Single-wavelength airborne sounding laser radar sea-land waveform classification method

Abstract

The invention discloses a single-wavelength airborne sounding laser radar sea-land waveform classification method, which belongs to the technical field of remote sensing detection and signal processing, is used for sea-land waveform classification, and comprises the following steps: obtaining original waveform data, and carrying out preprocessing by adopting a wavelet soft threshold and continuous wavelet transform to obtain a one-dimensional echo waveform effective signal and a two-dimensional time-frequency graph; and a dual-path multi-modal feature fusion network is constructed, the one-dimensional echo waveform effective signal and the two-dimensional time-frequency graph are used as input, a sea-land waveform classification result is output, and the dual-path multi-modal feature fusion network comprises an attention convolution residual network module, a one-dimensional convolution neural network time feature extraction module and a multi-modal data fusion module. According to the method, the time-frequency representation is introduced to enhance the discrimination information, the feature content is enriched, the two types of complementary information of the time-domain waveform and the time-frequency representation are utilized to construct a joint feature learning and fusion mechanism, the waveform confusion in a single feature space is reduced, and the classification stability of the near-shore land and water boundary region is improved.

Description

Technical Field

[0001] This invention discloses a method for classifying land and sea waveforms using a single-wavelength airborne depth sounding lidar, belonging to the field of remote sensing and signal processing technology. Background Technology

[0002] Currently, algorithms for classifying sea and land waveforms using airborne ALB can be broadly categorized into two types: the first is a waveform decomposition and manual feature extraction method based on physical models, and the second is an end-to-end modeling method based on deep learning. The core idea of ​​the former is "interpret the waveform first, then classify it." Specifically, a parametric model (e.g., representing echo components as several Gaussian or generalized Gaussian shapes) is used to fit the observed waveform curve, decomposing a complex waveform into multiple components with more defined physical meanings. In aquatic scenarios, these components often correspond to surface reflections and bottom reflections. After fitting, features that distinguish between sea and land are extracted from the fitting result and the original waveform. These features are then input into a traditional machine learning classifier (such as random forest or support vector machine) to output categories such as water, land, or "fuzzy / uncertain." This method often requires iterative optimization calculations for waveform fitting and decomposition, resulting in significant time consumption when dealing with massive amounts of waveform data. Furthermore, the fitting and peak detection processes are sensitive to settings such as initial values, thresholds, and the number of components, often requiring manual intervention and repeated parameter tuning in engineering, limiting automation. Especially in extremely shallow water environments, surface and bottom echoes may overlap or merge on the time axis, leading to unstable decomposition, feature confusion, and consequently, decreased classification reliability. The core idea of ​​the second method is to "let the model learn discriminative information from the waveform itself," reducing manual feature design and tedious waveform analysis. Typically, a one-dimensional full waveform sequence is directly input into a one-dimensional convolutional neural network or its improved structure. Local morphological features are automatically extracted through convolutional layers, and the classification head outputs the sea / land category. Some studies also construct multi-channel convolutional networks (e.g., using multiple receiving channels or waveforms in multiple bands simultaneously), fusing multi-channel information through the network structure to improve discriminative ability. This method eliminates the process of fitting waveforms one by one and manual feature selection, resulting in faster inference speeds after training. While fast and easier to batch process and automate large-scale data, this approach relies on system configuration and available data types. When the system provides multi-wavelength or multi-channel information, the network can utilize additional discriminative clues such as the differences in land and sea reflections across different wavebands, making classification easier. However, when only single-wavelength waveform information is available, the lack of auxiliary spectral information makes waveforms more prone to "mixing and similarity," especially in shallow water where surface and bottom echoes overlap. Water surface waveforms may be highly morphologically similar to those of land or artificial targets (such as aquaculture rafts, shoreline fragments, and wetlands), increasing the risk of misclassification in boundary areas. In other words, insufficient input information and shallow water waveform aliasing under single-wavelength conditions are key factors affecting the stability of end-to-end models, making the achievement of refined land and sea waveform classification based solely on single-wavelength information a persistent challenge. Summary of the Invention

[0003] The purpose of this invention is to provide a sea and land waveform classification method for single-wavelength airborne depth sounding lidar, in order to solve the problems of low computational efficiency, strong dependence on manual parameter tuning, and unstable decomposition in shallow water in existing physical model-based methods, while single-wavelength end-to-end methods based on deep learning have insufficient feature discrimination ability and low classification reliability under waveform aliasing conditions.

