Land space interference element identification method for cascade safety bottom line scene classification

Abstract

The invention discloses a territorial space interference element identification method for cascade safety bottom line scene classification, and the method comprises the steps: obtaining and constructing a safety bottom line scene detection network BaseNet, carrying out the iterative training through an enhanced safety bottom line scene semantic recognition sample set, obtaining a safety bottom line scene detection model, and carrying out the reasoning, obtaining a safety bottom line target vector result set; constructing a territorial space interference element instance segmentation network InterfNet, and performing iterative training by using the enhanced interference element instance segmentation sample set to obtain a territorial space interference element instance segmentation network model; and based on the safety bottom line target vector result set and the post-phase high-resolution remote sensing image, utilizing the territorial space interference element instance segmentation network model to obtain a final territorial space interference element target set under the control of the safety bottom line. The method has the advantages that the technical problem in a traditional method is solved, and good data support and technical guarantee are provided for constructing a monitoring network for territorial space planning and guaranteeing high-quality development of territorial space.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing land space interference element target detection, and particularly relates to a land space interference element identification method based on cascaded safety bottom line scene classification. BACKGROUND

[0002] With the successful development and completion of the first round of land space planning system construction and land space planning compilation and approval work, the focus of planning management work has gradually shifted to implementation supervision, and the effective implementation of land space planning supervision is mainly to quickly and accurately monitor various elements (land space interference elements) that affect the implementation of land space planning, efficiently identify various human activities, and changes in the earth's ecological environment and natural conditions, and to promote high-quality development of land space and improve the effectiveness of land space planning implementation.

[0003] With the rapid development of remote sensing technology, the advantages of non-contact, low cost and wide monitoring are rapidly promoting the efficient identification and monitoring of land space interference elements. However, traditional identification methods mainly use GIS overlay analysis and static threshold determination, which are difficult to adapt to the needs of dynamic interference element identification in complex geographical environments. Domestic and foreign research shows that about 67% of land space conflicts are caused by unclear identification of interference elements in multi-bottom line crossing areas, which not only fails to fully utilize the advantages of remote sensing applications, but also seriously affects the construction process of China's land space planning implementation network.

[0004] The rapid development of artificial intelligence technology has also rapidly promoted the revolutionary reform of remote sensing and land space interference element identification. Based on the construction of a deep change detection network model, multi-temporal land space interference elements are quickly identified. However, due to the scattered characteristics of remote sensing targets of interference elements, the features affecting the interference category cannot be well learned and trained, resulting in only basic general interference element detection in complex safety bottom line scenarios. Meanwhile, due to the interference of complex natural backgrounds, the accuracy cannot meet the requirements of engineering applications.

[0005] Therefore, in view of the above problems, directly using the conventional deep change detection network framework for interference element target detection ignores the constraints and control of safety bottom line scenarios, and it is difficult to utilize the feature enhancement capability under the coupling of scene knowledge. It is necessary to further cascade the land interference element activity space constraint capability and integrate it into the land space interference element target state identification process to overcome complex backgrounds and specialize in feature state learning, further improving the boundary accuracy and type refinement identification of interference elements. SUMMARY

[0006] The purpose of the present invention is to provide a method for identifying land space interference elements by cascading safety bottom line scenario classification, thereby solving the aforementioned problems existing in the prior art.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A method for identifying interference elements in land space by cascading safety bottom line scenario classification includes the following steps:

[0009] S1. Acquisition of enhanced safety bottom line scene semantic recognition sample set: Based on the previous phase high-resolution remote sensing image I Pre and subsequent high-resolution remote sensing images I Post Obtaining a secure hybrid image data source I 安全底线 , using hybrid image data sources based on security baseline I 安全底线 The image data source I′ obtained to accurately identify the safety bottom line 安全底线 and safety baseline scenario labeling data VEC scene Create a safety bottom line scene recognition sample set S base_label , identify the sample set S for the safety bottom line scenario base_label Perform enhancement processing to generate enhanced safety bottom line scene semantic recognition sample set S′ base_label ;

[0010] S2. Safety baseline scenario detection model construction: Construct a convolutional self-attention safety baseline scenario detection network BaseNet, and use the enhanced safety baseline scenario semantic recognition sample set S′ base_label Perform iterative training to obtain the safety bottom line scenario detection model M Base And reasoning to obtain the safety bottom line target vector result set R in the demonstration area Base ;

[0011] S3. Obtaining the enhanced interference factor instance segmentation sample set: Using the post-phase high-resolution remote sensing image I Post and interference factor target vector label data VEC Inf Make interference factor instance segmentation sample set S ele_label ; Segment the sample set S for the interference factor instance ele_label Perform enhancement processing to generate enhanced interference factor instance segmentation sample set S′ ele_label ;

[0012] S4. Construction of the network model for instance segmentation of national space interference elements: Construct the national space interference element instance segmentation network InterfNet based on the deep visual basic model, and use the enhanced interference element instance segmentation sample set S′ ele_label Perform iterative training to obtain the national land space interference element instance segmentation network model M Inf ;

[0013] S5, obtaining a target set of land space interference elements under the safety bottom line control: based on the safety bottom line target vector result set R Base and a post-phase high-resolution remote sensing image I Post using a land space interference element instance segmentation network model M Inf performing reasoning and labeling to obtain a safety bottom line target vector result set R Base a set of interference element semantic elements Rset in a constrained scenario Inf and based on the set of interference element semantic elements Rset Inf obtaining the final target set of land space interference elements under the safety bottom line control Rset' Inf .

