Ground feature classification method based on satellite-borne laser radar waveform data

Abstract

The invention discloses a ground feature classification method based on satellite-borne laser radar waveform data, and relates to the field of remote sensing technology and artificial intelligence, and the method comprises the following application steps: S1, obtaining laser echo waveform data of a to-be-classified region through a satellite-borne laser radar; s2, preprocessing the waveform data of the laser echo; the preprocessing comprises quality evaluation and standardization processing; s3, inputting the preprocessed laser echo waveform data into a pre-trained ground feature classification model, and outputting a ground feature classification result; the terrain classification result comprises one or more terrain types in a paddy field and a dry land and one or more terrain types in a broad-leaved forest and a coniferous forest; the ground feature classification model is a dense connection neural network model. The method has the advantages that the ground feature classification precision is improved, and ground feature types with similar spectral features and different structural features are effectively distinguished.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing technology and artificial intelligence, specifically to a method for classifying ground features based on satellite-borne lidar waveform data. Background Technology

[0002] Traditional remote sensing image classification methods primarily rely on spectral information for land cover identification. However, this approach has significant limitations when dealing with complex terrain and similar land cover types. Spectral information struggles to effectively distinguish between "same object, different spectral signature" and "different object, same spectral signature." For example, paddy fields and dry land may have similar spectral characteristics but are actually different land cover types; similarly, different vegetation types may exhibit similar spectral features in specific spectral bands. This limitation results in classification accuracy that fails to meet practical application requirements, especially in scenarios demanding fine-grained classification.

[0003] Full-waveform lidar (LiDAR) technology provides a new dimension for land cover classification by acquiring vertical structural information of ground features. This technology can record the entire process of laser pulse interaction with ground features, containing rich three-dimensional structural feature information. However, effectively extracting and utilizing these waveform features for accurate classification remains a challenge. Traditional waveform processing methods often rely on manual feature extraction and simple classification algorithms, making it difficult to fully mine the deep feature information contained in the waveform data.

[0004] Deep learning technology, especially DenseNet, has demonstrated powerful feature extraction capabilities in image processing. This network achieves feature reuse through dense connections, effectively extracting multi-level features. However, applying this technology to satellite-borne LiDAR waveform data processing requires addressing a series of technical challenges, including waveform data preprocessing, feature extraction, and model training. Existing methods still have shortcomings in waveform data quality assessment, standardization, and model generalization ability, making it difficult to meet the practical needs of large-scale, multi-type land cover classification.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] In view of this, the present invention provides a ground feature classification method based on satellite-borne lidar waveform data, which has the advantages of improving ground feature classification accuracy and effectively distinguishing ground feature types with similar spectral features but different structural features.

[0007] This invention provides a method for land cover classification based on spaceborne lidar waveform data, comprising the following application steps: S1. Acquire laser echo waveform data of the area to be classified using a spaceborne lidar; S2. Preprocess the laser echo waveform data; the preprocessing includes quality assessment and standardization. S3. Input the preprocessed laser echo waveform data into the pre-trained land cover classification model and output the land cover classification results; the land cover classification results include at least one land cover type among paddy fields and dry land, or at least one land cover type among broad-leaved forests and coniferous forests; the land cover classification model is a densely connected neural network model.

[0008] In one optional implementation, the land cover classification model is obtained through the following training steps: The laser echo waveform data and corresponding surface elevation data of multiple single land cover types areas were obtained by spaceborne lidar, and high-resolution remote sensing images of the corresponding areas were obtained as reference data. Based on the reference data, the original waveform data is labeled with land cover categories through visual interpretation to construct a sample dataset with category labels; The laser echo waveform data in the sample dataset are subjected to quality assessment and standardization to form a standardized sample set; A densely connected neural network model is constructed, and the model is trained using the normalized sample set. The model parameters are iteratively optimized until the model converges, resulting in a well-trained land cover classification model.

[0009] In one optional implementation, the quality assessment specifically includes: Calculate the signal-to-noise ratio and standard deviation of the laser echo waveform data, and remove laser echo waveform data with a signal-to-noise ratio lower than a preset threshold or an abnormal standard deviation. The formula for calculating the signal-to-noise ratio is: ; in, The maximum value of the laser echo waveform intensity. The mean of the background noise. The standard deviation of the background noise; The standardization process specifically includes: The corresponding saturation threshold is determined based on the laser gain parameters; The laser echo waveform data is normalized based on the saturation threshold. The calculation formula for the normalization process is as follows: ; in, These are the sampling points for the laser echo waveform data. The minimum value in the laser echo waveform data. This is the saturation threshold corresponding to the laser gain parameter. y These are the sampling points for the normalized laser echo waveform data.

[0010] In one alternative implementation, the saturation threshold The threshold is determined by a predefined gain-saturation threshold mapping table, which contains a continuous correspondence from threshold = 30 when gain value ≤ 8 to threshold = 239 when gain value ≥ 28.

[0011] In one optional implementation, the densely connected neural network model employs a densely connected structure, specifically including: Feature extraction module: Composed of convolutional layers, normalization layers, activation functions and pooling layers, used to extract multi-level features from the input laser echo waveform data; Feature fusion module: It consists of multiple dense blocks and transition layers alternately. Each dense block contains multiple convolutional units and uses dense connection to achieve feature reuse. Classification decision module: includes a global feature pooling layer and a fully connected classification layer, used to output the final land cover classification result.

[0012] In one optional implementation, in the feature fusion module, each dense block consists of several convolutional units, and the output feature map of each convolutional unit is concatenated with the input feature maps of all subsequent layers in the channel dimension. Transition layers are placed between dense blocks, and feature map size and number of channels are controlled through convolution and pooling operations.

[0013] In one alternative implementation, during the training step, a cross-entropy loss function is used to measure the difference between the predicted result and the true label. The adaptive moment estimation algorithm is used to optimize the model parameters, and a weight decay regularization term is set to prevent overfitting. An early stopping mechanism is used to monitor the performance of the validation set, and the training process is terminated when the performance no longer improves.

[0014] In one optional implementation, after the land cover classification model has been trained, the performance of the land cover classification model is evaluated using sample datasets from different geographical regions to verify the generalization ability and adaptability of the land cover classification model. The performance evaluation was conducted using a multi-dimensional indicator system, including overall accuracy, Kappa coefficient, precision, recall, and F1 score.

[0015] In one optional implementation, the method is capable of distinguishing land cover types with similar spectral characteristics but different structural characteristics, including: Based on the differences in echo energy characteristics caused by humidity differences, dry land and paddy fields can be distinguished. Based on the differences in multi-peak / single-wave echo characteristics caused by differences in canopy structure, broad-leaved forests and coniferous forests can be distinguished.

