A ground feature classification method based on spaceborne lidar waveform data

Abstract

The application discloses a ground object classification method based on spaceborne laser radar waveform data, relates to the technical field of remote sensing and artificial intelligence, and comprises the following application steps: S1, acquiring laser echo waveform data of a region to be classified through a spaceborne laser radar; S2, pre-processing the laser echo waveform data; the pre-processing comprises quality evaluation and standardization processing; S3, inputting the pre-processed laser echo waveform data into a pre-trained ground object classification model, and outputting a ground object classification result; the ground object classification result comprises one or more ground object types in paddy fields and dry lands, one or more ground object types in broad-leaved forests and coniferous forests; and the ground object classification model is a densely connected neural network model. The application has the advantages of improving ground object classification precision and effectively distinguishing ground object types with similar spectral characteristics but different structural characteristics.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of remote sensing technology and artificial intelligence, and particularly relates to a ground object classification method based on spaceborne LiDAR waveform data. BACKGROUND

[0002] Traditional remote sensing image classification methods mainly rely on spectral information for ground object recognition. This method has obvious limitations in dealing with complex terrain and similar ground objects. Spectral information is difficult to effectively distinguish the "same object with different spectra" and "different objects with the same spectrum" phenomena. For example, paddy fields and dry lands may be similar in spectral characteristics, but the actual ground object types are different; while different vegetation types may exhibit similar spectral characteristics in certain bands. This limitation makes it difficult to meet the actual application requirements of classification accuracy, especially in scenarios that require fine classification.

[0003] Full waveform LiDAR technology provides a new dimension for ground object classification by obtaining vertical structure information of ground objects. This technology can record the entire process of laser pulse interaction with ground objects, including rich three-dimensional structural feature information. However, how to effectively extract and utilize these waveform features for accurate classification still faces challenges. Traditional waveform processing methods rely on manual feature extraction and simple classification algorithms, making it difficult to fully exploit the deep feature information contained in waveform data.

[0004] Deep learning technology, especially DenseNet, has shown strong feature extraction capabilities in the field of image processing. This network achieves feature reuse through dense connection mechanism, which can effectively extract multi-level features. However, when applying this technology to spaceborne LiDAR waveform data processing, a series of technical problems need to be solved, such as waveform data preprocessing, feature extraction, and model training. Existing methods still have deficiencies in waveform data quality evaluation, standardization processing, and model generalization ability, making it difficult to meet the actual needs of large-scale, multi-type ground object classification.

[0005] In view of the above problems, the existing technology needs to be improved. SUMMARY

[0006] Therefore, the present application provides a ground object classification method based on spaceborne LiDAR waveform data, which has the advantages of improving ground object classification accuracy and effectively distinguishing ground object types with similar spectral characteristics but different structural characteristics.

[0007] The present application provides a ground object classification method based on spaceborne LiDAR waveform data, comprising the following application steps:

[0008] S1, obtaining laser echo waveform data of the region to be classified by spaceborne LiDAR;

[0009] S2, preprocessing the laser echo waveform data; the preprocessing includes quality evaluation and standardization processing;

[0010] S3, inputting the preprocessed laser echo waveform data into a pre-trained ground feature classification model to output a ground feature classification result; the ground feature classification result at least includes one or more ground feature types of paddy field, dry land, broadleaf forest and coniferous forest; the ground feature classification model is a densely connected neural network model.

[0011] In an optional embodiment, the ground feature classification model is obtained through the following training steps:

[0012] Obtain laser echo waveform data and corresponding ground elevation data of a plurality of single ground feature type regions by a spaceborne laser radar, and obtain high-resolution remote sensing images of the corresponding regions as reference data respectively;

[0013] Based on the reference data, visually interpret the original waveform data to label the ground feature categories, and construct a sample data set with category labels;

[0014] Perform quality evaluation and standardization processing on the laser echo waveform data in the sample data set to form a standardized sample set;

[0015] Construct a densely connected neural network model, train the densely connected neural network model using the standardized sample set, and optimize the model parameters through iteration until the model converges to obtain a trained ground feature classification model.

[0016] In an optional embodiment, the quality evaluation is specifically:

[0017] Calculate the signal-to-noise ratio and standard deviation of the laser echo waveform data, and remove the laser echo waveform data with a signal-to-noise ratio lower than a preset threshold or an abnormal standard deviation;

[0018] The formula for calculating the signal-to-noise ratio is:

[0019]

[0020] Wherein, is the maximum value of the laser echo waveform intensity, is the mean value of the background noise, is the standard deviation of the background noise;

[0021] The standardization processing is specifically:

[0022] Determine the corresponding saturation threshold according to the laser gain parameter;

[0023] According to the saturation threshold, the laser echo waveform data is normalized;

[0024] The calculation formula of the normalization processing is:

[0025]

[0026] wherein, is a sampling point of the laser echo waveform data, is a minimum value in the laser echo waveform data, is a saturation threshold corresponding to the laser gain parameter, is a sampling point of the normalized laser echo waveform data.

[0027] In an optional embodiment, the saturation threshold is determined by a predefined gain-saturation threshold mapping table, which contains a continuous corresponding relationship from threshold = 30 when the gain value is less than or equal to 8, to threshold = 239 when the gain value is greater than or equal to 28.

[0028] In an optional embodiment, the deep neural network classification model adopts a dense connection structure, specifically comprising:

[0029] a feature extraction module composed of a convolution layer, a normalization layer, an activation function and a pooling layer, for extracting multi-level features from the input laser echo waveform data;

[0030] a feature fusion module composed of multiple dense blocks and transition layers, each dense block containing multiple convolution units, and the feature reuse is realized by a dense connection method;

[0031] a classification decision module including a global feature pooling layer and a fully connected classification layer, for outputting the final ground object classification result.

[0032] In an optional embodiment, in the feature fusion module, each dense block is composed of several convolution units, and the output feature map of each convolution unit is spliced with the input feature map of all subsequent layers in the channel dimension;

[0033] a transition layer is arranged between the dense blocks, and the feature map size and the number of channels are controlled by convolution and pooling operations.

[0034] In an optional embodiment, in the training step, a cross-entropy loss function is used to measure the difference between the prediction result and the true label;

[0035] an adaptive moment estimation algorithm is used to optimize the model parameters, and a weight decay regularization term is set to prevent overfitting;

[0036] 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.

