A method, system and storage medium for identifying and classifying island and reef topography

Abstract

The present invention provides a method, system and storage medium for identifying and classifying island and reef landforms, and relates to the technical field of remote sensing data analysis. The method comprises: acquiring remote sensing data of an island and reef study area; predicting using remote sensing data through a water depth inversion model to obtain water depth data of the island and reef study area; inputting water depth data and remote sensing data into a landform and landform identification and classification model, and obtaining a classification result of the landform and landform of the island and reef study area according to the output of the landform and landform identification and classification model; wherein, pre-training an initial classification model using preset water depth data and preset remote sensing data through comparative learning, and training the pre-trained initial classification model according to the preset water depth data and preset remote sensing data to obtain a landform and landform identification and classification model. The present invention realizes intelligent identification and classification of island and reef landforms with high reliability, and improves the classification effect of island and reef landforms.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing data analysis, and in particular to a method, system and storage medium for identifying and classifying island and reef topography. Background Art

[0002] As small pieces of land surrounded by the sea, islands and reefs provide bases and fulcrums for marine development. In order to obtain and classify the topographic information of islands and reefs, the topography of islands and reefs is generally classified by analyzing remote sensing data. At the same time, in order to improve the classification efficiency, traditional machine learning or deep learning is usually used to replace manual classification of the topography of islands and reefs.

[0003] In the prior art, when classifying the topography of islands and reefs through machine learning or deep learning, the use of a simpler deep learning model and the application of a single island and reef image data for classification will result in inaccurate classification results of the island and reef topography. Based on this, if the accuracy is to be improved, it is necessary to rely on a large amount of island and reef topography data, but the island and reef topography data is difficult to obtain, which makes the classification too difficult. Summary of the invention

[0004] The technical problem solved by the present invention is how to improve the classification effect of island and reef topography.

[0005] The present invention provides a method for identifying and classifying island and reef topography, comprising:

[0006] Obtain remote sensing data of island and reef study areas;

[0007] The remote sensing data is used to predict the water depth of the island reef study area through a water depth inversion model to obtain water depth data;

[0008] Inputting the water depth data and the remote sensing data into a topographic and geomorphic recognition and classification model, and obtaining a classification result of the topography and geomorphic of the island and reef study area according to the output of the topographic and geomorphic recognition and classification model;

[0009] Among them, the initial classification model is pre-trained using preset water depth data and preset remote sensing data through comparative learning, and the pre-trained initial classification model is trained according to the preset water depth data and the preset remote sensing data to obtain the terrain recognition classification model.

[0010] Optionally, the predicting using the remote sensing data through a water depth inversion model to obtain water depth data of the island reef study area includes:

[0011] Enhance the remote sensing data through a preprocessing operation to obtain enhanced remote sensing data;

[0012] Input the enhanced remote sensing data into the water depth inversion model, and extract the image features of the enhanced remote sensing data through the convolutional layer of the water depth inversion model;

[0013] Input the image features into the pooling layer of the water depth inversion model for feature dimensionality reduction to obtain the dimension-reduced image features;

[0014] Input the dimension-reduced image features into the fully connected layer of the water depth inversion model for linear regression prediction to obtain the water depth data.

[0015] Optionally, input the water depth data and the remote sensing data into the terrain and landform recognition and classification model, and obtain the classification result of the terrain and landform of the island reef research area according to the output of the terrain and landform recognition and classification model, including:

[0016] Input the water depth data and the remote sensing data into the terrain and landform recognition and classification model, and respectively segment the water depth data and the remote sensing data through the image segmentation layer of the terrain and landform recognition and classification model to obtain multiple image blocks of the water depth data and the remote sensing data;

[0017] Input the image blocks into the image processing structure of the terrain and landform recognition and classification model to obtain remote sensing data features and water depth features; wherein, the image processing structure includes multiple cascaded image processing sub-structures, and the output data of one image processing sub-structure is the input data of the next image processing sub-structure; use the output data of the last image processing sub-structure as the remote sensing data features and the water depth features;

[0018] Obtain the classification result according to the remote sensing data features and the water depth features.

[0019] Optionally, the obtaining the classification result according to the remote sensing data features and the water depth features includes:

[0020] Perform feature fusion on the remote sensing data features and the water depth features to obtain the fusion features of the island reef research area;

[0021] Classify according to the fusion features to obtain the classification result of the terrain and landform of the island reef research area.

[0022] Optionally, the performing feature fusion on the remote sensing data features and the water depth features to obtain the fusion features of the island reef research area includes:

[0023] First, apply six Swin Transformer modules to perform feature fusion on the remote sensing data features and the water depth features, and then use a pyramid pooling module to further extract multi-scale features to obtain the fusion features.

[0024] Optionally, the classification according to the fusion features to obtain the classification results of the topography of the island reef research area includes:

[0025] Inputting the fusion features into the decoder of the MSTupnet network model, and obtaining the predicted value of the island reef study area through the decoder;

[0026] The classification result is obtained according to the predicted value.

[0027] Optionally, the pre-training of the initial classification model using preset water depth data and preset remote sensing data through comparative learning includes:

[0028] Using a Swin Transformer network to extract features from the preset water depth data and the preset remote sensing data, and then performing a linear transformation on the extracted water depth features and remote sensing data features, so as to map the water depth features and remote sensing data features into corresponding normalized low-dimensional representations;

[0029] Calculating the image-water depth similarity and the water depth-image similarity according to the normalized low-dimensional representation corresponding to the preset water depth feature and the preset image feature;

[0030] The initial classification model is pre-trained by using the image-water depth similarity and the water depth-image similarity to obtain the pre-trained initial classification model.

