Water network structure identification method, model training method device, storage medium and terminal
By using a water network structure recognition model training method, and utilizing a water surface segmentation subnetwork and a water network structure vectorization subnetwork, the problem of occlusion in urban water network structure recognition was solved, achieving higher accuracy and more consistent water body recognition results.
Patent Information
- Application Number
- CN202211493086.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-25
- Publication Date
- 2025-11-18
- Estimated Expiration
- 2042-11-25
AI Technical Summary
In existing technologies, the identification of urban water network structures is easily affected by the surrounding environment, resulting in inconsistent identification results and poor accuracy, especially the problem of bridge and highway obstruction.
A water network structure recognition model training method is adopted, including a water surface segmentation subnetwork and a water network structure vectorization subnetwork. Spatial-spectral features of satellite remote sensing images are extracted through a 3D convolutional network, and vectorized labels are used for secondary training to eliminate the influence of obstructions and achieve coherent recognition of water bodies.
It effectively eliminates the influence of obstructions such as bridges and highways, achieves continuous identification of water bodies, and improves the accuracy and completeness of water network structure identification.
Smart Images
Figure CN115909071B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of water network structure, and particularly relates to a water network structure identification method, a model training method device, a storage medium and a terminal. BACKGROUND
[0002] Identifying the water network structure of a city can provide useful information for water resource management, water environment monitoring and natural disaster prevention of the city. With the development of remote sensing technology, a large amount of remote sensing data can be obtained for high-precision water network structure identification and mapping, because the remote sensing technology can repeatedly collect scenes with high spatial and temporal resolution in a large area.
[0003] Because the water network structure inside a city is relatively complex, the water network structure in the identification result is not coherent when identifying rivers using satellite remote sensing images, because the rivers are often blocked by bridges and highways. In addition, a large number of water body identification studies only use the RGB band of satellite remote sensing images and ignore the useful information of the multi-spectral band, resulting in poor accuracy of water network identification.
[0004] How to avoid the influence of the surrounding environment and obtain clear and coherent water network structure inside the city has become a problem that needs to be solved by the existing technical personnel. SUMMARY
[0005] The technical problem to be solved by the present application is that the existing identification of the water network structure inside a city is easily affected by the surrounding environment, resulting in insufficient clarity of the water network structure in the identification result and the phenomenon of incoherence.
[0006] In order to solve the above technical problems, the present application provides a water network structure identification model training method, comprising:
[0007] obtaining a training data set, the training data set comprising a plurality of satellite remote sensing image data and a segmentation label image and a vectorization label image corresponding to each satellite remote sensing image data;
[0008] training a water network structure identification model based on the training data set to obtain a trained water network structure identification model;
[0009] The water network structure identification model comprises a water surface segmentation subnetwork and a water network structure vectorization subnetwork.
[0010] The water surface segmentation subnetwork is used to extract spatial information and dimensional information of the satellite remote sensing image data, and fuse the spatial information and the dimensional information into a probability feature map.
[0011] The water network structure vectorization subnetwork is used to vectorize the waterway structure in the probability feature map to obtain a sample image identification result set.
[0012] Preferably, the water network structure recognition model is trained based on the training data set to obtain a trained water network structure recognition model, and the step of training the water network structure recognition model based on the training data set comprises:
[0013] The water body surface segmentation sub-network is trained based on a segmentation training data set to obtain a trained water body surface segmentation sub-network, wherein the segmentation training data set comprises a plurality of satellite remote sensing image data and a segmentation label image corresponding to each satellite remote sensing image data.
[0014] The satellite remote sensing image data in the segmentation training data set is input into the trained water body surface segmentation sub-network to obtain a probability feature map set based on an output probability feature map, and a vectorization training data set is formed based on the probability feature map set and a vectorization label image corresponding to the satellite remote sensing image data.
[0015] The water network structure vectorization sub-network is trained based on the vectorization training data set to obtain a trained water network structure vectorization sub-network.
[0016] The trained water body surface segmentation sub-network and the trained water network structure vectorization sub-network are connected in series to form a trained water network structure recognition model.
[0017] Preferably, the water body surface segmentation sub-network comprises a backbone sub-network and a 3D convolution sub-network, both of which comprise a down-sampling path and an up-sampling path, and the vertically corresponding layers of the up-sampling path of the backbone sub-network and the up-sampling path of the 3D convolution sub-network are connected through a skip connection operation.
