Multi-feature attention cloud top height estimation algorithm based on satellite-borne laser radar

Abstract

The invention relates to the technical field of atmospheric remote sensing, in particular to a satellite-borne laser radar-based multi-feature attention cloud top height estimation algorithm, which comprises the following steps of: full-waveform laser radar data acquisition and preprocessing: acquiring an original echo signal through a satellite-borne full-waveform laser radar system; performing time domain alignment, noise suppression and saturated signal elimination on the signals; a median filtering and wavelet threshold denoising algorithm is combined to eliminate high-frequency noise, and background radiation interference is eliminated through dynamic baseline correction; according to the method, a multi-stage noise reduction strategy combining sliding window median filtering and self-adaptive wavelet denoising is introduced in the waveform feature extraction stage, pulse interference and signal oscillation are effectively restrained, meanwhile, through a terrain matching error compensation mechanism, the detection accuracy is improved, and the detection accuracy is improved. The physical consistency of the elevation residual error under the real terrain background is enhanced, and the authenticity and the checkability of the point elevation return are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of atmospheric remote sensing, and specifically to a multi-feature attention cloud top height estimation algorithm based on spaceborne lidar. Background Technique

[0002] In global climate change monitoring, tropospheric dynamics analysis, and weather forecasting model construction, cloud top height, as an important parameter characterizing the macroscopic physical structure of clouds, directly affects the simulation accuracy of radiation transfer models, precipitation evolution models, and cloud-aerosol interaction processes. With the wide deployment of spaceborne lidar (such as CALIOP) and ground-based scanning lidar, a large amount of multi-scale cloud echo data has been gradually accumulated, providing a rich data basis for high-resolution cloud top identification. However, it also poses higher requirements for feature analysis, model adaptation, and spatio-temporal consistency.

[0003] In traditional implementation methods, the extraction of cloud top height mostly relies on empirical threshold methods, fixed template matching methods, or peak detection algorithms based on a single physical parameter. Most methods use the strongest reflection point in the laser waveform as the basis for cloud top judgment, lacking a comprehensive modeling of waveform structure details, background noise status, and the influence of terrain changes, and it is difficult to handle complex scenarios such as multi-layer cloud superposition, weak signals of thin clouds in the boundary layer, and strong ground object background reflection interference. In addition, some algorithms use fixed spatial grids and single-scale discrimination criteria, ignoring the non-linear response of signal attenuation patterns and laser emission angles at different cloud heights, resulting in a significant increase in recognition errors in mid-high latitudes, high altitudes, or low-penetration cloud bodies, seriously limiting the wide-area applicability and accuracy consistency of the algorithms.

[0004] Therefore, we propose a multi-feature attention cloud top height estimation algorithm based on spaceborne lidar Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] The purpose of the present invention is to provide a multi-feature attention cloud top height estimation algorithm based on spaceborne lidar to solve the problems raised in the above background technique.

[0007] (2) Technical Solutions

[0008] To achieve the above purpose, the present invention provides the following technical solutions:

[0009] The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar includes the following steps:

[0010] (1) Full-wave lidar data acquisition and preprocessing: The original echo signal is obtained through the spaceborne full-wave lidar system, and the signal is subjected to time-domain alignment, noise suppression, and saturation signal rejection; the median filter combined with the wavelet threshold denoising algorithm is used to eliminate high-frequency noise, and the background radiation interference is eliminated through dynamic baseline correction; the invalid data segments are marked, and only the waveform data within the effective detection interval is retained;

[0011] (2) Multi-dimensional waveform feature extraction: The following features are extracted from the preprocessed waveform: ① waveform statistical features, including signal-to-noise ratio, kurtosis, skewness, and waveform energy distribution; ② spatial distribution features, including peak position, full width at half maximum, adjacent peak spacing, and interlayer distance; ③ time-domain correlation features, calculating the waveform local similarity and continuity indicators through a sliding window; ④ terrain matching features, combining high-resolution digital elevation model (DEM) data to calculate the residual between the laser foot point elevation and the terrain height;

[0012] (3) Multi-feature adaptive attention fusion: A fusion model combining a bidirectional long short-term memory network (Bi-LSTM) and a multi-head attention mechanism is constructed to dynamically allocate the weights of each feature; different attention coefficients are set for different altitude layers, where the low-altitude layer (<5 km) focuses on terrain matching features, the middle and high-altitude layers (5 - 10 km) focus on spatial distribution features, and the stratosphere (>10 km) focuses on time-domain correlation features; non-linear interaction between features is achieved through a learnable parameter matrix;

[0013] (4) Intelligent identification of cloud top height: Based on the fused features, a probability density function is generated, and an improved peak search algorithm is used to determine the candidate clouds; multiple criteria are defined: ① interlayer distance criterion, requiring the distance between adjacent clouds ≥30 m; ② signal-to-noise ratio criterion, the main peak signal-to-noise ratio ≥5 dB; ③ terrain consistency criterion, the residual between the cloud bottom height and the DEM elevation ≤20 m; the clouds that meet the criteria are sorted by height, and the upper boundary height of the highest effective layer is taken as the cloud top height;

[0014] (5) Iterative optimization and verification: The iterative closest point (ICP) algorithm with terrain constraints is introduced to perform spatial matching between the recognition results and the high-precision surface model, and the abnormal points are reclassified; a three-dimensional verification framework is established, and through cross-verification of adjacent orbit data and comparison with satellite visible light images, the cloud top height recognition accuracy indicators are calculated, including root mean square error (RMSE) ≤15 m and correlation coefficient ≥0.95.