[0004] A method for classifying land and sea waveforms using a single-wavelength airborne depth sounding lidar system includes: S1. Obtain the original waveform data and perform preprocessing, including using the wavelet soft thresholding algorithm to denoise the original waveform to obtain the one-dimensional echo waveform effective signal, mapping the one-dimensional echo waveform effective signal to a two-dimensional time-frequency representation through continuous wavelet transform, and performing amplitude normalization processing to obtain a two-dimensional time-frequency diagram. S2. Construct a dual-path multimodal feature fusion network. Use the one-dimensional echo waveform effective signal and the two-dimensional time-frequency diagram as input to train the neural network. Evaluate the classification accuracy of the dual-path multimodal feature fusion network in each training cycle until the classification accuracy reaches the expected value and output the sea and land waveform classification results. The dual-path multimodal feature fusion network includes an attention-based convolutional residual network module, a one-dimensional convolutional neural network temporal feature extraction module, and a multimodal data fusion module. The attention-based convolutional residual network module is a residual neural network embedded with an attention convolution module. It takes a two-dimensional time-frequency map as input and a time-frequency spatial feature vector as output. The residual neural network consists of stacked residual blocks. The attention convolution module includes a channel attention module and a spatial attention module. The one-dimensional convolutional neural network temporal feature extraction module consists of stacked one-dimensional convolutional blocks. It takes a one-dimensional echo waveform effective signal as input and a temporal feature vector as output. The time-frequency spatial feature vector and the temporal feature vector are input to the multimodal data fusion module to generate fused features and output the sea-land waveform classification result.

[0005] S1 includes, S1.1, using a wavelet soft thresholding algorithm to denoise the original waveform and selecting the tail of the waveform. Each sampling point is used to set a waveform threshold, and the process is iterated through. A sampling point is used, and points larger than the waveform threshold are selected as the valid signal of the one-dimensional echo waveform. : ; In the formula, Let the number of effective signals in the one-dimensional echo waveform be denoted as . For the effective signal index of the one-dimensional echo waveform, For the first A one-dimensional echo waveform is an effective signal.

[0006] S1 includes, S1.2, and will Represented as a continuous-time signal Selecting the mother wavelet function Through the Perform scaling and time shifting to generate wavelet basis functions. : ; ; In the formula, For continuous time variables, For scale parameters, For translation parameters, This is the normalization factor.

[0007] S1 includes, S1.3, utilizing right Perform a continuous wavelet transform to obtain the corresponding continuous wavelet transform coefficients. : ; In the formula, The differential symbol, It is the conjugate of the mother wavelet function; S1 includes S1.4, selecting a set of discrete scales. With Discrete Translation Set ,Will Mapped to a two-dimensional matrix : ; In the formula, The number of discrete scale parameters. The number of translation scale parameters, For discrete scale parameter indexes. , Discrete translation parameter index , For the first A discrete scale value, For the first Each discrete translation value; Will Generate a two-dimensional amplitude graph by taking the amplitude values: ; In the formula, It is the absolute value symbol. The amplitude time-frequency matrix, ; Using global minimum and maximum normalization pairs Normalization is performed to obtain a two-dimensional time-frequency diagram. : ; In the formula, , , To prevent extremely small positive numbers with a denominator of zero.

[0008] S2 includes, S2.1, and Input a one-dimensional convolutional neural network temporal feature extraction module, including... Two one-dimensional convolutional blocks are input sequentially. Each one-dimensional convolutional block consists of a one-dimensional convolutional layer and a pooling layer. The output of each one-dimensional convolutional block is first generated by the one-dimensional convolutional layer. ; In the formula, The feature map output by the one-dimensional convolutional layer. For convolution kernel index, For the first Each convolutional kernel weight vector For convolution operations, For activation functions; S2 includes, S2.2, and will The input pooling layer performs downsampling; after processing through two one-dimensional convolutional blocks, the output is a fixed-dimensional temporal feature vector. .

[0009] S2 includes S2.3, which involves inputting the two-dimensional time-frequency map into the front-end feature embedding layer of the attention convolutional residual network, and mapping the two-dimensional time-frequency map to the initial features through two-dimensional convolution operations. The residual neural network consists of four residual blocks, each of which is configured to learn a residual function. and through jump connections and Add: ; ; In the formula, For the target mapping function to be learned directly, This is an index of the number of trainable parameter layers. For the first Layer trainable parameters, This is the residual output.