[0018] And a multi-scale pooling module is further arranged between the down-sampling path and the up-sampling path of the backbone sub-network to encode context information in different scale ranges.
[0019] Preferably, each band of the satellite remote sensing image data has been fused to unify the resolution of the satellite remote sensing image data.
[0020] Preferably, before the satellite remote sensing image data is input into the water body surface segmentation sub-network, the satellite remote sensing image data, the segmentation label image and the vectorization label image are further cropped in the same cropping manner to make the cropped sample images meet the input size requirements of the water network structure recognition model.
[0021] To solve the above technical problems, the application further provides a water network structure recognition method, comprising:
[0022] Obtaining satellite remote sensing image data of a target region;
[0023] input the satellite remote sensing image data of the target region into the preset water network structure recognition model trained to obtain a sample image recognition result of the target region;
[0024] The trained water network structure recognition model is obtained by the water network structure recognition model training method in any one of claims 1-5.
[0025] Preferably, the method further comprises:
[0026] The sample image recognition results of the target region are superimposed to obtain a complete water network structure recognition result of the target region.
[0027] To solve the above technical problems, the application further provides a preset water network structure recognition model training device, characterized by comprising a data set acquisition module and a training module.
[0028] The data set acquisition module is configured to acquire a training data set, wherein the training data set comprises a plurality of satellite remote sensing image data, and a segmentation label image and a vectorization label image corresponding to each satellite remote sensing image data.
[0029] The training module is configured to train a preset water network structure recognition model based on the training data set to obtain a trained preset water network structure recognition model.
[0030] The preset water network structure recognition model comprises a water body surface segmentation subnetwork and a water network structure vectorization subnetwork.
[0031] The water body surface segmentation subnetwork is configured to extract spatial information and dimensional information of the satellite remote sensing image data, and fuse the spatial information and the dimensional information into a probability feature map.
[0032] The water network structure vectorization subnetwork is configured to vectorize a waterway structure in the probability feature map to obtain a sample image recognition result.
[0033] To solve the above technical problems, the application further provides a storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the preset water network structure recognition model training method or the preset water network structure recognition method.
[0034] To solve the above technical problems, the application further provides a terminal, comprising a processor and a memory, wherein the memory is in communication connection with the processor.
[0035] The memory is used for storing a computer program, and the processor is used for executing the computer program stored by the memory, so that the terminal executes the preset water network structure identification model training method or realizes the preset water network structure identification method.
[0036] Compared with the prior art, one or more embodiments in the above scheme can have the following advantages or beneficial effects:
[0037] The water network structure identification model training method provided by the embodiment of the application can fully extract spatial-spectral features of the spectral dimension of the satellite remote sensing image through a 3D convolutional network, and objects that are easily confused with water bodies can also be correctly classified. The main network used in the application can also achieve smooth segmentation of pixels at the water-land boundary, so as to correctly identify water bodies in a complex background. The water network structure vectorization sub-network uses vectorization labels for secondary training on the basis of the water body surface segmentation sub-network, perceives the context in a larger receptive field, can ignore some small bridges and road obstructions on small rivers, and can connect the water bodies in the identification result, and the water network structure vectorization sub-network can also eliminate some scattered misclassified pixels in the background area, so as to obtain a more optimal overall segmentation effect.
[0038] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and obtained by the structure particularly pointed out in the specification, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS
[0039] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the application embodiments, are used to explain the application, and do not constitute a limitation to the application. In the drawings:
[0040] Figure 1 A flowchart of a water network structure identification model training method according to an embodiment of the application is shown;
[0041] Figure 2 A structure diagram of a water network structure identification model according to an embodiment of the application is shown;
[0042] Figure 3 A comparison diagram of satellite remote sensing image data, a segmentation label image and a vectorization label image in an embodiment of the application is shown;
[0043] Figure 4 A comparison diagram of satellite remote sensing image data, a segmentation label image, a vectorization label image, a water body surface segmentation sub-network processing result and a water network structure vectorization sub-network processing result in an embodiment of the application is shown;
[0044] Figure 5 A flowchart of a water network structure identification method according to an embodiment of the present application is shown.
[0045] Figure 6 A result diagram of water network structure identification of Shanghai City using the water network structure identification method is shown.
[0046] Figure 7 A structural diagram of a water network structure identification model training device according to an embodiment of the present application is shown.