[0015] As a preferred technical solution, the specific implementation of waveform feature extraction in step (2) includes:

[0016] (2.1) Pulse noise is eliminated by using median filtering in a sliding window, and the window length is dynamically adjusted according to the signal-to-noise ratio. A 5-point window is used when the signal-to-noise ratio <10 dB, and a 3-point window is used when the signal-to-noise ratio ≥10 dB;

[0017] (2.2) Build a denoising model based on the db4 wavelet. The decomposition level is adaptively selected according to the number of waveform sampling points. When the number of sampling points exceeds 1000, 6-layer decomposition is adopted; otherwise, 4-layer decomposition is adopted.

[0018] (2.3) Design a multi-scale peak detection algorithm. In the first round of detection, the fixed threshold method is used to determine the candidate peaks, and the threshold is set to 3 times the standard deviation of the noise. In the second round of detection, the local adaptive threshold method is used. A 200m window is constructed centered on the candidate peak, and the mean value of the signal in the window plus 2 times the standard deviation is calculated as the new threshold.

[0019] (2.4) Introduce a terrain matching correction module. Convert the laser foot point coordinates to the WGS84 coordinate system, perform spatial interpolation matching with the ASTER GDEM V3 data, and consider the satellite attitude angle error compensation when calculating the elevation residual.

[0020] (2.5) Establish a feature standardization pipeline. Use Z-score standardization for statistical features, use the maximum-minimum normalization for spatial features, and use interval standardization based on elevation bands for terrain features.

[0021] As a preferred technical solution, in the peak detection algorithm in step (2.3), it is based on a two-stage screening structure. The primary stage uses the fixed threshold method, and the threshold is set to 3 times the standard deviation of the full waveform noise segment. The secondary stage uses the local adaptive threshold method. A 200m sliding window is constructed centered on each candidate peak, and the sum of the mean value of the signal in the window and 2 times the standard deviation is calculated as the local detection threshold.

[0022] The terrain matching correction module in step (2.4) calculates the height residual:

[0023] Δh = Δθ·H·tan(α):

[0024] Compensate for the attitude angle error, and the corrected elevation residual is used as the terrain consistency feature and input into the fusion model. Here, Δθ represents the attitude angle error, H is the lidar orbit height, and α is the laser emission angle. This formula quantifies the vertical elevation error caused by the attitude deviation.

[0025] As a preferred technical solution, the specific implementation of the multi-feature adaptive attention fusion in step (3) includes:

[0026] (3.1) Build a hierarchical feature encoder. In the first layer, a convolutional neural network is used to extract local features, and the convolutional kernel size is set to a multi-scale structure of [3, 5, 7]. In the second layer, a bidirectional LSTM is used to capture the temporal dependence relationship, and the number of hidden layer units is set to 128.

[0027] (3.2) Design space-channel dual attention mechanism. The spatial attention module generates a weight map of the attention region through deformable convolution, and the channel attention module uses the SE-Net structure to calculate the importance of feature channels.

[0028] (3.3) Achieve highly adaptive weight allocation and establish a mapping function between height intervals and attention coefficients.

[0029] (3.4) Introduce an adversarial training strategy. Construct a discriminator network to distinguish between manually annotated data and model outputs, and improve the robustness of feature fusion by minimizing the Wasserstein distance.

[0030] (3.5) Adopt knowledge distillation technology. Use the feature distribution of the teacher model (ResNet-34) as a supervision signal to guide the feature representation learning of the student model (Bi-LSTM).

[0031] As a preferred technical solution, the height-adaptive weight allocation function in step (3.3) is:

[0032] ω i =σ(a·h + b), where h is the absolute height corresponding to the current laser waveform, a and b are trainable parameters, and σ represents the Sigmoid activation function.

[0033] Use the function to map the absolute height h of each waveform point to the fusion weight ω i , where σ is the Sigmoid activation function. The parameters a and b are co-optimized with the feature encoder during training, enabling the model to automatically generate attention weights matching the spatial hierarchical structure at different height layers, improving the ability to distinguish high and low clouds and the fusion accuracy.

[0034] In step (3.4), an adversarial training strategy is introduced. By constructing a Wasserstein discriminator network to distinguish between manually annotated features and model-generated features, the optimization objective is to minimize the Wasserstein distance W(P r and the generated distribution P g ), improving the structural consistency and physical credibility of the model's fusion output, and thus enhancing the recognition robustness for complex multi-layer cloud structures. r ,P g )

[0035] As a preferred technical solution, the optimization method for cloud top height recognition in step (4) includes:

[0036] (4.1) Design a multi-constraint peak screening algorithm and define a comprehensive scoring function:

[0037] S = 0.4×SNR + 0.3×Width + 0.2×Isolation + 0.1×TerrainConsistency:

[0038] where SNR is the normalized signal-to-noise ratio, Width is the normalized reciprocal of the peak width, Isolation is the distance between adjacent peaks, and TerrainConsistency is the terrain matching degree;

[0039] (4.2) Implement height interval stratification processing. Divide the height range from 0 to 20 km into 200 intervals, calculate the characteristic statistics independently for each interval, and use cubic spline interpolation to correct the boundaries for cloud bodies across layers;

[0040] (4.3) Establish a cloud phase discrimination module. Distinguish ice clouds and water clouds by analyzing the attenuation ratio of the 532 nm and 1064 nm dual-wavelength signals. The attenuation ratio threshold for ice clouds is set to 1.2 - 1.5, and for water clouds is set to 0.8 - 1.1;

[0041] (4.4) Introduce time continuity constraints, perform Kalman filter smoothing on the cloud top height change rate of 10 consecutive detection points, and eliminate mutation values.