[0010] The attention convolution module receives the output of the front-end feature embedding layer of the attention convolution residual network, and the input features of the attention convolution module are... The value is .

[0011] The attention convolutional module receives the output of the residual block, and the input features of the attention convolutional module are... The value is .

[0012] S2 includes S2.4, for input features First, input the channel attention module, and then... The spatial dimension is subjected to average pooling and max pooling, and then a channel attention map is generated through a multilayer perceptron. ,Will and Element-wise multiplication yields intermediate features : ; In the formula, The symbol for element-wise multiplication; S2 includes, S2.5, and will The input spatial attention module performs average pooling and max pooling along the channel dimension and concatenates them, then performs convolution to generate a spatial attention map. ,Will and Element-wise multiplication yields the final enhanced feature. ; ; S2 includes, S2.6, and will Perform equivalent mapping to a fixed-dimensional time-frequency space feature vector .

[0013] S3 includes, S3.1, and and The input multimodal data fusion module uses a two-stage summation process. The first stage uses a fully connected layer to... and By projecting each feature onto the same dimension and summing the results, the fused features are obtained. ; S3 includes S3.2, the second level will The input is a classifier consisting of a fully connected layer and a softmax function, and the output is a classification result of land and sea waveforms. .

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention enhances the discrimination information by introducing time-frequency representation, enriches the feature content, and at the same time utilizes two complementary types of information, time-domain waveform and time-frequency representation, to construct a joint feature learning and fusion mechanism, reduce waveform confusion under a single feature space, and improve the classification stability of nearshore land-water boundary areas. Attached Figure Description

[0015] Figure 1 These are the experimental results of the land-sea classification method of this invention;

[0016] Figure 2 These are infrared laser point cloud images acquired at the same time;

[0017] Figure 3 It is a comparison of land and sea classification results using fuzzy clustering;

[0018] Figure 4 This is the result of the land-sea classification comparison using a support vector machine;

[0019] Figure 5 It is a comparison of land and sea classification results from a one-dimensional convolutional neural network;

[0020] Figure 6 This is the result of the comparison between land and sea classification using the method of this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] A method for classifying land and sea waveforms using a single-wavelength airborne depth sounding lidar system includes: S1. Obtain the original waveform data and perform preprocessing, including using the wavelet soft thresholding algorithm to denoise the original waveform to obtain the one-dimensional echo waveform effective signal, mapping the one-dimensional echo waveform effective signal to a two-dimensional time-frequency representation through continuous wavelet transform, and performing amplitude normalization processing to obtain a two-dimensional time-frequency diagram. S2. Construct a dual-path multimodal feature fusion network. Use the one-dimensional echo waveform effective signal and the two-dimensional time-frequency diagram as input to train the neural network. Evaluate the classification accuracy of the dual-path multimodal feature fusion network in each training cycle until the classification accuracy reaches the expected value and output the sea and land waveform classification results. The dual-path multimodal feature fusion network includes an attention-based convolutional residual network module, a one-dimensional convolutional neural network temporal feature extraction module, and a multimodal data fusion module. The attention-based convolutional residual network module is a residual neural network embedded with an attention convolution module. It takes a two-dimensional time-frequency map as input and a time-frequency spatial feature vector as output. The residual neural network consists of stacked residual blocks. The attention convolution module includes a channel attention module and a spatial attention module. The one-dimensional convolutional neural network temporal feature extraction module consists of stacked one-dimensional convolutional blocks. It takes a one-dimensional echo waveform effective signal as input and a temporal feature vector as output. The time-frequency spatial feature vector and the temporal feature vector are input to the multimodal data fusion module to generate fused features and output the sea-land waveform classification result.

[0023] S1 includes, S1.1, using a wavelet soft thresholding algorithm to denoise the original waveform and selecting the tail of the waveform. Each sampling point is used to set a waveform threshold, and the process is iterated through. A sampling point is used, and points larger than the waveform threshold are selected as the valid signal of the one-dimensional echo waveform. : ; In the formula, Let the number of effective signals in the one-dimensional echo waveform be denoted as . For the effective signal index of the one-dimensional echo waveform, For the first A one-dimensional echo waveform is an effective signal.

