing the shallow difference feature, the middle difference feature and the deep difference feature to obtain a difference feature map; and decoding the difference feature map to obtain the binary change map of the remote sensing image.

[0008] Preferably, the first shallow fusion feature, the first middle fusion feature and the first deep fusion feature are obtained by using the following formula:

[0009]

[0010] wherein, F s ′ 1 is the first shallow fusion feature, is the first middle fusion feature, F d ′ 1 is the first deep fusion feature, is the first shallow feature, is the first middle feature, is the first deep feature, P s is the preset shallow prompt vector, P m is the preset middle prompt vector, P d is the preset deep prompt vector, and a is a preset learnable scalar, σ is a Sigmoid function, Conv is a convolution operation, and MLP is a projection alignment dimension operation.

[0011] Preferably, the difference feature map is obtained by using the following formula:

[0012] ΔF = ConvAttn(|Up(ΔF s )||Up(ΔF m )||Up(ΔF d )|

[0013] Wherein, Delta F is the difference feature map, ConvAttn is a light convolution operation and an attention operation, Up is an up-sampling operation, Delta F s is the shallow difference feature, Delta F m is the middle difference feature, Delta F d is the deep difference feature.

[0014] Preferably, the land cover type change mode includes changing a building in the land cover type into bare land and changing a farmland in the land cover type into a greenhouse.

[0015] The specific technical scheme of the second embodiment of the application is: a remote sensing image change detection system, comprising: a remote sensing image acquisition module, a first feature extraction module, a second feature extraction module and a change identification module; the remote sensing image acquisition module is used to acquire double-time-phase remote sensing images of a region to be detected; the double-time-phase remote sensing image data comprises a first remote sensing image at a first time and a second remote sensing image at a second time; the first feature extraction module is used to extract a first feature set of the first remote sensing image; the first feature set comprises a first shallow feature, a first middle feature and a first deep feature; the shallow feature comprises an edge feature or a texture feature, the middle feature comprises a structure contour feature, and the deep feature comprises a semantic abstract feature, wherein the semantic abstract feature comprises a land cover type, a land cover type change mode, a city expansion direction and a farmland rotation rule; the second feature extraction module is used to extract a second feature set of the second remote sensing image; the second feature set comprises a second shallow feature, a second middle feature and a second deep feature; and the change identification module is used to obtain a remote sensing image binary change map of the double-time-phase remote sensing image according to the first feature set and the second feature set.

[0016] The specific technical scheme of the third embodiment of the application is: a remote sensing image change detection device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the method according to any one of the first embodiment of the application.

[0017] The specific technical scheme of the fourth embodiment of the application is: a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the method according to any one of the first embodiment of the application.

[0018] The implementation of the embodiments of the application has the following beneficial effects:

[0019] The shallow feature in the application can accurately capture subtle geometric changes such as building boundaries, effectively alleviate the boundary ambiguity problem, the middle feature enhances the continuity and clarity of the contour, and the deep feature significantly reduces the missed detection phenomenon of complex scenes such as farmland fragmentation area by identifying semantic features, so that the remote sensing image binary change graph is obtained according to the three features of the shallow feature, the middle feature and the deep feature, and the comprehensiveness and accuracy of the remote sensing image change detection are ensured. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 The step flow chart of the remote sensing image change detection method is shown in the figure.

[0022] Figure 2 The structural schematic diagram of the remote sensing image change detection system is shown in the figure.

[0023] Figure 3 The structural schematic diagram of the remote sensing image change detection system is shown in the figure.

[0024] Figure 4 The effect diagram of the binary change graph obtained by different methods is shown in the figure.

[0025] Figure 5 The internal structure diagram of the computer device is shown in the figure.