[0042] As a preferred technical solution, the optimization method for cloud top height identification in step (4) further includes:

[0043] (4.5) Construct an uncertainty quantification model, use the Monte Carlo Dropout method to calculate the confidence interval of the cloud top height, introduce uncertainty quantification and time continuity constraints to improve the stability and credibility of the cloud top height estimation result. Among them, uncertainty quantification uses the Monte Carlo Dropout mechanism, keeps some neurons in the network randomly inactivated during the inference stage, and statistically calculates the mean and confidence interval of the final estimation result after obtaining the output distribution through multiple forward propagations. The output form is defined as:

[0044] H top = μ ± 1.96σ:

[0045] where μ is the mean of multiple prediction results, σ is the standard deviation, and 1.96 is the scale coefficient at a 95% confidence level. This formula provides the confidence boundary for single-point cloud top height prediction, effectively captures the sources of model uncertainty, constructs a time smoothing module based on Kalman filter, dynamically estimates and eliminates anomalies for the height change rate of 10 consecutive laser detection points, and the Kalman state space model jointly adjusts the filter gain in the prior prediction and measurement update processes.

[0046] As a preferred technical solution, the specific implementation of the iterative optimization and verification in step (5) includes:

[0047] (5.1) ICP algorithm with terrain constraints designed;

[0048] (5.2) Establish a multi-source data verification framework, integrate CALIPSO satellite cloud products, Himawari-8 infrared cloud images and ground lidar observation data, and define verification indicators:

[0049] ① Horizontal matching accuracy: The spatial grid resolution is increased to 0.05°×0.05°;

[0050] ② Time synchronization error: ≤30 seconds;

[0051] ③ Vertical resolution: ≤10m;

[0052] (5.3) Implement dynamic error compensation and construct an error transfer model;

[0053] (5.4) Develop an online learning module, continuously optimize the classifier parameters through the incremental random forest algorithm, and update the model every time 1000 new samples are received;

[0054] (5.5) Design a visual verification interface to synchronously display lidar waveforms, feature heat maps, cloud profiles and 3D terrain mapping results, and support comparative analysis of manual annotation and model prediction.

[0055] As a preferred technical solution, the terrain-constrained iterative closest point matching algorithm designed in step (5.1) constructs a dual geometric consistency criterion between lidar point clouds and DEM elevation data, and forms a cost function by introducing the joint of elevation difference and gradient difference:

[0056] E = ∑(α·|h lidar - h DEM | + β·||▽h lidar - ▽h DEM ||):

[0057] Where h lidar represents the elevation value obtained by lidar detection, h DEM is the DEM terrain elevation at the corresponding position, ▽h

[0058] represents the terrain gradient vector at this position, α = 0.7 and β = 0.3 respectively represent the weights for elevation consistency and terrain structure consistency. By minimizing E, the spatial alignment of laser foot points to the terrain is realized, thereby effectively improving the terrain consistency verification ability of the cloud top position. As a preferred technical solution, in step (5.3), dynamic error compensation is implemented, and an error transfer model is constructed to correct the cloud top height offset introduced by platform attitude disturbance, time synchronization deviation and terrain data error. This mechanism quantifies and estimates the uncertainty sources by establishing an error transfer model, and its total error is defined as:

[0059]

[0060] Among them, δθ represents the platform pointing angle error, H is the ranging height, θ is the current laser incident angle, δt is the detection time synchronization error, c is the speed of light, and δ DEM represents the uncertainty of the DEM source elevation data;

[0061] In step (5.4), an online learning module is developed and a multi-source verification visualization system is constructed accordingly. By introducing the incremental random forest algorithm, a continuously optimized cloud recognition classifier is constructed. The windowed training strategy is adopted, and the model parameters are updated every time 1000 new samples are received, enabling the model to have adaptive learning ability. On this basis, a visualization verification interface is developed, integrating the capabilities of dynamic waveform signal drawing, feature heat map rendering, three-dimensional cloud structure reconstruction and terrain background fusion. The interface supports the loading of multi-source verification data measured by CALIPSO cloud products, Himawari-8 infrared images and ground lidar, and displays the differences between the manually marked results and the model predictions in a figure-data collaboration manner.

[0062] (III) Beneficial effects

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0064] 1. By introducing a multi-level noise reduction strategy that combines median filtering in a sliding window and adaptive wavelet denoising in the waveform feature extraction stage, pulse interference and signal oscillation are effectively suppressed, ensuring high confidence and positioning accuracy in peak detection. At the same time, through the terrain matching error compensation mechanism, the physical consistency of the elevation residuals in the real terrain background is enhanced, avoiding elevation offsets caused by attitude errors or tilt angle amplification effects, and improving the authenticity and verifiability of the point elevation feedback.