[0014] Preferably, step S1 specifically includes the following contents,

[0015] S11, performing low-frequency information preserving downsampling processing on the pre-phase high-resolution remote sensing image I Pre and the post-phase high-resolution remote sensing image I Post to obtain results I Pre4 and I Post4 after 4 times downsampling, to jointly construct a safety bottom line mixed image data source I 安全底线 ;

[0016] S12, fusing and analyzing the safety bottom line mixed image data source I 安全底线 and safety bottom line vector auxiliary data to eliminate false safety bottom line image content, forming an available accurate safety bottom line recognition image data source I' 安全底线 ;

[0017] S13, taking the available accurate safety bottom line recognition image data source I' 安全底线 and safety bottom line scene marking data VEC scene as data input, traversing the image and marking vector data with a certain spatial range and step to generate a safety bottom line scene recognition sample set S base_label ;

[0018] S14, performing enhancement processing on the safety bottom line scene recognition sample set S base_label to generate a strengthened safety bottom line scene semantic recognition sample set S' base_label .

[0019] Preferably, the safety bottom line vector auxiliary data is a planar polygon data range, and the content range it points to is the urban boundary, the rural boundary, and the factory land, which are the frequent activity ranges of human activities;

[0020] Step S12 is specifically to generate the safety bottom line vector auxiliary data and the safety bottom line mixed image data source I 安全底线 Auxiliary weights T of the same size 辅助权重 , by combining the security baseline mixed image data source I 安全底线 With auxiliary weight T 辅助权重 By superimposing and multiplying, we can obtain an image data source I′ that can accurately identify the safety bottom line. 安全底线 ;

[0021]

[0022] I′ 安全底线 =I 安全底线 °T 辅助权重

[0023] Among them, when the safety bottom line vector auxiliary data is located within the auxiliary polygon, the auxiliary weight T 辅助权重 The value of is 0.85; otherwise, the auxiliary weight T 辅助权重 The value of is 0.15.

[0024] Preferably, the enhancement process includes single-sample random enhancement and multi-sample mosaic semantic enhancement;

[0025] The single sample random enhancement includes geometric transformation and radiation transformation; geometric transformation includes random rotation around the sample center, longitudinal or transverse axis symmetry transformation; radiation transformation includes blurring, brightness, contrast, sharpening, and adding noise;

[0026] Obtain the enhanced safety baseline scene recognition sample set S′ base_label The geometric transformation involved in the single-sample random enhancement is to randomly rotate the sample image counterclockwise by a certain angle θ starting from the center point of the sample image itself, and perform the rotation enhancement processing on the sample image at intervals of 30 degrees, thereby preserving the original size. The 360 ​​degrees are divided into 12 parts, and one of the angles θ is randomly selected for each enhancement.

[0027]

[0028] θ∈[0,30,60,90,120,150,180]

[0029] Obtain the enhanced interference factor instance segmentation sample set S′ ele_label The geometric transformation involved in the single-sample random enhancement is to randomly rotate the sample image counterclockwise by a certain angle α starting from the center point of the sample image itself, and perform the rotation enhancement processing on the sample image at intervals of 15 degrees, thereby preserving the original size. The 360 ​​degrees are divided into 24 parts, and one of the angles α is randomly selected for each enhancement.

[0030]

[0031] a e [0, 15, 30, 45, 60, 75, 90, 105, 120, 135, 150, 165, 180]

[0032] wherein, S img is the sample picture matrix to be enhanced; θ and a are the rotation angles; S' img is the sample picture matrix after rotation enhancement;

[0033] The multi-sample mosaic semantic enhancement includes three modes of enhancement, i.e., 1*3, 2*2 and 3*3; the three fused enhanced samples are arranged in a random manner, and the part without real semantics is filled with background 0.

[0034] Preferably, the backbone network part of the convolutional self-attention safety baseline scene detection network BaseNet is jointly constructed by the networks of ResNet18 and VIT-Base; the classification head network adopts a 4-class semantic segmentation network, forming a specific semantic segmentation model suitable for safety baseline scene detection;

[0035] In the convolutional self-attention safety baseline scene detection network BaseNet, after the input data is subjected to 8-layer feature extraction by ResNet18, the feature map is converted into an Embedding feature block consistent with the input dimension of the VIT network through 2-dimensional convolution, and then is connected to the Transformer Encoder module, and finally is subjected to the MLP network to output the final classification scene semantic single-channel 4-class scene value feature probability map, completing the forward calculation process of the sample data in the network.

[0036] The process of obtaining the safety baseline scene detection model M Base During the iteration training of the convolutional self-attention safety baseline scene detection network BaseNet, the total training iteration number epoch is 7, the initial learning rate is 0.001, and the learning rate is optimized and adjusted in a linear preheating and cosine annealing manner as the iteration training proceeds.

[0037] Preferably, the land space interference element instance segmentation network InterfNet adopts Swin-Large as the feature backbone network to extract sample image features, which includes three head networks for detecting the spatial position, category and semantic mask information of the interference element.

[0038] The instance segmentation target category is set to 5, and the semantic segmentation graph in each positioning frame is a binary segmentation feature graph, which extracts the target semantic information of the interference element.

[0039] Preferably, the safety baseline scene recognition sample set S base_label and the safety baseline target vector result set R BaseAll of them have the safety bottom line scenario category attribute, with mark 1 representing the safety bottom line of cultivated land, 2 representing the safety bottom line of ecological range, 3 representing the safety bottom line of flood risk, and 0 representing the background being unmarked;

[0040] The interference factor instance segmentation sample set S ele_label It has interference element category attributes, with mark 0 representing roads, mark 1 representing large-scale construction land, mark 2 representing regular artificial lakes, mark 3 representing large-scale bare land construction, and mark 4 representing other typical structures.

[0041] Preferably, step S5 specifically includes the following contents:

[0042] S51, the safety bottom line target vector result set R in the demonstration area base As a constraint, and the high-resolution image I Post Jointly, use the national land space interference element instance segmentation network model M Inf Perform reasoning and labeling to obtain the safety bottom line target vector result set R base Interference factor semantic feature set Rset in the constraint scenario Inf ;

[0043] S52, add interference factor semantic element set Rset Inf The safety baseline type field is constructed by assigning each interference factor semantic element set Rset Inf The target frame positioning information and the safety bottom line target vector result set R base Perform spatial association to obtain the safety bottom line target type and fill it into the type field. Combined with its own target type, the final land space interference element target set Rset′ under the safety bottom line control is obtained. Inf .