[0026] Among them, 201, remote sensing image acquisition module; 202, first feature extraction module; 203, second feature extraction module; 204, change recognition module. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0028] The terms "first", "second", and the like in the description and in the claims of the present application and in the drawings refer to different objects without necessarily implying a specific order. Also, the terms "comprising", "having", "including", and the like in the description and in the claims of the present application are to be construed open-ended, unless otherwise indicated. For example, a process, method, object, or apparatus that comprises a list of steps or modules is not necessarily limited to those steps or modules which are recited. Rather, the process, method, object, or apparatus optionally includes those steps or modules in addition to those which are listed, and further includes other steps or modules which are not expressly listed or included.

[0029] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase "in an embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. It is expressly understood that any of the embodiments described herein can be combined with any of the other embodiments unless specifically noted otherwise.

[0030] Reference is made to Figure 1 A flowchart of the steps of a remote sensing image change detection method according to a first embodiment of the application is shown in Figure 1 to ensure the comprehensiveness and accuracy of remote sensing image change detection. The method comprises the following steps:

[0031] In step 101, a dual-time remote sensing image of a region to be detected is obtained. The dual-time remote sensing image data comprises a first remote sensing image at a first time and a second remote sensing image at a second time.

[0032] In step 102, a first feature set of the first remote sensing image is extracted. The first feature set comprises a first shallow feature, a first middle feature, and a first deep feature. The shallow feature comprises an edge feature or a texture feature, the middle feature comprises a structural contour feature, and the deep feature comprises a semantic abstract feature. The semantic abstract feature comprises a land cover type, a land cover type change pattern, a city expansion direction, and a farmland rotation rule.

[0033] In step 103, a second feature set of the second remote sensing image is extracted. The second feature set comprises a second shallow feature, a second middle feature, and a second deep feature.

[0034] In step 104, a remote sensing image binary change map of the dual-time remote sensing image is obtained according to the first feature set and the second feature set.

[0035] Specifically, first, the dual-time remote sensing data of the region to be detected is obtained through a satellite or a drone platform, including a first remote sensing image at a first time and a second remote sensing image at a second time. The two images need to meet the same spatial resolution, the same coverage range, and have been radiometrically corrected and geometrically registered to ensure the accuracy of subsequent feature matching. For the first remote sensing image, a convolutional neural network (CNN) is used for feature interpretation: the first shallow feature is extracted through a shallow convolution kernel, including basic geometric information such as edge gradient and texture direction; the first middle layer feature is captured by using a middle layer network module, focusing on analyzing structural elements such as building outlines and road topologies; the first deep layer feature is mined based on a deep layer semantic encoder, outputting advanced semantic information such as land cover types (such as forest land, water body, and construction land), land cover type change patterns (such as vegetation degradation and water expansion), urban expansion directions (identified through spatial clustering analysis to identify the main growth axis), and farmland rotation rules (classified based on time-series NDVI curves). Similarly, the same operation is performed on the second remote sensing image to obtain a second feature set. Finally, a change map is generated based on the feature differences in the first feature set and the second feature set, and a remote sensing image binary change map is output using the difference generation change map after threshold segmentation and morphological filtering processing, where the white area represents the pixels that have changed significantly, and the black area represents the unchanged area.

[0036] The shallow features in the method can accurately capture subtle geometric changes such as building boundaries, effectively alleviating the boundary ambiguity problem, the middle layer features enhance the continuity and clarity of the outlines, and the deep layer features significantly reduce the missed detection phenomenon in complex scenes such as farmland fragmentation by identifying semantic features. Therefore, the remote sensing image binary change map is obtained based on the three features of shallow features, middle layer features, and deep layer features, ensuring the comprehensiveness and accuracy of remote sensing image change detection.