[0016] As can be seen from the above, the ground feature classification method based on satellite-borne lidar waveform data provided in this application obtains waveform data through satellite-borne lidar and uses densely connected neural networks for feature extraction and classification, which solves the technical problem that traditional methods cannot effectively distinguish spectrally similar ground feature types, and has the advantages of improving classification accuracy and enhancing model generalization ability. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for classifying ground features based on satellite-borne lidar waveform data according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the training and application phases of a ground feature classification method based on satellite-borne lidar waveform data according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of DenseNet, a densely connected neural network model according to an embodiment of the present invention; Figure 4 This is a classification confusion matrix diagram of the DenseNet densely connected neural network model according to an embodiment of the present invention for paddy field and dry land datasets; Figure 5 This is a classification confusion matrix diagram of the DenseNet densely connected neural network model according to an embodiment of the present invention for broadleaf forest and coniferous forest datasets. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In existing technologies, remote sensing image classification mainly relies on spectral information for land cover identification, but this has significant limitations in classifying land cover in complex terrain areas and with similar spectral characteristics. For example, paddy fields and dry land have similar spectral reflectance characteristics due to differences in water content, and broadleaf forests and coniferous forests have highly similar spectral responses in specific bands, making it difficult for traditional methods to accurately distinguish them. Although full-waveform lidar technology can acquire vertical structure information of land cover, existing classification methods have failed to effectively combine waveform features with deep learning models, resulting in insufficient extraction of subtle differences hidden in waveform data and limiting the improvement of classification accuracy.

[0021] To address the aforementioned issues, the inventors discovered that traditional classification methods cannot effectively utilize the vertical dimension information contained in lidar waveform data when processing land cover types with similar spectra but different structural features. By analyzing the differences in waveform characteristics among different land cover types—for example, the difference between the single-peak waveform of paddy fields due to surface water cover and the multi-peak waveform of dry land due to rough surfaces, as well as the difference in echo characteristics between dense broadleaf forest canopies and layered coniferous forests—a technical approach combining waveform data preprocessing with deep feature learning is proposed. To address the problem of waveform data being affected by environmental noise and equipment parameters, a dynamic standardization processing method is designed; and to address the insufficient feature reuse of traditional convolutional neural networks, a dense connection structure is introduced to enhance the hierarchical extraction of waveform features.

[0022] Therefore, as Figure 1 As shown, this invention proposes a method for land cover classification based on spaceborne lidar waveform data, including the following application steps: Step S1: Obtain laser echo waveform data of the area to be classified using a spaceborne lidar.

[0023] Step S2: Preprocess the laser echo waveform data; preprocessing includes quality assessment and standardization.

[0024] Step S3: Input the preprocessed laser echo waveform data into the pre-trained land cover classification model and output the land cover classification results; the land cover classification results include at least one land cover type in paddy fields and dry land, or at least one land cover type in broad-leaved forests and coniferous forests; the land cover classification model is a densely connected neural network model.

[0025] Among them, spaceborne lidar refers to optical remote sensing equipment mounted on a satellite platform. It obtains three-dimensional spatial information by emitting laser pulses and receiving surface reflection signals. Specifically, it can be implemented using photon counting type or full waveform type lidar systems to obtain raw waveform data containing the vertical structural features of ground objects.

[0026] Preferably, GLA01 (laser echo waveform data) generated by GLAS on ICESat satellite can be used as a typical data source. This system is a full-waveform lidar that can completely record the change process of echo energy of laser pulses interacting with the ground over time. It is especially suitable for large-scale, high-precision extraction of information such as vegetation vertical profiles, ice sheet topography and sea surface elevation. Its data structure and accuracy characteristics can provide reliable support for the development and verification of related remote sensing inversion algorithms.

[0027] Quality assessment refers to screening valid data by calculating signal-to-noise ratio and standard deviation. Specifically, this can be achieved by using a sliding window to calculate background noise statistics, thus eliminating the impact of atmospheric scattering and equipment noise on data quality. Standardization involves dynamically adjusting the data normalization range based on laser operating parameters. This can be achieved by using a gain-saturation threshold mapping table to determine the normalization baseline value, thus eliminating data scale inconsistencies caused by differences in equipment parameters under different observation conditions. The DenseNet neural network model is a deep neural network classification model based on a density-connected structure. Dense connections are a topological connection method for cross-layer feature reuse in neural networks. Specifically, this can be achieved by cascading the output feature map channels of each convolutional layer within a dense block, enhancing the model's ability to capture local morphological features of waveform curves.

[0028] Specifically, after acquiring raw echo waveform data of the target area using a lidar sensor mounted on a satellite platform, the raw data is first assessed for quality, such as calculating the signal-to-noise ratio (SNR) of each waveform sequence. Data with an SNR below a set threshold is considered invalid and discarded. Then, a saturation threshold is determined by querying a predefined mapping table based on the laser gain parameters. The valid waveform data is then normalized to unify waveform data acquired under different observation conditions to the same numerical range. The preprocessed waveform data is input into a densely connected neural network model (DenseNet). The model extracts local undulation features of the waveform curve through convolutional layers and utilizes the dense connection structure to reuse and fuse multi-level features. Finally, a classification layer outputs the land cover type identification result. For example, when classifying paddy fields and dry land, the model can identify differences in surface roughness based on the number of waveform peaks and attenuation slope; when distinguishing between broadleaf and coniferous forests, the model can identify differences in canopy density based on the waveform backscattering intensity distribution.

[0029] Compared to existing technologies, traditional methods rely solely on spectral information for land cover classification, making it difficult to distinguish between land cover types with similar reflectance characteristics but different structural features. This invention, by introducing full-waveform lidar data and combining it with a densely connected neural network model (DenseNet), can simultaneously utilize both the spectral reflectance characteristics and vertical structural features of land cover for classification decisions. For example, in classifying paddy fields and dry land with similar spectral characteristics, traditional methods are prone to confusion, while this invention can accurately identify differences in surface moisture by analyzing the peak distribution characteristics of the waveform curves. In classifying broadleaf forests and coniferous forests, traditional methods rely on leaf spectral characteristics and are easily affected by seasonal changes, while this invention maintains classification stability through canopy structure feature analysis.

[0030] Through the above technical solution, this invention can effectively distinguish land cover types with similar spectral features but significant differences in vertical structure. For example, it can distinguish between paddy fields and dry land based on waveform energy attenuation characteristics, and between broad-leaved forests and coniferous forests based on echo peak distribution characteristics. This method solves the problem of insufficient accuracy of traditional classification methods in complex terrain areas and the identification of similar land covers. By integrating waveform data preprocessing and deep feature learning techniques, it improves the accuracy and environmental adaptability of land cover classification.