[0024] S1 includes, S1.2, and will Represented as a continuous-time signal Selecting the mother wavelet function Through the Perform scaling and time shifting to generate wavelet basis functions. : ; ; In the formula, For continuous time variables, For scale parameters, For translation parameters, This is the normalization factor.

[0025] S1 includes, S1.3, utilizing right Perform a continuous wavelet transform to obtain the corresponding continuous wavelet transform coefficients. : ; In the formula, The differential symbol, It is the conjugate of the mother wavelet function; S1 includes S1.4, selecting a set of discrete scales. With Discrete Translation Set ,Will Mapped to a two-dimensional matrix : ; In the formula, The number of discrete scale parameters. The number of translation scale parameters, For discrete scale parameter indexes. , Discrete translation parameter index , For the first A discrete scale value, For the first Each discrete translation value; Will Generate a two-dimensional amplitude graph by taking the amplitude values: ; In the formula, It is the absolute value symbol. The amplitude time-frequency matrix, ; Using global minimum and maximum normalization pairs Normalization is performed to obtain a two-dimensional time-frequency diagram. : ; In the formula, , , To prevent extremely small positive numbers with a denominator of zero.

[0026] S2 includes, S2.1, and Input a one-dimensional convolutional neural network temporal feature extraction module, including... Two one-dimensional convolutional blocks are input sequentially. Each one-dimensional convolutional block consists of a one-dimensional convolutional layer and a pooling layer. The output of each one-dimensional convolutional block is first generated by the one-dimensional convolutional layer. ; In the formula, The feature map output by the one-dimensional convolutional layer. For convolution kernel index, For the first Each convolutional kernel weight vector For convolution operations, For activation functions; S2 includes, S2.2, and will The input pooling layer performs downsampling; after processing through two one-dimensional convolutional blocks, the output is a fixed-dimensional temporal feature vector. .

[0027] S2 includes S2.3, which involves inputting the two-dimensional time-frequency map into the front-end feature embedding layer of the attention convolutional residual network, and mapping the two-dimensional time-frequency map to the initial features through two-dimensional convolution operations. The residual neural network consists of four residual blocks, each of which is configured to learn a residual function. and through jump connections and Add: ; ; In the formula, For the target mapping function to be learned directly, This is an index of the number of trainable parameter layers. For the first Layer trainable parameters, This is the residual output.

[0028] The attention convolution module receives the output of the front-end feature embedding layer of the attention convolution residual network, and the input features of the attention convolution module are... The value is .

[0029] The attention convolutional module receives the output of the residual block, and the input features of the attention convolutional module are... The value is .

[0030] S2 includes S2.4, for input features First, input the channel attention module, and then... The spatial dimension is subjected to average pooling and max pooling, and then a channel attention map is generated through a multilayer perceptron. ,Will and Element-wise multiplication yields intermediate features : ; In the formula, The symbol for element-wise multiplication; S2 includes, S2.5, and will The input spatial attention module performs average pooling and max pooling along the channel dimension and concatenates them, then performs convolution to generate a spatial attention map. ,Will and Element-wise multiplication yields the final enhanced feature. ; ; S2 includes, S2.6, and will Perform equivalent mapping to a fixed-dimensional time-frequency space feature vector .

[0031] S3 includes, S3.1, and and The input multimodal data fusion module uses a two-stage summation process. The first stage uses a fully connected layer to... and By projecting each feature onto the same dimension and summing the results, the fused features are obtained. ; S3 includes S3.2, the second level will The input is a classifier consisting of a fully connected layer and a softmax function, and the output is a classification result of land and sea waveforms. .

[0032] The following description, in conjunction with the embodiments and accompanying drawings, further illustrates the land-sea classification verification. A nearshore sea area was used for verification. The water depth was relatively shallow, typically approximately 2 to 5 meters, with a relatively high sediment content. Aquaculture activities were observed within the sea, with numerous floating aquaculture rafts and other facilities visible, exhibiting clear characteristics of complex nearshore targets. Experimental data was acquired using the Optech CZMIL bathymetry system. The raw data contained 344,107 laser points, manually divided into training and test sets. Simultaneously, the acquired data included corresponding infrared laser point clouds for evaluating the land-sea classification results.

