CN-121982514-A - Coastal zone remote sensing image multi-classification model under road segmentation constraint

CN121982514ACN 121982514 ACN121982514 ACN 121982514ACN-121982514-A

Abstract

The invention discloses a multi-classification model of a coastal zone remote sensing image under the constraint of road segmentation, which comprises a road identification module, a mask module, a multi-class ground object classification module and a result merging module. The method can obviously reduce the cross misjudgment of the miscible categories of roads, hardened pavements, tidal furrows, water bodies and the like under the complex background of the coastal zone, improves the geometric communication and edge form maintaining capability of the roads, and improves the overall classification precision and model robustness.

Inventors

LIU XIAODING
CHEN LEI
YANG LEI
AI BIN
LUO XIAOMEI
HUANG HUAMEI
QU CHENGYI
LIANG YUCHEN
YANG FAN
LIN XIAOBO
LEI LI

Assignees

广东省国土资源测绘院
中山大学
自然资源部南海发展研究院(自然资源部南海遥感技术应用中心)

Dates

Publication Date: 20260505
Application Date: 20251212

Claims (10)

1. The coastal zone remote sensing image multi-classification model under the constraint of road segmentation is characterized by comprising a road identification module, a mask module, a multi-class ground object classification module and a result merging module, wherein: the road recognition module is used for inputting the remote sensing image of the coastal zone to be recognized into the road recognition model and outputting a road probability map; The mask module is used for carrying out mask operation on the original coastal zone remote sensing image by utilizing the road probability map to obtain a mask image; the multi-category ground object classification module is used for inputting the mask image into the multi-category ground object classification model and outputting a classification result; and the result merging module is used for merging the classification result with the road probability map to obtain a final recognition result.
2. The multi-classification model of coastal zone remote sensing images under road segmentation constraint of claim 1, wherein the road recognition model constructs a road priority recognition model for a modified D-LinkNet framework, and comprises a first convolution layer, a first maximum pooling layer, a first residual block, a first channel attention block, a second maximum pooling layer, a second residual block, a second channel attention block, a third maximum pooling layer, a third residual block, a third channel attention block, a fourth maximum pooling layer, a fourth residual block, a convolution expansion block, a first upsampling block, a second upsampling block, a third upsampling block, a fourth upsampling block, a fifth upsampling block, a first channel addition block, a second channel addition block, a third channel addition block, a fourth channel addition block, a fifth channel addition block, a convolution block, and a 1 x 1 convolution block, and inputs the coastal zone remote sensing images to be recognized into the road recognition model, and outputs a road probability map, comprising the steps of: inputting the coastal zone remote sensing image into a first convolution layer to obtain a preliminary feature map; inputting the preliminary feature map into a first maximum pooling layer to obtain a first maximum pooling feature map; Inputting the first maximum pooling feature map into a first residual block to obtain a first residual feature map; Inputting the first residual characteristic diagram into a first channel attention block to obtain a first attention optimizing characteristic diagram; inputting the first attention optimizing feature map into a second maximum pooling layer to obtain a second maximum pooling feature map; inputting the second maximum pooling feature map into a second residual block to obtain a second residual feature map; inputting the second residual characteristic diagram into a second channel attention block to obtain a second attention optimizing characteristic diagram; inputting the second attention optimizing feature map into a third maximum pooling layer to obtain a third maximum pooling feature map; Inputting the third maximum pooling feature map into a third residual block to obtain a third residual feature map; inputting the third residual characteristic diagram into a third channel attention block to obtain a third attention optimizing characteristic diagram; Inputting the third attention optimizing feature map into a fourth maximum pooling layer to obtain bottleneck features; Inputting bottleneck characteristics into a convolution expansion block, inputting the output of each expansion rate branch and the residual direct connection branch into a first channel addition block, and outputting the bottleneck characteristics; inputting the bottleneck output into a first upsampling block to obtain a first upsampling feature map, and inputting the first upsampling feature map and a third attention optimizing feature map into a second channel adding block to obtain a first jump connection fusion feature map; Inputting the first jump connection fusion feature map into a second up-sampling block to obtain a