An impervious surface extraction method based on a spatial information semantic segmentation model

Abstract

The application belongs to the technical field of natural science, and discloses an impervious surface extraction method based on a spatial information semantic segmentation model, which comprises the following steps: step 1, remote sensing images are acquired, the remote sensing images are preprocessed to obtain a remote sensing image dataset, the dataset is cut to obtain an image slice dataset, and the image slice dataset is divided to obtain a remote sensing image slice dataset and a ground truth label dataset; step 2, an OCR_CLHRnet model is constructed; step 3, impervious surface results are predicted to obtain semantic segmentation results; step 4, the semantic segmentation results are reclassified to obtain impervious surface classification results of the image slice data; and step 5, the impervious surface classification results of the image slice data are subjected to a slicing operation for restoration and splicing, and the distribution results of the impervious surface of the region are obtained after splicing. The application improves the efficiency of model training, can realize large-scale and fine impervious surface monitoring, and improves the precision of the model in extracting the impervious surface.

Description

Technical Field

[0001] This invention belongs to the field of natural science and technology, specifically a method for extracting impermeable surfaces based on a spatial information semantic segmentation model. Background Technology

[0002] Our country is implementing new urbanization and building people-oriented cities. Building people-oriented cities requires implementation in all aspects of urban spatial layout, and necessitates full consideration of urban density, temperature, precision, breadth, and greenness. Impermeable surfaces in cities are a good indicator of the urbanization process, and their dynamic changes can provide effective data for urban environmental monitoring, spatial planning, and urban land use analysis.

[0003] Traditional methods for acquiring impermeable surface data mainly include four types: obtaining data from land survey databases, obtaining data from planning documents, manually interpreting construction drawings, and ground surveying and mapping. These methods all suffer from limitations such as the inability to acquire the latest data in real time, low levels of automation, and inefficiency in extracting information accurately. However, using satellites equipped with advanced multi-spectral remote sensing devices allows for a transition from surface descriptions of various phenomena to software analysis and quantitative exploration, providing a robust and objective data foundation for the monitoring and management of urban environmental systems.

[0004] The following problems still exist in the current impermeable surface extraction based on remote sensing images: (1) If the image resolution is lower than the spatial resolution, it is difficult to detect ground objects in the image unless the ground objects have unique brightness or color or unique edge and texture features; (2) Ground objects are diverse and constantly changing, with different types of ground objects interspersed and exhibiting nearly uniform spectral features. In addition, the same type of ground objects often exhibit heterogeneity in different scenes, which may lead to spectral confusion in the extraction of spectral coefficients; (3) The accuracy of the impermeable surface extraction model is limited by the spatial resolution of the remote sensing image, while high-resolution remote sensing images such as sub-meter level remote sensing images are expensive and difficult to obtain, and high-precision deep learning models are computationally complex and have long operation time, requiring high computer hardware performance; (4) Impermeable surface extraction based on semantic segmentation models often results in phenomena such as ground object holes, blurred boundaries, and ground object misclassification, which do not fully express the spatial information of buildings and roads in the impermeable surface and are prone to overfitting. Therefore, based on medium-resolution imagery, it is of great significance to study models that have more accurate interpretation capabilities for buildings, roads, etc. in cities and can extract higher-precision information on impermeable surfaces. Summary of the Invention

[0005] To address the issues of feature confusion and boundary blurring in current methods for extracting impervious surfaces based on remote sensing images, and the computational complexity and high computational cost of semantic segmentation models, this invention provides an impervious surface extraction method based on a spatial information semantic segmentation model. Considering the complexity of ground features and the diversity of scenes in remote sensing images, CLHRnet is introduced as the backbone network on the basis of the OCR semantic segmentation model to construct a multi-scale, multi-dimensional global context model. An improved spatial information module is constructed by combining ShuffleNet and coordinate attention mechanisms, achieving the goal of enhancing the extraction capability of spatial information of ground features without significantly increasing the computational cost. The DropBlock standardization of the convolutional network structure improves the overfitting problem and enhances the model accuracy.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] This invention is a method for extracting impermeable surfaces based on a spatial information semantic segmentation model. The method includes the following steps:

[0008] Step 1: Acquire remote sensing images, preprocess the remote sensing images to obtain a remote sensing image dataset, cut the remote sensing image dataset to obtain an image tile dataset, and divide the image tile dataset to obtain a remote sensing image tile dataset and a ground truth label dataset.

