Atmospheric laser radar data denoising method and system based on self-supervised learning
By employing a self-supervised learning blind spot network method, combined with logarithmic transformation and vertical smoothing constraints, the problem of removing structural noise in lidar signals was solved. This achieved efficient denoising and atmospheric structure preservation even without real, noise-free data, thereby improving lidar signal quality and model generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-03-13
- Publication Date
- 2026-07-24
AI Technical Summary
Existing lidar denoising methods struggle to effectively remove structural noise from atmospheric lidar signals while preserving detailed features of cloud boundaries and aerosol layers. In particular, traditional methods lack generalization ability when no real, noise-free reference data is available.
A self-supervised learning-based approach is adopted. The dynamic range of the data is compressed by logarithmic transformation, and a vertical smoothing prior constraint and transfer learning strategy are introduced to construct a blind spot network for denoising. The blind spot network is used to learn the vertical continuity and neighborhood features of the signal to suppress stripe noise.
In the absence of reference data, it significantly improves the noise suppression effect of lidar signals, preserves the detailed features of atmospheric structure, and enhances the generalization ability and denoising accuracy of the model, making it suitable for different observation scenarios and instrument conditions.
Smart Images
Figure CN121831736B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of lidar signal processing, and in particular to a method and system for denoising atmospheric lidar data based on self-supervised learning. Background Technology
[0002] Atmospheric lidar, by emitting laser pulses and receiving atmospheric backscattered signals, can invert the vertical structure of atmospheric components such as clouds and aerosols. However, in actual observations, lidar signals are nonlinear, non-stationary, and weak signals, easily affected by stray light noise, dark current noise, and thermal noise, exhibiting continuous amplitude but random phase noise characteristics. This leads to a significant deterioration in the quality of the original echo signal, and some data may even be unusable for scientific analysis. Therefore, effective denoising of lidar data is usually required before performing atmospheric inversion or layer structure identification. Existing lidar denoising methods mainly include: (1) Filtering methods based on local statistical characteristics: such as moving average, mean filtering, Gaussian filtering, etc., suppress noise by averaging the neighborhood. The method is simple and efficient, but it will cause blurring and boundary shift in details such as cloud boundaries and aerosol layer tops.
[0003] (2) Transform domain methods: For example, wavelet denoising uses the multi-resolution characteristics of wavelets to separate signals and noise through thresholding, but it relies on fixed basis functions and is difficult to adapt to the non-stationary characteristics of lidar signals. Fourier filtering is based on separating signals and noise in the frequency domain, but in actual lidar, the spectrum of signals and noise often overlaps, making this assumption difficult to hold.
[0004] (3) Empirical Mode Decomposition (EMD / VMD) methods: Although they have adaptive capabilities, they often suffer from mode aliasing. An intrinsic mode may contain multiple time-scale components, lacks clear physical meaning, and the denoising effect depends on the selection of empirical parameters.
[0005] In addition, some studies have attempted to combine multiple methods for denoising (such as wavelet + VMD, morphological filtering + EMD, etc.) in order to achieve complementary effects. However, such methods lack solid mathematical theoretical support, have poor generalization ability, and have limited adaptability to different noise conditions and instrument types.
[0006] With the development of deep learning, data-driven denoising methods have received widespread attention. However, most deep learning denoising algorithms rely on supervised learning, requiring a large amount of paired "noisy-clean" data for training. However, in actual LiDAR observations, it is impossible to obtain truly noise-free reference data, making supervised methods difficult to apply. Furthermore, supervised models are highly dependent on the distribution of training data; when observation conditions, instrument type, or noise levels change, the model's generalization ability significantly decreases, necessitating retraining.