[0031] In one alternative implementation, such as Figure 2 As shown, the land cover classification model is obtained through the following training steps: The laser echo waveform data and corresponding surface elevation data of multiple single land cover types areas were obtained by spaceborne lidar, and high-resolution remote sensing images of the corresponding areas were obtained as reference data. Based on reference data, the original waveform data is labeled with land cover categories through visual interpretation to construct a sample dataset with category labels; The laser echo waveform data in the sample dataset are subjected to quality assessment and standardization to form a standardized sample set; A densely connected neural network model (DenseNet) is constructed, and the model is trained using a normalized sample set. The model parameters are iteratively optimized until the model converges, resulting in a well-trained land cover classification model.

[0032] Among them, a single land cover type region refers to a geographical area with a single surface cover type and continuous spatial distribution. This can be verified through field surveys or high-resolution image interpretation to ensure the purity of the sample data categories. Quality assessment refers to the quantitative analysis of the signal-to-noise ratio and standard deviation of laser echo waveform data. Specifically, a preset threshold screening mechanism can be used to remove low-quality data to improve the overall quality of the sample set. Standardization processing refers to dynamically adjusting the normalization parameters according to the laser gain parameters. This can be achieved by establishing a gain-saturation threshold mapping table to eliminate the influence of equipment parameter differences on waveform amplitude. The deep neural network classification model refers to a deep learning architecture with a densely connected structure. Specifically, it can achieve multi-level feature learning by constructing feature extraction modules, feature fusion modules, and classification decision modules.

[0033] In one alternative implementation, the spaceborne lidar data includes ICESat / GLAS GLA01 (laser echo waveform data) and GLA14 (surface elevation data); the ground reference data includes high-resolution satellite remote sensing data (such as historical Google Earth imagery).

[0034] As an example, spaceborne lidar data can be provided by the U.S. National Snow and Ice Data Center. First, a region with a single land cover type is selected as the study area, and its latitude and longitude range is determined. Then, using this latitude and longitude range, GLA01 and GLA14 data points covering the study area are selected from the U.S. National Snow and Ice Data Center for further analysis and processing. Ground reference imagery data primarily uses historical imagery from Google Earth, selecting images with minimal time difference from the spaceborne lidar data as ground reference data. Next, visual interpretation methods are used to determine the land cover category (e.g., buildings, woodland, bare land, water bodies, etc.) to which the spaceborne lidar waveform data belongs. These interpretation results will serve as the training and validation sample sets for a densely connected neural network model, supporting subsequent model training and validation.

[0035] It should be noted that for flat and smooth surfaces (such as bare land and water bodies), echo signals typically exhibit an ideal Gaussian waveform with good symmetry. However, on smooth surfaces with a certain slope, due to the difference in the incident and reflection angles, the echo signal shows significant broadening, and this broadening effect becomes more pronounced with increasing surface slope. For complex surfaces, such as buildings and vegetation, echo signals usually consist of multiple peaks. When the laser beam interacts with a building, the echo signal is often more significant and has a larger amplitude, but the broadening effect is not obvious. In vegetated areas, branches and leaves make the echo signal difficult to distinguish, and due to low reflectivity, the echo amplitude is usually smaller. Based on these characteristics, the sample dataset can be optimized to improve the training effect and classification accuracy of deep learning models.

[0036] Specifically, in the sample construction phase, a multi-source data support system is formed by simultaneously acquiring laser echo waveforms, surface elevation, and high-resolution imagery data. Elevation data provides a terrain correction benchmark for waveform interpretation, while high-resolution imagery provides spatial detail reference for visual interpretation. This multi-source collaborative mechanism effectively solves the problem of insufficient spatial resolution of single waveform data. In the data preprocessing stage, a dual screening mechanism of signal-to-noise ratio calculation and standard deviation analysis is used to remove abnormal samples with severe noise interference from the data source. Simultaneously, a dynamic normalization method based on gain parameters is employed to eliminate the influence of equipment parameter differences on waveform amplitude, forming a standardized input data format. The densely connected neural network model achieves feature reuse through a densely connected structure. In the feature extraction stage, convolutional operations capture local waveform features; in the feature fusion stage, cross-layer connections between dense blocks integrate multi-scale feature information; and finally, global pooling and fully connected layers complete the classification decision. During training, regularization techniques and early stopping mechanisms are used to balance model complexity and generalization ability, ensuring stable convergence of the model parameter optimization process.

[0037] Compared to existing technologies, traditional methods typically rely on single waveform data and lack rigorous quality control processes, making them susceptible to noise interference that leads to decreased classification accuracy. The shallow classification models used in existing technologies struggle to effectively extract deep features from waveform data, exhibiting insufficient feature representation capabilities when handling complex land cover types. This proposed solution enhances sample reliability through a multi-source data collaborative annotation mechanism, ensures data quality through a rigorous quality assessment process, and employs a deep, densely connected network to enhance feature representation capabilities, significantly improving classification accuracy while maintaining model generalization ability.

[0038] Through the above technical solution, this invention effectively solves the problem of limited classification performance caused by noise interference and insufficient feature representation in waveform data. The multi-source data collaborative annotation mechanism ensures the accuracy of sample labels, quality assessment and standardization improve the signal-to-noise ratio and consistency of input data, and the deep dense connection network enhances the model's ability to capture subtle structural features through feature reuse mechanisms. This technical solution can stably process waveform data obtained with different gain parameters and exhibits stronger adaptability and robustness in complex terrain regions and spectrally similar land cover classification scenarios.

[0039] In one alternative implementation, the quality assessment specifically involves: Calculate the signal-to-noise ratio (SNR) and standard deviation of the laser echo waveform data, and remove laser echo waveform data with a SNR lower than a preset threshold or an abnormal standard deviation; The formula for calculating the signal-to-noise ratio is: ; in, The maximum value of the laser echo waveform intensity. The mean of the background noise. The standard deviation of the background noise; The standardization process specifically involves: The corresponding saturation threshold is determined based on the laser gain parameters; The laser echo waveform data is normalized based on the saturation threshold. The formula for normalization is: ; in, These are the sampling points for the laser echo waveform data. The minimum value in the laser echo waveform data. This is the saturation threshold corresponding to the laser gain parameter. y These are the sampling points for the normalized laser echo waveform data.

[0040] The signal-to-noise ratio (SNR) is the ratio of the peak signal intensity to the background noise statistic. It is calculated by subtracting the mean background noise from the maximum intensity value and then dividing by the noise standard deviation. This quantifies the effective information content of the waveform data. The standard deviation is a statistical indicator of the waveform data's dispersion, obtained by calculating the square root of the variance of the intensity values ​​at each sampling point. This is used to identify waveform samples with abnormal shapes. The saturation threshold is the upper limit of the signal intensity corresponding to the laser gain parameter. It is dynamically determined using a predefined gain-saturation threshold mapping table to adapt to data characteristics under different gain conditions. Normalization is the process of linearly transforming the original waveform intensity values ​​to a preset range. This is achieved by subtracting the minimum value from the sampling point intensity and then dividing by the difference between the saturation threshold and the minimum value. This eliminates dimensional differences caused by different gain settings.