[0009] Step 2: Construct the OCR_CLHRnet model, including building the CLHRnet model and optimizing the OCR model. Obtain high-resolution remote sensing images, input the high-resolution remote sensing images into the constructed OCR_CLHRnet model for training, and obtain a self-trained pre-trained model. On the basis of the OCR_CLHRnet model, load the self-trained pre-trained model, input the image tile dataset for training, and obtain the optimal model, which is the impermeable surface prediction model.

[0010] Step 3: Predict the impermeable surface result: Input the remote sensing image tile dataset and ground truth label dataset obtained in Step 1 into the impermeable surface prediction model obtained in Step 2 to obtain the semantic segmentation result;

[0011] Step 4: Reclassify the semantic segmentation results obtained in Step 3 to obtain the impermeable surface results, and obtain the impermeable surface classification results of the image slice data;

[0012] Step 5: Restore and stitch together the impermeable surface classification results of the image slice data obtained in Step 4 according to the image slice operation of the remote sensing image dataset in Step 1. After stitching, the distribution results of impermeable surfaces in the region are obtained.

[0013] A further improvement of the present invention is that step 2, in which the OCR_CLHRnet model is constructed, specifically includes the following steps:

[0014] Step 2-1: Construct the first-stage spatial convolutional network Bottleneck: 1×1 convolutional layer to obtain features, then through two 3×3 convolutional layers, followed by a downsampling layer, and finally through a Channel Shuffle layer to obtain the initial feature information of the first stage;

[0015] Step 2-2: Based on the spatial convolutional network Bottleneck constructed in Step 2-1, repeat the stacking of Bottleneck, modify the number of channels to 64, stack 4 branches for output, and the output result is the feature information of the first stage;

[0016] Steps 2-3: Construct the Spatial Information Module Basic Block: Construct two branches. The first branch includes a 5×5 depthwise convolutional layer, a 1×1 convolutional layer, and a coordinate attention mechanism layer. The second branch includes a 1×1 convolutional layer, a 5×5 depthwise convolutional layer, and a downsampling layer. Add the two branches together and input the result into Channel Shuffle to complete the fusion of feature information.

[0017] Steps 2-4: Obtain feature maps at different resolutions: The spatial information module BasicBlock constructed in Step 2-3 serves as the base network. A parallel network is built by repeatedly stacking the spatial information module BasicBlock and changing the feature size to retain multi-scale information. In the second stage, 8 branches are output, in the third stage, 16 branches are output, and in the fourth stage, 64 branches are output. The number of branches represents the resolution of the feature map, that is, the feature information at different resolutions. This results in a semantic segmentation network structure that achieves multi-scale information fusion, parallel processing of high and low resolution information, and spatial information sensitivity, namely CLHRnet.

[0018] Steps 2-5: To fully combine the advantages of CLHRnet and the OCR model, the CLHRnet built in Steps 2-4 is used as the backbone network. The feature information obtained in the fourth stage is input into the OCR model. The implementation of the OCR model mainly includes three stages: soft object region, object region representation, and final object context feature representation. The soft object region operation includes feature extraction and multi-scale feature representation, namely, feature extraction based on the attention mechanism SE (Spatial Excitation) module and feature extraction based on multi-level pyramid pooling ASPP (Atrous Spatial Pyramid). The Pooling module performs multi-scale feature representation. The operations for object region representation include feature extraction, bounding box localization, multi-scale feature fusion, multi-directional feature encoding, and contextual feature enhancement. The operations for object contextual feature representation include object feature extraction, contextual feature encoding, multi-scale feature fusion, object classification, and recognition. When the object contextual feature representation is concatenated with the feature representation of the deepest input layer of the network, it becomes the contextual information enhanced feature representation. Based on the enhanced feature representation, the semantic category of each pixel is predicted. That is, the OCR model calculates a set of feature representations of object regions. According to the similarity between the feature representations of object regions and the feature representations of pixels, these object region feature representations are transmitted to each pixel. DropBlock, a structured form of dropout, is introduced in the object contextual feature representation stage to randomly mask units in adjacent feature regions.

[0019] Steps 2-6: After processing the soft object region, the category region features are obtained, and the object region representation stage is entered, where the region is located and features are extracted.