[0007] Therefore, self-supervised denoising methods that do not require clean data have potential advantages in the LiDAR field. (Refer to...) Figure 1 Noise2Void (N2V), with its blind spot learning mechanism, has achieved significant denoising effects in multiple fields (such as medical imaging and image restoration). However, the Noise2Void method is based on the fundamental assumption of "noise independence," and the blind spot network forcibly excludes the information of the blind spot pixel itself when predicting the blind spot pixel, relying only on the statistical features of the neighborhood to complete the estimation. When the noise is random and independently distributed, this mechanism can effectively suppress noise; however, some structural noise appears in LiDAR data. Due to its significant correlation and structural characteristics in the time dimension, this noise no longer meets the N2V assumption. As a result, the original N2V network struggles to separate structural noise from the real signal, and may produce artifacts in gradient-sensitive areas such as cloud boundaries. Therefore, existing N2V methods still struggle to simultaneously achieve noise suppression and preserve atmospheric structure (cloud / aerosol) details in LiDAR data.
[0008] Furthermore, lidar noise exhibits continuous variations over time, which is the fundamental reason why existing N2V methods cannot handle it. This manifests as a distinct structural feature in the lidar's two-dimensional "time-height" color map, a characteristic that traditional one-dimensional methods based on a single profile struggle to identify. In actual observations, in addition to random noise such as photon shot noise and dark current noise, electronic systems (including detector preamplifiers, ADCs / acquisition cards, etc.) may produce baseline shifts, gain inconsistencies, or quantization errors. These errors are strongly correlated in the time dimension and typically appear as continuous or quasi-periodic horizontal stripe noise in the daily two-dimensional color map. This type of stripe noise varies very little in the vertical direction but maintains a consistent shift in the time direction, representing typical directional structural noise.
[0009] Based on the above problems, there is an urgent need for a self-supervised denoising method that can combine the physical characteristics of atmospheric lidar signals and is applicable to both noise removal and preservation of atmospheric structure details. Summary of the Invention
[0010] Based on this, it is necessary to provide a method and system for denoising atmospheric lidar data based on self-supervised learning to address the above-mentioned technical problems. By using logarithmic transformation, the dynamic range of the data can be effectively compressed, making the distribution of signals of different intensity levels more balanced in the feature space. Then, a vertical smoothing prior constraint term is introduced to force the network output to maintain reasonable continuity in the height dimension, avoiding artifacts such as edge distortion that may be caused by blind spot networks, and enhancing the model's ability to preserve the vertical gradient structure. Finally, a transfer learning strategy is introduced to achieve significant suppression of stripe noise.
[0011] Firstly, this application provides a method for denoising atmospheric lidar data based on self-supervised learning, including: The raw lidar echo data is acquired and preprocessed. In the preprocessed raw lidar echo data, some pixels are randomly masked as blind spots to generate random blind spot sample data. The preprocessing includes performing a logarithmic transformation on the raw lidar echo data. A blind spot network is constructed based on the Unet architecture. Sample data is input into the blind spot network to generate pixel prediction values. A mean squared error loss function is constructed based on the pixel prediction values and the actual pixel values. A vertical smoothing prior constraint term is introduced to generate the total loss function. Train the blind spot network using the total loss function until the total loss function is minimized, and generate the target blind spot network. The lidar echo data to be processed is input into the target blind spot network, and the target blind spot network is used to perform noise reduction processing on the lidar echo data to be processed.
[0012] In one embodiment, the preprocessing of raw lidar echo data includes: After removing background light from the original lidar echo data, range square correction is performed. The lidar echo data after range square correction is then logarithmically transformed and linearly normalized.
[0013] In one embodiment, the preprocessed raw lidar echo data is atmospheric lidar height-time two-dimensional echo intensity data. The row dimension of this data is the observation time series, and the column dimension is the detection height series. Each pixel corresponds to the lidar echo intensity value at a certain observation time and a certain detection height.
[0014] In one embodiment, randomly masking a portion of pixels in the preprocessed raw LiDAR echo data as blind spots to generate random blind spot sample data includes: The preprocessed raw lidar echo data is segmented to generate raw local block data of multiple target sizes; A set number of non-overlapping pixels are randomly selected within the original local block as blind spots. The selected blind spot pixels are then masked to generate local block data containing blind spots. Using local block data containing blind spots as input samples, and the actual pixel values of the blind spot locations in the corresponding original local block data as supervision labels, random blind spot sample data is constructed.