[0031] Optionally, the pre-training the initial classification model by using the image-water depth similarity and the water depth-image similarity includes:

[0032] Obtaining the probability distribution of the image-water depth similarity and the probability distribution of the water depth-image similarity according to the image-water depth similarity and the water depth-image similarity through a normalization function;

[0033] The initial classification model is pre-trained by cross entropy calculation, according to the probability distribution of the image-water depth similarity and the image-water depth true similarity label and the probability distribution of the water depth-image similarity and the water depth-image true similarity label.

[0034] The present invention also provides an island reef topography and landform identification and classification system, comprising:

[0035] Data acquisition unit, used to obtain remote sensing data of the island and reef research area;

[0036] A prediction unit, used to use the remote sensing data to make predictions through a water depth inversion model to obtain water depth data of the island reef study area;

[0037] A classification unit, used for inputting the water depth data and the remote sensing data into a topography and landform recognition classification model, and obtaining a classification result of the topography and landform of the island and reef research area according to the output of the topography and landform recognition classification model;

[0038] Among them, the initial classification model is pre-trained using preset water depth data and preset remote sensing data through comparative learning, and the pre-trained initial classification model is trained according to the preset water depth data and the preset remote sensing data to obtain the terrain recognition classification model.

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, it implements any of the above-mentioned methods for identifying and classifying island and reef terrain and landforms.

[0040] The method, system and storage medium for identifying and classifying the topography and landforms of islands and reefs of the present invention obtain remote sensing data of the island and reef research area, analyze the remote sensing data using a water depth inversion model, predict large-scale, high-precision water depth data, enhance the dimension of the data, and provide key depth information for the identification of topography and landforms. At the same time, by analyzing the remote sensing data and water depth data through the topography and landform identification classification model, the deep-level features in the data can be mined, and the subtle differences in the topography and landforms can be captured. The obtained classification results are obtained based on richer feature information, which improves the classification accuracy. For the topography and landform identification classification model, contrast learning is used to pre-train with preset water depth data and preset remote sensing data. The pre-training effectively mines the self-supervisory information in the preset water depth data and preset remote sensing data, reduces the dependence on a large number of samples, reduces the difficulty of small sample training, and enables the model to more effectively learn useful information from limited data. The initial classification model after pre-training is then formally trained to obtain a topography and landform identification classification model. It integrates information from different data sources and enhances the model's ability to recognize complex terrain and landforms, thereby achieving highly reliable intelligent recognition and classification of island and reef terrain and landforms. While improving the recognition accuracy of the terrain and landform recognition and classification model for island and reef terrain and landforms, it uses comparative learning for pre-training, thereby eliminating the need for a large amount of island and reef terrain and landform data, improving the classification effect of island and reef terrain and landforms, and providing an important scientific basis for the detection and analysis of islands and reefs. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of a flow chart of a method for identifying and classifying island and reef topography in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the structure of the Swin Transformer network in an embodiment of the present invention;

[0043] Figure 3 Schematic diagram of the structure of two consecutive Swin Transformer modules in an embodiment of the present invention;

[0044] Figure 4 A schematic diagram of the structure of a pre-trained network in an embodiment of the present invention;

[0045] Figure 5 It is a schematic diagram of the structure of the MSTUpnet network in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0047] Combination Figure 1 As shown, the present invention provides a method for identifying and classifying island and reef landforms, comprising:

[0048] Obtain remote sensing data of the island and reef study area.

[0049] Specifically, in a preferred embodiment of the present invention, photon counting lidar and multispectral remote sensing images are used as remote sensing data. The photon counting lidar can provide high-precision terrain data, while the multispectral images can provide rich spectral information.

[0050] The water depth data of the island reef study area is obtained by predicting the remote sensing data using a water depth inversion model.

[0051] Specifically, a one-dimensional convolutional neural network is used to construct an active and passive fusion water depth inversion model. The input layer receives the pixel value of the remote sensing image, the convolution layer extracts features, the pooling layer performs feature dimensionality reduction, and the fully connected layer performs linear regression prediction. The photon data is filtered, denoised and refraction corrected to obtain accurate water depth data. In a preferred embodiment of the present invention, water depth inversion is achieved using a basic one-dimensional convolutional neural network, which consists of three parts, including an input layer, a hidden layer (convolutional layer and pooling layer) and an output layer. The input layer inputs the pixel value of the remote sensing data, the convolution layer is used to extract the features of the remote sensing data, the pooling layer mainly performs feature dimensionality reduction, and finally a linear regression prediction is performed through the fully connected layer, and the output is the predicted water depth data.

[0052] The water depth data and the remote sensing data are input into a terrain and landform recognition and classification model, and the classification result of the terrain and landform of the island and reef study area is obtained according to the output of the terrain and landform recognition and classification model.

[0053] Specifically, the remote sensing images and water depth data are input into the topography and landform recognition classification model, and the topography and landform classification results of the island and reef study area are obtained through analysis by the topography and landform recognition classification model. By analyzing the remote sensing data and water depth data through the topography and landform recognition classification model, the deep-level features in the data can be mined, and the subtle differences in the topography and landforms can be captured. The classification results obtained contain richer feature information.

[0054] Among them, the initial classification model is pre-trained using preset water depth data and preset remote sensing data through comparative learning, and the pre-trained initial classification model is trained according to the preset water depth data and the preset remote sensing data to obtain the terrain recognition classification model.