[0041] Specifically, the quality assessment phase first involves statistically analyzing the background noise of the original waveform data, calculating its mean and standard deviation as a noise benchmark. Then, the maximum waveform intensity is extracted, and the signal-to-noise ratio (SNR) is calculated based on the noise statistics. When the SNR is below a preset threshold, it indicates that the waveform is severely affected by noise and is therefore discarded. Simultaneously, the overall standard deviation of the waveform data is calculated; when the standard deviation exceeds the normal range, it is identified as an abnormal waveform and filtered out.

[0042] By estimating the background noise of the reflected echo and calculating its signal-to-noise ratio (SNR) and standard deviation, invalid waveform data from the sample dataset, which may be affected by cloud cover, instrument malfunction, or other interference, can be filtered out. Generally, a smaller standard deviation and a larger SNR indicate less noise in the signal, higher waveform quality, and more distinct characteristics. As an example, the SNR and standard deviation of the laser echo waveform data can be estimated by calculating the background noise of the first and last 50 sampling points.

[0043] To improve data retrieval accuracy and classification reliability, laser echo waveform data needs to be normalized (i.e., standardized). In the standardization stage, the effective range of the current data is determined based on the laser's real-time gain parameters and a preset gain-saturation threshold correspondence. Based on this range, a linear transformation is performed on the retained waveform data, mapping the original intensity values ​​to a uniform numerical interval, ensuring the comparability of waveform data acquired under different gain conditions.

[0044] Compared to existing technologies, traditional methods typically use only the signal-to-noise ratio (SNR) as a single metric for data screening, failing to effectively identify samples with abnormal waveform morphology. This proposed solution employs a dual evaluation mechanism of SNR and standard deviation, enabling simultaneous detection of both noise interference and waveform distortion. Regarding standardization, existing technologies often use fixed thresholds for normalization, which cannot adapt to range differences caused by variations in gain parameters. This solution utilizes a dynamic mapping relationship between gain and saturation threshold to achieve adaptive standardization of waveform data under different acquisition parameters, resolving the data scale inconsistency problem caused by neglecting differences in equipment parameters in traditional methods.

[0045] Through the above technical solution, this invention effectively solves the problem of noise interference and gain differences in laser echo waveform data affecting classification accuracy. The quality assessment stage can screen out valid data with acceptable signal-to-noise ratio and normal waveform morphology, while the standardization process eliminates range differences caused by different gain settings, enabling subsequent classification models to receive input data with uniform scale and reliable quality. This technical solution improves data consistency under different acquisition conditions while ensuring data validity, providing a stable and reliable data foundation for deep learning models.

[0046] In one alternative implementation, the saturation threshold The threshold is determined by a predefined gain-saturation threshold mapping table, which contains a continuous correspondence from threshold = 30 when gain value ≤ 8 to threshold = 239 when gain value ≥ 28.

[0047] The gain-saturation threshold mapping table refers to a quantized correspondence table formed in advance by experimentally calibrating the optimal saturation threshold for different gain parameters. Specifically, it can be implemented using a combination of laboratory calibration and linear interpolation, by establishing the gain parameters... With physical saturation threshold The mathematical relationship model is used to achieve parameter matching.

[0048] Among them, the continuous correspondence refers to the uninterrupted linear or nonlinear functional relationship between the gain parameter and the saturation threshold. Specifically, it can be achieved by using piecewise linear functions or polynomial fitting methods. Mathematical modeling ensures a smooth transition in the threshold adjustment process when the gain parameter changes.

[0049] Specifically, the photoelectric conversion saturation point of the laser at different gain values ​​is calibrated in the laboratory, and the physical relationship between the gain parameter and the corresponding saturation threshold is recorded, forming gain-saturation threshold data pairs covering the typical operating range. During the data processing stage, based on the current laser gain parameter value, the corresponding saturation threshold is automatically matched by querying a mapping table, and then normalization is performed. When the gain parameter is in the range of 8 to 28, a linear interpolation method is used to calculate the intermediate threshold, enabling continuous threshold adjustment under dynamic gain parameter changes.

[0050] As an example, when normalizing laser echo waveform data, it is necessary to consider the differences in the energy receiving threshold of the GLAS satellite-borne laser altimeter system. The GLAS satellite has a dynamic threshold, setting different saturation thresholds for different gain values, as shown in Table 1.

[0051] Table 1 ;

[0052] Compared to existing technologies, traditional methods typically use fixed thresholds or manually set thresholds, which cannot adapt to working scenarios where laser gain parameters are dynamically adjusted, leading to the accumulation of data standardization errors. This solution combines experimental calibration with mathematical modeling to establish a precise correspondence between gain parameters and physical saturation thresholds, eliminating the subjective bias of manually set thresholds while ensuring data standardization consistency under different gain conditions.

[0053] Through the above technical solution, the present invention effectively solves the problem of saturation threshold determination error caused by dynamic changes in the gain parameters of spaceborne lidar. By using a predefined mapping table and continuous correspondence, the invention achieves automatic threshold matching and smooth transition, ensuring the scientific nature and consistency of waveform data normalization processing under different working conditions, and providing standardized input data quality assurance for classification models.

[0054] In one alternative implementation, such as Figure 3 As shown, the densely connected neural network model adopts a densely connected structure, specifically including a feature extraction module, a feature fusion module, and a classification decision module.

[0055] The feature extraction module (i.e., the initial convolutional layer) consists of convolutional layers, normalization layers, activation functions, and pooling layers, and is used to extract multi-level features from the input laser echo waveform data.

[0056] Specifically, the initial convolutional layer consists of a 7×1 convolutional layer (64 output channels), batch normalization (BatchNorm1d), ReLU activation function, and a 3×1 max pooling layer (stride 2) to initially extract features and reduce sequence length.

[0057] The feature fusion module (i.e., the dense connection module) consists of multiple dense blocks and transition layers, with each dense block containing multiple convolutional units (i.e., dense layers). It uses dense connections to achieve feature reuse. Specifically, in the feature fusion module, each dense block consists of several convolutional units (i.e., dense layers), and the output feature map of each convolutional unit is concatenated with the input feature maps of all subsequent layers in the channel dimension. Transition layers are placed between dense blocks, and feature map size and number of channels are controlled through convolution and pooling operations.