[0037] In an optional implementation, after the training of the ground object classification model is completed, a sample data set of different geographic regions is used to evaluate the performance of the ground object classification model to verify the generalization ability and adaptability of the ground object classification model.

[0038] The performance evaluation is completed by a multi-dimensional index system of overall accuracy, Kappa coefficient, precision, recall, and F1 score.

[0039] In an optional implementation, the method can distinguish ground object types with similar spectral characteristics but different structural characteristics, including:

[0040] Based on the echo energy feature difference caused by humidity difference, dry land and paddy field are distinguished.

[0041] Based on the multi-peak / single-wave echo feature difference caused by crown structure difference, broadleaf forest and coniferous forest are distinguished.

[0042] As can be seen from the above, the ground object classification method based on spaceborne laser radar waveform data provided by the application acquires waveform data by spaceborne laser radar and extracts features and classifies by using a densely connected neural network, solves the technical problem that the traditional method cannot effectively distinguish similar spectral ground object types, and has the advantages of improving classification accuracy and enhancing model generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the description of the specific embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0044] Figure 1 is a flowchart of a ground object classification method based on spaceborne laser radar waveform data according to an embodiment of the present application;

[0045] Figure 2 is a flowchart of the training and application stage of a ground object classification method based on spaceborne laser radar waveform data according to an embodiment of the present application;

[0046] Figure 3 is a structural diagram of a densely connected neural network model DenseNet according to an embodiment of the present application;

[0047] Figure 4 is a classification confusion matrix diagram of the densely connected neural network model DenseNet for paddy field and dry land data sets according to an embodiment of the present application;

[0048] Figure 5 Fig. 4 is a diagram of a confusion matrix of a classification of a broadleaf forest and a coniferous forest dataset according to a dense connected neural network model DenseNet of an embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0050] In the prior art, remote sensing image classification mainly relies on spectral information for ground object identification, but there are obvious limitations in complex terrain areas and ground object classification with similar spectral characteristics. For example, due to the difference in water content, the spectral reflection characteristics of paddy fields and dry lands are similar, and the spectral response of broadleaf forests and coniferous forests in specific bands is highly similar, so it is difficult for traditional methods to accurately distinguish them. Although the full waveform lidar technology can obtain the vertical structure information of ground objects, the existing classification methods fail to effectively combine the waveform characteristics and deep learning models, resulting in insufficient extraction of subtle difference features hidden in the waveform data, which restricts the improvement of classification accuracy.

[0051] In order to solve the above problems, the inventors found that the traditional classification method cannot effectively utilize the vertical dimension information contained in the lidar waveform data when dealing with ground objects with similar spectra but different structural characteristics. By analyzing the waveform feature differences of different ground object types, such as the difference between the single-peak waveform of paddy fields due to surface water coverage and the multi-peak waveform of dry land due to rough ground surface, and the echo feature difference between the dense canopy of broadleaf forests and the layered structure of coniferous forests, a technical route combining waveform data preprocessing and deep feature learning is proposed. In view of the problem that the waveform data is affected by environmental noise and device parameters, a dynamic standardization processing method is designed; in view of the defect of insufficient feature reuse of traditional convolutional neural network, a dense connection structure is introduced to strengthen the hierarchical extraction of waveform features.

[0052] Therefore, as shown in Figure 1 The present application proposes a ground object classification method based on spaceborne lidar waveform data, which comprises the following application steps:

[0053] Step S1, acquiring laser echo waveform data of a region to be classified by a spaceborne lidar.

[0054] Step S2, preprocessing the laser echo waveform data; the preprocessing includes quality evaluation and standardization processing.

[0055] Step S3, inputting the preprocessed laser echo waveform data into a pre-trained feature classification model to output a feature classification result; the feature classification result at least includes one or more feature types of paddy field, dry land, broadleaf forest and coniferous forest; and the feature classification model is a dense neural network model.

[0056] The spaceborne laser radar refers to an optical remote sensing device carried on a satellite platform, which obtains three-dimensional spatial information by emitting laser pulses and receiving ground surface reflection signals, and can be implemented by using a photon counting type or full waveform type laser radar system to obtain original waveform data containing vertical structure features of features.

[0057] Preferably, GLA01 (laser echo waveform data) generated by the GLAS carried by the ICESat satellite can be used as a typical data source, and the system belongs to a full waveform type laser radar, which can record the change process of echo energy with time of laser pulse and ground surface interaction, and is particularly suitable for large-scale and high-precision extraction of vegetation vertical profile, ice cap topography and sea surface height information. The data structure and precision characteristics can provide reliable support for related remote sensing inversion algorithm development and verification.

[0058] Quality assessment refers to screening effective data by calculating signal-to-noise ratio and standard deviation indicators, and can be implemented by using a sliding window to calculate background noise statistics, which is used to eliminate the influence of atmospheric scattering and device noise on data quality. Standardization processing refers to dynamically adjusting the data normalization range according to the working parameters of the laser, which can be implemented by using a gain-saturation threshold mapping table to determine the normalization reference value, which is used to eliminate the problem of inconsistent data scales caused by device parameter differences under different observation conditions. The dense neural network model (DenseNet) is a deep neural network classification model based on a dense connection structure, and the dense connection structure is a topological connection method for cross-layer feature reuse in neural networks, which can be implemented by using a dense block (DenseBlock) to cascade the output feature maps of each convolution layer, which is used to enhance the ability of the model to capture local morphological features of the waveform curve.

[0059] Specifically, after the satellite platform carries the laser radar sensor to obtain the original echo waveform data of the target area, first, the original data is quality evaluated, for example, the signal-to-noise ratio of each waveform sequence is calculated, and when the signal-to-noise ratio is lower than the set threshold, it is determined as invalid data and is rejected. Subsequently, according to the laser gain parameter, the saturation threshold is determined by querying the pre-defined mapping table, and the effective waveform data is normalized, and the waveform data obtained under different observation conditions is unified to the same numerical range. The pre-processed waveform data is input into a dense connection neural network model (DenseNet), the model extracts the local fluctuation features of the waveform curve through the convolution layer, realizes the reuse and fusion of multi-level features through the dense connection structure, and finally outputs the ground object type recognition result through the classification layer. For example, when processing the classification of paddy field and dry land, the model can recognize the difference in surface roughness through the waveform peak value number and decay slope characteristics; when distinguishing broadleaf forest and coniferous forest, the model can recognize the difference in canopy density according to the waveform backscattering intensity distribution.