[0055] Specifically, this step is to establish a terrain and landform recognition classification model. The present invention uses preset water depth data and preset remote sensing data and adopts a contrastive learning method to pre-train the initial classification model. By shortening the distance between positive samples and increasing the distance between negative samples, the self-supervisory information in the data is effectively mined, and the unlabeled data is used to learn a more discriminative feature representation, which reduces the difficulty of training small samples. While improving the recognition accuracy of the terrain and landform recognition classification model for island and reef terrain and landforms, contrastive learning is used for pre-training, thereby not requiring a large amount of island and reef terrain and landform data. In addition, the pre-trained initial classification model is formally trained according to the preset water depth data and the preset remote sensing data, and then the terrain and landform recognition classification model is obtained. The present application combines feature information of different scales together by pyramid pooling to increase the diversity and richness of features, so that the model can better understand the semantic information of different scales of the image, thereby improving the performance and generalization ability of the model. In a preferred embodiment of the present invention, label data can be downloaded from a website, and the label data can be converted from a vector file to a raster file using ArcGIS software, and then the remote sensing data, water depth data and label data are clipped respectively.

[0056] In a preferred embodiment of the present invention, when the initial classification model is pre-trained and formally trained, the preset remote sensing data in the data set used specifically includes remote sensing image data; in the stage of model training preparation, the data used in the embodiment of the present invention is remote sensing image data in the remote sensing data, and in addition to remote sensing image data, photon counting laser radar data is also used in the entire process. The island and reef terrain and landform recognition and classification method of the present invention significantly improves the recognition accuracy and reliability of island and reef terrain and landforms by combining remote sensing image data and photon counting laser radar data. Photon counting laser radar can provide high-precision terrain data, measure the height change of terrain by laser reflection, and thus obtain accurate three-dimensional terrain information; it is crucial for identifying complex terrain and landform features. Multispectral imaging can capture light of different wavelengths, thereby providing rich spectral information; it helps to identify the spectral characteristics of different substances and enhance the recognition ability of terrain and landforms. By comprehensively utilizing multiple data sources, the recognition accuracy and reliability of island and reef terrain and landforms are significantly improved.

[0057] The method for identifying and classifying the topography and landforms of islands and reefs of the present invention obtains remote sensing data of the island and reef research area, analyzes the remote sensing data using a water depth inversion model, predicts large-scale, high-precision water depth data, enhances the dimension of the data, and provides key depth information for the identification of topography and landforms. At the same time, by using remote sensing data and water depth data for prediction, deep-level features in the data can be mined, and subtle differences in topography and landforms can be captured. The obtained classification results contain richer feature information. For the classification model for identifying topography and landforms, contrastive learning is used to pre-train using preset water depth data and preset remote sensing data. Pre-training effectively mines the self-supervisory information in the preset water depth data and preset remote sensing data, reduces the dependence on a large number of samples, reduces the difficulty of training small samples, and enables the model to more effectively learn useful information from limited data. The initial classification model after pre-training is then formally trained to obtain a classification model for identifying topography and landforms. It integrates information from different data sources and enhances the model's ability to recognize complex terrain and landforms, thereby achieving highly reliable intelligent recognition and classification of island and reef terrain and landforms. While improving the recognition accuracy of the terrain and landform recognition and classification model for island and reef terrain and landforms, it uses comparative learning for pre-training, thereby eliminating the need for a large amount of island and reef terrain and landform data, improving the classification effect of island and reef terrain and landforms, and providing an important scientific basis for the detection and analysis of islands and reefs.

[0058] Optionally, the predicting using the remote sensing data through a water depth inversion model to obtain water depth data of the island reef study area includes:

[0059] Enhance the remote sensing data through a preprocessing operation to obtain enhanced remote sensing data;

[0060] Inputting the enhanced remote sensing data into the water depth inversion model, and extracting image features of the enhanced remote sensing data through the convolution layer of the water depth inversion model;

[0061] Inputting the image features into the pooling layer of the water depth inversion model to perform feature dimension reduction, thereby obtaining the image features after dimension reduction;

[0062] The reduced-dimensional image features are input into the fully connected layer of the water depth inversion model for linear regression prediction to obtain the water depth data.

[0063] Specifically, the remote sensing data is enhanced through preprocessing operations to obtain enhanced remote sensing data, wherein, in a preferred embodiment of the present invention, the credibility of the remote sensing data can be enhanced through preprocessing operations such as radiation calibration, atmospheric correction, flare elimination, and water-land separation. The enhanced remote sensing data is input into the water depth inversion model, and the image features of the enhanced remote sensing data are extracted through the convolution layer of the water depth inversion model.

[0064] In a preferred embodiment of the present invention, a one-dimensional convolutional neural network structure can be used to construct a water depth inversion model to obtain water depth data, wherein the water depth inversion model is constructed by combining Figure 2 As shown in the Swin Transformer network structure, this network model is the backbone of the Multi-Source Transformer Upernet (MSTUpnet) overall network model, and is used as an encoder to extract remote sensing data features and water depth features. STAGE1 consists of a linear embedding layer and a Swin Transformer module, and other STAGEs consist of an image block merging layer with downsampling function and a repeatedly stacked Swin Transformer module. Figure 3 As shown in the figure, the structure diagram of two consecutive Swin Transformer modules is shown. This part is used to extract the image features of the enhanced remote sensing data. There are two types of this module, and they are used in pairs continuously. The difference is that the multi-head self-attention modules are different. The odd positions are window multi-head self-attention modules, and the even positions are shift window multi-head self-attention modules. The former effectively reduces the amount of calculation by limiting the attention to the specified window, but there is a lack of communication between windows; the latter effectively solves this problem and realizes cross-window connection through window sliding operation, so that the information between adjacent windows can be effectively utilized to better model global semantic information.