[0020] Steps 2-7: After completing the feature information extraction in the object region representation stage of Step 2-6, we proceed to the final object context information extraction network and prediction network, thus completing the construction of the OCR_CLHRnet model.

[0021] Steps 2-8: To accelerate model training and improve model performance, train the pre-trained model: acquire high-resolution remote sensing images, create the input model dataset according to the steps in Part 1, input the constructed OCR_CLHRnet model for training, set the pre-trained model training parameters, and after training, obtain the self-trained pre-trained model.

[0022] Step 2-9: Load the pre-trained model obtained in step 2-8 into the OCR_CLHRnet model in step 2-7, and train it according to the parameter configuration of the prediction model to obtain the optimal model, which is the impermeable surface prediction model.

[0023] A further improvement of this invention lies in the following: In steps 2-3, the construction of the coordinate attention mechanism layer specifically involves: to enhance the model's ability to perceive spatial information, a self-attention mechanism that strengthens the spatial relationships of features is introduced, namely, the coordinate attention mechanism. First, the input feature layer passes through two convolutional layers to obtain coordinate features and feature channels, respectively. The coordinate features are normalized to obtain a coordinate feature map. The feature channels and the coordinate feature map are multiplied to obtain a feature representation that enhances spatial information. The obtained features are then output through a convolution operation.

[0024] A further improvement of the present invention is that steps 2-6 specifically involve:

[0025] Step 2-6-1, Feature Extraction: The FCN model is used to segment the text regions and obtain their respective text border information;

[0026] Step 2-6-2, Border Location: The text border is located using the RPN model to obtain more accurate text border information;

[0027] Step 2-6-3, Multi-scale feature fusion: The text features are layered and extracted using pyramid pooling, which makes the network more expressive at different scales;

[0028] Step 2-6-4 Multi-directional feature encoding: The features extracted in step 2-6-3 are encoded using the direction-aware ROIAlign operation to obtain feature representations in different directions;

[0029] Step 2-6-5, Contextual Feature Enhancement: The DSS module is used to interact with the local and global features from Step 2-6-4 to enhance the contextual features of the text region.

[0030] A further improvement of the present invention is that steps 2-7 specifically include the following steps:

[0031] Step 2-7-1: Use a deep residual network-based object feature extraction module to extract features from the input image. Step 2-7-2: Use a context feature encoding module based on the SE mechanism to encode and enhance object features.

[0032] Step 2-7-3: Use the ASPP module to fuse the features encoded in Step 2-7-2 to obtain multi-scale feature representations;

[0033] Step 2-7-4: Use the RPN module to perform a sliding window search on the feature representation obtained in step 2-7-3 to determine the position and size of the object;

[0034] Step 2-7-5: Use a recurrent neural network-based object recognition module to classify and recognize the objects from step 2-7-4;

[0035] Step 2-7-6: The non-maximum suppression (NMS) algorithm is used to filter candidate boxes to obtain the final object detection results. The OCR model achieves accurate and robust detection and recognition of objects in the image.

[0036] A further improvement of the present invention is that step 1 specifically includes the following steps:

[0037] Step 1-1: Obtain publicly available medium-resolution remote sensing images, and perform atmospheric correction, radiometric calibration, and image cropping preprocessing on the initial remote sensing images to obtain the preprocessed initial remote sensing image dataset.

[0038] Step 1-2: Perform regular tiling on the initial remote sensing image dataset after preprocessing in Step 1-1. Cut the image into 256×256 pixels with a repeat rate of 0.1 to obtain the image tile dataset.

[0039] Steps 1-3: Divide the image tile dataset obtained in Steps 1-2 into training set, validation set and test set in a ratio of 6:2:2. Use Labelme tool to create ground truth labels for the training set and validation set data through manual visual interpretation. The labels are divided into 5 categories: background, other, vegetation, road and building. This results in the remote sensing image tile dataset and the ground truth label dataset.

[0040] Steps 1-4: To fully verify the generalization ability of the model, data augmentation processing was performed on the image slice dataset obtained in Steps 1-2. Specifically, operations including resizing, image cropping, horizontal flipping, vertical flipping, Gaussian blurring, color and brightness perturbation, and normalization were performed to enhance the diversity of the data and improve the robustness and generalization ability of the model.