[0044] Preferably, when reasoning and marking are performed in step S51, the safety bottom line target vector result set R base Post-temporal high-resolution image I under spatial mask Post As the image to be inferred, the semantic geographic information of the interference factors is obtained by using block multi-threaded parallel reasoning, and the interference factor semantic feature set Rset is obtained by spatial vectorization. Inf Spatial range, index its category information according to the semantic label of the interference element and fill in the interference information category attribute field, and then obtain the safety bottom line target vector result set R base Interference factor semantic feature set Rset in the constraint scenario Inf .

[0045] Preferably, step S52 specifically includes firstly performing the interference factor semantic element set Rset InfIncrease the safety bottom line type field, secondly define the target frame positioning information of each element as the minimum circumscribed rectangle, and obtain the four-position range of the specific interference element, thirdly, for each interference element, associate the safety bottom line target vector result of the corresponding previous phase through the four-position space range, obtain the corresponding safety bottom line target type, and fill in the current element safety bottom line target type field, and then combine the target type itself to obtain the final safety bottom line controlled land space interference element target set Rset' Inf .

[0046] The beneficial effects of the present application are: 1. The present application adopts a deep cascade network coupling safety bottom line scene discrimination and interference element instance segmentation for automatic identification of land space interference elements, which has higher identification accuracy and anti-safety bottom line background interference advantage than conventional single change detection network model. 2. The present application firstly locks the safety bottom line actual scene range of the previous and subsequent phases by using auxiliary data and safety bottom line scene recognition framework, and then extracts specific interference element information in the real actual scene, so that the spatial constraint of the safety bottom line scene is considered during model deduction, and the image range of deduction is reduced, which not only reduces the false alarm rate of interference elements caused by false safety bottom line targets, but also speeds up the deduction speed of interference element targets. 3. The safety bottom line scene semantic recognition network using the compound convolutional neural network with down-sampling of the previous and subsequent phases and self-attention features further enhances the safety bottom line target recognition capability while maintaining good model convergence on the basis of a small amount of iterative training. 4. The present application uses geographic space correlation algorithm to reversely track the safety bottom line scene type of the interference element through the four-position space range of the interference element, reduces the error source determination problem of the interference element in the multi-class safety bottom line scene, improves the safety bottom line dependent relationship identification accuracy of the interference element, realizes high-precision determination and source basis of the interference element, and can quickly locate and better assist field investigation. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a flowchart of the identification method in the embodiment of the present application;

[0048] Figure 2 is a schematic diagram of three modes in the multi-sample mosaic semantic enhancement in the embodiment of the present application;

[0049] Figure 3 is a structural schematic diagram of the safety bottom line scene detection network BaseNet in the embodiment of the present application;

[0050] Figure 4 is the original high-resolution remote sensing image (A) and the low-resolution remote sensing image (B) after ordinary 4 times down-sampling used in the embodiment of the present application, and C is the safety bottom line low-frequency information reserved remote sensing image after 4 times down-sampling;

[0051] Figure 5 This is a remote sensing data set that can be used for training in an embodiment of the present invention, obtained by removing false safety bottom line image targets in the image by combining auxiliary data; A is the later phase data, and B is the earlier phase data;

[0052] Figure 6 It is a training sample set (C) generated by combining the safety bottom line vector data (A) and the safety bottom line remote sensing data set (B) in an embodiment of the present invention.

[0053] Figure 7 The results of the enhanced remote sensing safety baseline sample set in the embodiment of the present invention are shown in Figure 1. A is the original sample, B is the sample result after radiation enhancement, C is the sample result after a random 45° counterclockwise rotation around the image center, and D is the sample result after 1*3 mosaic semantic enhancement.

[0054] Figure 8 It is the result of inference of the safety bottom line training model using the later-phase remote sensing image in the embodiment of the present invention. A is the original remote sensing image, and the dark linear range in B is the range of the inferred cultivated land red line.

[0055] Figure 9 Figures 1 and 2 are samples of interference factors and their enhancement effects in an embodiment of the present invention; Figures A and B show the sample image of the later phase and the instance sample mark of the interference factor located therein, respectively; Figures B and B show the result of randomly rotating the sample image of the later phase by 30° counterclockwise around the center and the instance sample mark of the interference factor located therein after the same rotation; Figures C and C show the sample image after the interference factor is enhanced with a 2*2 mosaic sample and the corresponding instance mark sample result, respectively;

[0056] Figure 10 is the result of inference of interference factors of remote sensing images in the later phase under the safety bottom line constraint in the embodiment of the present invention (B), and A is the case where the result is fitted to the previous phase;

[0057] Figure 11 It is the final identification result of the interference factors of the post-temporal remote sensing image in the embodiment of the present invention; A is the geographical correlation between the spatial range of the interference factors and the safety bottom line constraints, and B is the content of the attribute table of the final interference factors. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] Example 1

[0060] In order to overcome the limitations of single change detection interference element network identification, the embodiment provides a land space interference element identification method of cascaded security bottom line scene classification. The method faces the identification of land space interference elements in the complex background of remote sensing images. First, the real scene identification model is constructed by using the low-resolution large-scale remote sensing image in the previous phase within the demonstration area and combining the known auxiliary data such as cultivated land and ecological red line, to obtain the real and effective spatial range of the security bottom line in the monitoring area. Secondly, within the bottom line scene range, the high-resolution remote sensing image in the demonstration area is used to carry out small-scale interference element status instance segmentation sample making, target detection network training and reasoning, to obtain the target positioning and semantic information of the interference element. Finally, the interference element target corresponding to the spatial position of the corresponding phase is spatially erased and associated with the security bottom line range to form the interference element target change monitoring process in the demonstration area. The overall algorithm process further cascades the precise identification of interference elements at a small scale by coupling the spatial knowledge of the security bottom line and the previous phase attributes, forming a land space interference element target identification algorithm with multi-task scene cascade and geographical attribute spatial association. As shown in Figure 1 The method specifically includes the following parts,