second up-sampling feature map, and inputting the second up-sampling feature map and the second attention optimizing feature map into a third channel addition block to obtain a second jump connection fusion feature map; inputting the second jump connection fusion feature map into a third up-sampling block to obtain a third up-sampling feature map, and inputting the third up-sampling feature map and the first attention optimizing feature map into a fourth addition block to obtain a third jump connection fusion feature map; Inputting the third jump connection fusion characteristic diagram into a fourth up-sampling block to obtain a fourth up-sampling characteristic diagram, and inputting the fourth up-sampling characteristic diagram and the preliminary characteristic diagram into a fifth channel addition block to obtain a fourth jump connection fusion characteristic diagram; inputting the fourth jump connection fusion characteristic diagram into a transposed convolution block to obtain a transposed convolution characteristic diagram; And inputting the transposed convolution feature map into a 1 multiplied by 1 convolution block to obtain a final road probability map.
3. The multi-classification model of coastal zone remote sensing images under road segmentation constraint of claim 2 wherein the first, second and third channel attention blocks each comprise a global average pooling unit, a global maximum pooling unit, a stitching unit, a two-layer fully connected network, a 1 x1 convolution alignment unit and a channel-by-channel multiplication unit, wherein the two-layer fully connected network comprises a ReLU activation unit and a Sigmoid activation unit, wherein the attention block is utilized to extract an attention optimization feature map comprising the steps of: Inputting a feature map F to be extracted into a global average pooling unit and a global maximum pooling unit respectively to generate channel description vectors GAP (F) and GMP (F), wherein F epsilon R H×W×C , R represents a real number domain, H represents the height of the feature map, W represents the width of the feature map, and C represents the channel number; GAP (F) and GMP (F) are input into a splicing unit for vector splicing to obtain splicing vectors; inputting the spliced vector into a two-layer fully-connected network, wherein the first layer is compressed to C/16 dimension through a ReLU (remote control unit) activation unit, and the second layer is restored to C dimension through a Sigmoid activation unit to generate a channel weight vector Wc epsilon R 1×1×C ; Wc and F are input into a channel-by-channel multiplication unit, an attention optimizing feature map Fout is output, and the expression is as follows: 。
4. the multi-classification model of coastal zone remote sensing image under road segmentation constraint of claim 2 wherein the first, second, third and fourth residual blocks each comprise a first 3 x 3 convolution block, a second 3 x 3 convolution block and a channel addition block, and wherein the residual block is used to extract a residual feature map comprising the steps of: inputting the input features into a first 3×3 convolution block to obtain a first convolution result; inputting the first convolution result into a second 3×3 convolution block to obtain a second convolution result; mapping the input features by identity mapping or projection to obtain a branch feature map; And inputting the branch characteristic diagram and the second convolution result into a channel addition block to obtain a residual characteristic diagram.
5. The multi-classification model of coastal zone remote sensing images under road segmentation constraint according to claim 2, wherein the convolution expansion block comprises a residual block, a 3 x 3 hole convolution sequence and a channel addition block, the expansion rate n e {1,2,4,8}, and the bottleneck characteristics are output by the convolution expansion block, comprising the steps of: Inputting the bottleneck characteristic B0 into a residual error block and outputting a residual error characteristic diagram R; Inputting R into a3 multiplied by 3 cavity convolution block, wherein the expansion rate is 1, and outputting a first expansion characteristic diagram B1; B1 is input into a 3 multiplied by 3 cavity convolution block, the expansion rate is 2, and a second expansion characteristic diagram B2 is output; Inputting B2 into a 3 multiplied by 3 cavity convolution block, wherein the expansion rate is 4, and outputting a third expansion characteristic diagram B3; b3 is input into a 3 multiplied by 3 cavity convolution block, the expansion rate is 8, and a fourth expansion characteristic diagram B4 is output; b0, R, B, B2, B3 and B4 are input into the channel addition block to output bottleneck characteristics.
6. The multi-classification model of coastal zone remote sensing images under the constraint of road segmentation according to claim 1, wherein the masking module performs masking operation on the original remote sensing images by using a road probability map, and the method comprises the following steps: Using road probability maps Generating a binary mask by combining a preset threshold T The expression is as follows: Wherein, the Representing coordinates of points in the probability map; Mask the mask Performing pixel-by-pixel multiplication with the original remote sensing image I to obtain a mask image The expression is as follows: Where b represents the band index of the image.