[0033] Experiments were conducted on the collected data using the method of this invention, and settings were configured. The expected classification accuracy is 97%. The neural network model of this invention is trained with a waveform threshold calculated as the average of the last 100 sampling points of the waveform plus three times the standard deviation. The preprocessed one-dimensional green echo waveform and its corresponding two-dimensional time-frequency representation are used as input. The network automatically performs feature extraction and category discrimination. The classification results are as follows: Figure 1 As shown, the infrared laser point cloud obtained at the same time is as follows: Figure 2 As shown, the method of the present invention accurately identifies the overall sea and land areas, and the classification map shows good spatial continuity with relatively smooth boundary transitions. It can better maintain the true distribution pattern of land features. Especially in the sea surface aquaculture raft area, the proposed scheme can clearly delineate its range and has a strong ability to characterize and distinguish the subtle waveform differences between the shallow water aquaculture area and the surrounding water surface.

[0034] To further illustrate the experimental results of the method of this invention, fuzzy clustering, support vector machine, and one-dimensional convolutional neural network methods were used to classify the land-sea waveforms in the experimental area. Comparison of results from different schemes is also provided. Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown, fuzzy clustering exhibits significant misclassification, primarily manifested in a large number of sea areas being misclassified as land, and poor identification of aquaculture raft areas, making it difficult to form complete and stable target boundaries. Support vector machines and one-dimensional convolutional neural networks show some improvement in classification results compared to fuzzy clustering, but misclassifications are still concentrated at the sea-land boundary, and both tend to over-identify aquaculture raft areas. In contrast, the method of this invention has fewer misclassified points near the land-water boundary, a smoother and more continuous classification map, and a lower misclassification rate within the aquaculture raft area.

[0035] The classification performance of different methods was quantitatively evaluated by classifying accuracy. The overall classification accuracy and kappa coefficient of fuzzy clustering were 79.16% and 0.4598, respectively; the overall classification accuracy and kappa coefficient of support vector machine were 95.50% and 0.8617, respectively; the overall classification accuracy and kappa coefficient of one-dimensional convolutional neural network were 95.26% and 0.8590, respectively; and the overall classification accuracy and kappa coefficient of the method of this invention were 97.33% and 0.9168, respectively. This demonstrates that the method of this invention has the advantages of high automation and high accuracy.

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying land and sea waveforms using a single-wavelength airborne depth-sounding lidar, characterized in that, include: S1. Obtain the original waveform data and perform preprocessing, including using the wavelet soft thresholding algorithm to denoise the original waveform to obtain the one-dimensional echo waveform effective signal, mapping the one-dimensional echo waveform effective signal to a two-dimensional time-frequency representation through continuous wavelet transform, and performing amplitude normalization processing to obtain a two-dimensional time-frequency diagram. S2. Construct a dual-path multimodal feature fusion network. Use the one-dimensional echo waveform effective signal and the two-dimensional time-frequency diagram as input to train the neural network. Evaluate the classification accuracy of the dual-path multimodal feature fusion network in each training cycle until the classification accuracy reaches the expected value and output the sea and land waveform classification results. The dual-path multimodal feature fusion network includes an attention-based convolutional residual network module, a one-dimensional convolutional neural network temporal feature extraction module, and a multimodal data fusion module. The attention-based convolutional residual network module is a residual neural network with an embedded attention convolutional module. The attention-based convolutional residual network module takes a two-dimensional time-frequency map as input and a time-frequency spatial feature vector as output. The residual neural network consists of stacked residual blocks. The attention convolutional module includes a channel attention module and a spatial attention module. The one-dimensional convolutional neural network temporal feature extraction module consists of stacked one-dimensional convolutional blocks, takes a one-dimensional echo waveform effective signal as input, and a temporal feature vector as output. The time-frequency spatial feature vector and the time-series feature vector are input into the multimodal data fusion module to generate fused features and output the sea-land waveform classification results.

2. The sea-land waveform classification method of a single-wavelength airborne depth-sounding lidar according to claim 1, characterized in that, S1 includes, S1.1, using a wavelet soft thresholding algorithm to denoise the original waveform and selecting the tail of the waveform. Each sampling point is used to set a waveform threshold, and the process is iterated through. A sampling point is used, and points larger than the waveform threshold are selected as the valid signal of the one-dimensional echo waveform. : ； In the formula, Let the number of effective signals in the one-dimensional echo waveform be denoted as . For the effective signal index of the one-dimensional echo waveform, For the first A one-dimensional echo waveform is an effective signal.