[0065] In this embodiment, the quality of remote sensing data is significantly improved, noise and error are reduced, and the accuracy of water depth prediction is improved through preprocessing operations such as radiation calibration, atmospheric correction, flare elimination, and water-land separation. At the same time, a one-dimensional convolutional neural network is used to build a water depth inversion model to effectively extract and model global semantic information. The multi-head self-attention mechanism of Swin Transformer extracts features, which enhances the model's ability to understand complex scenes.

[0066] Optionally, the step of inputting the water depth data and the remote sensing data into a terrain and landform recognition and classification model, and obtaining a classification result of the terrain and landform of the island and reef study area according to an output of the terrain and landform recognition and classification model, comprises:

[0067] Inputting the water depth data and the remote sensing data into the terrain and landform recognition and classification model, and segmenting the water depth data and the remote sensing data respectively through the image segmentation layer of the terrain and landform recognition and classification model to obtain multiple image blocks of the water depth data and the remote sensing data;

[0068] Input the image block into the image processing structure of the terrain recognition and classification model to obtain remote sensing data features and water depth features; wherein the image processing structure includes a plurality of cascaded image processing substructures, the output data of one image processing substructure is the input data of the next image processing substructure; the output data of the last image processing substructure is used as the remote sensing data features and the water depth features;

[0069] The classification result is obtained according to the remote sensing data characteristics and the water depth characteristics.

[0070] Specifically, first, the water depth data and remote sensing data are input into the terrain recognition and classification model; the image segmentation layer in the terrain recognition and classification model divides the input data into multiple image blocks; the large image data is decomposed into smaller parts to facilitate more detailed analysis and processing by the model. The segmented image blocks are input into the image processing structure of the terrain recognition and classification model. This structure consists of multiple image processing substructures, and the output of each substructure becomes the input of the next substructure. Through the image processing structure, the model can extract remote sensing data features and water depth features. The data output by the last image processing substructure is used as the final remote sensing data features and water depth features.

[0071] For example, it is necessary to classify the topography of an island reef group to determine which areas are suitable for building an ocean observation station. First, high-resolution satellite images, underwater topography data, and other geographic information data of the island reef group are collected. The collected data is preprocessed, including data cleaning and standardization, to meet the input requirements of the topography recognition and classification model. The preprocessed data is trained using the topography recognition and classification model to automatically learn the characteristics of the island reef topography. The image segmentation layer of the model divides the data into multiple image blocks, and then extracts the features of each image block through the image processing structure. Finally, these features are used to classify the topography of the island reef. According to the classification results, the areas in the island reef group suitable for building an ocean observation station are identified, such as areas with flat terrain, moderate water depth, and less vegetation coverage. It can not only improve the accuracy of the classification of island reef topography, but also improve the efficiency and automation of the classification process.

[0072] In this embodiment, the input water depth data and remote sensing data are segmented through the image segmentation layer of the terrain recognition classification model to divide the large image or data set into smaller image blocks. Image segmentation can reduce the computational complexity and help the model analyze the data more carefully and improve the accuracy of feature extraction. Figure 2 As shown, Figure 2This is the SwinTransformer network structure diagram. In this embodiment, the image is first input into the image segmentation layer (which can be the PatchPartition module) for block division, that is, each 4×4 adjacent pixels is a Patch, and then flattened in the channel direction. Assuming that the input is an RGB three-channel image, then each patch has 4×4=16 pixels, and each pixel has three values ​​of R, G, and B, so after flattening it is 16×3=48. Therefore, after Patch Partition, the image shape changes from [H, W, 3] to [H / 4, W / 4, 48]. Then, through the Linear Embedding layer, the channel data of each pixel is linearly transformed from 48 to C, that is, the image shape changes from [H / 4, W / 4, 48] to [H / 4, W / 4, C]. In fact, in the source code, Patch Partition and Linear Embedding are directly implemented through a convolutional layer. Then, feature maps of different sizes are constructed through four STAGEs. Except for STAGE1, which first passes through a Linear Embedding layer, the remaining three STAGEs all pass through a Patch Merging layer for downsampling. Then, Swin TransformerBlock is repeatedly stacked. Note that there are actually two structures of Block here. Finally, for the classification network, a LayerNorm layer, a global pooling layer, and a fully connected layer are added to obtain the final output.

[0073] The overall architecture is divided into multiple stages. The input image size is H*W*3 (the default is 224*224*3); the input image is divided into multiple patches after the patchpartition operation; patch size = 4*4, then the image size after the operation is 56*56*48; stage1: includes a Linear Embedding operation and 2 Swin Transformer blocks. C represents a hyperparameter, that is, the value acceptable to the Transformer. For the Swin-Tiny network, C defaults to 96, and the dimension becomes 56*56*96 (the swin transformer block does not change the image size); the patch partition operation and Linear Embedding operation here are equivalent to the Linear Projection operation of ViT, and in the code, it can be completed with a single convolution operation; stage2: includes Patch Merging and 2 blocks, and the dimension is 28*28*192 at this time; for Patch In general, the merging operation halves the spatial dimension and doubles the number of channels, which is completely equivalent to the convolutional neural network; stage 3: repeat stage 2, with a dimension of 14*14*384; stage 4: repeat stage 2, with a dimension of 7*7*768; In a preferred embodiment of the present invention, there are two consecutive Swin Transformer Blocks, and a Swin Transformer Block consists of a shifted window based MSA with two layers of MLP. A Layer Norm (LN) layer is used before each MSA module and each MLP, and a residual connection is used after each MSA and MLP. The two layers belong to one combination, so the number of blocks in the stage is even.