[0041] The beneficial effects of this invention are: This invention proposes an impermeable surface extraction method based on a spatial information semantic segmentation model, which can achieve high-precision impermeable surface extraction using medium-resolution images, and the model training speed is fast, enabling large-scale and refined impermeable surface monitoring.

[0042] The OCR_CLHRnet model proposed in this invention has a high ability to perceive spatial information. It introduces a spatial attention mechanism on the basis of the HRnet model to improve the ability to extract feature information. It combines ShuffleNet to make the network lightweight. It introduces DropBlock canonical convolutional network on the basis of OCR model to effectively reduce the overfitting problem of the model.

[0043] This method can improve the extraction of boundary information, spatial information, and feature information of impermeable landforms, enhance the efficiency of model training, and improve the accuracy of model extraction of impermeable surfaces. Attached Figure Description

[0044] Figure 1 This is a technical flowchart of an embodiment of the present invention.

[0045] Figure 2 This is a structural diagram of the OCR_CLHRnet model of the present invention.

[0046] Figure 3 This is a structural diagram of the spatial information module of the present invention.

[0047] Figure 4 This is a preprocessed remote sensing image of Nanjing City, as described in this invention example.

[0048] Figure 5 This is a map showing the distribution of impermeable surfaces in Nanjing City, obtained through model prediction in an example of this invention. Detailed Implementation

[0049] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not essential.

[0050] like Figure 1 As shown, this invention is a method for extracting impermeable surfaces based on a spatial information semantic segmentation model. The method includes the following steps:

[0051] Step 1: Obtain remote sensing image data from an open-source platform, such as the European Space Agency (ESA) data platform (https: / / scihub.copernicus.eu / dhus / # / home[2020-0616]) to obtain Sentinel-2 data. Preprocess the remote sensing images to obtain a remote sensing image dataset. Segment the remote sensing image dataset to obtain an image tile dataset. Divide the image tile dataset into a remote sensing image tile dataset and a ground truth label dataset. The specific implementation process includes the following steps:

[0052] Step 1-1: Obtain publicly available medium-resolution remote sensing images, such as Sentinel-2 data, and perform atmospheric correction, radiometric calibration, and image cropping preprocessing on the initial remote sensing images to obtain the preprocessed initial remote sensing image dataset.

[0053] Step 1-2: Perform regular tiling on the initial remote sensing image dataset after preprocessing in Step 1-1. Cut the image into 256×256 pixels with a repeat rate of 0.1 to obtain the image tile dataset.

[0054] Steps 1-3: Divide the image tile dataset obtained in Steps 1-2 into training set, validation set and test set in a ratio of 6:2:2. Use Labelme tool to create ground truth labels for the training set and validation set data through manual visual interpretation. The labels are divided into 5 categories: background, other, vegetation, road and building. This results in the remote sensing image tile dataset and the ground truth label dataset.

[0055] Steps 1-4: To fully verify the generalization ability of the model, data augmentation processing was performed on the image slice dataset obtained in Steps 1-2. Specifically, operations including resizing, image cropping, horizontal flipping, vertical flipping, Gaussian blurring, color and brightness perturbation, and normalization were performed to enhance the diversity of the data and improve the robustness and generalization ability of the model.

[0056] Step 2: Construct the OCR_CLHRnet model, including building the CLHRnet model and optimizing the OCR model. Obtain high-resolution remote sensing images, input the high-resolution remote sensing images into the constructed OCR_CLHRnet model for training, and obtain a self-trained pre-trained model. On the basis of the OCR_CLHRnet model, load the self-trained pre-trained model, input the image tile dataset for training, and obtain the optimal model, which is the impermeable surface prediction model.

[0057] The network structure of the CLHRnet model is shown in Table 1:

[0058] Table 1 CLHRnet Network Structure Table

[0059]

[0060]

[0061] Specifically, the steps for building the OCR_CLHRnet model are as follows:

[0062] Step 2-1: Construct the first-stage spatial convolutional network Bottleneck: 1×1 convolutional layer to obtain features, then through two 3×3 convolutional layers, followed by a downsampling layer, and finally through a Channel Shuffle layer to obtain the initial feature information of the first stage;

[0063] In HRNet, 3×3 convolutional layers are primarily used for feature extraction, standardization, and enhancing the non-linearity of features. Specifically, they can perform convolution operations by sliding across feature maps to identify different features in an image, such as edges and textures, and convert them into high-level features.