[0065] 2. The solution constructs a hierarchical multi-modal feature fusion framework that integrates CNN, Bi-LSTM, spatial-channel attention mechanism and height-sensitive weighted mapping function, which can effectively extract the signal morphology and time series evolution law at the local scale, and realize the reconstruction of differential attention regions in different cloud height intervals. The specially constructed mapping function dynamically adjusts the attention coefficient, enabling the model to have stronger discrimination ability in the feature expression of high and low clouds, and improving the recognition effect of ice clouds and water clouds, boundary layer clouds and high-level cirrus clouds.

[0066] 3. Aiming at the complex interference factors in the process of cloud top height extraction, the solution introduces multi-constraint peak screening and temporal Kalman filtering smoothing processing, making the cloud boundary recognition more stable and continuous. Through the dual-mechanism discrimination of cubic spline boundary correction and attenuation ratio spectral analysis, mixed cloud types and multi-layer cloud structures are effectively separated, enhancing the discrimination credibility under complex weather conditions.

[0067] 4. The solution introduces an adversarial strategy and a knowledge distillation structure in the model training and deployment phases. The teacher model guides the student model to learn the cross-layer semantic feature distribution, and at the same time, the Wasserstein distance is used to optimize the distribution matching process, improving the model generalization ability and robustness. Finally, in the model verification phase, a joint uncertainty control strategy based on Kalman filtering and Monte Carlo Dropout is constructed, and a multi-source data fusion verification platform and an online learning mechanism are built to enable the system to have the ability of continuous update and dynamic parameter adjustment, supporting the application deployment on long time scales and across regional ranges. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0069] Figure 1 Schematic diagram of the altitude distribution of cloud structures and terrain profiles in different latitudes of the present invention;

[0070] Figure 2 Schematic diagram of the cloud structure and terrain profile of latitude-iteration height in the present invention;

[0071] Figure 3 Schematic diagram of the coupling relationship between cloud top height and terrain in the mid-high latitudes of the present invention;

[0072] Figure 4 Comparison diagram of cross-latitude cloud layer distribution and terrain profile in the present invention;

[0073] Figure 5 Schematic diagram of the cloud structure and terrain elevation distribution near the polar region in the present invention;

[0074] Figure 6 Comprehensive analysis diagram of table type and cloud layer characteristics in the present invention;

[0075] Figure 7 Three-dimensional verification schematic diagram of the cloud top height recognition result in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0076] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0077] According to the attached Figures 1-7 As shown, the embodiments of the present invention provide a multi-feature attention cloud top height estimation algorithm based on spaceborne lidar, including the following steps:

[0078] (1) Acquisition and preprocessing of full-wave lidar data: Obtain the original echo signal through a spaceborne full-wave lidar system, perform time-domain alignment, noise suppression, and saturation signal removal on the signal; use median filtering combined with wavelet threshold denoising algorithm to eliminate high-frequency noise, and eliminate background radiation interference through dynamic baseline correction; mark invalid data segments and only retain the waveform data within the effective detection interval;

[0079] (2) Extraction of multi-dimensional waveform features: Extract the following features from the preprocessed waveform: ① waveform statistical features, including signal-to-noise ratio, kurtosis, skewness, and waveform energy distribution; ② spatial distribution features, including peak position, full width at half maximum, adjacent peak spacing, and layer distance; ③ time-domain correlation features, calculate waveform local similarity and continuity indicators through a sliding window; ④ terrain matching features, combine high-resolution digital elevation model (DEM) data, and calculate the residual between the laser footprint elevation and the terrain height;

[0080] (3) Multi-feature adaptive attention fusion: Construct a fusion model combining a bidirectional long short-term memory network (Bi-LSTM) and a multi-head attention mechanism to dynamically allocate the weights of each feature; set different attention coefficients for different altitude layers, where the low altitude layer (<5km) focuses on terrain matching features, the middle and high altitude layers (5-10km) focus on spatial distribution features, and the stratosphere (>10km) focuses on time-domain correlation features; realize non-linear interaction between features through a learnable parameter matrix;

[0081] (4) Intelligent identification of cloud top height: Generate a probability density function based on the fusion features, and use an improved peak search algorithm to determine candidate clouds; define multi-layer criteria: ① layer distance criterion, requiring the distance between adjacent clouds ≥30m; ② signal-to-noise ratio criterion, the main peak signal-to-noise ratio ≥5dB; ③ terrain consistency criterion, the residual between the cloud base height and the DEM elevation ≤20m; sort the clouds that meet the criteria by height, and take the upper boundary height of the highest effective layer as the cloud top height;

[0082] (5) Iterative optimization and verification: Introduce an iterative closest point (ICP) algorithm with terrain constraints, perform spatial matching of the recognition results with a high-precision surface model, and perform secondary classification on abnormal points; establish a three-dimensional verification framework, calculate the cloud top height recognition accuracy indicators through cross-verification of adjacent orbit data and comparison with satellite visible light images, including root mean square error (RMSE) ≤15m, correlation coefficient ≥0.95.

[0083] As a further limitation of this solution, the specific implementation of waveform feature extraction in step (2) includes:

[0084] (2.1) Use moving window median filtering to eliminate impulse noise. The window length is dynamically adjusted according to the signal-to-noise ratio. When the signal-to-noise ratio < 10dB, a 5-point window is used; when the signal-to-noise ratio ≥ 10dB, a 3-point window is used.