3. The sea-land waveform classification method of a single-wavelength airborne depth-sounding lidar according to claim 2, characterized in that, S1 includes, S1.2, and will Represented as a continuous-time signal Selecting the mother wavelet function Through the Perform scaling and time shifting to generate wavelet basis functions. : ； ； In the formula, For continuous time variables, For scale parameters, For translation parameters, This is the normalization factor.

4. The sea-land waveform classification method of a single-wavelength airborne depth-sounding lidar according to claim 3, characterized in that, S1 includes, S1.3, utilizing right Perform a continuous wavelet transform to obtain the corresponding continuous wavelet transform coefficients. : ； In the formula, The differential symbol, It is the conjugate of the mother wavelet function; S1 includes S1.4, selecting a set of discrete scales. With Discrete Translation Set ,Will Mapped to a two-dimensional matrix : ； In the formula, The number of discrete scale parameters. The number of translation scale parameters, For discrete scale parameter indexes. , Discrete translation parameter index , For the first A discrete scale value, For the first Each discrete translation value; Will Generate a two-dimensional amplitude graph by taking the amplitude values: ; In the formula, It is the absolute value symbol. The amplitude time-frequency matrix, ; Using global minimum and maximum normalization pairs Normalization is performed to obtain a two-dimensional time-frequency diagram. : ； In the formula, , , To prevent extremely small positive numbers with a denominator of zero.

5. The sea-land waveform classification method of a single-wavelength airborne depth-sounding lidar according to claim 4, characterized in that, S2 includes, S2.1, and Input a one-dimensional convolutional neural network temporal feature extraction module, including... Two one-dimensional convolutional blocks are input sequentially. Each one-dimensional convolutional block consists of a one-dimensional convolutional layer and a pooling layer. The output of each one-dimensional convolutional block is first generated by the one-dimensional convolutional layer. ； In the formula, The feature map output by the one-dimensional convolutional layer. For convolution kernel index, For the first Each convolutional kernel weight vector For convolution operations, For activation functions; S2 includes, S2.2, and will The input pooling layer performs downsampling; after processing through two one-dimensional convolutional blocks, the output is a fixed-dimensional temporal feature vector. .

6. The sea-land waveform classification method of a single-wavelength airborne depth-sounding lidar according to claim 5, characterized in that, S2 includes S2.3, which involves inputting the two-dimensional time-frequency map into the front-end feature embedding layer of the attention convolutional residual network, and mapping the two-dimensional time-frequency map to the initial features through two-dimensional convolution operations. The residual neural network consists of four residual blocks, each of which is configured to learn a residual function. and through jump connections and Add: ； ； In the formula, For the target mapping function to be learned directly, This is an index of the number of trainable parameter layers. For the first Layer trainable parameters, This is the residual output.

7. The sea-land waveform classification method of a single-wavelength airborne depth-sounding lidar according to claim 6, characterized in that, The attention convolution module receives the output of the front-end feature embedding layer of the attention convolution residual network, and the input features of the attention convolution module are... The value is .

8. A method for classifying land and sea waveforms using a single-wavelength airborne depth-sounding lidar according to claim 6, characterized in that, The attention convolutional module receives the output of the residual block, and the input features of the attention convolutional module are... The value is .

9. A method for classifying sea and land waveforms using a single-wavelength airborne depth-sounding lidar according to claim 7 or 8, characterized in that, S2 includes S2.4, for input features First, input the channel attention module, and then... The spatial dimension is subjected to average pooling and max pooling, and then a channel attention map is generated through a multilayer perceptron. ,Will and Element-wise multiplication yields intermediate features : ； In the formula, The symbol for element-wise multiplication; S2 includes, S2.5, and will The input spatial attention module performs average pooling and max pooling along the channel dimension and concatenates them, then performs convolution to generate a spatial attention map. ,Will and Element-wise multiplication yields the final enhanced feature. ; ； S2 includes, S2.6, and will Perform equivalent mapping to a fixed-dimensional time-frequency space feature vector .

10. A method for classifying land and sea waveforms using a single-wavelength airborne depth-sounding lidar according to claim 9, characterized in that, S3 includes, S3.1, and and The input multimodal data fusion module uses a two-stage summation process. The first stage uses a fully connected layer to... and By projecting each feature onto the same dimension and summing the results, the fused features are obtained. ; S3 includes S3.2, the second level will The input is a classifier consisting of a fully connected layer and a softmax function, and the output is a classification result of land and sea waveforms. .

Images