[0037] In specific embodiments, obtaining the remote sensing image binary change map of the dual-time remote sensing image based on the first feature set and the second feature set further includes: fusing a preset shallow prompt vector with the first shallow feature and the second shallow feature respectively to obtain first shallow fusion features and second shallow fusion features; fusing a preset middle layer prompt vector with the first middle layer feature and the second middle layer feature respectively to obtain first middle layer fusion features and second middle layer fusion features; fusing a preset deep layer prompt vector with the first deep layer feature and the second deep layer feature respectively to obtain first deep layer fusion features and second deep layer fusion features; and obtaining the remote sensing image binary change map based on the first shallow fusion features, the second shallow fusion features, the first middle layer fusion features, the second middle layer fusion features, the first deep layer fusion features, and the second deep layer fusion features.

[0038] Specifically, for shallow feature fusion, a preset shallow prompt vector is designed, which contains an edge direction distribution template or a texture frequency weight map. The shallow prompt vector is fused with the first shallow feature (edge gradient, texture direction) and the second shallow feature through a channel-by-channel convolution operation to generate the first shallow fusion feature and the second shallow fusion feature. The fusion process adopts an attention weighting mechanism, which gives higher weight to the area where the edge response is more prominent, while suppressing noise interference. For mid-level feature fusion, a preset mid-level prompt vector is used to dynamically fuse the mid-level prompt vector with the first mid-level feature (building contour, road topology) and the second mid-level feature to generate the first mid-level fusion feature and the second mid-level fusion feature. The fusion process emphasizes the closure and direction continuity of the contour, solving the structure fracture problem caused by changes in viewing angle. In the deep feature fusion stage, a preset deep prompt vector is used to fuse the deep prompt vector with the first deep feature (land cover type, change pattern) and the second deep feature to generate the first deep fusion feature and the second deep fusion feature. The fusion process highlights areas with strong semantic consistency and suppresses false positives caused by seasonal differences. The six groups of fusion features are input into a multi-scale decision fusion network, and cross-level feature interaction is realized through 3D convolution. Finally, the difference generation change map is output through the Sigmoid activation function, and the Otsu algorithm is used to automatically determine the threshold to generate the remote sensing image binary change map. By injecting domain knowledge through prompt vectors, the performance of change detection in complex scenes is effectively improved.

[0039] In specific embodiments, the remote sensing image binary change map is obtained according to the first shallow fusion feature, the second shallow fusion feature, the first mid-level fusion feature, the second mid-level fusion feature, the first deep fusion feature and the second deep fusion feature, comprising: obtaining a shallow difference feature according to the first shallow fusion feature and the second shallow fusion feature; obtaining a mid-level difference feature according to the first mid-level fusion feature and the second mid-level fusion feature; obtaining a deep difference feature according to the first deep fusion feature and the second deep fusion feature; obtaining the remote sensing image binary change map according to the shallow difference feature, the mid-level difference feature and the deep difference feature.

[0040] Specifically, the absolute value of the difference between the first shallow fusion feature and the second shallow fusion feature is the shallow difference feature, the absolute value of the difference between the first mid-level fusion feature and the second mid-level fusion feature is the mid-level difference feature, and the absolute value of the difference between the first deep fusion feature and the second deep fusion feature is the deep difference feature. The shallow, mid-level and deep difference feature maps are input into a decision network and subjected to multi-scale fusion and threshold segmentation to finally output the remote sensing image binary change map.

[0041] In specific embodiments, the obtaining the remote sensing image binary change map according to the shallow difference feature, the middle difference feature and the deep difference feature comprises: splicing and fusing the shallow difference feature, the middle difference feature and the deep difference feature to obtain a difference feature map; and decoding the difference feature map to obtain the remote sensing image binary change map.