[0085] (2.2) Construct a denoising model based on db4 wavelet. The decomposition level is adaptively selected according to the number of waveform sampling points. When the number of sampling points exceeds 1000, 6-layer decomposition is used; otherwise, 4-layer decomposition is used.

[0086] (2.3) Design a multi-scale peak detection algorithm. In the first round of detection, the fixed threshold method is used to determine candidate peaks, and the threshold is set to 3 times the standard deviation of the noise. In the second round of detection, the local adaptive threshold method is used. A 200m window is constructed centered on the candidate peak, and the mean value of the signal in the window plus 2 times the standard deviation is calculated as the new threshold.

[0087] (2.4) Introduce a terrain matching correction module to convert the laser foot point coordinates to the WGS84 coordinate system, perform spatial interpolation matching with ASTER GDEM V3 data, and consider satellite attitude angle error compensation when calculating the elevation residual.

[0088] (2.5) Establish a feature standardization pipeline. Use Z-score standardization for statistical features, min-max normalization for spatial features, and interval normalization based on elevation bands for terrain features.

[0089] As a further limitation of this solution, the peak detection algorithm in step (2.3) is based on a two-stage screening structure. The primary stage uses the fixed threshold method, and the threshold is set to 3 times the standard deviation of the full waveform noise segment. The secondary stage uses the local adaptive threshold method. A 200m sliding window is constructed centered on each candidate peak, and the sum of the mean value of the signal in the window and 2 times the standard deviation is calculated as the local detection threshold.

[0090] The terrain matching correction module in step (2.4) calculates the height residual:

[0091] Δh = Δθ·H·tan(α):

[0092] Compensate for the attitude angle error, and the corrected elevation residual is used as the terrain consistency feature and input into the fusion model. Here, Δθ represents the attitude angle error, H is the lidar orbit height, and α is the laser emission angle. This formula quantifies the vertical elevation error caused by attitude deviation.

[0093] As a further limitation of this solution, the specific implementation of multi-feature adaptive attention fusion in step (3) includes:

[0094] (3.1) Construct a hierarchical feature encoder. In the first layer, a convolutional neural network is used to extract local features, and a multi-scale structure with convolutional kernel sizes of [3, 5, 7] is set; in the second layer, a bidirectional LSTM is used to capture temporal dependencies, and the number of hidden layer units is set to 128;

[0095] (3.2) Design a spatial-channel dual attention mechanism. The spatial attention module generates a weight map of the attention region through deformable convolution, and the channel attention module uses the SE-Net structure to calculate the importance of feature channels;

[0096] (3.3) Implement highly adaptive weight allocation, and establish a mapping function between the height interval and the attention coefficient;

[0097] (3.4) Introduce an adversarial training strategy, construct a discriminator network to distinguish between manually annotated data and model outputs, and improve the robustness of feature fusion by minimizing the Wasserstein distance;

[0098] (3.5) Adopt knowledge distillation technology, use the feature distribution of the teacher model (ResNet-34) as a supervision signal to guide the feature representation learning of the student model (Bi-LSTM).

[0099] As a further limitation of this solution, the height adaptive weight allocation function in step (3.3) is:

[0100] ω i = σ(a·h + b), where h is the absolute height corresponding to the current laser waveform, a and b are trainable parameters, and σ represents the Sigmoid activation function;

[0101] Use the function to map the absolute height h of each waveform point to the fusion weight ω i , where σ is the Sigmoid activation function, and the parameters a and b are co-optimized with the feature encoder during training, enabling the model to automatically generate attention weights matching the spatial hierarchical structure at different height layers, improving the ability to distinguish high and low clouds and the fusion accuracy;

[0102] In step (3.4), an adversarial training strategy is introduced. By constructing a Wasserstein discriminator network to distinguish between manually annotated features and model-generated features, the optimization objective is to minimize the Wasserstein distance W(P r and the generated distribution P g ), improving the structural consistency and physical credibility of the model's fusion output, and thus enhancing the recognition robustness for complex multi-layer cloud structures. r ,P g )

[0103] As a further limitation to this solution, the optimization method for cloud top height identification in step (4) includes:

[0104] (4.1) Design a multi-constraint peak screening algorithm and define a comprehensive scoring function:

[0105] S = 0.4×SNR + 0.3×Width + 0.2×Isolation + 0.1×TerrainConsistency:

[0106] where SNR is the normalized signal-to-noise ratio, Width is the reciprocal normalized peak width, Isolation is the adjacent peak spacing, and TerrainConsistency is the terrain matching degree;

[0107] (4.2) Implement height interval stratification processing. Divide the height range of 0 - 20 km into 200 intervals, calculate the characteristic statistics independently for each interval, and use cubic spline interpolation to correct the boundaries for cloud bodies spanning multiple layers;

[0108] (4.3) Establish a cloud phase discrimination module. Distinguish ice clouds from water clouds by analyzing the attenuation ratio of 532 nm and 1064 nm dual-wavelength signals. The attenuation ratio threshold for ice clouds is set to 1.2 - 1.5, and for water clouds is set to 0.8 - 1.1;

[0109] (4.4) Introduce time continuity constraints. Perform Kalman filter smoothing on the cloud top height change rate of 10 consecutive detection points to eliminate mutation values.