[0064] Furthermore, convolutional layers can be normalized using Batch Normalization layers, a process that makes the neural network more stable during training and accelerates network convergence. Simultaneously, the ReLU activation function can perform a non-linear transformation on features, enabling the neural network to better model complex patterns and relationships.

[0065] The downsampling layer uses Multi-Scale Fusion, which fuses multiple feature maps with varying resolutions from high to low. A 1×1 convolution is performed on the lower-resolution feature map, increasing the number of channels. This layer primarily reduces the resolution of the feature maps, thereby improving computational efficiency.

[0066] Channel shuffle groups input features along the channel dimension and then shuffles the channels within each group. This shuffling operation helps break down correlations between channels, enhances feature diversity and robustness, and thus improves the model's generalization ability.

[0067] Step 2-2: Based on the spatial convolutional network Bottleneck constructed in Step 2-1, repeat the stacking of Bottleneck, modify the number of channels to 64, stack 4 branches for output, and the output result is the feature information of the first stage;

[0068] Steps 2-3: Construct the Spatial Information Module Basic Block, with the following structure: Figure 3 As shown: Two branches are constructed. The first branch includes a 5×5 depthwise convolutional layer, a 1×1 convolutional layer, and a coordinate attention mechanism layer. The second branch includes a 1×1 convolutional layer, a 5×5 depthwise convolutional layer, and a downsampling layer. The two branches are added together and input into ChannelShuffle to complete the fusion of feature information. The 3×3 convolutional layer in the first stage is expanded to a 5×5 convolutional layer to increase the receptive field, reduce overfitting, and improve accuracy. Feature information is fused through a 1×1 convolutional layer. The coordinate attention mechanism considers global information and strengthens spatial relationships to improve the model's expressive power.

[0069] In steps 2-3, the construction of the coordinate attention mechanism layer specifically involves: to enhance the model's ability to perceive spatial information, a self-attention mechanism that strengthens the spatial relationships of features is introduced, namely the coordinate attention mechanism. First, the input feature layer passes through two convolutional layers to obtain coordinate features and feature channels, respectively. The coordinate features are normalized to obtain a coordinate feature map. The feature channels and the coordinate feature map are multiplied to obtain a feature representation that enhances spatial information. The obtained features are then output through a convolution operation.

[0070] Step 2-4: Obtain feature maps at different resolutions: The spatial information module BasicBlock constructed in Step 2-3 serves as the base network. A parallel network is built by repeatedly stacking the spatial information module BasicBlock and changing the feature size to retain multi-scale information. In the second stage, 8 branches are output, in the third stage, 16 branches are output, and in the fourth stage, 64 branches are output. The number of branches represents the resolution of the feature map, that is, the feature information at different resolutions. The above steps 2-1 to 2-4 construct a semantic segmentation network structure that realizes multi-scale information fusion, parallel processing of high and low resolution information, and spatial information sensitivity, namely CLHRnet.

[0071] Steps 2-5: To fully combine the advantages of CLHRnet and the OCR model, the CLHRnet built in Steps 2-4 is used as the backbone network. The feature information obtained in the fourth stage is input into the OCR model. The implementation of the OCR model mainly includes three stages: soft object region, object region representation, and final object context feature representation. The soft object region operation includes feature extraction and multi-scale feature representation, namely, feature extraction based on the attention mechanism SE (Spatial Excitation) module and feature extraction based on multi-level pyramid pooling ASPP (Atrous Spatial Pyramid). The pooling module performs multi-scale feature representation. Object region representation operations include feature extraction, bounding box localization, multi-scale feature fusion, multi-directional feature encoding, and contextual feature enhancement. Object contextual feature representation operations include object feature extraction, contextual feature encoding, multi-scale feature fusion, object classification, and recognition. The object contextual feature representation is concatenated with the feature representation input from the deepest layer of the network to form the context-enhanced feature representation. Based on the enhanced feature representation, the semantic category of each pixel is predicted. That is, the OCR model calculates a set of object region feature representations and transmits these object region feature representations to each pixel based on the similarity between the object region feature representation and the pixel feature representation. DropBlock, a structured form of dropout, is introduced in the object contextual feature representation stage to randomly mask units in adjacent feature regions, improving model accuracy and robustness to hyperparameter selection. The feature information obtained through CLHRNet is input into the feature extraction network of the OCR model, i.e., into the soft object region. This part uses the attention mechanism SE (SpatialExcitation) module and multi-layer pyramid pooling ASPP (Atrous Spatial Pyramid). The Pooling module is used to capture the contextual information of a region and fuse it with the region's features, thereby reducing the impact of ambiguous boundary information.