[0058] Furthermore, the alternating structure of dense blocks and transition layers contains a total of 4 dense blocks, each of which (except for the last one) is connected to a transition layer; Dense Block: Composed of multiple dense layers, each of which extracts features through 1×1 convolutions (128 intermediate channels) and 3×1 convolutions (32 output channels, i.e., growth rate), and concatenates them with the input features along the channel dimension (dense connectivity). The four dense blocks have 6, 12, 24, and 16 layers respectively.

[0059] Transition layer: Composed of batch normalization, ReLU activation function, 1×1 convolution (halving the number of channels) and 2×1 average pooling, used to compress the number of feature channels and further reduce the sequence length.

[0060] The classification decision module (i.e., the classification layer) includes a global feature pooling layer and a fully connected classification layer, which are used to output the final land cover classification results.

[0061] Specifically, the classification layer uses adaptive average pooling (compressing the sequence length to 1) to obtain global features.

[0062] The classification results are output through a two-layer fully connected network (the number of channels in the middle layer is half the number of feature channels, including ReL activation and 50% Dropout).

[0063] Among them, dense connection structure refers to concatenating the output feature map of each convolutional unit with the input feature maps of all subsequent layers in the channel dimension. Specifically, it can be implemented by channel concatenation operation. By preserving the original feature information of each layer's output, it promotes the cross-layer reuse of features from different layers.

[0064] The feature extraction module consists of convolutional layers, normalization layers, activation functions, and pooling layers. Specifically, it can use a one-dimensional convolutional layer with a kernel size of 3 to extract local waveform features, combine it with a batch normalization layer to stabilize the data distribution, introduce nonlinear transformation capabilities through the ReLU activation function, and use the max pooling layer to compress the feature dimension, forming the basis for multi-level feature expression.

[0065] The feature fusion module consists of multiple dense blocks and transition layers alternately. Each dense block contains multiple convolutional units, specifically a dense block structure containing 4 convolutional units. Feature reuse is achieved through dense connections. The transition layer compresses the number of channels through 1×1 convolution and combines it with an average pooling operation with a stride of 2 to control the feature map size.

[0066] Specifically, in the laser echo waveform data processing, the feature extraction module captures local waveform morphological features through convolutional layers, normalization layers eliminate data distribution offsets, activation functions enhance the model's nonlinear expressive power, and pooling layers filter key features and reduce dimensionality. The feature fusion module achieves cross-layer fusion of shallow detailed features and deep semantic features through dense connections of multiple convolutional units within dense blocks. The transition layer balances computational complexity and feature preservation requirements through channel compression and spatial downsampling. The classification decision module utilizes global pooling layers to eliminate spatial location sensitivity, and fully connected layers integrate information from all feature channels to complete the classification decision.

[0067] Compared to existing technologies, traditional convolutional neural networks rely solely on layer-by-layer propagation mechanisms during feature fusion, which can easily lead to the loss of shallow, detailed features. Densely connected structures, through cross-layer feature reuse, enable the network to simultaneously utilize low-level waveform details and high-level semantic features, enhancing its ability to capture subtle waveform differences. The introduction of transition layers preserves effective information while avoiding the increased model complexity caused by a surge in the number of feature map channels.

[0068] Through the above technical solution, this invention solves the problem of insufficient multi-level feature extraction from laser echo waveform data. It achieves efficient reuse of waveform detail features through a dense connection mechanism, improving the model's ability to distinguish between humidity differences between paddy fields and dry land, and between broadleaf and coniferous forest canopy structures. The transition layer structure reduces computational resource consumption while maintaining classification accuracy, and the global feature pooling layer enhances the model's robustness to spatial changes in waveform data, ultimately achieving accurate classification of land cover types.

[0069] In one alternative implementation, during the training step, a cross-entropy loss function is used to measure the difference between the predicted result and the true label. The adaptive moment estimation algorithm is used to optimize the model parameters, and a weight decay regularization term is set to prevent overfitting. An early stopping mechanism is used to monitor the performance of the validation set, and the training process is terminated when the performance no longer improves.

[0070] As an example, the training parameter settings include: cross-entropy loss function, Adam optimizer (learning rate = 0.0001, weight decay = 0.00001), 100 iterations (early stopping mechanism: termination if validation set accuracy does not improve after 5 iterations), and batch size = 64. During model training, the preprocessed normalized sample set is divided into training and validation sets in an 8:2 ratio. The model is trained using the training set, and the model performance is monitored in real time using the validation set. The final training iterations are determined based on the early stopping mechanism to obtain the trained DenseNet classification model.

[0071] The cross-entropy loss function is a mathematical tool used to measure the difference between the model's predicted probability distribution and the true label distribution. Specifically, it can be implemented using the multi-class cross-entropy formula, calculating the logarithmic loss between the predicted class probability and the true label to provide gradient direction for model parameter optimization. The adaptive moment estimation algorithm is an optimization algorithm combining momentum methods and adaptive learning rate adjustment. Specifically, it can be implemented using the Adam optimizer, dynamically adjusting the learning rate by calculating the first and second moment estimates of each parameter to balance the update magnitude of different parameters. The weight decay regularization term adds an L2 norm penalty term to the model parameters in the loss function. This can be implemented by setting a decay coefficient in the optimizer configuration, constraining model complexity to prevent overfitting to the training data. The early stopping mechanism continuously monitors validation set performance metrics during training. Specifically, it can use the validation set loss value or classification accuracy as monitoring metrics. Training is automatically terminated when the metrics fail to improve for several consecutive training epochs, preventing the model from overlearning on the training set.

[0072] Specifically, during model training, the cross-entropy loss function first compares the class probability distribution output by the model with the true labels, and calculates gradient information through the backpropagation algorithm. The adaptive moment estimation algorithm dynamically adjusts the learning rate based on the gradient history of each parameter, using a smaller learning rate to update parameters with frequent large gradients and a larger learning rate to update parameters with sparse gradients, effectively improving the learning efficiency of complex waveform features. The weight decay regularization term applies L2 norm constraints with each parameter update, suppressing excessive growth of network layer weights and maintaining the model's generalization ability. The early stopping mechanism monitors the performance of the independent validation set, promptly saving the optimal parameter state and terminating training when the model begins to show signs of overfitting, ensuring that the final model is in the best generalization state. These three techniques form a closed-loop optimization system: cross-entropy loss provides the core optimization objective for the classification task, adaptive moment estimation achieves efficient search of the parameter space, and weight decay and early stopping mechanisms provide dual protection against overfitting.