[0015] In one embodiment, the formula for the total loss function is:
[0016] in, This is the network's predicted value; The actual value; mask iThe mask is marked with 0 or 1. 1 indicates that the i-th position is the area where the loss needs to be calculated, i.e., the blind spot position, and 0 indicates that the area where the loss does not need to be calculated, i.e., the non-blind spot position. and These are the predicted values for adjacent height pairs at the same observation time; The parameter is to be adjusted, and its value range is [0,1].
[0017] In one embodiment, the process of training the blind spot network further includes: introducing a transfer learning strategy, using unstructured noise LiDAR data to pre-train the blind spot network, so that the network can fully learn the physical continuity of the signal in the vertical direction, the layered structure distribution and typical gradient features, and then fine-tuning it on data with strong stripe noise.
[0018] Secondly, this application also provides an atmospheric lidar data denoising system based on self-supervised learning, comprising: The training sample construction module acquires and preprocesses the original LiDAR echo data. After preprocessing, it randomly masks some pixels in the original LiDAR echo data as blind spots to generate random blind spot sample data. The preprocessing includes performing a logarithmic transformation on the original LiDAR echo data. The first model building module constructs a blind spot network based on the Unet architecture, inputs sample data into the blind spot network to generate pixel prediction values, constructs a mean squared error loss function based on the pixel prediction values and the pixel true values, and introduces a vertical smoothing prior constraint term to generate the total loss function. The second model building module trains the blind spot network using the total loss function until the total loss function is minimized, thereby generating the target blind spot network. The noise reduction module is used to input the LiDAR echo data to be processed into the target blind spot network, and then use the target blind spot network to perform noise reduction processing on the LiDAR echo data to be processed.
[0019] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps: The raw lidar echo data is acquired and preprocessed. Then, some pixels are randomly masked in the preprocessed raw lidar echo data to generate random blind spot sample data. A blind spot network is constructed based on the Unet architecture. The blind spot network is trained using random blind spot sample data to generate the target blind spot network. The lidar echo data to be processed is input into the target blind spot network, and the target blind spot network is used to perform noise reduction processing on the lidar echo data to be processed.
[0020] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps: The raw lidar echo data is acquired and preprocessed. Then, some pixels are randomly masked in the preprocessed raw lidar echo data to generate random blind spot sample data. A blind spot network is constructed based on the Unet architecture. The blind spot network is trained using random blind spot sample data to generate the target blind spot network. The lidar echo data to be processed is input into the target blind spot network, and the target blind spot network is used to perform noise reduction processing on the lidar echo data to be processed.
[0021] This application employs the aforementioned method and system for denoising atmospheric lidar data based on self-supervised learning, which has the following beneficial effects: 1. Considering that atmospheric echo signals have clear physical continuity in the vertical direction, a very large dynamic range, and noise consisting of "random + structured stripe noise," a logarithmic transformation is used to effectively compress the dynamic range of the data, making the distribution of signals of different intensity levels more balanced in the feature space. Then, a vertical smoothing prior constraint is introduced to force the network output to maintain reasonable continuity in the height dimension, avoiding artifacts such as edge distortion that may be caused by blind spot networks, and enhancing the model's ability to preserve the vertical gradient structure. Finally, a transfer learning strategy is introduced to achieve significant suppression of stripe noise.
[0022] 2. A self-supervised learning framework based on the blind spot mechanism is adopted. Random blind spot samples are constructed from the original lidar echo data. The random blind spot samples are used to train a blind spot network structure with neighborhood prediction capability, so that the model can retain detailed features in the atmospheric profile such as cloud boundaries and aerosol gradients, and achieve a balance between noise suppression and signal fidelity.
[0023] 3. A self-supervised learning framework based on the blind spot mechanism is adopted, which enables the network to complete model training without reference data. This mechanism gets rid of the dependence of traditional deep learning denoising methods on "noisy-noise-free" paired training samples, solves the fundamental problem of not being able to obtain real noise-free signals in lidar observation, and significantly improves the usability of the algorithm in actual business.