[0074] The image segmentation layer is used to segment and obtain multiple image blocks, which are processed by four STAGEs in turn, and finally output to obtain remote sensing data features and water depth features. In the preferred embodiment of the present invention, the Swin Transformer network model is used as an encoder, and the output features of the Swin Transformer module of STAGE4 are used as the final remote sensing data features and water depth features.

[0075] In this embodiment, the image block is processed by multiple STAGEs in sequence. This multi-stage processing method allows the model to extract and integrate features at different levels, thereby obtaining richer information. Using the Swin Transformer network model as an encoder, the structured features in the image block can be effectively learned. Swin Transformer is an efficient neural network model that can capture long-distance dependencies.

[0076] Optionally, obtaining the classification result according to the remote sensing data feature and the water depth feature includes:

[0077] Performing feature fusion on the remote sensing data features and the water depth features to obtain fusion features of the island reef study area;

[0078] Classification is performed based on the fusion features to obtain the classification results of the topography and landforms of the island reef study area.

[0079] Specifically, six Swin Transformer modules can be applied to complete the fusion of these two types of features. In the process of classifying the fused features, the classification results of the topography of the island reef study area are obtained by decoding the fused feature results. In a preferred embodiment of the present invention, for the classification of coral reef topography, there is often a large difference in the number of different topography labels, that is, there is a class imbalance problem. In order to alleviate this problem, the loss function can be calculated by using Cross Entropy Loss and Dice Loss weighted by weight. Cross Entropy Loss is used to measure the loss between the model prediction and the actual result. When using Cross Entropy Loss, different weights are assigned according to different types of labels.

[0080] In this embodiment, the model's ability to understand remote sensing data and water depth data is improved through feature fusion and multi-scale feature extraction, thereby improving the overall performance of the model.

[0081] Optionally, the performing feature fusion on the remote sensing data features and the water depth features to obtain the fusion features of the island reef research area includes:

[0082] First, six Swin Transformer modules are applied to fuse the remote sensing data features and the water depth features, and then a pyramid pooling module is used to further extract multi-scale features to obtain the fused features.

[0083] Specifically, the Swin Transformer is used as the encoder to extract the features of remote sensing data and water depth features, and then six Swin Transformer modules are applied to complete the fusion of these two types of features. After feature fusion, a decoding operation needs to be performed to obtain the final classification result. The decoder selected is the UPerNet model. The feature pyramid network in the UPerNet framework is a commonly used feature extraction method that fuses feature information at different scales through pyramid pooling. This method can increase the diversity and richness of features, enabling the model to better understand the semantic information of different scales of the image, thereby improving the performance and generalization ability of the model. Combined with Figure 5 the MSTUpnet network structure diagram shown in, the encoder of the MSTUpnet network model is the Swin Transformer network model. After the Swin

[0084] Transformer network extracts the features of remote sensing data and water depth features, six Swin

[0085] Transformer modules are applied for feature fusion, and then the pyramid pooling module is beneficial for further extracting multi-scale features as fusion features. The pyramid pooling module in this network structure is mainly used to obtain multi-scale high-level features, thereby obtaining more context information and being beneficial for improving the classification accuracy.

[0086] In this embodiment, through the encoder of the MSTUpnet network model, the features of remote sensing data and water depth data are respectively extracted, and then six Swin Transformer modules are applied to complete the fusion of these two types of features. Then, the pyramid pooling module is used to further extract features to obtain multi-scale features as fusion features, which helps to integrate information from different sources and scales and enhance the model's understanding of complex scenes. Feature fusion and multi-scale feature extraction improve the model's understanding ability of remote sensing data and water depth data, thereby improving the overall performance of the model.

[0087] Optionally, classifying according to the fusion features to obtain the classification result of the topography and geomorphology of the island reef research area includes:

[0088] Inputting the fusion features into the decoder of the MSTUpnet network model, and obtaining the predicted value of the island reef research area through the decoder;

[0089] Obtaining the classification result according to the predicted value.

[0090] Specifically, the fused features are input into the decoder of the MSTUPnet network model, and the predicted values ​​of the island reef study area are obtained through the decoder; the classification results are obtained according to the predicted values. In a preferred embodiment of the present invention, for the classification of coral reef topography and geomorphology, there is often a large difference in the number of different topography and geomorphology labels, that is, there is a class imbalance problem. In order to alleviate this problem, the loss function used in the MSTUPnet network model is calculated by weighting Cross Entropy Loss and Dice Loss. Cross Entropy Loss is used to measure the loss between the model prediction and the actual result. When using Cross Entropy Loss, different weights are assigned according to different types of labels; Dice Loss is used to evaluate the overall similarity between the real samples and the predicted results, which can also effectively alleviate the class imbalance problem. The calculation formula of Cross Entropy Loss is:

[0091]

[0092] The Dice Loss calculation formula is:

[0093]

[0094] Among them, Loss ce (y,y pred ) is the loss value between the predicted value and the true value, Loss Dice (y,y pred ) is the loss value between the predicted value and the true value, y represents the true value, y pred Represents the predicted value, n represents the number of classifications, Lossce represents the cross entropy loss function, and LossDice represents the Dice similarity coefficient loss function.

[0095] In this embodiment, the fused features are input into the decoder, which makes full use of the multi-scale and multi-source feature information extracted and fused in the previous steps, and provides a rich data basis for the accurate classification of terrain and landforms.