[0073] Compared to existing technologies, traditional model training methods often employ stochastic gradient descent with a fixed learning rate, lacking dynamic adjustment of parameter update magnitude and prone to getting trapped in local optima or oscillations. Conventional regularization methods mostly use single L2 regularization or dropout techniques, failing to incorporate dynamic monitoring mechanisms during training. Existing early stopping strategies are often based on empirically set fixed training epochs, lacking real-time evaluation of model generalization performance. This solution constructs a training framework with adaptive adjustment capabilities through the synergistic optimization of cross-entropy loss and adaptive moment estimation, combined with the combined effect of weight decay and early stopping mechanisms, overcoming the contradiction between optimization efficiency and generalization performance inherent in traditional methods.

[0074] Through the above technical solutions, this invention effectively solves the problem of inaccurate prediction bias quantification during model training, and accurately captures the differences in probability distributions across multiple categories through the cross-entropy loss function. The adaptive moment estimation algorithm improves the convergence speed and stability of the parameter optimization process, avoiding the manual parameter tuning defects of traditional optimization methods. The combined application of weight decay regularization and early stopping mechanism significantly suppresses the model's tendency to overfit to training data noise, ensuring the generalization performance of the classification model in different geographical regions.

[0075] In one alternative implementation, after the land cover classification model has been trained, the performance of the land cover classification model is evaluated using sample datasets from different geographical regions to verify the generalization ability and adaptability of the land cover classification model. Performance evaluation was conducted using a multi-dimensional metric system that included overall accuracy, Kappa coefficient, precision, recall, and F1 score.

[0076] The sample datasets from different geographical regions refer to independent validation datasets containing multiple climate zones, topographic features, and vegetation cover types. Specifically, this can be achieved using sample data from temperate monsoon regions, subtropical humid regions, and plateau cold desert regions with cross-latitudinal gradient distributions. By covering regional data with different environmental characteristics, geographical heterogeneity in practical applications can be simulated. The multi-dimensional indicator system refers to an evaluation set composed of classification accuracy, class consistency, prediction reliability, and comprehensive balance indicators. Specifically, it can be implemented by using overall precision to reflect the overall classification accuracy, the Kappa coefficient to eliminate random consistency interference, precision to measure the accuracy of prediction results, recall to assess class coverage, and the F1 score to comprehensively balance precision and recall. This multi-indicator synergy reveals the performance fluctuations of the model in different scenarios.

[0077] Among them, precision is the ratio of the number of correctly predicted samples of a certain type of land cover to the total number of predicted samples of that type, reflecting the accuracy of the prediction results; recall is the ratio of the number of correctly predicted samples of a certain type of land cover to the actual total number of samples of that type, reflecting the coverage of actual samples; the F1 score is the harmonic mean of precision and recall, comprehensively measuring the balance between the two indicators. Overall precision is the ratio of the total number of correctly classified samples across all categories to the total number of samples, reflecting the overall classification effect of the model; the Kappa coefficient assesses the reliability of classification precision by comparing the consistency between actual classification results and random classification results. Its calculation logic is: based on the total number of samples, the actual number of samples in each category, and the predicted number of samples, the consistency coefficient is obtained by eliminating the influence of random factors. The higher the value, the better the consistency between the classification results and the actual situation.

[0078] Specifically, after the model training phase is complete, new sample datasets are collected from independent geographical regions that did not participate in the training. These regions have different environmental characteristics from the training areas. The standardized new sample data is input into the trained model for classification prediction, and the prediction results are then compared with the human-interpreted true labels. Overall precision calculates the proportion of correctly classified samples out of the total samples, reflecting the model's overall classification ability; the Kappa coefficient, by eliminating the influence of random consistency, assesses the consistency between the model's classification results and the true labels; precision calculates the proportion of correctly predicted numbers out of the total number of predictions for each land cover category, identifying whether the model has overpredictive issues; recall calculates the proportion of correctly predicted numbers out of the number of true samples in that category, detecting whether the model has missed detection risks; the F1 score, by combining precision and recall using the harmonic mean, provides a balanced evaluation at the category level. This multi-indicator evaluation method can expose the problem of decreased classification accuracy caused by factors such as terrain undulation and seasonal changes in vegetation when the model is applied across regions.

[0079] Compared to existing technologies, traditional methods typically use test data from a single geographic region or rely on a single accuracy metric for model validation, failing to effectively identify the model's adaptability deficiencies in unknown environments. The same-region validation approach used in existing technologies easily leads to model overfitting to local features, while single-metric evaluation cannot comprehensively reflect multiple dimensions of classification performance. This solution, through the synergistic application of cross-regional test data and a multi-dimensional metric system, achieves a systematic verification of the model's generalization ability, overcoming the limitations of traditional methods in regional adaptability assessment.

[0080] Through the above technical solution, this invention can effectively detect the performance degradation of land cover classification models in cross-regional applications and pinpoint the sources of classification errors through multi-dimensional indicators. For example, when the model's F1 score in the high-altitude cold desert region is significantly lower than that in the training region, it can be identified that the model's adaptability to high-altitude terrain is insufficient; when the recall rate of the paddy field category decreases in the subtropical humid region, it indicates that the model is not sensitive enough to changes in paddy field characteristics caused by seasonal changes. This diagnostic capability provides a clear direction for model optimization, such as enhancing the model's environmental adaptability by increasing training samples in high-altitude areas or introducing time-series waveform features, thereby improving the reliability of the land cover classification model in practical applications.

[0081] In one optional embodiment, the method of the present invention is capable of distinguishing land cover types with similar spectral characteristics but different structural characteristics, including: Based on the differences in echo energy characteristics caused by humidity differences, dry land and paddy fields can be distinguished. Based on the differences in multi-peak / single-wave echo characteristics caused by differences in canopy structure, broad-leaved forests and coniferous forests can be distinguished.

[0082] The difference in echo energy characteristics refers to the difference in energy attenuation caused by laser pulses reflected from surfaces with different humidity levels. This can be achieved by calculating the ratio of the mean waveform intensity to the saturation threshold, which quantifies the degree to which soil moisture content absorbs laser energy. The difference in multi-peak / single-wave echo characteristics refers to the influence of the vertical structure of the vegetation canopy on the laser pulse penetration path. This can be achieved by extracting the number of waveform peaks and the distance between peaks. These parameters characterize the density and hierarchical structure of branches and leaves within the canopy.

[0083] Specifically, when a laser pulse irradiates the surface of a paddy field, the water layer and saturated soil absorb some of the laser energy, resulting in an overall decrease in the intensity of the reflected echo. In this case, the normalized mean waveform intensity is typically in the range of 0.3-0.5. Dryland, due to its dry soil, has a higher reflectivity, and its mean waveform intensity can reach the range of 0.6-0.8. For broadleaf forests, the multi-layered leaf structure causes multiple reflections during laser penetration, resulting in a waveform with 2-3 distinct peaks and a peak spacing of 3-5 sampling points. Coniferous forests, with their vertically pyramidal leaf distribution, have a single laser penetration path, resulting in a waveform with only a single main peak and a narrow peak width.