[0060] Compared with the prior art, the traditional method only uses spectral information for ground object classification, and it is difficult to distinguish ground object types with similar reflection characteristics but different structural characteristics. The present application can simultaneously use the spectral reflection characteristics and vertical structural characteristics of the ground object for classification decision by introducing full waveform lidar data and combining a dense connection neural network model (DenseNet). For example, in the classification of paddy field and dry land with the same spectral characteristics, the traditional method is prone to confusion, while the present application can accurately identify the difference in surface humidity by analyzing the peak distribution characteristics of the waveform curve; in the classification of broadleaf forest and coniferous forest, the traditional method relies on leaf spectral characteristics and is easily affected by seasonal changes, and the present application can maintain the stability of classification by analyzing the canopy structure characteristics.

[0061] Through the above technical solution, the present application can effectively distinguish ground object types with similar spectral characteristics but significant vertical structure differences, for example, distinguishing paddy field and dry land according to the waveform energy decay characteristics, and distinguishing broadleaf forest and coniferous forest according to the echo peak value distribution characteristics. This method solves the problem of insufficient accuracy of traditional classification methods in complex terrain areas and similar ground object recognition, and improves the accuracy and environmental adaptability of ground object classification by fusing waveform data preprocessing and deep feature learning technology.

[0062] In an optional embodiment, as shown in Figure 2 The ground object classification model is obtained through the following training steps:

[0063] Obtain laser echo waveform data and corresponding ground elevation data of a plurality of single ground object type regions by a spaceborne lidar, and obtain high-resolution remote sensing images of the corresponding regions as reference data;

[0064] Based on the reference data, the original waveform data is labeled by visual interpretation of ground object categories, and a sample data set with category labels is constructed.

[0065] The laser echo waveform data in the sample data set is quality evaluated and standardized to form a standardized sample set.

[0066] A dense connection neural network model (DenseNet) is constructed, the dense connection neural network model is trained using the standardized sample set, the model parameters are iteratively optimized until the model converges, and a trained ground object classification model is obtained.

[0067] Among them, the single ground object type area refers to a geographical area with a single land cover type and continuous spatial distribution, which can be verified by field investigation or high-resolution image interpretation to ensure the category purity of the sample data. Quality evaluation refers to quantitative analysis of the signal-to-noise ratio and standard deviation of the laser echo waveform data, which can be implemented by using a pre-set threshold filtering mechanism 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, which can be achieved by establishing a gain-saturation threshold mapping table to eliminate the influence of device parameter differences on waveform amplitude. The deep neural network classification model refers to a deep learning architecture with a dense connection structure, which can be implemented by constructing a feature extraction module, a feature fusion module, and a classification decision module to achieve multi-level feature learning.

[0068] In an alternative embodiment, the spaceborne lidar data includes GLA01 (laser echo waveform data) and GLA14 (ground elevation data) of ICESat / GLAS; the ground reference data includes high-resolution satellite remote sensing data (such as Google Earth historical images).

[0069] As an example, the spaceborne lidar data can be provided by the National Snow and Ice Data Center. First, select an area with a single ground object type as the research area and determine the latitude and longitude range of the area. Then, using the latitude and longitude range, filter out the GLA01 and GLA14 data of the research area from the National Snow and Ice Data Center for further analysis and processing. The ground reference image data mainly uses the historical images of Google Earth, and selects the images with a small time difference from the spaceborne lidar data as the ground reference data. Next, using the method of visual interpretation, determine the ground object category (such as buildings, forests, bare land, water bodies, etc.) to which the spaceborne lidar waveform data belongs. These interpretation results will be used as the training sample set and verification sample set of the dense connection neural network model, providing support for subsequent model training and verification.

[0070] It should be noted that for flat and smooth ground surfaces (such as bare ground and water bodies), the echo signal usually presents an ideal Gaussian waveform with good symmetry; while smooth ground surfaces with a certain slope, due to the difference between the incident angle and the reflection angle, the echo signal will appear obvious broadening phenomenon, and with the increase of the ground slope, the broadening effect is more and more significant. For complex ground surfaces, such as buildings and vegetation, the echo signal is usually composed of multiple peaks. When the laser beam interacts with the building, the echo signal is often more significant and has a larger amplitude, but the broadening effect is not obvious; while in the vegetation area, branches and leaves will cause the echo signal to be difficult to distinguish, and due to the low reflectivity, the echo amplitude is usually small. According to these characteristics, the sample data set can be optimized to improve the training effect and classification accuracy of the deep learning model.

[0071] Specifically, in the sample construction phase, a multi-source data support system is formed by synchronously collecting laser echo waveform, ground elevation and high-resolution image data. The elevation data provides a terrain correction reference for waveform interpretation, and the high-resolution image provides a 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 link, through the double screening mechanism of signal-to-noise ratio calculation and standard deviation analysis, abnormal samples with serious noise interference are removed from the data source, and a dynamic normalization method based on gain parameter is used to eliminate the influence of device parameter difference on waveform amplitude, forming a standardized input data format. The dense connection neural network model realizes feature reuse through dense connection structure, captures local waveform features through convolution operation in the feature extraction stage, integrates multi-scale feature information through cross-layer connection between dense blocks in the feature fusion stage, and finally completes the classification decision through global pooling and fully connected layer. In the training process, the regularization technique and early stopping mechanism are used to balance the model complexity and generalization ability, and to ensure the stable convergence of the model parameter optimization process.

[0072] Compared with the prior art, the traditional method usually relies on single waveform data and lacks strict quality control process, and is easily affected by noise interference, resulting in decreased classification accuracy. The shallow classification model used in the prior art cannot effectively extract deep features of waveform data, and has the defect of insufficient feature expression ability when dealing with complex ground objects. The present scheme improves the reliability of samples through a multi-source data collaborative labeling mechanism, guarantees the data quality through a strict quality evaluation process, and enhances the feature expression ability through a deep dense connection network, while maintaining the model generalization ability and significantly improving the classification accuracy.