[0061] I. Obtain the enhanced security bottom line scene semantic identification sample set

[0062] Based on the previous phase high-resolution remote sensing image I Pre and the later phase high-resolution remote sensing image I Post , obtain the security bottom line mixed image data source I 安全底线 , use the accurate identification of the security bottom line image data source I′ 安全底线 obtained based on the security bottom line mixed image data source I 安全底线 and the security bottom line scene marking data VEC scene to make the security bottom line scene identification sample set S base_label , and perform enhancement processing on the security bottom line scene identification sample set S base_label to generate the enhanced security bottom line scene semantic identification sample set S′ base_label . Specifically,

[0063] 1.1, perform downsampling processing based on low-frequency information preservation on the previous phase high-resolution remote sensing image I Pre and the later phase high-resolution remote sensing image I Post to obtain the results after 4 times downsampling as I Pre4 and I Post4 , to jointly construct the security bottom line mixed image data source I 安全底线 .

[0064] I 安全底线 includes the previous and later phase high-resolution remote sensing images based on a 3*3 size filter Fl After completing the low-frequency information retention of the safety bottom line, the I is obtained by 4 times downsampling. Pre4 and I Post4 The mixed data composed of F l The expression is,

[0065]

[0066] The 4-fold downsampling method used is the average of the maximum pooling downsampling and the average pooling downsampling with a kernel size of 2 and a step size of 2. Pre Taking 4 times downsampling as an example, the result I Pre4 The expression is,

[0067] I Pre4 =(Maxpooling(2,2)(I Pre )+Avepooling(2,2)(I Pre )) / 2.

[0068] 1.2. Put I 安全底线 Fusion analysis with the safety bottom line vector auxiliary data eliminates the false safety bottom line image content and forms an image data source I′ that can accurately identify the safety bottom line. 安全底线 .

[0069] Among them, the safety bottom line vector auxiliary data is usually a surface polygon data range, and the content range it points to is the range of human activities with frequent activities such as urban boundaries, rural boundaries and factory sites.

[0070] The fusion analysis in this step removes the false safety bottom line image content, which is mainly completed by locking the high weight parameter algorithm. The specific execution process is as follows:

[0071] Combine the safety bottom line vector auxiliary data generation with I 安全底线 Auxiliary weights T of the same size 辅助权重 , the security bottom line mixed image data source I 安全底线 With auxiliary weight T 辅助权重 Superposition and multiplication (here superposition and multiplication uses matrix element position multiplication) to obtain the image data source I′ that can accurately identify the safety bottom line. 安全底线 The specific setting method of auxiliary weight is:

[0072]

[0073] I′ 安全底线 =I 安全底线 °T 辅助权重

[0074] Among them, when the safety bottom line vector auxiliary data is located within the auxiliary polygon, the auxiliary weight T 辅助权重 The value of is 0.85; otherwise, the auxiliary weight T 辅助权重 The value of is 0.15.

[0075] 1.3. Change I′ 安全底线 and safety baseline scenario labeling data VEC scene As data input, the image and marker vector data are traversed with a certain spatial range and step size to generate a safety bottom line scene recognition sample set S base_label .

[0076] Safety bottom line scene recognition sample set S base_label The size is 512px*512px, and the safety bottom line scenarios corresponding to the type labels are shown in Table 1.

[0077] Table 1. Safety baseline scenarios corresponding to type labels

[0078] Serial number Label value Safety bottom line scenario name 1 0 Background 2 1 Arable land range safety bottom line 3 2 Ecological range safety bottom line 4 3 Flood risk safety bottom line

[0079] 1.4. Identify sample set S for safety bottom line scenarios base_label Perform enhancement processing to generate the enhanced safety baseline scenario semantic recognition sample set S′ base_label .

[0080] By S base_label Generate enhanced safety baseline scene recognition sample set S′ base_label The enhancement algorithms involved are divided into two aspects: single-sample random enhancement and multi-sample mosaic semantic enhancement.

[0081] (1) The random enhancement of a single sample includes geometric transformation (random rotation around the sample center, symmetric transformation around the vertical or horizontal axis) and radiation transformation (blurring, brightness, contrast, sharpening, adding noise, etc.). The principle of rotation transformation in geometric space is to randomly rotate the sample image counterclockwise by a certain angle θ with the sample image center as the starting point, and rotate the sample to retain the original size at intervals of 30 degrees. Then, 360 degrees is divided into 12 parts, and one of the angles is randomly selected for enhancement each time. The calculation formula is as follows.

[0082]

[0083] θ∈[0,30,60,90,120,150,180]

[0084] Among them, S img is the sample image matrix to be enhanced, θ is the rotation angle; S′ imgRotate the enhanced sample picture matrix. Since no translation is involved, the third column of the rotation matrix is 0, and the rotation angle has the angular symmetry property, so only the 0-180° range needs to be considered.

[0085] (2) Multi-sample mosaic semantic enhancement, including 1*3, 2*2 and 3*3 three modes of enhancement, as shown in Figure 2 Unlike the traditional 1*3 mode, the three fused enhanced samples are not in the traditional horizontal arrangement, but in a random arrangement, and the part without real semantics is filled with background 0.

[0086] II. Construction of safety baseline scene detection model

[0087] Construct the convolutional self-attention safety baseline scene detection network BaseNet, and use the enhanced safety baseline scene semantic recognition sample set S' base_label Iterative training to obtain the safety baseline scene detection model M Base And reasoning, to get the safety baseline target vector result set R Base .