[0047] Figure 8 A structural diagram of a terminal according to an embodiment of the present application is shown. DETAILED DESCRIPTION
[0048] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and embodiments, so that how the present application applies technical means to solve technical problems and achieves technical effects can be fully understood and implemented. It should be noted that, as long as there is no conflict, each embodiment in the present application and each feature in each embodiment can be combined with each other, and the technical solutions formed thereby are all within the protection scope of the present application.
[0049] In a normal city water body identification task, the water network structure identification is easily affected by the surrounding complex environment shielding. Because the water network structure in the city is relatively complex, when using satellite remote sensing images to identify rivers, the water network structure in the identification result is often not coherent due to the shielding of bridges and highways. In addition, a large number of water body identification researches only use the RGB band of satellite remote sensing images and ignore the useful information of multispectral bands, resulting in poor accuracy of water network identification.
[0050] Example One
[0051] To solve the technical problems existing in the prior art, an embodiment of the present application provides a water network structure identification model training method.
[0052] Figure 1 A flowchart of a water network structure identification model training method according to an embodiment of the present application is shown. Referring to Figure 1 The water network structure identification model training method according to the embodiment of the present application includes the following steps.
[0053] In step S101, a training data set is obtained, and the training data set includes a plurality of satellite remote sensing image data and a segmentation label image and a vectorized label image corresponding to each satellite remote sensing image data.
[0054] Specifically, a plurality of satellite remote sensing image data are acquired by downloading or other reasonable means, and then each satellite remote sensing image data is made into a segmentation label and a vectorization label by visual interpretation to obtain a segmentation label image and a vectorization label image of each satellite remote sensing image data. Figure 3 A comparison diagram of satellite remote sensing image data, segmentation label image and vectorization label image in the embodiment of the present application is shown; refer to Figure 3 As shown, the water body in the segmentation label image is marked in gray, and the background part is black; and in the vectorization label image, the pixels belonging to the bridge or road background class across the river are also marked as water body class.
[0055] Each satellite remote sensing image data and the corresponding segmentation label image and vectorization label image of the satellite remote sensing image data form an original sample pair, and all the original sample pairs are collected as a training data set for training the water network structure recognition model. Further, the satellite remote sensing image data in the original sample pair and the corresponding segmentation label image can also form a segmentation sample pair, and all the segmentation sample pairs form a segmentation sample set as a training set of a water surface segmentation subnetwork in the water network structure recognition model. The vectorization label image in the original sample pair can also form a vectorization sample pair with the probability feature map output by the trained water surface segmentation subnetwork when the corresponding satellite remote sensing image data is input into the water surface segmentation subnetwork, to serve as a training set of a water network structure vectorization subnetwork in the water network structure recognition model.
[0056] The satellite remote sensing image data can be RGB influence data, multispectral image data or hyperspectral image data, etc.; preferably, since the 3D convolutional network in the water surface segmentation subnetwork of the present embodiment has the best spectral dimension information extraction effect on multispectral image data, the multispectral image data are preferably selected as the training data in the water network structure recognition model training method of the present embodiment.
[0057] Further, in order to improve the training effect, the bands of all satellite remote sensing image data need to be fused before training to unify the resolution of the satellite remote sensing image data. For example, assuming that the satellite remote sensing image data is Sentinel 2 multispectral image data, the bands of the Sentinel 2 multispectral image data need to be fused to 10m before training to unify the resolution of the Sentinel 2 multispectral image data. The band fusion process of the Sentinel 2 multispectral image data can be implemented by using the Sen2Res tool in the SNAP software. Further, in order to avoid processing invalid data, the 10th band of the Sentinel 2 used for cirrus cloud detection can be discarded before fusing the Sentinel 2 multispectral image data.
[0058] Step S102, training the water network structure recognition model based on the training data set to obtain a trained water network structure recognition model.
[0059] The water network structure recognition model specifically includes a water surface segmentation sub-network and a water network structure vectorization sub-network, and both the water surface segmentation sub-network and the water network structure vectorization sub-network are encoder-decoder structures.