[0073] By the technical scheme, the problem that the classification performance is limited due to the waveform data noise interference and insufficient feature expression is effectively solved. The multi-source data collaborative labeling mechanism ensures the accuracy of sample labels, the quality evaluation and standardization processing improves the signal-to-noise ratio and consistency of input data, and the deep dense connection network enhances the capturing ability of the model to subtle structural features through the feature reuse mechanism. The technical scheme can stably process the waveform data obtained by different gain parameters, and has stronger adaptability and robustness in complex terrain areas and spectral similar ground object classification scenes.

[0074] In an optional implementation, the quality evaluation is specifically:

[0075] The signal-to-noise ratio (SNR) and the standard deviation of the laser echo waveform data are calculated, and the laser echo waveform data with a signal-to-noise ratio lower than a preset threshold or an abnormal standard deviation is removed;

[0076] The calculation formula of the signal-to-noise ratio is:

[0077]

[0078] wherein, is the maximum value of the laser echo waveform intensity, is the mean value of the background noise, is the standard deviation of the background noise;

[0079] The standardization processing is specifically:

[0080] According to the laser gain parameter, a corresponding saturation threshold is determined;

[0081] According to the saturation threshold, the laser echo waveform data is normalized;

[0082] The calculation formula of the normalization processing is:

[0083]

[0084] wherein, is the sampling point of the laser echo waveform data, is the minimum value in the laser echo waveform data, is the saturation threshold corresponding to the laser gain parameter, is the sampling point of the normalized laser echo waveform data.

[0085] The signal-to-noise ratio refers to the ratio of the signal peak intensity to the background noise statistics, and is specifically calculated by subtracting the background noise mean from the maximum intensity value and then dividing by the noise standard deviation, for quantifying the effective information content of the waveform data. The standard deviation refers to a statistical indicator of the degree of dispersion of the waveform data, and is specifically obtained by calculating the square root of the variance of the sampling point intensity value, for identifying morphologically abnormal waveform samples. The saturation threshold refers to the upper limit value of the signal intensity corresponding to the laser gain parameter, and is specifically dynamically determined by using a predefined gain-saturation threshold mapping table, which can adapt to the data characteristics under different gain conditions. The normalization process refers to the process of linearly converting the original waveform intensity value to a preset interval, which is specifically implemented by subtracting the minimum value from the sampling point intensity and dividing by the difference between the saturation threshold and the minimum value, which can eliminate the dimensional differences caused by different gain settings.

[0086] Specifically, the quality evaluation stage first performs background noise statistics on the original waveform data, calculates the mean and standard deviation thereof as noise reference. Then, the maximum value of the waveform intensity is extracted, and the signal-to-noise ratio index is calculated in combination with the noise statistics. When the signal-to-noise ratio is lower than the preset threshold, it indicates that the waveform is severely disturbed by noise and is rejected. At the same time, the overall standard deviation of the waveform data is calculated, and when the standard deviation exceeds the normal range, it is determined as an abnormal waveform for filtering.

[0087] By estimating the background noise of the reflected echo and calculating the signal-to-noise ratio (SNR) and standard deviation of the laser echo waveform based thereon, invalid waveform data in the sample data set disturbed by cloud cover, instrument failure, etc. is screened out. Generally, a smaller standard deviation and a larger signal-to-noise ratio mean that there is less noise component in the signal, the waveform quality is higher, and the features are more obvious. As an example, the signal-to-noise ratio 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 of the laser echo waveform data.

[0088] In order to improve the data retrieval accuracy and classification reliability, the laser echo waveform data needs to be normalized (i.e., standardized). In the standardization stage, according to the real-time gain parameter of the laser, the effective range of the current data is determined by the pre-set gain-saturation threshold correspondence. Based on the range, the retained waveform data is linearly transformed, and the original intensity value is mapped to a unified numerical interval, ensuring that the waveform data collected under different gain conditions is comparable.

[0089] Compared with the prior art, the traditional method usually only adopts a single signal-to-noise ratio index for data screening, and cannot effectively identify waveform morphology abnormal samples. The scheme can detect two types of quality problems of noise interference and waveform distortion through the double evaluation mechanism of signal-to-noise ratio and standard deviation. In the aspect of standardization processing, the prior art usually adopts a fixed threshold for normalization, and cannot adapt to the range difference caused by gain parameter change. The scheme realizes adaptive standardization of waveform data under different acquisition parameters through the dynamic mapping relationship of gain-saturation threshold, and solves the problem of inconsistent data scale caused by ignoring the parameter difference of the device in the traditional method.

[0090] Through the above technical scheme, the influence of noise interference and gain difference in the laser echo waveform data on the classification accuracy is effectively solved. The quality evaluation link can screen the effective data with up-to-standard signal-to-noise ratio and normal waveform morphology, and the standardization processing link eliminates the range difference caused by different gain settings, so that the subsequent classification model can receive input data with uniform scale and reliable quality. The technical scheme guarantees the effectiveness of the data while improving the consistency of the data under different acquisition conditions, and provides a stable and reliable data basis for the deep learning model.

[0091] In an optional implementation, the saturation threshold value is determined through a predefined gain-saturation threshold mapping table, and the mapping table includes a continuous corresponding relationship from threshold value = 30 when gain value <= 8 to threshold value = 239 when gain value >= 28.

[0092] The gain-saturation threshold mapping table refers to a quantitative corresponding relationship table formed by pre-calibrating the best saturation threshold corresponding to different gain parameters through experiments, and can be realized by using a laboratory calibration method combined with a linear interpolation method. The gain parameter and the physical saturation threshold value are matched through a mathematical relationship model.

[0093] The continuous corresponding relationship refers to a continuous linear or nonlinear function relationship formed between the gain parameter and the saturation threshold value, and can be realized by using a piecewise linear function or a polynomial fitting method. The mathematical modeling ensures smooth transition of the threshold value adjustment process when the gain parameter changes.

[0094] Specifically, the saturation points of the laser photoelectric conversion under different gain values are calibrated in the laboratory, the physical relationship between the gain parameter and the corresponding saturation threshold value is recorded, and a gain-saturation threshold data pair covering the typical working range is formed. In the data processing stage, the corresponding saturation threshold value is automatically matched through the mapping table according to the current laser gain parameter value, and then the normalization operation is performed. When the gain parameter is in the interval of 8 to 28, the linear interpolation method is used to calculate the intermediate threshold value, and the continuous adjustment of the threshold value in the dynamic change scenario of the gain parameter is realized.