[0084] In some specific implementations, echo energy difference analysis can be performed using waveform integral area calculation, and the classification threshold can be set by comparing the distribution of integral values ​​between paddy field and dryland samples. Canopy structure analysis can be performed using the waveform derivative method to detect the number of peaks; when the number of zero-crossing points of the first derivative is greater than 2, it is determined to be a broadleaf forest characteristic.

[0085] Compared to existing technologies, traditional remote sensing classification methods rely on differences in spectral reflectance of multispectral data, making it difficult to distinguish between paddy fields and dry land, which have similar spectral characteristics. This scheme establishes classification indicators directly related to the structural characteristics of land features by analyzing the differences in physical properties contained in lidar waveform data, thus overcoming the limitation of spectral similarity on classification accuracy.

[0086] Through the above technical solution, this invention effectively solves the problem of misjudgment caused by spectral confusion in traditional remote sensing classification, and achieves accurate differentiation of land cover types such as paddy fields and dry land, broad-leaved forests and coniferous forests. This technical solution constructs a classification basis directly corresponding to the physical attributes of land cover by quantifying waveform energy attenuation characteristics and morphological features, significantly improving the accuracy of land cover identification under complex terrain conditions.

[0087] To verify the technical effectiveness of this invention, the classification confusion matrix of the DenseNet neural network in the original dataset is shown in Table 2: Table 2 ;

[0088] For the building category, the DenseNet model achieved an accuracy of 89.2%, correctly classifying 892 out of 1000 samples. In forest classification, the accuracy was 92.7%, correctly identifying 927 out of 1000 samples. For desert classification, the accuracy was 97.0%, correctly classifying 970 out of 1000 samples. In water body classification, the accuracy was 99.5%, correctly identifying 995 out of 1000 samples. The overall accuracy of the DenseNet model, calculated from the confusion matrix data, was 0.932, and the Kappa coefficient was 0.843. These metrics not only confirm the model's high accuracy but also reflect its high consistency and reliability in land cover classification tasks, indicating its potential for wide application in remote sensing and geographic information systems.

[0089] The classification results using the DenseNet neural network in the original dataset are evaluated as shown in Table 3.

[0090] Table 3 ;

[0091] In the original dataset, the DenseNet model demonstrated robust classification capabilities across four categories: buildings, woodland, bare land, and water. For buildings, it achieved a precision of 0.931, a recall of 0.892, and an F1 score of 0.911, showcasing an excellent balance between accuracy and error control. In woodland classification, it achieved a precision of 0.964 and a recall of 0.927 (F1 score: 0.945), highlighting its significant consistency in detecting vegetation patterns. Notably, bare land classification achieved the highest recall (0.970) and an F1 score of 0.947 (precision: 0.924), indicating minimal omission errors in this category. The model performed particularly well in water classification, with near-perfect recall (0.995), precision of 0.970, and an F1 score of 0.982, confirming its reliability in identifying hydrological features. Overall, these results highlight the model’s trade-off between precision and recall, as well as its adaptability to complex land cover spectral variations.

[0092] To verify the generalization ability of the model, a dense neural network (DenseNet) trained on the original dataset was used to classify the new study area dataset. The classification confusion matrix is ​​shown in Table 4. Table 4 ;

[0093] In the new study area sample set, the DenseNet model performed well in building category identification, accurately classifying 87.9% of the samples. However, 5.8% were misclassified as woodland, 4.9% as bare land, and 1.3% as water bodies. Within the woodland category, the model demonstrated high accuracy, correctly identifying 93.5% of the samples. Confusion with bare land (0.4%) and water bodies (0.1%) was minimal, although 5.9% were mislabeled as buildings. Bare land classification showed exceptional specificity, with a correct classification rate of 96.4%, while the woodland misclassification rate was near zero (0.0%). However, 3.2% of the samples were confused with water bodies. Water bodies achieved near-perfect discrimination with an accuracy of 98.9%, with only 1.0% misclassified as bare land. The model's overall accuracy was 94.3%, with a Kappa coefficient of 0.924, highlighting its strong consistency and adaptability to spectral overlap between built-up areas and natural landscapes. These results demonstrate that the DenseNet model not only achieves high accuracy but also high consistency in classification tasks on the new dataset.

[0094] Next, the classification results were evaluated using multidimensional evaluation indicators, and the evaluation results are shown in Table 5: Table 5 ;

[0095] The DenseNet model exhibits near-perfect differentiation for water bodies (precision: 97.0%, recall: 98.9%, F1 score: 98.2%) and minimal cross-class confusion (<3.2% misclassified as bare land). It demonstrates the highest recall (96.4%) and robust precision (92.4%, F1 score: 94.7%) on bare land, despite subtle spectral overlap with water bodies. Woodland classification shows high consistency (precision: 96.4%, recall: 92.7%, F1 score: 94.5%), with the main errors stemming from confusion with buildings (5.8%). While achieving high precision (93.1%) for built structures, its recall is only 89.2% (F1 score: 91.1%) due to spectral ambiguity with natural landscapes (6.0% misclassified as woodland, 4.9% misclassified as bare land). The model achieved an overall accuracy of 94.3% and a Kappa coefficient of 0.924, highlighting its strong consistency and adaptability to spectral overlap between built-up areas and natural landscapes.

[0096] For paddy fields and dry land, DenseNet's classification confusion matrix for the dataset is as follows: Figure 4 As shown; The DenseNet model achieved 1669 true positives in dryland classification and 330 false positives in paddy field classification, with only 158 false positives in the dryland category. The overall accuracy calculated from the confusion matrix was 0.878, and the kappa coefficient was 0.756, indicating that the model's classification results are highly consistent with the actual distribution of land features, effectively overcoming the interference of "different distributions of the same object," and demonstrating reliable performance in classification tasks for both dryland and paddy fields.

[0097] The classification evaluation of DenseNet is shown in Table 6: Table 6 ;

[0098] For dryland classification, the precision reached 0.914, meaning that approximately 91.4% of the samples predicted as dryland were actually dryland, demonstrating high reliability of the prediction results. The recall was 0.835, indicating that approximately 83.5% of the actual dryland samples could be effectively identified by the model, demonstrating good coverage. The harmonic mean of both yielded an F1-score of 0.873, comprehensively validating the balanced performance of dryland classification in terms of "accurate identification" and "comprehensive coverage." For paddy field classification, the precision was 0.848, with a prediction accuracy of 84.8% for actual paddy field samples. The recall reached 0.921, with a capture rate of over 92% for actual paddy fields. The F1-score was 0.883, highlighting the excellent balance between "accuracy" and "completeness" in paddy field classification.