[0060] Figure 4 This diagram illustrates a comparison of satellite remote sensing image data, segmented labeled images, vectorized labeled images, the processing results of the water surface segmentation sub-network, and the processing results of the water network structure vectorized sub-network in Embodiment 1 of the present invention; (Refer to...) Figure 4 As shown, the water surface segmentation subnetwork is mainly used to extract spatial and dimensional information from satellite remote sensing image data, and then fuse the extracted spatial and dimensional information to form a probabilistic feature map. Figure 2 A schematic diagram of the water network structure identification model according to Embodiment 1 of the present invention is shown; Reference Figure 2 As shown, the water surface segmentation subnetwork includes a backbone network and a 3D convolutional integral network. Both the backbone network and the 3D convolutional integral network include downsampling and upsampling paths. Adjacent horizontal corresponding layers within the downsampling and upsampling paths are connected by skip connections. Simultaneously, vertical corresponding layers within the upsampling paths of the backbone network and the 3D convolutional integral network are also connected by skip connections to achieve the fusion of spatial and dimensional information from satellite remote sensing image data. The backbone network is a classic segmentation network, U-Net, and a multi-scale pooling module is incorporated between the downsampling and upsampling paths. This multi-scale pooling module is located at the bottom of the backbone network to encode contextual information within different scale ranges, thereby improving the performance of the water network structure recognition model. The 3D convolutional integral network also has a U-shaped structure.
[0061] refer to Figure 4 As shown, the water network structure vectorization subnetwork is used to vectorize the waterway structure in the probabilistic feature map to obtain the sample image recognition results. (Reference) Figure 2 As shown, the water network structure vectorization subnetwork is also a classic segmentation network, U-Net, including downsampling and upsampling paths. Adjacent horizontal layers within both downsampling and upsampling paths are connected by skip connections. A multi-scale pooling module is also incorporated between the downsampling and upsampling paths, located at the bottom of the backbone subnetwork, to encode contextual information across different scales, thereby improving the performance of the water network structure recognition model. Because the two subnetworks have similar structures, and the water network structure vectorization subnetwork performs convolutional encoding and other operations based on the segmentation prediction of the water surface segmentation subnetwork, it has less background noise and achieves the water network structure vectorization task by perceiving a wider range of data. Furthermore, to prevent overfitting, the water network structure vectorization subnetwork omits the 3D convolutional network, resulting in a relatively smaller backbone network.
[0062] The specific process of training the water network structure identification model includes the following sub-steps.
[0063] In step S1021, the water surface segmentation sub-network is trained based on the segmentation training data set to obtain the trained water surface segmentation sub-network.
[0064] Specifically, the water surface segmentation sub-network is trained by the segmentation training data set. For example, the Adam optimization algorithm can be used, the initial learning rate is 0.001, the default hyperparameters β1=0.9 and β2=0.999, and the batch size is 2. The training state of the network is evaluated and adjusted through the validation set. If the loss on the validation set does not improve within three training periods, the learning rate will be reduced by 0.5 times; in addition, if the loss on the validation set does not improve within 20 training periods, the training of the network will be stopped to prevent overfitting, and the model parameters related to the minimum validation loss will be saved. Finally, the trained water surface segmentation sub-network is obtained. The segmentation training data set includes a plurality of segmentation sample pairs, each segmentation sample pair including a satellite remote sensing image data and a segmentation label image corresponding to the satellite remote sensing image data.
[0065] In step S1022, the satellite remote sensing image data in the segmentation training data set is input into the trained water surface segmentation sub-network to obtain a set of probability feature maps based on the output probability feature maps.
[0066] Specifically, the satellite remote sensing image data in the segmentation training data set is sequentially input into the trained water surface segmentation sub-network, and the trained water surface segmentation sub-network outputs the corresponding probability feature maps, all of which form a set of probability feature maps. Then, based on the set of probability feature maps and the vectorized label image corresponding to the satellite remote sensing image data, a vectorized training data set is formed; further, the probability feature maps obtained by inputting the satellite remote sensing image data into the trained water surface segmentation sub-network, and the vectorized label image corresponding to the satellite remote sensing image data form a vectorized sample pair, and all vectorized sample pairs form a vectorized training data set.
[0067] In step S1023, the water network structure vectorization sub-network is trained based on the vectorized training data set to obtain the trained water network structure vectorization sub-network.
[0068] Specifically, the water network structure vectorization sub-network is trained by the vectorized training data set. The training process of the water network structure vectorization sub-network is the same as that of the water surface segmentation sub-network, which will not be described here.