[0095] As an example, when normalizing laser echo waveform data, the energy receiving threshold difference of the GLAS satellite laser altimetry system needs to be considered. The GLAS satellite has a dynamic threshold, and different saturation thresholds are set for different gain values, as shown in Table 1.

[0096] Table 1

[0097]

[0098] Compared with the prior art, the conventional method usually adopts a fixed threshold or a threshold set by artificial experience, and cannot adapt to the working scene of dynamic adjustment of the gain parameter of the laser, resulting in accumulation of data standardization errors. The scheme combines experimental calibration and mathematical modeling to establish an accurate correspondence between the gain parameter and the physical saturation threshold, eliminate the subjective deviation of the threshold set by the artificial, and at the same time, guarantee the consistency of data standardization under different gain conditions.

[0099] Through the above technical scheme, the present application effectively solves the problem of saturation threshold determination error caused by the dynamic change of the gain parameter of the satellite-borne laser radar, realizes automatic matching and smooth transition of the threshold through the predefined mapping table and continuous correspondence, and ensures the scientificity and consistency of the waveform data normalization processing under different working conditions, thereby providing quality guarantee for the standardized input data of the classification model.

[0100] In an alternative embodiment, as shown in Figure 3 The deep neural network classification model adopts a dense connection structure, and specifically includes a feature extraction module, a feature fusion module and a classification decision module.

[0101] The feature extraction module (i.e., the initial convolution layer) is composed of a convolution layer, a normalization layer, an activation function and a pooling layer, and is used to extract multi-level features from the input laser echo waveform data.

[0102] Specifically, the initial convolution layer includes a 7x1 convolution layer (with 64 output channels), a BatchNorm1d, a ReLU activation function, and a 3x1 max pooling layer (with a step of 2), which is used to preliminarily extract features and reduce the sequence length.

[0103] The feature fusion module (i.e., the dense connection module) is alternately composed of multiple dense blocks (DenseBlock) and transition layers (Transition), each dense block includes multiple convolution units (i.e., dense layers), and the feature reuse is realized in a dense connection manner;

[0104] Specifically, in the feature fusion module, each dense block is composed of several convolution units (i.e., DenseLayer), and the output feature map of each convolution unit is spliced with the input feature map of all subsequent layers in the channel dimension.

[0105] The transition layer is arranged between the dense blocks, and the feature map size and the number of channels are controlled through convolution and pooling operations.

[0106] Further, the dense block (DenseBlock) and the transition layer (Transition) are alternately arranged, and each dense block is connected with a transition layer after it (except for the last one).

[0107] The dense block (DenseBlock) is composed of multiple dense layers (DenseLayer), each of which extracts features through 1x1 convolution (intermediate channel number 128) and 3x1 convolution (output channel number 32, i.e., growth rate), and splices (dense connection feature) with the input feature in the channel dimension. The number of layers of the four dense blocks is 6, 12, 24, and 16, respectively.

[0108] The transition layer (Transition) is composed of batch normalization, ReLU activation function, 1x1 convolution (half of the number of channels), and 2x1 average pooling, which is used to compress the number of feature channels and further reduce the sequence length.

[0109] The classification decision module (i.e., classification layer) includes a global feature pooling layer and a fully connected classification layer, which is used to output the final ground object classification result.

[0110] Specifically, the classification layer is subjected to adaptive average pooling (compressing the sequence length to 1) to obtain global features.

[0111] The classification result is output through a two-layer fully connected network (half of the number of channels in the intermediate layer is the number of feature channels, containing ReL activation and 50% Dropout).

[0112] The dense connection structure refers to the output feature map of each convolution unit being spliced with the input feature map of all subsequent layers in the channel dimension, which can be implemented by using a channel splicing operation, thereby preserving the original feature information of each layer output and promoting cross-layer reuse of different level features.

[0113] The feature extraction module is composed of convolution layers, normalization layers, activation functions, and pooling layers, which can extract local waveform features by using one-dimensional convolution layers with a convolution kernel size of 3, stabilize data distribution by combining batch normalization layers, introduce non-linear transformation capabilities through ReLU activation functions, and compress feature dimensions by using maximum pooling layers to form a base for multi-level feature expression.

[0114] The feature fusion module is composed of multiple dense blocks and transition layers, each dense block contains multiple convolution units, and a dense block structure containing four convolution units can be used, feature reuse is achieved through dense connection, and the transition layer controls the feature map size by 1*1 convolution compression of the channel number and combination of the average pooling operation with a step of 2.

[0115] Specifically, in the laser echo waveform data processing process, the feature extraction module captures the local shape features of the waveform through the convolution layer, the normalization layer eliminates the data distribution deviation, the activation function enhances the nonlinear expression ability of the model, and the pooling layer selects the key features and reduces the dimension. The feature fusion module realizes the cross-layer fusion of shallow detail features and deep semantic features through the dense connection of multiple convolution units in the dense block, and the transition layer balances the calculation complexity and feature preservation requirements through channel compression and spatial down-sampling. The classification decision module uses the global pooling layer to eliminate the spatial position sensitivity, and the fully connected layer integrates all feature channel information to complete the classification decision.

[0116] Compared with the prior art, the traditional convolutional neural network only relies on the layer-by-layer transmission mechanism in feature fusion, which easily causes the loss of shallow detail features. The dense connection structure enables the network to utilize low-level waveform details and high-level semantic features at the same time through cross-layer feature reuse, thereby enhancing the ability to capture subtle waveform differences. The introduction of the transition layer avoids the problem of rising model complexity caused by the explosive growth of feature map channels while preserving effective information.

[0117] Through the above technical solutions, the present application solves the problem of insufficient multi-level feature extraction of laser echo waveform data, efficiently reuses waveform detail features through the dense connection mechanism, and improves the ability of the model to distinguish between paddy fields and dry land humidity, and between broadleaf forest and coniferous forest canopy structure. The transition layer structure reduces the consumption of computing resources while maintaining classification accuracy, the global feature pooling layer enhances the robustness of the model to spatial position changes of the waveform data, and finally realizes accurate classification of ground object types.

[0118] In an optional embodiment, in the training step, a cross-entropy loss function is used to measure the difference between the predicted results and the true labels.

[0119] An adaptive moment estimation algorithm is used to optimize the model parameters, and a weight decay regularization term is set to prevent overfitting.