[0024] 4. A blind spot network structure with neighborhood prediction capability was designed. Through signal preprocessing and reasonable network receptive field settings, the model can effectively suppress random components such as background light noise, dark current noise, and photon statistical noise. At the same time, because the network training relies on the self-learning mechanism of the neighborhood structure, it can fully preserve the detailed features in the atmospheric profile such as cloud boundaries and aerosol gradients, achieving a balance between noise suppression and signal fidelity, which is significantly better than traditional denoising methods.
[0025] 5. A transfer learning strategy is introduced, using unstructured noise-free LiDAR data to pre-train the blind spot network. This allows the network to fully learn the physical continuity of the signal in the vertical direction, its layered structure distribution, and typical gradient features, thus forming a "vertical structure priority" representation capability. Fine-tuning is then performed on data with strong stripe noise. Because the network has learned vertical structure priority, it can automatically ignore horizontal noise inconsistent with the vertical structure during the transfer phase, achieving significant suppression of stripe noise.
[0026] 6. The network is pre-trained using biased random noise lidar data to learn the prior vertical structure of real atmospheric echoes. Horizontal fringe noise exhibits an approximately constant offset in the horizontal direction, with its gradient remaining almost unchanged over time, showing significant statistical orthogonality to the vertical gradient structure of the real atmospheric echo signal. In other words, the real signal contains obvious vertical variation patterns, while fringe noise lacks structural variation along the height direction. During the transfer learning phase, the network automatically identifies fringes as anomalous structures and suppresses them. This mechanism achieves cross-noise type transfer learning from point-like random noise to fringe noise suppression, significantly contributing to the enhancement of horizontal fringe noise suppression capabilities in atmospheric lidar data—a feature not found in traditional N2V methods.
[0027] 7. The model in this application does not rely on specific noise distributions or instrument characteristics. Through the local structure learning mechanism of the blind spot network and vertical prior constraints, it can maintain consistent denoising performance under conditions of day-night variations, multi-layered cloud structures, weak high-altitude signals, and varying instrument noise. Experiments show that the method maintains stable performance across different observation scenarios and data sources, demonstrating strong generalization ability.
[0028] 8. The method described in this application does not require manual setting of complex thresholds, has weak parameter dependence, and is simple to implement. It can be directly embedded into the data preprocessing links of ground-based, airborne, and spaceborne lidar. This method can significantly improve the inversion accuracy of atmospheric structure and is applicable to meteorological detection, environmental monitoring, atmospheric physics research, and other fields, possessing significant engineering application value and scientific significance. Attached Figure Description
[0029] Figure 1 This is a defect image after traditional N2V denoising in one embodiment; wherein, Figure 1 (a) in the diagram is a schematic diagram of actual lidar signal data; Figure 1 (b) in the figure represents the result of traditional N2V denoising; Figure 2 This is a schematic diagram of structured noise data from an atmospheric lidar in one embodiment; Figure 3 A concept diagram of a blind spot provided in one embodiment; Figure 4 Here is a diagram of the blind spot network structure in one embodiment; Figure 5 This is a schematic diagram illustrating the noise reduction effect analysis in one embodiment; wherein, Figure 5 (a) in the diagram is a schematic diagram of actual lidar signal data; Figure 5 (b) in the figure represents the denoising result of the method in this application; Figure 6 This is a schematic diagram of a robustness test provided in one embodiment; wherein, Figure 6 (a) in the diagram is a schematic diagram of actual lidar signal data; Figure 6 (b) in the figure represents the denoising result of the method in this application; Figure 7 This is a schematic diagram of single-profile denoising analysis provided in one embodiment; Figure 8 This is a schematic diagram comparing the backscattering intensity after denoising using different lidar denoising algorithms in one embodiment; Figure 9 This is a comparison chart of the denoising results of the conventional N2V method and the denoising results of the method of this application in one embodiment; wherein, Figure 9 (a) in the figure represents the denoising result of the traditional N2V method; Figure 9 (b) in the figure represents the denoising result of the method in this application; Figure 10 This is a schematic diagram of the denoising result of the method of this application in one embodiment; wherein, Figure 10 (a) in the diagram is a schematic diagram of actual lidar signal data; Figure 10 (b) in the figure represents the denoising result of the method in this application. Detailed Implementation
[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0031] Firstly, this application provides a method for denoising atmospheric lidar data based on self-supervised learning, including: S100: Acquire the original lidar echo data and preprocess it. Randomly mask some pixels in the preprocessed original lidar echo data as blind spots to generate random blind spot sample data. The preprocessing includes performing a logarithmic transformation on the original lidar echo data.