[0069] Step S1024, connect the trained water body surface segmentation sub-network and the trained water network structure vectorization sub-network in series to form a trained water network structure identification model.
[0070] It should be noted that the training of the water body surface segmentation sub-network and the training of the water network structure vectorization sub-network belong to a set of coherent processes, and the data permutation and the like can be automatically completed in sequence under the water network structure identification model integrated with the water body surface segmentation sub-network and the water network structure vectorization sub-network, without human intervention or adjustment.
[0071] It should be noted that since the water network structure identification model cannot realize the input of the entire satellite remote sensing image data, the satellite remote sensing image data also needs to be cropped before being input into the water body surface segmentation sub-network. Preferably, the satellite remote sensing image data can be cut into multiple sample images at a certain repetition rate by a sliding cropping method, so that the cropped sample images can be used as the input of the water network structure identification model, that is, the cropped sample images need to meet the input size requirements of the water network structure identification model. For example, the repetition rate can be selected as 0.01 or 0.3, and the sample image size can be cut to 256x256. In order to expand the data set and avoid overfitting, the cut sample images can also be data enhanced by vertical and horizontal flipping.
[0072] Similarly, the segmentation label image and the vectorization label image also need to be cropped in the same way as the satellite remote sensing image data.
[0073] The water network structure identification model training method provided by the embodiment of the present application can fully extract spatial-spectral features of the spectral dimension of the satellite remote sensing image through the 3D convolutional network, and can correctly classify objects that are easily confused with water bodies, and the main network used can also achieve smooth segmentation of pixels at the water-land junction to correctly identify water bodies in a complex background. The water network structure vectorization sub-network is trained again using the vectorization label on the basis of the water body surface segmentation sub-network, perceives the context in a larger receptive field, can ignore some small bridges and highways and the like on the river to connect the water bodies in the identification result, and can also eliminate some scattered incorrectly classified pixels in the background area to obtain a more optimal overall segmentation effect.
[0074] Embodiment Two
[0075] To solve the technical problems in the prior art, the embodiment of the present application provides a water network structure identification model training method.
[0076] Figure 5 The flowchart of the water network structure identification method of the embodiment two of the present application is shown;Figure 5 As shown, the water network structure identification method of the embodiment of the present application comprises the following steps.
[0077] Step S201, acquiring satellite remote sensing image data of a target area.
[0078] Specifically, the satellite remote sensing image data of the target area is acquired by downloading or other reasonable means, and each band of the satellite remote sensing image data is fused to unify the resolution of the satellite remote sensing image data. Then, the satellite remote sensing image data of the target area is cropped according to the cropping method of the satellite remote sensing image data in Embodiment One to obtain a plurality of sample images. For example, the satellite remote sensing image data of the entire target area can be cut into a plurality of 256x256 sample images according to an overlap rate of 0.3, and at this time, the overlapping pixels between every two adjacent sample images are about 78.
[0079] Step S202, inputting the satellite remote sensing image data of the target area into the trained water network structure identification model to obtain the sample image recognition result of the target area.
[0080] Specifically, all the sample images cropped and acquired are input into the trained water network structure identification model, and the trained water network structure identification model outputs a plurality of sample image recognition results of the target area. The trained water network structure identification model is obtained by training the water network structure identification model training method disclosed in Embodiment One.
[0081] Step S203, superimposing all the sample image recognition results to obtain the complete water network structure recognition result of the target area.
[0082] Specifically, when all the sample image recognition results are spliced together, the superimposition is performed in the superimposition mode using addition operation. Through the above superimposition operation, the large image splicing traces can be effectively eliminated, and the error classification caused by the lack of edge information of each small sample can be avoided, and after splicing, the complete water network structure recognition result of the target area is obtained. For example Figure 6 A result schematic diagram of identifying the water network structure of Shanghai City by using the water network structure identification method is shown.
[0083] The water network structure identification method provided by the embodiment of the present application can fully extract spatial-spectral features of the spectral dimension of the satellite remote sensing image by the water body surface segmentation subnetwork in the trained water network structure identification model, and can correctly classify objects that are easily confused with the water body in appearance, and the main branch subnetwork can also achieve relatively smooth segmentation of the pixels at the water-land junction to correctly identify the water body in the complex background; the water network structure vectorization subnetwork can perceive the context in a larger receptive field range, can ignore some small bridges and road obstructions on the river to connect the water bodies in the identification results, and can also eliminate some scattered misclassified pixels in the background area to obtain a more optimal overall segmentation effect.