[0120] 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.

[0121] As an example, the training parameter settings include: using a cross-entropy loss function, an Adam optimizer (learning rate = 0.0001, weight decay = 0.00001), iterating 100 times (early stopping mechanism: terminate if the validation set accuracy does not improve for 5 rounds), and batch size = 64. During the model training process, the pre-processed normalized sample set is divided into a training set and a validation set in a ratio of 8:2, the training set is used to train the model, the model performance is monitored in real time through the validation set, the final training round is determined according to the early stopping mechanism, and the trained DenseNet classification model is obtained.

[0122] The cross-entropy loss function refers to a mathematical tool for measuring the difference between the model prediction probability distribution and the true label distribution, which can be implemented by using a multi-class cross-entropy formula, and the log loss value of the predicted class probability and the true label is calculated to provide the gradient direction for model parameter optimization. The adaptive moment estimation algorithm refers to an optimization algorithm that combines the momentum method and adaptive learning rate adjustment, which can be implemented by using an Adam optimizer to dynamically adjust the learning rate by calculating the first and second moment estimates of each parameter, and balance the update amplitude of different parameters. The weight decay regularization term refers to adding an L2 norm penalty term of the model parameters in the loss function, which can be implemented by setting the decay coefficient in the optimizer configuration to constrain the model complexity to prevent overfitting to the training data. The early stopping mechanism refers to continuously monitoring the performance indicators of the validation set during the training process, which can use the validation set loss value or classification accuracy as the monitoring indicator, and automatically terminate the training when the indicator does not improve for multiple training rounds, to avoid overfitting to the training set.

[0123] Specifically, during the model training process, the cross-entropy loss function first compares the class probability distribution output by the model with the true label, and calculates the gradient information through the backpropagation algorithm. The adaptive moment estimation algorithm dynamically adjusts the learning rate according to the gradient history of each parameter, and updates the parameters with frequent large gradients with a smaller learning rate, and updates the parameters with sparse gradients with a larger learning rate, effectively improving the learning efficiency of complex waveform features. The weight decay regularization term applies L2 norm constraint at each parameter update, suppresses the excessive growth of network layer weight values, and maintains the generalization ability of the model. The early stopping mechanism monitors the performance of the independent validation set, saves the optimal parameter state and terminates the training when the model starts to show overfitting trend, and ensures that the final model is in the best generalization state. These three techniques form a closed-loop optimization system, the cross-entropy loss provides the core optimization target for the classification task, the adaptive moment estimation realizes efficient search in the parameter space, and the weight decay and early stopping mechanism form a double protection against overfitting.

[0124] Compared with the prior art, the conventional model training method adopts a random gradient descent method with a fixed learning rate, lacks dynamic adjustment of the parameter update amplitude, and is prone to fall into local optimum or produce oscillation. The conventional regularization method mainly adopts single L2 regularization or dropout technology, and fails to combine a dynamic monitoring mechanism of the training process. The existing early stopping strategy is mainly based on the experience setting of a fixed training round, and lacks real-time evaluation of the model generalization performance. The scheme cooperates the adaptive moment estimation and the cross-entropy loss, and combines the joint action of the weight decay and the early stopping mechanism to construct a training framework with adaptive adjustment capability, and overcomes the contradiction between the optimization efficiency and the generalization performance of the conventional method.

[0125] Through the above technical scheme, the application effectively solves the problem of inaccurate quantification of prediction bias in the model training process, accurately captures the difference in multi-class probability distribution 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 adjustment defect of the learning rate of the traditional optimization method. The joint application of the weight decay regularization and the early stopping mechanism significantly suppresses the overfitting tendency of the model to the training data noise, and guarantees the generalization performance of the classification model in different geographical regions.

[0126] In an optional embodiment, after the completion of the training of the ground feature classification model, the performance of the ground feature classification model is evaluated by using sample data sets of different geographical regions to verify the generalization ability and adaptability of the ground feature classification model.

[0127] The performance evaluation is completed by a multi-dimensional index system of overall accuracy, Kappa coefficient, precision, recall and F1 score.

[0128] The sample data sets of different geographical regions refer to independent verification data sets containing various climate zones, topography and vegetation coverage types, and can be realized by using sample data of a temperate monsoon zone, a subtropical humid zone and a highland cold desert zone distributed in a cross-latitude gradient. The geographical heterogeneity in actual application is simulated by covering regional data with different environmental characteristics. The multi-dimensional index system refers to an evaluation set composed of classification accuracy, category consistency, prediction reliability and comprehensive balance index. The overall accuracy can be used to reflect the overall classification accuracy, the Kappa coefficient can be used to eliminate the interference of random consistency, the precision can be used to measure the accuracy of the prediction result, the recall can be used to evaluate the category coverage, and the F1 score can be used to balance the precision and recall. The performance fluctuation of the model in different scenarios is revealed by the synergistic effect of multiple indexes.

[0129] Among them, the precision is the ratio of the number of samples of a certain type of surface feature correctly predicted to the total number of samples predicted of that type, reflecting the accuracy of the prediction results; the recall rate is the ratio of the number of samples of a certain type of surface feature correctly predicted to the total number of actual samples of that type, reflecting the coverage ability of the actual samples; the F1 score is the harmonic mean of the precision and recall, which comprehensively measures the balance level of the two indicators. The overall accuracy is the ratio of the total number of samples correctly classified to the total number of samples, reflecting the overall classification effect of the model; the Kappa coefficient evaluates the reliability of the classification accuracy by comparing the consistency of the actual classification results and the random classification results, and its calculation logic is: based on the total number of samples, the actual number of samples and the predicted number of samples of each class, the consistency coefficient is obtained by eliminating the influence of random factors, and the higher the value, the better the consistency of the classification results and the true situation.

[0130] Specifically, after the model training phase is completed, new sample data sets that do not participate in training are collected from independent geographic areas, and these areas have different environmental characteristics from the training area. The standardized new sample data is input into the trained model for classification prediction, and then the prediction results are compared with the true labels interpreted by artificial. The overall accuracy calculates the proportion of correctly classified samples to total samples, reflecting the overall classification ability of the model; the Kappa coefficient eliminates the influence of random consistency and evaluates the consistency degree of the model classification results and the true label; the precision calculates the proportion of the number of correctly predicted samples to the total number of samples of each type of surface feature, to identify whether the model has over-prediction problem; the recall rate calculates the proportion of the number of correctly predicted samples to the number of actual samples of each type of surface feature, to detect whether the model has the risk of missing measurement; the F1 score integrates the precision and recall by harmonic mean to provide a balanced evaluation at the class level. This multi-index linkage evaluation method can expose the problem of classification accuracy decline caused by factors such as terrain undulation and vegetation seasonal change when the model is applied across regions.