[0110] As a further limitation to this solution, the optimization method for cloud top height identification in step (4) also includes:

[0111] (4.5) Construct an uncertainty quantification model. Use the Monte Carlo Dropout method to calculate the confidence interval of the cloud top height. Introduce uncertainty quantification and time continuity constraints to improve the stability and credibility of the cloud top height estimation results. Among them, uncertainty quantification adopts the Monte Carlo Dropout mechanism. Keep some neurons in the network randomly inactivated during the inference stage, and after obtaining the output distribution through multiple forward propagations, perform statistics on the mean and confidence interval of the final estimation result. The output form is defined as:

[0112] H top = μ ± 1.96σ:

[0113] Among them, μ is the mean of multiple prediction results, σ is the standard deviation, and 1.96 is the scale coefficient under a 95% confidence level. This formula provides the confidence boundary for the prediction of the single-point cloud top height, effectively capturing the sources of model uncertainty, constructing a time smoothing module based on Kalman filtering, dynamically estimating and abnormally eliminating the height change rate of 10 consecutive lidar detection points, and the Kalman state space model jointly adjusts the filtering gain in the prior prediction and measurement update processes.

[0114] As a further limitation of this solution, the specific implementation of the iterative optimization and verification in step (5) includes:

[0115] (5.1) Design an ICP algorithm with terrain constraints;

[0116] (5.2) Establish a multi-source data verification framework, integrate CALIPSO satellite cloud products, Himawari-8 infrared cloud images, and ground lidar observation data, and define verification indicators:

[0117] ① Horizontal matching accuracy: The spatial grid resolution is increased to 0.05°×0.05°;

[0118] ② Time synchronization error: ≤ 30 seconds;

[0119] ③ Vertical resolution: ≤ 10m;

[0120] (5.3) Implement dynamic error compensation and construct an error transfer model;

[0121] (5.4) Develop an online learning module, continuously optimize the classifier parameters through the incremental random forest algorithm, and update the model every time 1000 new samples are received;

[0122] (5.5) Design a visual verification interface to synchronously display lidar waveforms, feature heat maps, cloud profiles, and three-dimensional terrain mapping results, and support comparative analysis of manual annotation and model prediction.

[0123] As a further limitation of this solution, the terrain constraint iterative closest point matching algorithm designed in step (5.1) constructs a dual geometric consistency criterion between lidar point clouds and DEM elevation data, and forms a cost function by introducing the joint of elevation difference and gradient difference:

[0124] E = ∑(α·|h lidar - h DEM | + β·||▽h lidar - ▽h DEM ||):

[0125] Among them, h lidar represents the elevation value obtained by lidar detection, and h DEM is the DEM terrain elevation at the corresponding position. It represents the terrain gradient vector at this position. α = 0.7 and β = 0.3 respectively represent the weights for elevation consistency and terrain structure consistency. By minimizing E, the spatial alignment of the laser foot points with the terrain is achieved, thereby effectively improving the terrain consistency verification ability at the cloud top position.

[0126] As a further limitation to this solution, in step (5.3), dynamic error compensation is implemented, and an error transfer model is constructed to correct the cloud top height offset introduced by platform attitude perturbation, time synchronization deviation, and terrain data error. This mechanism quantifies and estimates the uncertainty sources by establishing an error transfer model, and its total error is defined as:

[0127]

[0128] where δθ represents the platform pointing angle error, H is the ranging height, θ is the current laser incident angle, δt is the detection time synchronization error, c is the speed of light, and δ DEM represents the uncertainty of the DEM source elevation data;

[0129] In step (5.4), an online learning module is developed and a multi-source verification visualization system is constructed accordingly. By introducing the incremental random forest algorithm, a continuously optimized cloud recognition classifier is constructed. The windowed training strategy is adopted, and the model parameters are updated every time 1000 new samples are received, enabling the model to have the ability of adaptive learning. On this basis, a visualization verification interface is developed, integrating the capabilities of dynamic waveform signal plotting, feature heatmap rendering, three-dimensional cloud structure reconstruction, and terrain background fusion. The interface supports the loading of multi-source verification data measured by CALIPSO cloud products, Himawari-8 infrared images, and ground lidar, and displays the differences between the manually marked results and the model predictions in a figure-data collaborative manner.