[0088] The backbone network part of the safety baseline scene detection network BaseNet is constructed by the network of ResNet18 and VIT-Base, and the classification head network adopts a 4-class semantic segmentation network, forming a specific semantic segmentation model suitable for safety baseline scene detection. On the basis of a small number of iterations in a large range, the model can quickly converge.

[0089] The specific BaseNet network structure is shown in Figure 3 , in which the input data is extracted by 8 layers of ResNet18, and then converted into Embedding feature blocks with the same input dimension as the Vit network through 2-dimensional convolution. Then it is connected to the Transformer Encoder module, and finally it is output through the MLP network. The final classification scene semantic single-channel 4-class scene value feature probability map is obtained, and the forward calculation process of the sample data in the network is completed.

[0090] The safety baseline scene detection model M Base obtained by iterative training of the safety baseline scene detection network BaseNet, the total training iteration number epoch is set to 7, and the initial learning rate is set to 0.001. With the iteration of training, the learning rate is optimized and adjusted in a linear preheating and cosine annealing manner.

[0091] The safety baseline scene detection model M Base is used to infer and recognize the safety baseline scene detection result R baseThe vector file format has a security bottom line scene category attribute (see Table 1), and is marked 1 for arable land security bottom line, 2 for ecological security bottom line, 3 for flood risk security bottom line, and background 0 is not marked.

[0092] III. Obtaining the enhanced interference element instance segmentation sample set

[0093] Using the post-phase high-resolution remote sensing image I Post and the interference element target vector marking data VEC Inf to make the interference element instance segmentation sample set S ele_label ; to the interference element instance segmentation sample set S ele_label Perform enhancement processing to generate the enhanced interference element instance segmentation sample set S' ele_label .

[0094] The label file in the interference element instance segmentation sample set S ele_label is of the txt type, and there are 5 categories in total, as shown in Table 2.

[0095] Table 2: Interference element type

[0096] Serial number Label value Interference element type 1 0 Road 2 1 Large-scale building site 3 2 Regulation artificial lake 4 3 Large-scale bare land construction 5 4 Other typical structures

[0097] The enhanced interference element instance segmentation sample set S' ele_label generated by S ele_label involves enhancement processing consistent with the enhancement processing in the first part, with the difference being that, in the rotation enhancement of the sample in the geometric transformation, considering that the spatial distribution of the interference element target is small and is easily disturbed by the background environment, the sample is rotated by 15 degrees as an interval in the rotation enhancement processing with the sample picture center as the reference, and then 360 degrees are divided into 24 parts, and each enhancement randomly selects one angle α for enhancement processing.

[0098]

[0099] Due to the point symmetry principle of the angle, the angle in the formula only needs to be controlled within the range of 0-180 degrees, that is, α∈[0, 15, 30, 45, 60, 75, 90, 105, 120, 135, 150, 165, 180].

[0100] IV. Construction of the land space interference element instance segmentation network model

[0101] Construct the land space interference element instance segmentation network InterfNet based on the deep visual basic model, and use the enhanced interference element instance segmentation sample set S' ele_label for iterative training to obtain the land space interference element instance segmentation network model M Inf .

[0102] The InterfNet adopts Swin-Large as a feature backbone network to extract sample image features, and three head networks are used to detect the spatial position, category and semantic mask information of the interference elements.

[0103] Among them, the instance segmentation target category is set to 5, and the semantic segmentation graph in each positioning box is a binary segmentation feature graph, which extracts the target semantic information of the interference elements.

[0104] Five, target set of land space interference elements under the control of the safety bottom line

[0105] Based on the safety bottom line target vector result set R Base and the post-phase high-resolution remote sensing image I Post , the land space interference element instance segmentation network model M Inf is used for reasoning and labeling to obtain the safety bottom line target vector result set R Base , the semantic element set Rset Inf of the interference elements in the constraint scene, and based on the semantic element set Rset Inf of the interference elements, the final target set of land space interference elements under the control of the safety bottom line Rset' Inf is obtained. Specifically,

[0106] 5.1, the safety bottom line target vector result set R base in the demonstration area is taken as a constraint condition, combined with the post-phase high-resolution image I Post , the land space interference element instance segmentation network model M Inf is used for reasoning and labeling to obtain the safety bottom line target vector result set R base , the semantic element set Rset Inf of the interference elements in the constraint scene.

[0107] When reasoning and labeling, the post-phase high-resolution image I base under the spatial mask of the safety bottom line target vector result set R Post is taken as the image to be reasoned, and the block multi-thread parallel reasoning is used (the number of parallel threads opened is consistent with the number of GPUs), and the obtained semantic geographic information of the interference elements is spatially vectorized to obtain the semantic element set Rset Inf of the interference elements, the category information of the interference elements is indexed according to the semantic label, and the category attribute field of the interference information is filled in, and then the safety bottom line target vector result set R base is obtained, the semantic element set Rset Inf of the interference elements in the constraint scene.

[0108] 5.2, increase the interference element semantic element set Rset Inf The security bottom line type field is obtained by spatially correlating the target box positioning information of each interference element semantic element set Rset Inf with the security bottom line target vector result set R base , and filling into the type field, and combining the target type of itself to obtain the final interference element target set Rset' under the control of the security bottom line Inf .

[0109] Specifically, first, the security bottom line type field is added to the interference element semantic element set Rset Inf , secondly, the target box positioning information of each element is defined as the minimum circumscribed rectangle, and the four-position range of the specific interference element is obtained, thirdly, for each interference element, the corresponding security bottom line target type is obtained by correlating the four-position space range with the security bottom line target vector result of the corresponding previous phase, and filling into the current element security bottom line target type field, and then combining the target type of itself to obtain the final interference element target set Rset' under the control of the security bottom line Inf .

[0110] Embodiment two

[0111] In this embodiment, the ZY-3 satellite remote sensing image data with a spatial resolution of 2 meters is used to perform the extraction experiment of the land space interference element by using the method of the present application, so as to better illustrate the execution process and advantages of the method of the present application.