[0042] Specifically, the shallow difference feature, the middle difference feature and the deep difference feature are spliced in the channel dimension to form a three-dimensional difference feature map. A progressive decoding network structure is designed. The decoder adopts an encoder-decoder symmetric architecture and is composed of three up-sampling modules in series: in the first stage, the resolution of the difference feature map is increased to 1 / 4 of the original image through transposed convolution, while the preset spatial detail information of the shallow encoder is fused; in the second stage, the receptive field is expanded by using the atrous convolution to combine the middle layer feature to strengthen the structural continuity; in the third stage, the original resolution is restored through bilinear interpolation, and the preset deep semantic constraint is introduced to ensure the consistency of the change type. Residual connection is used in each stage to alleviate the gradient disappearance problem and improve the feature propagation efficiency. After the probability map output by the decoder is mapped to the [0, 1] interval through the Sigmoid activation function, the adaptive Otsu algorithm is used to determine the best segmentation threshold. This algorithm automatically distinguishes the change / non-change areas by maximizing the inter-class variance, and finally outputs a high-precision remote sensing image binary change map, in which the white area represents the detected feature change and the black area represents the unchanged background.

[0043] In specific embodiments, the first shallow fusion feature, the first middle fusion feature and the first deep fusion feature are obtained by using the following formula:

[0044]

[0045] wherein, F s ′ 1 is the first shallow fusion feature, is the first middle fusion feature, F d ′ 1 is the first deep fusion feature, is the first shallow feature, is the first middle feature, is the first deep feature, P s is the preset shallow prompt vector, P m is the preset middle prompt vector, P d is the preset deep prompt vector, and a is a preset learnable scalar, σ is a Sigmoid function, Conv is a convolution operation, and MLP is a projection alignment dimension operation.

[0046] Specifically, preset shallow, middle and deep prompt vectors are introduced and fused with the corresponding level features. These prompt vectors can contain prior knowledge or relevant information of specific tasks. Through the fusion with the features, additional information can be injected into the model to guide the model to pay more attention to the key features related to change detection, which helps to improve the accuracy and pertinence of change detection. In the calculation of the middle layer fusion feature, convolution operation and Sigmoid function are used. Convolution operation can extract local features and enhance the correlation between features, and Sigmoid function can map the output to the interval [0, 1], which plays a role in normalization and noise suppression. This nonlinear transformation can explore the complex relationship between features, enhance useful feature information, and reduce irrelevant information interference, improving the quality and expression ability of features. In the calculation of the deep layer fusion feature, the projection alignment dimension operation MLP is used. This operation can ensure that the features of different levels can be effectively fused in the dimension. By projecting the deep prompt vector to a suitable dimension space, it can be smoothly added to the deep feature , ensuring the feasibility and stability of feature fusion, and avoiding calculation errors or information loss caused by dimension mismatch. Using the formula in this embodiment, the final difference feature map can more accurately reflect the change information between the two-phase remote sensing images, thereby laying a good foundation for subsequent generation of high-precision remote sensing image binary change map, effectively improving the overall performance of change detection, including improving the accuracy, recall rate and F1 value of change detection, and better meeting the demand for remote sensing image change detection in actual application.

[0047] In specific embodiments, the difference feature map is obtained using the following formula:

[0048] ΔF=ConvAttn(|Up(ΔF s )||Up(ΔF m )||Up(ΔF d )|)

[0049] Where ΔF is the difference feature map, ConvAttn is the lightweight convolution operation and attention operation, Up is the up-sampling operation, ΔF s is the shallow difference feature, ΔF m is the middle difference feature, and ΔF d is the deep difference feature. The combination of lightweight convolution and attention mechanism reduces the model complexity while ensuring the feature expression ability, which is suitable for real-time processing of large-scale remote sensing data.

[0050] In specific embodiments, the land cover type change pattern includes changing a building in the land cover type to bare land, and changing a farmland in the land cover type to a greenhouse. Specifically, to enhance the detection capability of specific land cover type change patterns, a targeted optimization mechanism is introduced in the semantic abstract feature extraction stage. Among them, in the deep feature extraction module, special feature encoders are designed for two typical change patterns of building changing to bare land and farmland changing to greenhouse: building changing to bare land passes through a spectral-texture joint encoder to capture spectral mutation and texture homogenization features after building demolition, such as the increase of infrared band reflectivity, and a pattern recognition model is trained in combination with historical demolition case data; farmland changing to greenhouse adopts a structure-vegetation double-branch encoder to extract regular geometric contour and vegetation index decline features of the greenhouse, such as rectangular or arc-shaped roof, and the spatial layout prior of greenhouse construction is injected by using transfer learning, which significantly improves the detection capability of typical land cover type changes such as building changing to bare land and farmland changing to greenhouse, and avoids the mis-detection problem caused by pattern generalization.