[0072] Steps 2-6: After processing the soft object region, the category region features are obtained, and the object region representation stage is entered. This involves bounding box localization and feature extraction, specifically including the following features:

[0073] Step 2-6-1, Feature Extraction: The FCN model is used to segment the text regions and obtain their respective text border information;

[0074] Step 2-6-2, Border Location: The text border is located using the RPN model to obtain more accurate text border information;

[0075] Step 2-6-3, Multi-scale feature fusion: The text features are layered and extracted using pyramid pooling, which makes the network more expressive at different scales;

[0076] Step 2-6-4 Multi-directional feature encoding: The features extracted in step 2-6-3 are encoded using the direction-aware ROIAlign operation to obtain feature representations in different directions;

[0077] Step 2-6-5, Contextual Feature Enhancement: The DSS module is used to interact with the local and global features from Step 2-6-4 to enhance the contextual features of the text region.

[0078] Steps 2-7: After completing the feature information extraction in the object region representation stage of Step 2-6, we proceed to the final object context information extraction network and prediction network, thus completing the construction of the OCR_CLHRnet model.

[0079] Specifically, the steps include the following:

[0080] Step 2-7-1: Use a deep residual network-based object feature extraction module to extract features from the input image. Step 2-7-2: Use a context feature encoding module based on the SE mechanism to encode and enhance object features.

[0081] Step 2-7-3: Use the ASPP module to fuse the features encoded in Step 2-7-2 to obtain multi-scale feature representations;

[0082] Step 2-7-4: Use the RPN module to perform a sliding window search on the feature representation obtained in step 2-7-3 to determine the position and size of the object;

[0083] Step 2-7-5: Use a recurrent neural network-based object recognition module to classify and recognize the objects from step 2-7-4;

[0084] Step 2-7-6: The non-maximum suppression (NMS) algorithm is used to filter candidate boxes to obtain the final object detection results. The OCR model achieves accurate and robust detection and recognition of objects in the image.

[0085] Steps 2-8: To accelerate model training and improve model performance, train the pre-trained model: Acquire high-resolution remote sensing imagery (1m resolution), create the input model dataset following the steps in Part 1, and input the constructed OCR_CLHRnet model for training. The batch size is 8, the number of iterations is 20,000, the optimizer is Adam, and the initial learning rate is 0.0001. The training parameters for the pre-trained model are shown in Table 2. After training, the self-trained pre-trained model is obtained.

[0086] Table 2 Model Parameter Configuration

[0087] Model Model batch size Number of iterations Algorithm optimizer Initial learning rate pre-trained model 8 20000 Adam 0.0001 Predictive Model 8 6000 Adam 0.0001

[0088] Pre-trained models can effectively improve training speed, accelerate model convergence, and enhance model performance, resulting in better outcomes. Considering the limitations of medium-resolution imagery in impermeable surface extraction, high-resolution imagery (1m resolution) is used as auxiliary data input to train the model structured in Part 3, thus obtaining a pre-trained model that can, to some extent, compensate for the accuracy deficiencies of medium-resolution imagery.

[0089] Step 2-9: Load the pre-trained model obtained in step 2-8 into the OCR_CLHRnet model in step 2-7, and train it according to the parameter configuration of the prediction model to obtain the optimal model, which is the impermeable surface prediction model.

[0090] Step 3: Predict the impermeable surface result: Input the remote sensing image slice dataset and ground truth label dataset obtained in Step 1 into the impermeable surface prediction model obtained in Step 2 to obtain the semantic segmentation result. The result at this time is the classification result obtained according to the label data category.

[0091] Step 4: Reclassify the semantic segmentation results obtained in Step 3 to obtain the impermeable surface results, and obtain the impermeable surface classification results of the image slice data;

[0092] Step 5: Restore and stitch together the impermeable surface classification results of the image slice data obtained in Step 4 according to the image slice operation of the remote sensing image dataset in Step 1. The input image size is 256×256 and the repetition rate is 0.1. After stitching, the distribution results of impermeable surfaces in the region are obtained.