[0096] Optionally, the pre-training of the initial classification model using preset water depth data and preset remote sensing data through comparative learning includes:

[0097] Using a Swin Transformer network to extract features from the preset water depth data and the preset remote sensing data, and then performing a linear transformation on the extracted water depth features and remote sensing data features, so as to map the water depth features and remote sensing data features into corresponding normalized low-dimensional representations;

[0098] Calculating the image-water depth similarity and the water depth-image similarity according to the normalized low-dimensional representation corresponding to the preset water depth feature and the preset image feature;

[0099] The initial classification model is pre-trained by using the image-water depth similarity and the water depth-image similarity to obtain the pre-trained initial classification model.

[0100] Specifically, since the topographic data of islands and reefs are difficult to obtain and the number of samples is small and inconvenient for training, the contrastive learning method is used for pre-training before the formal training of the model. This method constructs positive sample pairs and negative sample pairs, and during the training process, the distance between positive samples is shortened and the distance between negative samples is increased. This enables the model to effectively mine the self-supervisory information in the data, and use unlabeled data to learn more discriminative feature representations. It can extract as much information as possible from limited samples, thereby reducing dependence on a large number of samples and effectively reducing the difficulty of training small samples. Figure 4 This is a schematic diagram of the structure of the pre-training network of the present invention. The pre-training network uses the Swin Transformer network model as an encoder. First, the model receives image data and water depth data as input. The image data is processed by the encoder and converted into an internal representation that the model can understand. Similarly, the water depth data is also processed by the encoder to extract useful feature information. Among them, Swin Transformer is an effective visual model that can extract local and global features of images. These features include the structure, texture, color, etc. of the image. The momentum model is an exponential moving average version of the basic model parameters; momentum distillation is a self-training method that uses pseudo targets generated by the momentum model as additional supervisory signals to improve the pre-training of the model; momentum update is the process of updating the momentum model parameters. This update method enables the momentum model to smoothly track the changes in the basic model parameters and reduce the drastic fluctuations caused by noise or outliers in the training data. The negative samples in the model are relative to the positive samples, which can help the model learn features better. The model is trained by using a contrast loss function so that the model distinguishes between positive and negative samples and optimizes the decision boundary of the model by minimizing the difference between them. Combined with Figure 4 As shown, the contrastive learning pre-training method mainly refers to the ITC loss function part in the ALBEF model. The preset water depth data and the preset remote sensing data are respectively input into the corresponding Swin Transformer network model, and the corresponding water depth features and remote sensing data features are extracted. According to the preset water depth features and the preset image features, they are mapped into corresponding normalized low-dimensional representations, and the image-water depth similarity and water depth-image similarity are calculated. In the preferred embodiment of the present invention, the following formula can be used for calculation, and the calculation formula is as follows:

[0101] s=g v (v cls ) T g w (w cls );

[0102] Among them, the image-water depth similarity s(I,D) is:

[0103] s(I,D)=g v (v cls ) T g′ w (w′ cls );

[0104] The water depth-image similarity s(D,I) is:

[0105] s(D,I)=g w (w cls ) T g′ v (v′ cls );

[0106] Where s is the definition of similarity, and the corresponding S(I,D) is the image-water depth similarity and S(D,I) is the water depth-image similarity, which are all concretized according to the definition of s; g v , g w are the linear transformations in the basic model, g v ′, g′ w are the linear transformations in the momentum model; v cls 、w cls are the remote sensing data features and water depth features extracted from the basic model, v c ' ls 、w c ' ls They are the remote sensing data features and water depth features extracted from the momentum model respectively.

[0107] In this embodiment, contrastive learning, as a self-supervised learning method, can train the model without or with only a small amount of labeled data, and mine the intrinsic structure in the data. Through contrastive learning, the model can learn a discriminative feature representation, and can extract as much information as possible even when the number of samples is limited. By shortening the distance between positive sample pairs and increasing the distance between negative sample pairs, the feature representations learned by the model can be better aligned, which is helpful for subsequent feature fusion and classification tasks.

[0108] Optionally, the pre-training the initial classification model by using the image-water depth similarity and the water depth-image similarity includes:

[0109] Obtaining the probability distribution of the image-water depth similarity and the probability distribution of the water depth-image similarity according to the image-water depth similarity and the water depth-image similarity through a normalization function;

[0110] The initial classification model is pre-trained by cross entropy calculation, according to the probability distribution of the image-water depth similarity and the image-water depth true similarity label and the probability distribution of the water depth-image similarity and the water depth-image true similarity label.

[0111] Specifically, according to the image-water depth similarity and the water depth-image similarity, the probability distribution of the image-water depth similarity and the probability distribution of the water depth-image similarity are obtained through a normalization function, and the image-water depth and water depth-image similarities are subjected to Softmax normalization. The calculation formula is as follows:

[0112]

[0113] in, It represents the Softmax normalization result of image-water depth similarity. represents the Softmax normalization result of water depth-image similarity; τ is the learnable temperature coefficient, s(I,D m ) is the image-water depth similarity, s(D,I m ) is the water depth-image similarity, M is the size of the queue, and the queue is used to store the feature expression of the most recent M image water depth data pairs from the dynamic single-mode encoder; I represents the image data calculated in the basic model, I m represents the image data calculated in the momentum model; D represents the water depth data calculated in the basic model, D m Represents the water depth data calculated in the momentum model.