[0084] Embodiment three
[0085] To solve the technical problems in the prior art, the embodiment of the present application provides a water network structure identification model training device.
[0086] Figure 7 The structure of the water network structure identification model training device of the embodiment of the present application is shown, and the water network structure identification model training device of the embodiment of the present application includes a data set acquisition module and a training module. Figure 7
[0087] The data set acquisition module is used to acquire the training data set, and the training data set includes a plurality of satellite remote sensing image data and the segmentation label image and the vectorization label image corresponding to each satellite remote sensing image data.
[0088] The training module trains the water network structure identification model based on the training data set to obtain the trained water network structure identification model.
[0089] The water network structure identification model includes a water body surface segmentation subnetwork and a water network structure vectorization subnetwork.
[0090] The water body surface segmentation subnetwork is used to extract the spatial information and the dimension information of the satellite remote sensing image data, and to fuse the spatial information and the dimension information into a probability feature map.
[0091] The water network structure vectorization subnetwork is used to vectorize the waterway structure in the probability feature map to obtain a sample image identification result.
[0092] The water network structure recognition device provided by the embodiment of the present application can fully extract spatial-spectrum features of the spectral dimension of satellite remote sensing images through a 3D convolutional network, and can correctly classify objects that are easily confused with water bodies in appearance, and the main network used can also achieve relatively smooth segmentation of pixels at the water-land boundary to correctly identify water bodies in complex backgrounds; the water network structure vectorization sub-network can perform secondary training using vectorization labels on the basis of the water body surface segmentation sub-network, perceive the context in a larger receptive field range, can ignore some small bridges and road obstructions on small rivers, and can connect the water bodies in the recognition result, and the water network structure vectorization sub-network can also eliminate some scattered misclassified pixels in the background area to obtain a more optimal overall segmentation effect.
[0093] Example Four
[0094] To solve the above technical problems in the prior art, the embodiment of the present application further provides a storage medium storing a computer program, which can realize all steps of the water network structure recognition model training method in the embodiment one or the water network structure recognition method in the embodiment two when executed by a processor.
[0095] The specific steps of the water network structure recognition model training method in the embodiment one and the beneficial effects obtained by applying the readable storage medium provided by the embodiment of the present application are the same as those of the embodiment one, and the specific steps of the water network structure recognition model training method in the embodiment two and the beneficial effects obtained by applying the readable storage medium provided by the embodiment of the present application are the same as those of the embodiment two, and thus will not be described here.
[0096] It should be noted that the storage medium includes ROM, RAM, a magnetic disc or an optical disc and various media that can store program codes.
[0097] Example Five
[0098] To solve the above technical problems in the prior art, the embodiment of the present application further provides a terminal.
[0099] Figure 8 The embodiment five of the present application shows a terminal structure schematic diagram, and referring to Figure 8 The terminal of the embodiment includes a processor and a memory connected to each other; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal can realize all steps of the water network structure recognition model training method in the embodiment one or all steps of the water network structure recognition method in the embodiment two when executed.
[0100] The specific steps of the water network structure identification model training method and the beneficial effects of the terminal provided by the embodiment of the application are the same as those of embodiment one, and the specific steps of the water network structure identification method and the beneficial effects of the terminal provided by the embodiment of the application are the same as those of embodiment two, and details thereof will not be described herein.
[0101] It should be noted that the memory can include a random access memory (RAM) and can also include a non-volatile memory, such as at least one disk memory. Similarly, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.
[0102] Although the embodiments disclosed in the present application are as above, the content described is only for the purpose of facilitating understanding of the embodiments adopted by the present application, and is not intended to limit the present application. Any person skilled in the art of the present application can make any modification and change in the form and details without departing from the spirit and scope of the present application, but the protection scope of the present application shall be subject to the scope defined in the appended claims.