[0093] This invention can achieve high-precision impermeable surface extraction using medium-resolution imagery, improve the extraction of boundary information, spatial information, and feature information of impermeable surface features, enhance the efficiency of model training, realize large-scale and refined impermeable surface monitoring, and improve the accuracy of model extraction of impermeable surfaces.

[0094] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for extracting impermeable surfaces based on a spatial information semantic segmentation model, characterized in that: The method for extracting impermeable surfaces includes the following steps: Step 1: Acquire remote sensing images, preprocess the remote sensing images to obtain a remote sensing image dataset, cut the remote sensing image dataset to obtain an image tile dataset, and divide the image tile dataset to obtain a remote sensing image tile dataset and a ground truth label dataset. Step 2: Construct a spatial information semantic segmentation model (OCR_CLHRnet model), including the construction of the CLHRnet model and the optimization of the OCR model. Acquire high-resolution remote sensing images, input the high-resolution remote sensing images into the constructed OCR_CLHRnet model for training, and obtain a self-trained pre-trained model. On the basis of the OCR_CLHRnet model, load the self-trained pre-trained model, input the image tile dataset for training, and obtain the optimal model, which is the impermeable surface prediction model. Step 3: Predict the impermeable surface result: Input the remote sensing image tile dataset and ground truth label dataset obtained in Step 1 into the impermeable surface prediction model obtained in Step 2 to obtain the semantic segmentation result; Step 4: Reclassify the semantic segmentation results obtained in Step 3 to obtain the impermeable surface results, and obtain the impermeable surface classification results of the image slice data; Step 5: Restore and stitch together the impermeable surface classification results of the image slice data obtained in Step 4 according to the image slice operation of the remote sensing image dataset in Step 1. After stitching, the distribution results of impermeable surfaces in the region are obtained. Step 2, constructing the OCR_CLHRnet model, specifically includes the following steps: Step 2-1: Construct the first-stage spatial convolutional network Bottleneck: 1×1 convolutional layer to obtain features, then through two 3×3 convolutional layers, followed by a downsampling layer, and finally through a Channel Shuffle layer to obtain the initial feature information of the first stage; Step 2-2: Based on the spatial convolutional network Bottleneck constructed in Step 2-1, repeat the stacking of Bottleneck, modify the number of channels to 64, stack 4 branches for output, and the output result is the feature information of the first stage; Steps 2-3: Construct the Spatial Information Module Basic Block: Construct two branches. The first branch includes a 5×5 depthwise convolutional layer, a 1×1 convolutional layer, and a coordinate attention mechanism layer. The second branch includes a 1×1 convolutional layer, a 5×5 depthwise convolutional layer, and a downsampling layer. Add the two branches together and input the result into Channel Shuffle to complete the fusion of feature information. Steps 2-4: Obtain feature maps at different resolutions: The spatial information module Basic Block constructed in Step 2-3 serves as the base network. A parallel network is built by repeatedly stacking the spatial information module Basic Block and changing the feature size to retain multi-scale information. In the second stage, 8 branches are output, in the third stage, 16 branches are output, and in the fourth stage, 64 branches are output. The number of branches represents the resolution of the feature map, that is, the feature information at different resolutions. This results in a semantic segmentation network structure that achieves multi-scale information fusion, parallel processing of high and low resolution information, and spatial information sensitivity, namely CLHRnet. Steps 2-5: Using the CLHRnet built in Steps 2-4 as the backbone network, input the feature information obtained in the fourth stage into the OCR model. The implementation of the OCR model mainly includes three stages: soft object region, object region representation, and final object context feature representation. The soft object region operation includes feature extraction and multi-scale feature representation, namely, feature extraction based on the attention mechanism SE (Spatial Excitation) module and feature extraction based on multi-level pyramid pooling ASPP (Atrous Spatial Pyramid). The Pooling module performs multi-scale feature representation. The operations for object region representation include feature extraction, bounding box localization, multi-scale feature fusion, multi-directional feature encoding, and contextual feature enhancement. The operations for object contextual feature representation include object feature extraction, contextual feature encoding, multi-scale feature fusion, object classification, and recognition. When the object contextual feature representation is concatenated with the feature representation of the deepest input layer of the network, it becomes the contextual information enhanced feature representation. Based on the enhanced feature representation, the semantic category of each pixel is predicted. That is, the OCR model calculates a set of feature representations of object regions. According to the similarity between the feature representations of object regions and the feature representations of pixels, these object region feature representations are transmitted to each pixel. DropBlock, a structured form of dropout, is introduced in the object contextual feature representation stage to randomly mask units in adjacent feature regions. Steps 2-6: After processing the soft object region, the category region features are obtained, and the object region representation stage is entered, where the region is located and features are extracted. Steps 2-7: After completing the feature information extraction in the object region representation stage of Step 2-6, we proceed to the final object context information extraction network and prediction network, thus completing the construction of the OCR_CLHRnet model. Steps 2-8: To accelerate model training and improve model performance, train the pre-trained model: acquire high-resolution remote sensing images, create the input model dataset according to the steps in Part 1, input the constructed OCR_CLHRnet model for training, set the pre-trained model training parameters, and after training, obtain the self-trained pre-trained model. Step 2-9: Load the pre-trained model obtained in step 2-8 into the OCR_CLHRnet model in step 2-7, and train it according to the parameter configuration of the prediction model to obtain the optimal model, which is the impermeable surface prediction model.