[0051] Specifically, the detailed embodiments of the remote sensing image change detection method are as follows:

[0052] S1: Remote sensing image input

[0053] Dual-time remote sensing images and (size H x W x C). and denote dual-time input images (time phase 1 and time phase 2), H, W, and C are the height, width, and channel number of the image, respectively.

[0054] S2: Basic feature extraction

[0055] The HyperSIGMA backbone network with shared weights outputs three-stage features:

[0056] Shallow features (edge / texture);

[0057] Middle features (structural contour);

[0058] Deep features (semantically abstract);

[0059] wherein D s , D m , and D d are the channel numbers of the feature maps of each stage of the backbone network, and their values depend on the architecture design of HyperSIGMA, and are 64, 128, and 256 by default.

[0060] Deep features refer to the characteristic expression of high-level land cover categories and transformation relationships, for example:

[0061] Land cover types: semantic category distribution of buildings, farmland, bare land, etc.;

[0062] Change patterns: Buildings → Bare land (demolition); Farmland → Greenhouse (land use conversion);

[0063] Scene-level understanding: urban expansion direction, crop rotation patterns; using t-SNE clustering (a non-linear dimensionality reduction visualization technique used to project high-dimensional data (such as deep learning features) into two- or three-dimensional space while preserving local similarity relationships between data points) to display deep features separated by semantic categories (buildings / farmland / background).

[0064] S3: TriPrompt encoding: Shallow hints P s =MLP(p s ); Middle layer prompt P m =ConvBlock(p m ); Deeper hints P d =Transformer(p d ); where p s p m p d Initialize the learnable cue vectors for each layer, P s P m P d This refers to the prompt vectors output by each layer.

[0065] S4: Layered hint fusion:

[0066] Shallow layer: F s ′=α·F s +(1-α)·P s ;

[0067] Middle layer: β=σ(Conv(|F m ||P m |))→F′ m =β·F m +(1-β)·P m ;

[0068] Deep: F d ′=F d +MLP(P d );

[0069] S5: Stage Difference Calculation: ΔF i =|F i ′ 1 -F i ′ 2|, i∈{s, m, d}; where, F i 1 ,F i 2 Refers to the characteristics of s, m, d (shallow, middle, deep) of the images of phase 1, phase 2, and the superscripts 1 and 2 refer to the before and after phases.

[0070] S6: Multi-scale fusion: ΔF = ConvAttn(|Up(ΔF s )||Up(ΔF m )||Up(ΔF d )|).

[0071] S7: Output: The decoder generates a binary change map Y∈{0,1} H×W (Sigmoid activation). For comparison of different methods to obtain binary change maps, please refer to Figure 4 , and for comparison of the prediction accuracy of different models, please refer to Table 1. As can be seen from Table 1, the prediction accuracy of the model (RFHP-CD) after implementing the method in the embodiment is significantly improved.

[0072] OA (%) UA (%) PA (%) IoU (%) F1(％) STANet 97.1 86.29 90.46 78.69 88.13 HANet 97.31 86.77 90.69 79.27 88.97 Changer 97.85 92.66 86.73 80.15 89.31 BiT 97.92 92.7 88.44 81.23 90.11 SNUNet 97.21 87.62 92.71 81.1 89.75 SAM-CD 97.45 93.37 84.94 81.13 90.3 RFHP-CD 98.54 91.97 90.61 81.45 91.15