[0130] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and do not limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. Multi-feature attention cloud top height estimation algorithm based on spaceborne lidar, characterized in that It includes the following steps: (1) Acquisition and preprocessing of full-wave lidar data: The original echo signal is obtained through a spaceborne full-wave lidar system, and the signal is subjected to time-domain alignment, noise suppression, and saturation signal removal; median filtering combined with wavelet threshold denoising algorithm is used to eliminate high-frequency noise, and background radiation interference is eliminated through dynamic baseline correction; invalid data segments are marked, and only the waveform data within the effective detection interval is retained; (2) Extraction of multi-dimensional waveform features: The following features are extracted from the preprocessed waveform: ① Waveform statistical features, including signal-to-noise ratio, kurtosis, skewness, and waveform energy distribution; ② Spatial distribution features, including peak position, full width at half maximum, adjacent peak spacing, and interlayer distance; ③ Time-domain correlation features, calculating waveform local similarity and continuity indicators through a sliding window; ④ Terrain matching features, combining high-resolution digital elevation model (DEM) data to calculate the residual between the laser foot point elevation and the terrain height; (3) Multi-feature adaptive attention fusion: A fusion model combining a bidirectional long short-term memory network (Bi-LSTM) and a multi-head attention mechanism is constructed to dynamically allocate the weights of each feature; different attention coefficients are set for different altitude layers, with the low-altitude layer (<5 km) focusing on terrain matching features, the middle and high-altitude layers (5 - 10 km) focusing on spatial distribution features, and the stratosphere (>10 km) focusing on time-domain correlation features; non-linear interaction between features is achieved through a learnable parameter matrix; (4) Intelligent identification of cloud top height: A probability density function is generated based on the fusion features, and an improved peak search algorithm is used to determine candidate clouds; multiple-layer criteria are defined: ① Interlayer distance criterion, requiring the distance between adjacent clouds to be ≥30 m; ② Signal-to-noise ratio criterion, the main peak signal-to-noise ratio ≥5 dB; ③ Terrain consistency criterion, the residual between the cloud base height and the DEM elevation ≤20 m; the clouds that meet the criteria are sorted by height, and the upper boundary height of the highest effective layer is used as the cloud top height; (5) Iterative optimization and verification: The iterative closest point (ICP) algorithm with terrain constraints is introduced to perform spatial matching between the recognition result and the high-precision surface model, and secondary classification of abnormal points is carried out; a three-dimensional verification framework is established, and through cross-verification of adjacent orbit data and comparison with satellite visible light images, the cloud top height recognition accuracy indicators are calculated, including root mean square error (RMSE) ≤15 m and correlation coefficient ≥0.

95.

2. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 1, wherein: The specific implementation of waveform feature extraction in step (2) includes: (2.1) Pulse noise is eliminated by using median filtering in a sliding window, and the window length is dynamically adjusted according to the signal-to-noise ratio. A 5-point window is used when the signal-to-noise ratio <10 dB, and a 3-point window is used when the signal-to-noise ratio ≥10 dB; (2.2) A denoising model based on db4 wavelet is constructed, and the decomposition level is adaptively selected according to the waveform sampling points. When the sampling points exceed 1000, 6-layer decomposition is used, otherwise 4-layer decomposition is used; (2.3) A multi-scale peak detection algorithm is designed. In the first round of detection, the fixed threshold method is used to determine candidate peaks, and the threshold is set to 3 times the noise standard deviation; in the second round of detection, the local adaptive threshold method is used. A 200m window is constructed centered on the candidate peak, and the signal mean plus 2 times the standard deviation within the window is calculated as the new threshold; (2.4) Introduce a terrain matching correction module to convert the lidar foot point coordinates to the WGS84 coordinate system, perform spatial interpolation matching with ASTER GDEM V3 data, and consider satellite attitude angle error compensation when calculating the elevation residual; (2.5) Establish a feature normalization pipeline, perform Z-score normalization on statistical features, min-max normalization on spatial features, and interval normalization based on elevation bands on terrain features.

3. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 2, wherein: In the peak detection algorithm in step (2.3), a two-level screening structure is adopted. The primary level uses a fixed threshold method with the threshold set to 3 times the standard deviation of the full waveform noise segment. The secondary level uses a local adaptive threshold method. A 200m sliding window is constructed at the center of each candidate peak, and the sum of the signal mean and 2 times the standard deviation within the window is calculated as the local detection threshold; The terrain matching correction module in step (2.4) compensates for the attitude angle error by calculating the height residual: Δh = Δθ · H · tan(α): The compensated elevation residual is used as the terrain consistency feature and input into the fusion model. Here, Δθ represents the attitude angle error, H is the lidar orbit height, and α is the laser emission angle. This formula quantifies the vertical elevation error caused by the attitude deviation.

4. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 1, characterized in that: The specific implementation of multi-feature adaptive attention fusion in step (3) includes: (3.1) Construct a hierarchical feature encoder. In the first layer, a convolutional neural network is used to extract local features with a multi-scale structure of convolutional kernel sizes [3, 5, 7]. In the second layer, a bidirectional LSTM is used to capture temporal dependencies, and the number of hidden layer units is set to 128; (3.2) Design a spatial-channel dual attention mechanism. The spatial attention module generates a weight map of the attention region through deformable convolution, and the channel attention module uses the SE-Net structure to calculate the importance of feature channels; (3.3) Implement height-adaptive weight allocation and establish a mapping function between height intervals and attention coefficients; (3.4) Introduce an adversarial training strategy, construct a discriminator network to distinguish between manually annotated data and model outputs, and improve the robustness of feature fusion by minimizing the Wasserstein distance; (3.5) Adopt knowledge distillation technology, use the feature distribution of the teacher model (ResNet-34) as a supervision signal to guide the feature representation learning of the student model (Bi-LSTM).

5. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 4, wherein: The height-adaptive weight allocation function in step (3.3) is: ω i = σ(a·h + b), where h is the absolute height corresponding to the current laser waveform, a and b are trainable parameters, and σ represents the Sigmoid activation function; Mapping the absolute height h of each waveform point to the fusion weight ω using a function i , where σ is the Sigmoid activation function, and the parameters a and b are co-optimized with the feature encoder during training, enabling the model to automatically generate attention weights matching the spatial hierarchical structure at different height levels, enhancing the ability to distinguish high and low clouds and the fusion accuracy; The adversarial training strategy is introduced in step (3.4) to distinguish the manually labeled features from the model generated features by constructing a Wasserstein discriminator network. The optimization goal is to minimize the true distribution P r and the generating distribution P g The Wasserstein distance W(P r ,P g ), improve the structural consistency and physical credibility of the model fusion output, and thus enhance the recognition robustness of complex multi-layer cloud structures.

6. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 1, wherein: The optimization method for cloud top height identification in step (4) includes: (4.1) Design a multi-constraint peak screening algorithm and define a comprehensive scoring function: S = 0.4 × SNR + 0.3 × Width + 0.2 × Isolation + 0.1 × TerrainConsistency: Where SNR is the normalized signal-to-noise ratio, Width is the normalized reciprocal of the peak width, Isolation is the adjacent peak spacing, and TerrainConsistency is the terrain matching degree; (4.2) Implement hierarchical processing for altitude intervals. Divide the altitude range of 0 - 20 km into 200 intervals, calculate the characteristic statistics for each interval independently, and use cubic spline interpolation to correct the boundaries for cross-layer cloud bodies. (4.3) Establish a cloud phase discrimination module. Distinguish ice clouds and water clouds by analyzing the attenuation ratio of the 532 nm and 1064 nm dual-wavelength signals. The attenuation ratio threshold for ice clouds is set to 1.2 - 1.5, and for water clouds is set to 0.8 - 1.

1. (4.4) Introduce time continuity constraints, and perform Kalman filter smoothing on the cloud top height change rate of 10 consecutive detection points to eliminate outliers.

7. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 6, wherein: The optimization method for cloud top height identification in step (4) further includes: (4.5) Construct an uncertainty quantification model. Use the Monte Carlo Dropout method to calculate the confidence interval of the cloud top height. Introduce uncertainty quantification and time continuity constraints to improve the stability and credibility of the cloud top height estimation results. Among them, uncertainty quantification uses the Monte Carlo Dropout mechanism, keeps some neurons in the network randomly inactivated during the inference stage, and statistically analyzes the mean and confidence interval of the final estimation result after obtaining the output distribution through multiple forward propagations. The output form is defined as: H top = μ ± 1.96σ: where μ is the mean of multiple prediction results, σ is the standard deviation, and 1.96 is the scale factor at a 95% confidence level. This formula provides the confidence boundary for the single-point cloud top height prediction, effectively capturing the sources of model uncertainty. Construct a time smoothing module based on Kalman filter to dynamically estimate and eliminate anomalies in the height change rate of 10 consecutive lidar detection points. The Kalman state space model jointly adjusts the filtering gain in the prior prediction and measurement update processes.

8. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 1, characterized in that: The specific implementation of iterative optimization and verification in step (5) includes: (5.1) Design an Iterative Closest Point (ICP) algorithm with terrain constraints. (5.2) Establish a multi-source data verification framework, integrate CALIPSO satellite cloud layer products, Himawari-8 infrared cloud images, and ground lidar observation data, and define verification metrics: ① Horizontal matching accuracy: The spatial grid resolution is increased to 0.05° × 0.05°. ② Time synchronization error: ≤ 30 seconds. ③ Vertical resolution: ≤ 10 m. (5.3) Implement dynamic error compensation and construct an error transfer model. (5.4) Develop an online learning module, continuously optimize the classifier parameters through the incremental random forest algorithm, and update the model every 1000 new samples received. (5.5) Design a visual verification interface to synchronously display the lidar waveform, feature heat map, cloud layer profile, and 3D terrain mapping results, and support the comparative analysis of manual annotation and model prediction.

9. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 8, characterized in that: The terrain-constrained Iterative Closest Point matching algorithm designed in step (5.1) constructs a dual geometric consistency criterion between the lidar point cloud and DEM elevation data by introducing a cost function jointly composed of elevation difference and gradient difference: where h lidar represents the elevation value obtained by laser detection, and h DEM is the DEM terrain elevation at the corresponding position. represents the terrain gradient vector at this position. α = 0.7 and β = 0.3 respectively represent the weights for elevation consistency and terrain structure consistency. By minimizing E, the spatial alignment of the laser foot points to the terrain is achieved, thereby effectively improving the terrain consistency verification ability at the cloud top position.

10. The multi-feature attention cloud top height estimation algorithm based on spaceborne lidar according to claim 8, characterized in that: In step (5.3), dynamic error compensation is implemented to construct an error transfer model for correcting the cloud top height offset introduced by platform attitude disturbance, time synchronization deviation, and terrain data error. This mechanism quantifies and estimates the uncertainty sources by establishing an error transfer model, and its total error is defined as: where δθ represents the platform pointing angle error, H is the ranging altitude, θ is the current laser incident angle, δt is the detection time synchronization error, c is the speed of light, and δ DEM represents the uncertainty of the DEM source elevation data; In step (5.4), an online learning module is developed and a multi-source verification visualization system is constructed accordingly. By introducing the incremental random forest algorithm, a continuously optimized cloud identification and classification model is built. With a windowed training strategy, the model parameters are updated every time 1000 new samples are received, enabling the model to have adaptive learning ability. On this basis, a visualization verification interface is developed, integrating the capabilities of dynamic waveform signal plotting, feature heatmap rendering, three-dimensional cloud structure reconstruction, and terrain background fusion. The interface supports the loading of multi-source verification data measured by CALIPSO cloud products, Himawari-8 infrared images, and ground lidar, and displays the differences between the manually annotated results and the model predictions in a figure-data collaborative manner.

Images