2. The method for extracting impermeable surfaces based on a spatial information semantic segmentation model according to claim 1, characterized in that: In steps 2-3, the construction of the coordinate attention mechanism layer is as follows: First, the input feature layer passes through two convolutional layers to obtain coordinate features and feature channels respectively. The coordinate features are normalized to obtain a coordinate feature map. The feature channels and the coordinate feature map are multiplied to obtain a feature representation that enhances spatial information. The obtained features are output through a convolution operation.

3. The method for extracting impermeable surfaces based on a spatial information semantic segmentation model according to claim 1, characterized in that: Steps 2-6 specifically refer to: Step 2-6-1, Feature Extraction: The FCN model is used to segment the text regions and obtain their respective text border information; Step 2-6-2, Border Location: The text border is located using the RPN model to obtain more accurate text border information; Step 2-6-3, Multi-scale feature fusion: The text features are layered and extracted using pyramid pooling, which makes the network more expressive at different scales; Step 2-6-4 Multi-directional feature encoding: The features extracted in step 2-6-3 are encoded using the direction-aware ROIAlign operation to obtain feature representations in different directions; Step 2-6-5, Contextual Feature Enhancement: The DSS module is used to interact with the local and global features from Step 2-6-4 to enhance the contextual features of the text region.

4. The method for extracting impermeable surfaces based on a spatial information semantic segmentation model according to claim 1, characterized in that: Steps 2-7 specifically include the following steps: Step 2-7-1: Use a deep residual network-based object feature extraction module to extract features from the input image. Step 2-7-2: Use a context feature encoding module based on the SE mechanism to encode and enhance object features; Step 2-7-3: Use the ASPP module to fuse the features encoded in Step 2-7-2 to obtain multi-scale feature representations; Step 2-7-4: Use the RPN module to perform a sliding window search on the feature representation obtained in step 2-7-3 to determine the position and size of the object; Step 2-7-5: Use a recurrent neural network-based object recognition module to classify and recognize the objects from step 2-7-4; Step 2-7-6: The non-maximum suppression (NMS) algorithm is used to filter candidate boxes to obtain the final object detection results. The OCR model achieves accurate and robust detection and recognition of objects in the image.

5. The method for extracting impermeable surfaces based on a spatial information semantic segmentation model according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1-1: Obtain publicly available medium-resolution remote sensing images, and perform atmospheric correction, radiometric calibration, and image cropping preprocessing on the initial remote sensing images to obtain the preprocessed initial remote sensing image dataset. Step 1-2: Perform regular tiling on the initial remote sensing image dataset after preprocessing in Step 1-1. Cut the image into 256×256 pixels with a repeat rate of 0.1 to obtain the image tile dataset. Steps 1-3: Divide the image tile dataset obtained in Steps 1-2 into training set, validation set and test set according to the proportion. Use Labelme tool to create ground truth labels for the training set and validation set data through manual visual interpretation. The labels are divided into 5 categories: background, other, vegetation, road and house. This results in the remote sensing image tile dataset and the ground truth label dataset. Steps 1-4: Perform data augmentation on the image slice dataset obtained in Steps 1-2. Specifically, operations include resizing, image cropping, horizontal flipping, vertical flipping, Gaussian blurring, color and brightness perturbation, and normalization to enhance data diversity and improve the robustness and generalization ability of the model.

Images