[0114] Define y i2d (I) and y d2i (D) is the corresponding true label similarity, then the IDC loss function is defined as:

[0115]

[0116] Where, L idc is the image depth contrast loss function; represents the cross entropy calculation, y i2d (I) is the image-water depth real similarity label, y d2i (D) is the water depth-image true similarity label; p i2d (I) is the probability distribution of the predicted image-water depth similarity, p d2i(D) is the probability distribution of the predicted water depth-image similarity. In addition, momentum distillation technology is applied on the existing basis to learn from the pseudo targets generated by the momentum model. The momentum model is an evolving teacher model composed of an exponential moving average version of a unimodal encoder. During the training process, the base model is trained to match its predictions with those of the momentum model. By using momentum distillation, the model can be more fully trained, which can effectively improve the model's predictive stability and robustness. After applying momentum distillation technology, the final loss function is:

[0117]

[0118] in, is the image depth contrast loss function with momentum distillation improvement; L idc is the image depth contrast loss function. α is the set weight value, KL() represents the KL divergence calculation, and p i2d (I) is the probability distribution of image-water depth similarity predicted by the basic model, p d2i (D) is the probability distribution of water depth-image similarity predicted by the basic model; q i2d (I) is the probability distribution of image-water depth similarity predicted by the momentum model, q d2i (D) is the probability distribution of water depth-image similarity predicted by the momentum model.

[0119] After pre-training, the similarity of paired image depth features can be made higher and higher, while the similarity of unpaired image depth features can be made lower and lower, thereby achieving the effect of feature alignment, which is beneficial to feature fusion and further improves classification accuracy.

[0120] In this embodiment, the similarity probability distributions of image-water depth and water depth-image are calculated by Softmax normalization, so that the model can evaluate the similarity on a unified scale.

[0121] The present invention also provides an island reef topography and landform recognition and classification system, comprising:

[0122] Data acquisition unit, used to obtain remote sensing data of the island and reef research area;

[0123] A prediction unit, used to use the remote sensing data to make predictions through a water depth inversion model to obtain water depth data of the island reef study area;

[0124] A classification unit, used for inputting the water depth data and the remote sensing data into a topography and landform recognition classification model, and obtaining a classification result of the topography and landform of the island and reef research area according to the output of the topography and landform recognition classification model;

[0125] Among them, the initial classification model is pre-trained using preset water depth data and preset remote sensing data through comparative learning, and the pre-trained initial classification model is trained according to the preset water depth data and the preset remote sensing data to obtain the terrain recognition classification model.

[0126] The island reef topography and landform recognition and classification system of the present invention obtains remote sensing data of the island reef study area, analyzes the remote sensing data using a water depth inversion model, predicts large-scale, high-precision water depth data, enhances the dimension of the data, and provides key depth information for the recognition of topography and landforms. At the same time, by using remote sensing data and water depth data for prediction, deep-level features in the data can be mined, and subtle differences in topography and landforms can be captured. The obtained classification results contain richer feature information. For the topography and landform recognition and classification model, contrast learning is used to use preset water depth data and preset remote sensing data for pre-training. Through pre-training, the self-supervision information in the preset water depth data and preset remote sensing data is effectively mined, which reduces the dependence on a large number of samples and reduces the difficulty of small sample training, so that the model can more effectively learn useful information from limited data. The initial classification model after pre-training is then formally trained to obtain a topography and landform recognition and classification model. It integrates information from different data sources and enhances the model's ability to recognize complex terrain and landforms, thereby achieving highly reliable intelligent recognition and classification of island and reef terrain and landforms. While improving the recognition accuracy of the terrain and landform recognition and classification model for island and reef terrain and landforms, it uses comparative learning for pre-training, thereby eliminating the need for a large amount of island and reef terrain and landform data, improving the classification effect of island and reef terrain and landforms, and providing an important scientific basis for the detection and analysis of islands and reefs.

[0127] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, it implements any of the above-mentioned methods for identifying and classifying island and reef terrain and landforms.

[0128] The computer-readable storage medium of the present invention obtains remote sensing data of the island reef study area, analyzes the remote sensing data using a water depth inversion model, predicts large-scale, high-precision water depth data, enhances the dimension of the data, and provides key depth information for the identification of topography and landforms. At the same time, by using remote sensing data and water depth data for prediction, deep-level features in the data can be mined, and subtle differences in topography and landforms can be captured. The obtained classification results contain richer feature information. For the classification model for topography and landform recognition, contrastive learning is used to pre-train with preset water depth data and preset remote sensing data. Pre-training effectively mines self-supervisory information in preset water depth data and preset remote sensing data, reduces dependence on a large number of samples, reduces the difficulty of small sample training, and enables the model to more effectively learn useful information from limited data. The initial classification model after pre-training is then formally trained to obtain a classification model for topography and landform recognition. It integrates information from different data sources and enhances the model's ability to recognize complex terrain and landforms, thereby achieving highly reliable intelligent recognition and classification of island and reef terrain and landforms. While improving the recognition accuracy of the terrain and landform recognition and classification model for island and reef terrain and landforms, it uses comparative learning for pre-training, thereby eliminating the need for a large amount of island and reef terrain and landform data, improving the classification effect of island and reef terrain and landforms, and providing an important scientific basis for the detection and analysis of islands and reefs.