[0073] Table 1: Accuracy of different models in implementation

[0074] In specific embodiments, please refer to Figure 2 and Figure 3 are a structural schematic diagram and a system framework schematic diagram of a remote sensing image change detection system in a second embodiment of the application, and the system comprises: a remote sensing image acquisition module 201, a first feature extraction module 202, a second feature extraction module 203, and a change identification module 204; the remote sensing image acquisition module 201 is configured to acquire double-time-phase remote sensing images of a region to be detected; the double-time-phase remote sensing image data comprises a first remote sensing image at a first time and a second remote sensing image at a second time; the first feature extraction module 202 is configured to extract a first feature set of the first remote sensing image; the first feature set comprises a first shallow feature, a first middle feature, and a first deep feature; the shallow feature comprises an edge feature or a texture feature, the middle feature comprises a structural contour feature, and the deep feature comprises a semantic abstract feature, and the semantic abstract feature comprises a land cover type, a land cover type change pattern, a city expansion direction, and a farmland rotation rule; the second feature extraction module 203 is configured to extract a second feature set of the second remote sensing image; the second feature set comprises a second shallow feature, a second middle feature, and a second deep feature; and the change identification module 204 is configured to obtain a remote sensing image binary change map of the double-time-phase remote sensing images according to the first feature set and the second feature set.

[0075] ​​The shallow features in the system can accurately capture subtle geometric changes such as building boundaries, effectively alleviate the boundary ambiguity problem, the middle features enhance the continuity and clarity of the contour, and the deep features significantly reduce the missed detection phenomenon of complex scenes such as farmland fragmentation area by identifying semantic features, so that the remote sensing image binary change graph is obtained according to the three features of the shallow features, the middle features and the deep features, and the comprehensiveness and accuracy of the remote sensing image change detection are ensured.

[0076] In specific embodiments, the third embodiment of the present application provides a remote sensing image change detection device, comprising a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the method in any one of the first embodiments of the present application.

[0077] In specific embodiments, the fourth embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the method in any one of the first embodiments of the present application.

[0078] Figure 5 The internal structure diagram of the computer device in an embodiment is shown. The computer device can be a terminal or a server. Please refer to Figure 5 , the computer device includes a processor, a memory and the like connected by a system bus. Among them, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system, and can also store a computer program, which is executed by the processor to make the processor realize the method in the embodiment. The internal memory can also store a computer program, which is executed by the processor to make the processor execute the method in the embodiment. Those skilled in the art can understand Figure 5 the structure shown in the figure, only the block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or less components than those shown in the figure, or combine certain components, or have different component arrangements.

[0079] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the patent scope of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

[0080] The above merely describes preferred embodiments of the present application, but is not intended to limit the present application to other forms, any person skilled in the art can make changes or modifications to the above disclosed technical contents into equivalent embodiments with equivalent changes, and apply to other fields, but any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solution content of the present application still belongs to the protection scope of the present application technical solution.

Claims

1. A method for detecting changes in remote sensing images, characterized in that, The method includes: Acquire dual-temporal remote sensing images of the area to be detected; the dual-temporal remote sensing image data includes a first remote sensing image at a first time and a second remote sensing image at a second time. Extract a first feature set from the first remote sensing image; the first feature set includes a first shallow feature, a first medium feature, and a first deep feature; the shallow feature includes edge features or texture features, the medium feature includes structural contour features, and the deep feature includes semantic abstract features, which include land cover type, land cover type change pattern, urban expansion direction, and crop rotation pattern; Extract a second feature set from the second remote sensing image; the second feature set includes a second shallow feature, a second middle feature, and a second deep feature. The binary transformation map of the dual-temporal remote sensing image is obtained based on the first feature set and the second feature set.