7. The multi-classification model of coastal zone remote sensing images under road segmentation constraint of claim 6, wherein the predetermined threshold is 0.7.
8. The multi-classification model of coastal zone remote sensing images under road segmentation constraint of claim 1, wherein the multi-class feature classification model is ResUnet model comprising a first, a second, a first, a third, a fourth, a second, a fifth, a sixth, a third, a seventh, an eighth, a fourth, a ninth, a tenth, a first, an eleventh, a twelfth, a second, a thirteenth, a fourteenth, a third, a fifteenth, a sixteenth, a fourth, a seventeenth, a1 x 1, a classification model of mask images using the multi-class feature classification model comprising the steps of: Inputting the mask image into a first convolution block and a second convolution block in sequence to obtain a first coding feature map; Inputting the first coding feature map into a first maximum pooling layer to obtain a first maximum pooling feature map; sequentially inputting the first maximum pooling feature map into a third convolution block and a fourth convolution block to obtain a second coding feature map; inputting the second coding feature map into a second maximum pooling layer to obtain a second maximum pooling feature map; Sequentially inputting the second maximum pooling feature map into a fifth convolution block and a sixth convolution block to obtain a third coding feature map; Inputting the third coding feature map into a third maximum pooling layer to obtain a third maximum pooling feature map; sequentially inputting the third maximum pooling feature map into a seventh convolution block and an eighth convolution block to obtain a fourth coding feature map; inputting the fourth coding feature map into a fourth maximum pooling layer to obtain a fourth maximum pooling feature map; Sequentially inputting the fourth maximum pooling feature map into a ninth convolution block and a tenth convolution block to obtain bottleneck features; the bottleneck feature is input into a first up-sampling convolution block to obtain a first up-sampling convolution feature map, and the fourth coding feature map is copied and cut and then spliced with the first up-sampling convolution feature map to obtain a first spliced feature map; sequentially inputting the first spliced feature map into an eleventh convolution block and a twelfth convolution block to obtain a first decoding feature map; the first decoding feature map is input into a second up-sampling convolution block to obtain a second up-sampling convolution feature map; sequentially inputting the second spliced characteristic map into a thirteenth convolution block and a fourteenth convolution block to obtain a second decoding characteristic map; The second decoding feature map is input into a third up-sampling convolution block to obtain a third up-sampling convolution feature map, and the second coding feature map is copied and cut and then spliced with the third up-sampling convolution feature map to obtain a third spliced feature map; sequentially inputting the third spliced characteristic diagram into a fifteenth convolution block and a sixteenth convolution block to obtain a third decoding characteristic diagram; The third decoding feature map is input into a fourth up-sampling convolution block to obtain a fourth up-sampling convolution feature map; Sequentially inputting the fourth spliced characteristic diagram into a seventeenth convolution block and an eighteenth convolution block to obtain a fourth decoding characteristic diagram; the fourth decoding feature map is input into a1×1 convolution block to obtain a classification map.
9. The multi-classification model of coastal zone remote sensing image under road segmentation constraint of claim 8 wherein the first and second, third and fourth, fifth and sixth, seventh and eighth convolution blocks are each replaced with a residual network comprising a first weight layer, a second weight layer, an identity mapping or projection mapping branch and a channel addition block, the extraction of the encoded feature map using the residual network comprising the steps of: Inputting the input feature map into a first weight layer, and outputting an intermediate feature map; inputting the intermediate feature map into a second weight layer, and outputting a main branch feature map; the input feature map is mapped through identity mapping or projection mapping paths to obtain a branch feature map; and inputting the main branch characteristic diagram and the branch characteristic diagram into a channel addition block to output a coding characteristic diagram.
10. The multi-classification model of coastal zone remote sensing images under the constraint of road segmentation according to claim 1, wherein the method of combining the road probability map and the classification result by the result combining module is combining by using a combining tool of GIS software.