[0129] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0130] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for identifying and classifying island and reef topography, characterized in that: include: Obtain remote sensing data of island and reef study areas; The remote sensing data is used to predict the water depth of the island reef study area through a water depth inversion model to obtain water depth data; Inputting the water depth data and the remote sensing data into a topographic and geomorphic recognition and classification model, and obtaining a classification result of the topography and geomorphic of the island and reef study area according to the output of the topographic and geomorphic recognition and classification model; The initial classification model is pre-trained using preset water depth data and preset remote sensing data through comparative learning, including: using a Swin Transformer network to extract features of the preset water depth data and the preset remote sensing data, and then performing a linear transformation on the extracted water depth features and image features of the remote sensing data, and mapping the water depth features and image features of the remote sensing data into corresponding normalized low-dimensional representations; Calculating the image-water depth similarity and the water depth-image similarity according to the normalized low-dimensional representation corresponding to the water depth feature and the image feature of the remote sensing data; Pre-training the initial classification model by using the image-water depth similarity and the water depth-image similarity to obtain the pre-trained initial classification model specifically includes: Obtaining the probability distribution of the image-water depth similarity and the probability distribution of the water depth-image similarity according to the image-water depth similarity and the water depth-image similarity through a normalization function; Pre-training the initial classification model by calculating the corresponding loss value according to the probability distribution of the image-water depth similarity and the image-water depth true similarity label and the probability distribution of the water depth-image similarity and the water depth-image true similarity label through cross entropy calculation; The pre-trained initial classification model is trained according to the preset water depth data and the preset remote sensing data to obtain the terrain recognition classification model.

2. The method for identifying and classifying island and reef topography according to claim 1, characterized in that: The water depth inversion model is used to predict the remote sensing data to obtain the water depth data of the island reef study area, including: Enhance the remote sensing data through a preprocessing operation to obtain enhanced remote sensing data; Inputting the enhanced remote sensing data into the water depth inversion model, and extracting image features of the enhanced remote sensing data through the convolution layer of the water depth inversion model; Inputting the image features into the pooling layer of the water depth inversion model to perform feature dimension reduction, thereby obtaining the image features after dimension reduction; The reduced-dimensional image features are input into the fully connected layer of the water depth inversion model for linear regression prediction to obtain the water depth data.

3. The method for identifying and classifying island and reef topography according to claim 1, characterized in that: The step of inputting the water depth data and the remote sensing data into a topographic and geomorphic recognition and classification model, and obtaining the classification result of the topography and geomorphic of the island and reef study area according to the output of the topographic and geomorphic recognition and classification model, comprises: Inputting the water depth data and the remote sensing data into the terrain and landform recognition and classification model, and segmenting the water depth data and the remote sensing data respectively through the image segmentation layer of the terrain and landform recognition and classification model to obtain multiple image blocks of the water depth data and the remote sensing data; The image block is input into the image processing structure of the terrain recognition and classification model to obtain the image features and water depth features of the remote sensing data; wherein the image processing structure includes a plurality of cascaded image processing substructures, the output data of one image processing substructure is the input data of the next image processing substructure; the output data of the last image processing substructure is used as the image features and water depth features of the remote sensing data; The classification result is obtained according to the image features of the remote sensing data and the water depth features.

4. The method for identifying and classifying island and reef topography according to claim 3 is characterized in that: The obtaining of the classification result according to the image features and the water depth features of the remote sensing data includes: Performing feature fusion on the image features of the remote sensing data and the water depth features to obtain fusion features of the island reef study area; Classification is performed based on the fusion features to obtain the classification results of the topography and landforms of the island reef study area.

5. The method for identifying and classifying island and reef topography according to claim 4 is characterized in that: The feature fusion of the image features and the water depth features of the remote sensing data to obtain the fusion features of the island reef research area includes: First, six Swin Transformer modules are applied to fuse the image features of the remote sensing data and the water depth features, and then a pyramid pooling module is used to further extract multi-scale features to obtain the fused features.

6. The method for identifying and classifying island and reef topography according to claim 5 is characterized in that: The classification according to the fusion features is performed to obtain the classification results of the topography and landforms of the island reef research area, including: Inputting the fusion features into a decoder of the MSTupnet network model, and obtaining the predicted value of the island reef study area through the decoder; The classification result is obtained according to the predicted value.

7. A system for identifying and classifying island and reef topography, characterized in that: include: Data acquisition unit, used to obtain remote sensing data of the island and reef research area; A prediction unit, used to use the remote sensing data to make predictions through a water depth inversion model to obtain water depth data of the island reef study area; A classification unit, used for inputting the water depth data and the remote sensing data into a topography and landform recognition classification model, and obtaining a classification result of the topography and landform of the island and reef research area according to the output of the topography and landform recognition classification model; The initial classification model is pre-trained using preset water depth data and preset remote sensing data through comparative learning, including: using a Swin Transformer network to extract features of the preset water depth data and the preset remote sensing data, and then performing a linear transformation on the extracted water depth features and image features of the remote sensing data, and mapping the water depth features and image features of the remote sensing data into corresponding normalized low-dimensional representations; Calculating the image-water depth similarity and the water depth-image similarity according to the normalized low-dimensional representation corresponding to the water depth feature and the preset image feature; Pre-training the initial classification model by using the image-water depth similarity and the water depth-image similarity to obtain the pre-trained initial classification model specifically includes: Obtaining the probability distribution of the image-water depth similarity and the probability distribution of the water depth-image similarity according to the image-water depth similarity and the water depth-image similarity through a normalization function; Pre-training the initial classification model by calculating the corresponding loss value according to the probability distribution of the image-water depth similarity and the image-water depth true similarity label and the probability distribution of the water depth-image similarity and the water depth-image true similarity label through cross entropy calculation; The pre-trained initial classification model is trained according to the preset water depth data and the preset remote sensing data to obtain the terrain recognition classification model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for identifying and classifying island and reef topography according to any one of claims 1 to 6 is implemented.

Images