[0112] I. Obtain the strengthened security bottom line scene semantic recognition sample set

[0113] In the demonstration area, select 4137px*3254px size of spatial resolution of 2 meters of front and rear phase true color satellite remote sensing image I Pre and I Post , after 4 times down-sampling processing algorithm based on low frequency information reservation, obtain I Pre4 and I Post4 , the image size becomes 1110px*873px, the spatial resolution becomes 8 meters, and I 安全底线 is constructed together. As shown in Figure 4 .

[0114] The I 安全底线 data is processed by using the city vector auxiliary data to lock the high weight parameter algorithm, the city building range in I 安全底线 is removed from the true value background, the pseudo security bottom line image part which affects the image security bottom line range recognition is weakened, I′ 安全底线 is obtained, at this time I′ 安全底线The size remains 1110px*873px and the spatial resolution becomes 8 meters. As shown in Figure 5 .

[0115] I′ 安全底线 and the real safety baseline scene marking vector data VEC scene to generate S base_label , since only arable land safety baseline in the present exemplary range, the sample marking value is 1, and the sample size is 512*512px, after increasing the step and expanding the copy, a total of 418 pairs of sample sets are generated. As shown in Figure 6 .

[0116] The obtained 418 pairs of safety baseline sample set S base_label Perform single-sample randomness enhancement (including geometric transformation: random rotation with sample center, longitudinal axis or horizontal axis symmetry transformation; blur, brightness, contrast, sharpening, adding noise, etc. Radiation changes) and multi-sample mosaic semantic enhancement method to generate enhanced safety baseline scene recognition sample set S′ base_label , the number of enhanced samples is expanded by 10 times to become 4180 pairs, and is divided into training set and validation set according to the ratio of 9:1, which are 3762 pairs and 418 pairs respectively, and the sample size remains 512px*512px. As shown in Figure 7 .

[0117] II. Safety baseline scene detection model construction

[0118] Using the obtained S′ base_label , the self-attention safety baseline scene detection network BaseNet is trained and constructed, the initial learning rate is set to 0.001, and a total of 7 epochs are designed. After training, it can quickly converge to a verification accuracy of 0.976, and obtain a safety baseline scene detection model file M Base , the model file size is 175MB, and is saved to the disk.

[0119] Using the obtained safety baseline scene detection model file M Base , infer the post-phase remote sensing image in the demonstration area to obtain the safety baseline target vector result set R base in the demonstration area. Since there is only arable land in the present demonstration area, the safety baseline recognition category is 1. As shown in Figure 8 .

[0120] III. Obtain enhanced interference factor instance segmentation sample set

[0121] Using the post-phase high-resolution remote sensing image I Post and the interference factor target vector marking data VEC Inf , the interference factor instance segmentation sample set S ele_label, 2080 pairs of interference element instance sample pairs are generated, with a sample size of 256px*256px, and through single-sample randomness enhancement (including geometric transformation: random rotation with sample center, vertical or horizontal axis symmetry transformation; blur, brightness, contrast, sharpening, noise addition, etc. Radiation changes) and multi-sample mosaic instance segmentation enhancement, an enhanced interference element instance segmentation sample set S' is generated ele_label At this time, the sample size expands to 18720 pairs, and according to the training verification ratio of 8:2, 14976 and 3744 pairs of interference element training set and validation set samples are respectively allocated. As shown in Figure 9

[0122] Four, construction of land space interference element instance segmentation network model

[0123] Using the obtained S' ele_label , the constructed land space interference element instance segmentation network InterfNet is trained, the initial learning rate is set to 0.01, the batch size is set to 8, and 120 epochs are set. Finally, the land space interference element instance segmentation network model M Inf with a validation set mAP of 0.85 is obtained. The model file is 853MB.

[0124] Five, acquisition of land space interference element target set under the control of safety bottom line

[0125] Using the obtained safety bottom line target vector result set R base in the demonstration area and the interference element instance segmentation network model file M Inf , the post-time image block reasoning under the constraint of R base is adopted to obtain the semantic element set Rset Inf of the post-time interference element. In this example, 5 interference elements are identified, all of which are large-scale building land interference targets with a label value of 1. As shown in Figure 10

[0126] Iterate through each interference element in the semantic element set Rset Inf , and fill in the established "safety bottom line" field by spatially associating the instance detection target quadrilateral range with the cultivated land safety bottom line. A interference element map in this demonstration area can be obtained, with the remark information "located in the cultivated land safety bottom line Large-scale building land interference", and the remark information of the other 4 interference elements is "not in any safety bottom line field, discard the building land interference" (as shown in Figure 11 B), and the most accurate interference element set Rset' Inf is obtained through safety bottom line scene constraint analysis. As shown in Figure 11 The constraint analysis of interference element results is shown in Table 3. ​​

[0127] Table 3 constraint analysis of interference element results

[0128]

[0129] The method faces the identification of land space interference elements in a complex background of remote sensing images. First, low-resolution large-scale remote sensing images in the previous phase are used in combination with known city, township and other activity area auxiliary data to obtain the real and effective spatial range of the safety bottom line in the monitoring area through intelligent discrimination of the safety bottom line scene model. Second, within this bottom line scene range, high-resolution remote sensing images in the later phase are used to carry out small-scale interference element status instance segmentation sample production, target detection network training and reasoning to obtain target positioning and semantic information of the interference elements. Finally, the spatial attribute association of the interference element targets corresponding to the spatial position of the corresponding phase and the safety bottom line range is carried out to form the monitoring process of the attribute transition change of the interference element targets in the demonstration area. The land space interference element target identification algorithm model formed by the multi-task scene cascade and geographical attribute spatial association of the application further refines the accurate identification of interference elements at a small scale by coupling the spatial knowledge of the safety bottom line and the previous phase attributes, breaks through the problems of weak spatial attribute association of the safety bottom line scene in traditional methods and the inability to judge the attribute change of interference elements, and provides good data support and technical support for the construction of land space planning implementation monitoring networks and the high-quality development of land space.