[0131] Compared with the prior art, the traditional method usually only uses test data of a single geographic area or relies on a single precision indicator for model verification, and cannot effectively identify the adaptability defects of the model in unknown environments. The same region verification method used in the prior art is prone to overfitting to local features, and a single indicator evaluation cannot fully reflect the multiple dimensions of classification performance. The present scheme realizes the systematic test of the generalization ability of the model through the cooperative application of cross-regional test data and multi-dimensional index system, and overcomes the limitations of traditional methods in regional adaptability evaluation.

[0132] By the technical solution, the application can effectively detect the performance degradation of the ground object classification model in cross-region application, and locate the source of classification error through multi-dimensional indexes. For example, when the F1 score of the model in the plateau cold desert area is significantly lower than that in the training area, it can be identified that the model has insufficient adaptability to high-altitude terrain; when the recall rate of the paddy field category decreases in the subtropical humid area, it indicates that the model has insufficient sensitivity to the feature change of the paddy field caused by seasonal change. This diagnostic capability provides a clear direction for model optimization, for example, by increasing the training samples in the plateau area or introducing time-series waveform features to enhance the environmental adaptability of the model, thereby improving the reliability of the ground object classification model in actual application.

[0133] In an optional embodiment, the method of the application can distinguish ground object types with similar spectral features but different structural features, including:

[0134] Based on the echo energy feature difference caused by humidity difference, dry land and paddy field are distinguished;

[0135] Based on the multi-peak / single-wave echo feature difference caused by crown structure difference, broadleaf forest and coniferous forest are distinguished.

[0136] Among them, the echo energy feature difference refers to the energy attenuation difference of laser pulse reflection on different humidity ground surfaces, which can be realized by calculating the ratio of waveform intensity mean value to saturation threshold value. The ratio can quantify the degree of laser energy absorption by soil moisture. The multi-peak / single-wave echo feature difference refers to the influence of vegetation canopy vertical structure on laser pulse penetration path, which can be realized by extracting waveform peak value number and peak distance parameters. These parameters can represent the branch and leaf distribution density and hierarchical structure inside the canopy.

[0137] Specifically, when the laser pulse irradiates the surface of the paddy field, the water layer and the saturated soil will absorb part of the laser energy, causing the intensity of the reflected echo to decrease as a whole, and at this time the normalized waveform intensity mean value is usually in the interval of 0.3-0.5. The dry soil of dry land has high reflectivity, and its waveform intensity mean value can reach the interval of 0.6-0.8. For broadleaf forest, the multi-layer leaf structure causes the laser to produce multiple reflections during penetration, and the waveform presents 2-3 obvious wave peaks with a peak distance of 3-5 sampling points. The branches and leaves of coniferous forest are vertically tower-shaped, the laser penetration path is single, and the waveform presents only a single main peak with a narrow peak width.

[0138] In some specific embodiments, the echo energy difference analysis can use waveform integral area calculation to set the classification threshold value by comparing the integral value distribution of paddy field and dry land samples. The crown structure analysis can use the waveform derivative method to detect the number of wave peaks, and when the number of zero-crossing points of the first derivative is greater than 2 times, it is determined as a broadleaf forest feature.

[0139] Compared with the prior art, the traditional remote sensing classification method relies on the spectral reflectance difference of multispectral data, and it is difficult to distinguish water fields and dry lands with similar spectral characteristics. The scheme breaks through the limitation of spectral similarity on classification accuracy by analyzing the physical characteristic difference contained in the laser radar waveform data and establishing a classification index directly related to the structure characteristics of the ground object.

[0140] Through the above technical scheme, the present application effectively solves the misjudgment problem caused by spectral confusion in traditional remote sensing classification, and realizes accurate distinction of ground object types such as water fields and dry lands, broad-leaved forests and coniferous forests. The technical scheme quantifies the waveform energy attenuation characteristics and morphological characteristics, and constructs a classification basis directly corresponding to the physical properties of the ground object, which significantly improves the ground object recognition accuracy under complex terrain conditions.

[0141] In order to verify the technical effect of the present application, the confusion matrix of DenseNet in the original data set is shown in Table 2.

[0142] Table 2

[0143]

[0144] For the building category, the accuracy of the DenseNet model reaches 89.2%, and 892 out of 1000 samples are correctly classified. In forest classification, the accuracy of the model is 92.7%, and 927 out of 1000 samples are correctly identified. For the desert category, the accuracy of the model reaches 97.0%, and 970 out of 1000 samples are correctly classified. In water body classification, the accuracy of the model reaches 99.5%, and 995 out of 1000 samples are correctly identified. The overall accuracy of the DenseNet model calculated according to the confusion matrix data is 0.932, and the Kappa coefficient is 0.843. These indicators not only confirm the high accuracy of the model, but also reflect its high consistency and reliability in the task of land cover classification, indicating its potential for wide application in remote sensing and geographic information systems.

[0145] In the original data set, the classification result evaluation using DenseNet is shown in Table 3.

[0146] Table 3

[0147]

[0148] In the original dataset, the DenseNet model demonstrated robust classification capabilities across the four classes (buildings, forest, bare land, and water). For buildings, the model achieved an accuracy of 0.931, a recall of 0.892, and an F1 score of 0.911, showcasing its excellent balance between identification accuracy and error control. In forest classification, it achieved an accuracy of 0.964, 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 (accuracy: 0.924), indicating minimal missed errors in this category. The model's performance on water classification was particularly outstanding, with a near-perfect recall of 0.995, an accuracy of 0.970, and an F1 score of 0.982, confirming its reliability in identifying hydrological features. Overall, these results highlight the model's balanced trade-off between accuracy and recall, as well as its adaptability to complex spectral variations in land cover.