[0099] Regarding broadleaf and coniferous forests, the classification confusion matrix of the DenseNet neural network is as follows: Figure 5 As shown; DenseNet performed well in coniferous forest classification, correctly classifying 767 samples with only 232 misclassifications. It also performed well in broadleaf forest classification, with only 384 misclassifications. The overall accuracy calculated from the confusion matrix was 0.692, and the kappa coefficient was 0.383. The model effectively overcomes the interference of "different species sharing the same spectrum" and possesses basic reliability in the detailed classification task between broadleaf and coniferous forests.

[0100] The classification evaluation of DenseNet is shown in Table 7 below; Table 7 ;

[0101] In the classification task of broadleaf forests and coniferous forests in this invention, the model performance is accurately measured based on the confusion matrix: For broadleaf forest classification, the precision reaches 0.666, meaning that approximately 66.6% of samples predicted as broadleaf forest are actually of the true category, ensuring prediction reliability; the recall is 0.768, meaning approximately 76.8% of the true broadleaf forest samples are effectively identified, demonstrating good coverage; the harmonic mean F1-score is 0.715, balancing precision and comprehensiveness. For coniferous forest classification, the precision is 0.726, with a true sample accuracy rate of 72.6%; the recall is 0.616, with a true coniferous forest capture rate exceeding 61.6%; the F1-score is 0.667, verifying the balance between precision and comprehensiveness. This invention's DenseNet model deeply mines the differences in lidar waveform features, enhancing feature reuse through dense connections, breaking through the traditional bottleneck of "different objects with the same spectrum" identification, and providing a reliable and detailed land cover classification scheme for scenarios such as ecological assessment and forestry monitoring.

[0102] This result demonstrates the reliability of the model in classifying ground features based on satellite-borne lidar waveform data for "same object, different spectra" and "different objects, same spectra".

[0103] Therefore, the land cover classification method proposed in this invention, based on satellite-borne lidar waveform data, deeply integrates the vertical structure information of land cover contained in satellite-borne lidar with the fine category labeling of ground reference data. This not only provides a new data source with both vertical dimensional features and high spatial resolution for land cover classification, but also addresses the pain point of insufficient accuracy in the classification of "same object with different spectra" (such as paddy fields and dry land) and "different objects with the same spectra" (such as bare land and water bodies) land cover by leveraging the dense feature reuse mechanism of the DenseNet model to accurately capture subtle waveform differences. Thus, a solution that balances classification accuracy and regional adaptability is constructed, providing a new technical path for improving the intelligence level of land cover classification.

[0104] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for land cover classification based on spaceborne lidar waveform data, characterized in that, The following application steps are included: S1. Acquire laser echo waveform data of the area to be classified using a spaceborne lidar; S2. Preprocess the laser echo waveform data; the preprocessing includes quality assessment and standardization. S3. Input the preprocessed laser echo waveform data into the pre-trained land cover classification model and output the land cover classification results; The land cover classification results include at least one land cover type from paddy fields and dry land, or at least one land cover type from broadleaf forests and coniferous forests; the land cover classification model is a densely connected neural network model.

2. The method according to claim 1, characterized in that, The land cover classification model is obtained through the following training steps: The laser echo waveform data and corresponding surface elevation data of multiple single land cover types areas were obtained by spaceborne lidar, and high-resolution remote sensing images of the corresponding areas were obtained as reference data. Based on the reference data, the original waveform data is labeled with land cover categories through visual interpretation to construct a sample dataset with category labels; The laser echo waveform data in the sample dataset are subjected to quality assessment and standardization to form a standardized sample set; A densely connected neural network model is constructed, and the model is trained using the normalized sample set. The model parameters are iteratively optimized until the model converges, resulting in a well-trained land cover classification model.

3. The method according to claim 1 or 2, characterized in that, The quality assessment specifically refers to: Calculate the signal-to-noise ratio and standard deviation of the laser echo waveform data, and remove laser echo waveform data with a signal-to-noise ratio lower than a preset threshold or an abnormal standard deviation. The formula for calculating the signal-to-noise ratio is: ； in, The maximum value of the laser echo waveform intensity. The mean of the background noise. The standard deviation of the background noise; The standardization process specifically includes: The corresponding saturation threshold is determined based on the laser gain parameters; The laser echo waveform data is normalized based on the saturation threshold. The calculation formula for the normalization process is as follows: ； in, These are the sampling points for the laser echo waveform data. The minimum value in the laser echo waveform data. This is the saturation threshold corresponding to the laser gain parameter. y These are the sampling points for the normalized laser echo waveform data.

4. The method according to claim 3, characterized in that, The saturation threshold The threshold is determined by a predefined gain-saturation threshold mapping table, which contains a continuous correspondence from threshold = 30 when gain value ≤ 8 to threshold = 239 when gain value ≥ 28.

5. The method according to claim 2, characterized in that, The densely connected neural network model adopts a densely connected structure, specifically including: Feature extraction module: Composed of convolutional layers, normalization layers, activation functions and pooling layers, used to extract multi-level features from the input laser echo waveform data; Feature fusion module: It consists of multiple dense blocks and transition layers alternately. Each dense block contains multiple convolutional units and uses dense connection to achieve feature reuse. Classification decision module: includes a global feature pooling layer and a fully connected classification layer, used to output the final land cover classification result.

6. The method according to claim 5, characterized in that, In the feature fusion module, each dense block consists of several convolutional units, and the output feature map of each convolutional unit is concatenated with the input feature maps of all subsequent layers in the channel dimension. Transition layers are placed between dense blocks, and feature map size and number of channels are controlled through convolution and pooling operations.

7. The method according to claim 2, characterized in that, In the training step, the cross-entropy loss function is used to measure the difference between the predicted result and the true label; The adaptive moment estimation algorithm is used to optimize the model parameters, and a weight decay regularization term is set to prevent overfitting. An early stopping mechanism is used to monitor the performance of the validation set, and the training process is terminated when the performance no longer improves.

8. The method according to claim 2, characterized in that, After the land cover classification model is trained, the performance of the land cover classification model is evaluated using sample datasets from different geographical regions to verify the generalization ability and adaptability of the land cover classification model. The performance evaluation was conducted using a multi-dimensional indicator system, including overall accuracy, Kappa coefficient, precision, recall, and F1 score.

9. The method according to claim 1, characterized in that, The method can distinguish between land cover types with similar spectral features but different structural features, including: Based on the differences in echo energy characteristics caused by humidity differences, dry land and paddy fields can be distinguished. Based on the differences in multi-peak / single-wave echo characteristics caused by differences in canopy structure, broad-leaved forests and coniferous forests can be distinguished.

Images