[0130] By adopting the above technical solutions disclosed in the application, the following beneficial effects are obtained:

[0131] The present invention provides a method for identifying interference factors in land space by cascading safety bottom line scene classification. The present invention adopts a deep cascade network that couples safety bottom line scene discrimination and interference factor instance segmentation to automatically identify interference factors in land space, which has higher recognition accuracy and anti-safety bottom line background interference advantages than conventional single change detection network models. The present invention first uses auxiliary data and the safety bottom line scene recognition framework to lock the actual scene range of the safety bottom line in the front and back phases, and then extracts specific interference factor information in the real actual scene. The spatial constraint effect of the safety bottom line scene is taken into account when the model is deduced, and the image range of reasoning is reduced at the same time, which not only reduces the false alarm rate of interference factors caused by erroneous safety bottom line targets, but also speeds up the reasoning speed of interference factor targets. The composite convolutional neural network with front and back phase downsampling and the safety bottom line scene semantic recognition network with self-attention features adopted by the present invention further enhance the safety bottom line target recognition ability while maintaining good model convergence on the basis of a small amount of iterative training. The present invention reversely traces the safety bottom line scenario type of the interference factors by four-position spatial range of the interference factors and using the geospatial association algorithm, thereby reducing the problem of erroneous source determination of interference factors in multiple types of safety bottom line scenarios, improving the accuracy of identifying the safety bottom line affiliation of the interference factors, and achieving high-precision determination and source basis of the interference factors, which can quickly locate and better assist field investigations.

[0132] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying land space interference elements based on cascaded safety baseline scenario classification, characterized by: The following steps are included: S1. Acquisition of enhanced safety bottom line scene semantic recognition sample set: Based on the previous phase high-resolution remote sensing image I Pre and subsequent high-resolution remote sensing images I Post Obtaining a secure hybrid image data source I 安全底线 , using hybrid image data sources based on security baseline I 安全底线 The image data source I′ obtained to accurately identify the safety bottom line 安全底线 and safety baseline scenario labeling data VEC scene Create a safety bottom line scene recognition sample set S base_label , identify the sample set S for the safety bottom line scenario base_label Perform enhancement processing to generate enhanced safety bottom line scene semantic recognition sample set S′ base_label ; S2. Safety baseline scenario detection model construction: Construct a convolutional self-attention safety baseline scenario detection network BaseNet, and use the enhanced safety baseline scenario semantic recognition sample set S′ base_label Perform iterative training to obtain the safety bottom line scenario detection model M Base And reasoning to obtain the safety bottom line target vector result set R in the demonstration area Base ; S3. Obtaining the enhanced interference factor instance segmentation sample set: Using the post-phase high-resolution remote sensing image I Post and interference factor target vector label data VEC Inf Make interference factor instance segmentation sample set S ele_label ; Segment the sample set S for the interference factor instance ele_label Perform enhancement processing to generate enhanced interference factor instance segmentation sample set S′ ele_label ; S4. Construction of the network model for instance segmentation of national space interference elements: Construct the national space interference element instance segmentation network InterfNet based on the deep visual basic model, and use the enhanced interference element instance segmentation sample set S′ ele_label Perform iterative training to obtain the national land space interference element instance segmentation network model M Inf ; S5. Obtaining the target set of land space interference elements under the safety bottom line control: Based on the safety bottom line target vector result set R in the demonstration area Base and subsequent high-resolution remote sensing images I Post Using the national land space interference element instance segmentation network model M Inf Perform reasoning and annotation to obtain the safety bottom line target vector result set R Base Interference factor semantic feature set Rset under constraint scenario Inf , and based on the interference factor semantic feature set Rset Inf Obtain the final target set of land space interference elements Rset′ under the control of the safety bottom line Inf .

2. The method for identifying land space interference elements based on cascaded safety baseline scenario classification according to claim 1 is characterized by: Step S1 specifically includes the following contents: S11, the previous phase high-resolution remote sensing image I in the demonstration area Pre and subsequent high-resolution remote sensing images I Post The downsampling process based on low-frequency information retention is performed, and the results after 4 times downsampling are respectively I Pre4 and I Post4 , to jointly build a safe bottom line hybrid image data source I 安全底线 ; S12. Mixing security baseline with image data source I 安全底线 Fusion analysis with the safety bottom line vector auxiliary data eliminates the false safety bottom line image content and forms an image data source I′ that can accurately identify the safety bottom line. 安全底线 ; S13, the available image data source I' that accurately identifies the safety bottom line 安全底线 and safety baseline scenario labeling data VEC scene As data input, the image and marker vector data are traversed with a certain spatial range and step size to generate a safety bottom line scene recognition sample set S base_label ; S14, identify sample set S for safety bottom line scenario base_label Perform enhancement processing to generate the enhanced safety baseline scenario semantic recognition sample set S′ base_label .

3. The method for identifying land and space interference elements based on cascaded safety baseline scenario classification according to claim 2 is characterized by: The safety bottom line vector auxiliary data is a polygonal data range, and the content range it points to is the city boundary, rural boundary and factory site, which are areas with frequent human activities; Step S12 is specifically to generate the safety bottom line vector auxiliary data and the safety bottom line mixed image data source I 安全底线 Auxiliary weights T of the same size 辅助权重 , by combining the security baseline mixed image data source I 安全底线 With auxiliary weight T 辅助权重 By superimposing and multiplying, we can obtain an image data source I′ that can accurately identify the safety bottom line. 安全底线 ; Among them, when the safety bottom line vector auxiliary data is located within the auxiliary polygon, the auxiliary weight T 辅助权重 The value of is 0.85; otherwise, the auxiliary weight T 辅助权重 The value of is 0.

15.