[0149] To validate the model's generalizability, the DenseNet trained on the original dataset was used to classify a new study area dataset, with the classification confusion matrix shown in Table 4:

[0150] Table 4

[0151]

[0152] In the new study area sample set, the DenseNet model performed well in identifying the building class, accurately classifying 87.9% of the samples. However, 5.8% were incorrectly classified as forest, 4.9% as bare land, and 1.3% as water. In the forest class, the model demonstrated high precision, correctly identifying 93.5% of the samples. There was minimal confusion with bare land (0.4%) and water (0.1%), although 5.9% were incorrectly labeled as buildings. Bare land classification exhibited exceptional specificity, with a correct classification rate of 96.4%, and near-zero misclassification rate for forest (0.0%). However, 3.2% were confused with water. Water achieved near-perfect discrimination, with an accuracy of 98.9%, and only 1.0% misclassified as bare land. The overall accuracy of the model was 94.3%, with a Kappa coefficient of 0.924, highlighting its strong consistency and adaptability to spectral overlaps between built-up areas and natural landscapes. These results indicate that the DenseNet model not only has high precision but also high consistency in the classification task on the new dataset.

[0153] Subsequently, the classification results were evaluated using multi-dimensional evaluation indicators, with the evaluation results shown in Table 5:

[0154] Table 5

[0155]

[0156] DenseNet model showed near perfect discrimination (accuracy: 97.0%, recall: 98.9%, F1 score: 98.2%) and minimal cross-class confusion (<3.2% misclassified as bare soil) for water. It showed the highest recall (96.4%) and robust precision (92.4%, F1 score: 94.7%) for bare soil, despite subtle spectral overlap with water. Forest had high consistency in classification (accuracy: 96.4%, recall: 92.7%, F1 score: 94.5%), with the main error coming from confusion with buildings (5.8%). Building structures, while achieving high precision (93.1%), had a recall of 89.2% (F1 score: 91.1%) due to spectral ambiguity with natural landscapes (6.0% misclassified as forest, 4.9% misclassified as bare soil). The overall accuracy of the model 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.

[0157] For paddy field and dry land, the classification confusion matrix of DenseNet for the dataset is shown in Table 5: Figure 4

[0158] The DenseNet model achieved 1669 true positives in dry land classification and 330 false positives in paddy field, with only 158 false positives in the dry land category. According to the confusion matrix, the overall precision was 0.878 and the kappa coefficient was 0.756, indicating that the model classification results had strong consistency with the true distribution of ground objects, effectively overcoming the "same object different spectrum" interference, and performed reliably in the classification task of dry land and paddy field.

[0159] The classification evaluation of DenseNet is shown in Table 6:

[0160] Table 6

[0161]

[0162] ​For dry land classification, the precision reached 0.914, that is, about 91.4% of the samples predicted as dry land are true dry land, which reflects the high reliability of the prediction result; the recall is 0.835, which indicates that about 83.5% of the true dry land samples can be effectively identified by the model, and the coverage ability is good; the F1-score calculated by the harmonic mean of the two is 0.873, which verifies the balanced performance of dry land classification in the dimensions of "accurate identification" and "comprehensive coverage". For paddy field classification, the precision is 0.848, and the true rate of the predicted paddy field sample reaches 84.8%; the recall reaches 0.921, and the capture rate of the true paddy field is more than 92%; the F1-score is 0.883, which highlights the excellent balance between "precision" and "completeness" in the classification of paddy field.

[0163] In terms of broadleaf forest and coniferous forest, the classification confusion matrix of DenseNet is as shown in Table 6: Figure 5

[0164] DenseNet performs well in coniferous forest classification, correctly classifying 767 samples and only 232 misclassifications, and also performs well in broadleaf forest classification, with 384 misclassifications. According to the confusion matrix, the overall precision is 0.692, and the kappa coefficient is 0.383, indicating that the model effectively overcomes the "same spectrum of different objects" interference and has basic reliability in the classification task of broadleaf forest and coniferous forest.

[0165] The classification evaluation of DenseNet is shown in Table 7:

[0166] Table 7

[0167]

[0168] In the classification task of broadleaf forest and coniferous forest of the present application, the model performance is accurately measured based on the confusion matrix: in the broadleaf forest classification dimension, the precision reaches 0.666, that is, about 66.6% of the samples predicted as broadleaf forest are true categories, ensuring the reliability of the prediction; the recall is 0.768, about 76.8% of the true broadleaf forest samples are effectively identified, and the coverage ability is good, and the harmonic average F1-score of the two is 0.715, balancing precision and comprehensiveness. In the coniferous forest classification dimension, the precision is 0.726, and the true rate of the predicted sample reaches 72.6%; the recall is 0.616, and the capture rate of the true coniferous forest is more than 61.6%; the F1-score is 0.667, verifying the balance between "precision" and "completeness". The DenseNet model of the present application deeply excavates the feature difference of laser radar waveform, strengthens feature reuse through dense connection, breaks through the traditional "same spectrum of different objects" recognition bottleneck, and provides a reliable subdivided ground object classification scheme for ecological evaluation, forestry monitoring and other scenes.

[0169] ​This result shows that the model based on the waveform data of the spaceborne lidar has reliability in the application of the same object with different spectra and different objects with the same spectrum.

[0170] Therefore, the land cover classification method based on the waveform data of the spaceborne lidar provided by the application deeply fuses the vertical structure information of the land cover contained in the spaceborne lidar and the fine category annotation of the ground reference data, not only provides a new data source with vertical dimension features and high spatial resolution for land cover classification, but also solves the pain points of the insufficient accuracy of the traditional method in the classification of the same object with different spectra (such as paddy field and dry land) and different objects with the same spectrum (such as bare land and water body), realizes the accurate capture of subtle waveform differences by means of the dense feature reuse mechanism of the DenseNet model, and thus constructs a solution considering the classification accuracy and regional adaptability, thereby providing a new technical path for improving the intelligent level of land cover classification.

[0171] Although the embodiments of the application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the application, and such modifications and changes 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. 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. These are the sampling points for the normalized laser echo waveform data; 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 or more land cover types from paddy fields, dry land, broad-leaved forests, and coniferous forests; the land cover classification model is a densely connected neural network model; 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. 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.

2. The method according to claim 1, 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.

3. The method according to claim 1, characterized in that, Deep neural network classification models employ 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.

4. The method according to claim 3, 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.

5. The method according to claim 1, 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.

6. The method according to claim 1, 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.

Images