Claims
1. A method for training a water network structure recognition model, comprising: Obtain a training dataset, which includes multiple satellite remote sensing image data and segmented label images and vectorized label images corresponding to each satellite remote sensing image data; In the vectorized labeled image, the pixels of the background section that runs across the river are labeled as water. The water network structure recognition model is trained based on the training dataset to obtain the trained water network structure recognition model. The water network structure identification model includes a water surface segmentation sub-network and a water network structure vectorization sub-network. The water surface segmentation sub-network is used to extract the spatial and dimensional information of the satellite remote sensing image data, and to fuse the spatial and dimensional information into a probabilistic feature map. The water network structure vectorization subnetwork is used to vectorize the waterway structure in the probability feature map to obtain the sample image recognition result; The water surface segmentation sub-network includes a backbone sub-network and a 3D convolutional network. Both the backbone sub-network and the 3D convolutional network include downsampling paths and upsampling paths. The vertically corresponding layers of the upsampling paths of the backbone sub-network and the upsampling paths of the 3D convolutional network are connected by a skip connection operation. Furthermore, a multi-scale pooling module is provided between the downsampling path and the upsampling path of the backbone sub-network to encode context information within different scale ranges.
2. The method according to claim 1, characterized in that, The steps for training the water network structure recognition model based on the training dataset to obtain the trained water network structure recognition model include: The water surface segmentation sub-network is trained based on the segmentation training dataset to obtain the trained water surface segmentation sub-network. The segmentation training dataset includes multiple satellite remote sensing image data and segmentation label images corresponding to each satellite remote sensing image data. The satellite remote sensing image data in the segmentation training dataset is input into the trained water surface segmentation sub-network to obtain a set of probability feature maps based on the output probability feature maps. A vectorized training dataset is formed based on the set of probability feature maps and the vectorized label images corresponding to the satellite remote sensing image data. The vectorized subnetwork of the water network structure is trained based on the vectorized training dataset to obtain the trained vectorized subnetwork of the water network structure. The trained water surface segmentation subnetwork and the trained water network structure vectorization subnetwork are connected in series to form a trained water network structure recognition model.
3. The method according to claim 1, characterized in that, The various bands of the satellite remote sensing image data have been fused to unify the resolution of the satellite remote sensing image data.
4. The method according to claim 1, characterized in that, Before inputting the satellite remote sensing image data into the water surface segmentation sub-network, the satellite remote sensing image data, segmentation label image, and vectorized label image need to be cropped in the same way so that the cropped sample image meets the input size requirements of the water network structure recognition model.
5. A method for identifying water network structures, comprising: Acquire satellite remote sensing image data of the target area; The satellite remote sensing image data of the target area is input into the trained water network structure recognition model to obtain the sample image recognition results of the target area. The trained water network structure recognition model is obtained by the water network structure recognition model training method described in any one of claims 1-4.
6. The method according to claim 5, characterized in that, Also includes: The recognition results of all the sample images of the target area are superimposed to obtain the complete water network structure recognition result of the target area.
7. A water network structure recognition model training device, characterized in that, Includes a dataset acquisition module and a training module; The dataset acquisition module is used to acquire a training dataset, which includes multiple satellite remote sensing image data and segmented label images and vectorized label images corresponding to each satellite remote sensing image data; in the vectorized label images, the background pixels that cross the river are labeled as water bodies. The training module trains the water network structure recognition model based on the training dataset to obtain the trained water network structure recognition model. The water network structure identification model includes a water surface segmentation sub-network and a water network structure vectorization sub-network. The water surface segmentation sub-network is used to extract the spatial and dimensional information of the satellite remote sensing image data, and to fuse the spatial and dimensional information into a probabilistic feature map. The water network structure vectorization subnetwork is used to vectorize the waterway structure in the probability feature map to obtain the sample image recognition result; The water surface segmentation sub-network includes a backbone sub-network and a 3D convolutional network. Both the backbone sub-network and the 3D convolutional network include downsampling paths and upsampling paths. The vertically corresponding layers of the upsampling paths of the backbone sub-network and the upsampling paths of the 3D convolutional network are connected by a skip connection operation. Furthermore, a multi-scale pooling module is provided between the downsampling path and the upsampling path of the backbone sub-network to encode context information within different scale ranges.
8. A storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the water network structure recognition model training method as described in any one of claims 1 to 4 or the water network structure recognition method as described in any one of claims 5 to 6.
9. A terminal, characterized in that, include: A processor and a memory, wherein the memory and the processor are communicatively connected; The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory, so that the terminal executes the water network structure recognition model training method as described in any one of claims 1 to 4 or implements the water network structure recognition method as described in any one of claims 5 to 6.
Citation Information
Patent Citations
Building vector profile drawing method, device and equipment and medium
CN114998379A