4. The method for identifying land and space interference elements based on cascading safety baseline scenario classification according to claim 1 is characterized by: The enhancement process includes single-sample random enhancement and multi-sample mosaic semantic enhancement; The single sample random enhancement includes geometric transformation and radiation transformation; geometric transformation includes random rotation around the sample center, longitudinal or transverse axis symmetry transformation; radiation transformation includes blurring, brightness, contrast, sharpening, and adding noise; Obtain the enhanced safety baseline scene recognition sample set S′ base_label The geometric transformation involved in the single-sample random enhancement is to randomly rotate the sample image counterclockwise by a certain angle θ starting from the center point of the sample image itself, and perform the rotation enhancement processing on the sample image at intervals of 30 degrees, thereby preserving the original size. The 360 ​​degrees are divided into 12 parts, and one of the angles θ is randomly selected for each enhancement. θ∈[0,30,60,90,120,150,180] Obtain the enhanced interference factor instance segmentation sample set S′ ele_label The geometric transformation involved in the single-sample random enhancement is to randomly rotate the sample image counterclockwise by a certain angle α starting from the center point of the sample image itself, and perform the rotation enhancement processing on the sample image at intervals of 15 degrees, thereby preserving the original size. The 360 ​​degrees are divided into 24 parts, and one of the angles α is randomly selected for each enhancement. α∈[0,15,30,45,60,75,90,105,120,135,150,165,180] Among them, S img is the sample image matrix to be enhanced; θ and α are the rotation angles; S′ img is the sample image matrix after rotation enhancement; The multi-sample mosaic semantic enhancement includes three mode enhancement modes: 1*3, 2*2 and 3*3; the three fused enhancement samples are randomly arranged, and the parts without real semantics are filled with background 0.

5. The method for identifying land space interference elements based on cascading safety baseline scenario classification according to claim 1 is characterized by: The backbone of the convolutional self-attention safety baseline scenario detection network, BaseNet, is constructed using a ResNet-18 and VIT-Base network. Its classification head network uses a 4-category semantic segmentation network to form a semantic segmentation model specifically suited for safety baseline scenario detection. In the convolutional self-attention safety baseline scene detection network BaseNet, input data is extracted through the 8-layer feature extraction of ResNet18. The feature map is then converted into an embedding feature block with the same input dimension as the VIT network through 2D convolution. This feature block is then connected to the Transformer Encoder module and finally passes through the MLP network to output the final classified scene semantics single-channel feature probability map of four scene values, completing the forward calculation process of the sample data in the network. The safety bottom line scene detection model M is obtained through iterative training of the convolutional self-attention safety bottom line scene detection network BaseNet. Base During the training process, the total number of training iterations is 7, the initial learning rate is 0.001, and the learning rate is optimized and adjusted with linear warm-up and cosine annealing as the training iterations progress.

6. The method for identifying land space interference elements based on cascading safety baseline scenario classification according to claim 1 is characterized by: Interference element instance segmentation network InterfNet uses Swin-Large as the feature backbone network to extract sample image features. It includes three head networks for detecting the spatial location, category and semantic mask information of interference elements respectively; The instance segmentation target category is set to 5, and the semantic segmentation map within each positioning box is a binary segmentation feature map to extract the target semantic information of the interference elements.

7. The method for identifying land space interference elements based on cascading safety baseline scenario classification according to claim 1 is characterized by: The safety bottom line scene recognition sample set S base_label and the safety bottom line target vector result set R Base All of them have the safety bottom line scenario category attribute, with mark 1 representing the safety bottom line of cultivated land, 2 representing the safety bottom line of ecological range, 3 representing the safety bottom line of flood risk, and 0 representing the background being unmarked; The interference factor instance segmentation sample set S ele_label It has interference element category attributes, with mark 0 representing roads, mark 1 representing large-scale construction land, mark 2 representing regular artificial lakes, mark 3 representing large-scale bare land construction, and mark 4 representing other typical structures.

8. The method for identifying land and space interference elements based on cascading safety baseline scenario classification according to claim 1 is characterized by: Step S5 specifically includes the following contents: S51, the safety bottom line target vector result set R in the demonstration area base As a constraint, and the high-resolution image I Post Jointly, use the national land space interference element instance segmentation network model M Inf Perform reasoning and labeling to obtain the safety bottom line target vector result set R base Interference factor semantic feature set Rset in the constraint scenario Inf ; S52, add interference factor semantic element set Rest Inf The safety baseline type field is constructed by assigning each interference factor semantic element set Rset Inf The target frame positioning information and the safety bottom line target vector result set R base Perform spatial association to obtain the safety bottom line target type and fill it into the type field. Combined with its own target type, the final land space interference element target set Rset′ under the safety bottom line control is obtained. Inf .

9. The method for identifying land space interference elements based on cascading safety baseline scenario classification according to claim 8 is characterized by: When reasoning and labeling are performed in step S51, the safety bottom line target vector result set R base Post-temporal high-resolution image I under spatial mask Post As the image to be inferred, the semantic geographic information of the interference factors is obtained by using block multi-threaded parallel reasoning, and the interference factor semantic feature set Rset is obtained by spatial vectorization. Inf Spatial range, index its category information according to the semantic label of the interference element and fill in the interference information category attribute field, and then obtain the safety bottom line target vector result set R base Interference factor semantic feature set Rset in the constraint scenario Inf .

10. The method for identifying land space interference elements by cascading safety baseline scenario classification according to claim 8 is characterized by: Step S52 specifically includes firstly performing the interference factor semantic element set Rset Inf Add the safety bottom line type field, then define the target frame positioning information of each element as the minimum enclosing rectangle, and obtain the four-position range of the specific interference element. Again, for each interference element, associate the safety bottom line target vector result of the corresponding previous phase through the four-position spatial range to obtain the corresponding safety bottom line target type, and fill it in the safety bottom line target type field of the current element. Then, combine it with its own target type to obtain the final land space interference element target set Rset′ under the safety bottom line control. Inf .

Images