2. The remote sensing image change detection method as described in claim 1, characterized in that, The step of obtaining the binary transformation map of the dual-temporal remote sensing image based on the first feature set and the second feature set further includes: The preset shallow cue vector is fused with the first shallow feature and the second shallow feature respectively to obtain the first shallow fused feature and the second shallow fused feature; The preset mid-level cue vector is fused with the first mid-level feature and the second mid-level feature respectively to obtain the first mid-level fused feature and the second mid-level fused feature; The preset deep cue vector is fused with the first deep feature and the second deep feature respectively to obtain the first deep fused feature and the second deep fused feature; The binary transformation map of the remote sensing image is obtained based on the first shallow fusion feature, the second shallow fusion feature, the first middle fusion feature, the second middle fusion feature, the first deep fusion feature, and the second deep fusion feature.

3. The remote sensing image change detection method as described in claim 2, characterized in that, The step of obtaining the binary transformation map of the remote sensing image based on the first shallow fusion feature, the second shallow fusion feature, the first mid-layer fusion feature, the second mid-layer fusion feature, the first deep fusion feature, and the second deep fusion feature includes: The shallow layer difference features are obtained based on the first shallow layer fusion features and the second shallow layer fusion features; The middle-layer difference features are obtained based on the first middle-layer fusion feature and the second middle-layer fusion feature; Deep difference features are obtained based on the first deep fusion feature and the second deep fusion feature; The binary transformation map of the remote sensing image is obtained based on the shallow layer difference features, the middle layer difference features, and the deep layer difference features.

4. The remote sensing image change detection method as described in claim 3, characterized in that, The step of obtaining the binary transformation map of the remote sensing image based on the shallow layer difference features, the middle layer difference features, and the deep layer difference features includes: The shallow layer difference features, the middle layer difference features, and the deep layer difference features are spliced ​​and fused to obtain a difference feature map; The difference feature map is decoded to obtain the binary transformation map of the remote sensing image.

5. The remote sensing image change detection method as described in claim 2, characterized in that, The first shallow layer fusion feature, the first middle layer fusion feature, and the first deep layer fusion feature are obtained using the following formula: in, This is the first shallow fusion feature. This is the first mid-layer fusion feature. This is the first deep fusion feature. This is the first shallow layer feature. This is the first middle-layer feature. For the first deep feature, P s For the preset shallow hint vector, P m P is the preset mid-level prompt vector. d Let α be the preset deep cue vector, σ be the preset learnable scalar, σ be the Sigmoid function, Conv be the convolution operation, and MLP be the projection-aligned dimension operation.

6. The remote sensing image change detection method as described in claim 4, characterized in that, The difference feature map is obtained using the following formula: ΔF=ConvAttn(|Up(ΔF s )||Up(ΔF m )||Up(ΔF d )|) Where ΔF is the differential feature map, ConvAttn is the lightweight convolution operation and attention operation, Up is the upsampling operation, and ΔF s For the shallow layer difference features, ΔF m For mid-layer differences, ΔF d This represents a deep-seated difference in characteristics.

7. The remote sensing image change detection method as described in claim 1, characterized in that, The land cover type change pattern includes changing buildings in the land cover type to bare land and changing farmland in the land cover type to greenhouses.

8. A remote sensing image change detection system, characterized in that, The system includes: a remote sensing image acquisition module, a first feature extraction module, a second feature extraction module, and a change recognition module; The remote sensing image acquisition module is used to acquire dual-temporal remote sensing images of the area to be detected; the dual-temporal remote sensing image data includes a first remote sensing image at a first time and a second remote sensing image at a second time. The first feature extraction module is used to extract a first feature set from the first remote sensing image; the first feature set includes a first shallow feature, a first medium feature, and a first deep feature; the shallow feature includes edge features or texture features, the medium feature includes structural contour features, and the deep feature includes semantic abstract features, which include land cover type, land cover type change pattern, urban expansion direction, and crop rotation pattern. The second feature extraction module is used to extract a second feature set from the second remote sensing image; the second feature set includes a second shallow feature, a second mid-level feature, and a second deep feature. The change recognition module is used to obtain a binary change map of the remote sensing image of the dual-temporal remote sensing image based on the first feature set and the second feature set.

9. A remote sensing image change detection device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images