Description

Coastal zone remote sensing image multi-classification model under road segmentation constraint Technical Field The invention belongs to the technical field of remote sensing, and particularly relates to a coastal zone remote sensing image multi-classification model under road segmentation constraint. Background The coastal zone serves as a key ecological transition zone of the sea Liu Jiaohu and carries dynamic balance of multiple human activities and natural processes such as port construction, wetland protection, aquaculture and the like. The high-precision classification of the land features of the coastal zone is not only a core data base for realizing quantitative monitoring of coastline transition (such as erosion and deposition rate calculation) and protection of ecological sensitive areas such as mangrove, but also a prerequisite for decision making such as national soil space planning, storm surge disaster risk assessment and the like of the coastal zone. However, the traditional model relies on spectral characteristics and visual interpretation, and has the problems of low efficiency and strong subjectivity. In recent years, end-to-end multi-classification models based on deep learning models such as convolutional neural networks (Convolutional Neural Networks, CNN), U-Net, deep LabV3+ and the like have become the mainstream. For multi-classification tasks of remote sensing images, the current technical scheme is to train sample data by manufacturing multi-class sample tags and selecting a proper network architecture, and a class prediction result is given at an output end. This approach belongs to a single-stage strategy, i.e. all the required prediction results can be output during the training phase of the model. Although the single-stage implementation process is simple, the method is highly sensitive to label quality, characteristic coupling is easy to cause multi-class systematic confusion, misjudgment is easy to generate particularly on linear ground features (such as roads) with insignificant geometric forms, meanwhile, the method is insufficient in geometric form retaining capability, common in edge saw teeth, topological fracture and other problems, and the overall practicability is limited. The existing scheme is constructed and evaluated around inland scenes, and is faced to a complex area with stronger multi-source interference and category interaction of a coastal zone, and the classification precision and the robustness are difficult to meet the requirements of high-precision drawing and updating. The invention patent with the publication number of CN111767810A in the prior art proposes a remote sensing image road extraction method based on D-LinkNet, which comprises the steps of S1, inputting a characteristic image into a D-LinkNet network, completing processing in an encoder sub-network based on a residual error network and transfer learning, S2, inputting the characteristic image output in the step S1 into a characteristic extraction sub-network based on an expansion convolution and convolution block attention module for characteristic extraction, and S3, inputting the characteristic image obtained after processing of the first two sub-networks into a decoder sub-network based on a transposition convolution to realize image recovery, wherein the scheme does not perform multi-category form retention and inter-category confusion inhibition, so that multi-category fine segmentation application is difficult to directly support. Disclosure of Invention The invention provides a multi-classification model of a remote sensing image of a coastal zone under the constraint of road segmentation, which aims to solve the problem of insufficient classification precision caused by insufficient geometric form retaining capability of the remote sensing image classification model of the coastal zone on linear features in the prior art. The primary purpose of the invention is to solve the technical problems, and the technical scheme of the invention is as follows: The invention provides a multi-classification model of a coastal zone remote sensing image under the constraint of road segmentation, which comprises a road identification module, a mask module, a multi-class ground object classification module and a result merging module, wherein: the road recognition module is used for inputting the remote sensing image of the coastal zone to be recognized into the road recognition model and outputting a road probability map; The mask module is used for carrying out mask operation on the original coastal zone remote sensing image by utilizing the road probability map to obtain a mask image; the multi-category ground object classification module is used for inputting the mask image into the multi-category ground object classification model and outputting a classification result; and the result merging module is used for merging the classification result with the road probability map to obtain a final rec