[0032] In one embodiment, the preprocessing of the raw lidar echo data includes: removing background light from the raw lidar echo data and performing range square correction; then performing logarithmic transformation on the range square-corrected lidar echo data and applying linear normalization.
[0033] Reference Figure 2 This application uses a time-height-intensity map from a single-channel lidar as the network input. In this two-dimensional image, the horizontal axis represents the time series, the vertical axis represents the detection height, and the pixel value corresponds to the backscattering intensity information at the corresponding spatiotemporal location. However, given the inherently large dynamic range of lidar data, directly applying linear normalization leads to two problems: first, weak signals are over-compressed, resulting in the loss of detailed information; second, the gradient distribution between strong and weak signals is severely unbalanced, which is detrimental to network training. Therefore, this paper performs a logarithmic transformation (log10) on the original data before normalization to effectively compress the dynamic range of the data, making the distribution of signals of different intensity levels more balanced in the feature space.
[0034] The formula for the logarithmic transformation mentioned above is:
[0035] in, This represents the original backscattering intensity. The minimum value in the data, which is also the basis for preventing numerical underflow.
[0036] Considering the physical magnitude characteristics of lidar data, this paper sets the data to 10, which can effectively avoid the numerical instability of log(0) and will not significantly affect the dynamic range of the effective signal.
[0037] In one embodiment, randomly masking a portion of pixels in the preprocessed raw LiDAR echo data as blind spots to generate random blind spot sample data includes: segmenting the preprocessed raw LiDAR echo data to generate multiple target-sized raw local block data; randomly selecting a set number of non-overlapping pixels within the raw local blocks as blind spots, masking the selected blind spot pixels to generate local block data containing blind spots; using the local block data containing blind spots as input samples, and using the actual pixel values at the blind spot positions in the corresponding raw local block data as supervision labels to construct random sample point data.
[0038] For example, if a set of original lidar echo data is 1024*1024 in size, it can be divided into 32*32 segments during the segmentation process.
[0039] In one embodiment, refer to Figure 2 After preprocessing, the original lidar echo data is atmospheric lidar height-time two-dimensional echo intensity data. The row dimension of this data is the observation time series, and the column dimension is the detection height series. Each pixel corresponds to the lidar echo intensity value at a certain observation time and a certain detection height.
[0040] During atmospheric lidar detection, the signal is prone to exhibiting both random noise and structured horizontal stripe noise due to factors such as photon shot noise, dark current noise, background light interference, and detector gain drift. This noise is particularly pronounced in weak signal regions at high altitudes, severely impacting the accuracy and resolvability of the backscattered signal. Traditional filtering methods, relying on fixed kernel functions or manually defined mathematical models, struggle to balance noise suppression and boundary preservation, and are ill-suited to the non-stationary characteristics and multi-source noise structure of lidar signals, leading to the loss of weak layer information or layer top blurring.
[0041] To overcome the above problems, this invention provides a self-supervised learning framework based on blind spot networks to achieve denoising of lidar signals under no-reference conditions. For example... Figure 3 As shown, unlike traditional networks, blind spot networks have a masked receptive field that excludes the central pixel and can learn to suppress noise by observing surrounding pixels. The core mechanism of blind spot learning is to randomly mask some pixels in the input data as "blind spots." When predicting these pixels, the network cannot use their true values but can only estimate them using the neighborhood structure. By minimizing the blind spot prediction error, the model can automatically learn the spatial correlation of LiDAR signals and suppress noise that does not conform to the statistical characteristics of the neighborhood.
[0042] S200 constructs a blind spot network based on the Unet architecture. Sample data is input into the blind spot network to generate pixel prediction values. A mean squared error loss function is constructed based on the pixel prediction values and the true pixel values. A vertical smoothing prior constraint term is introduced to generate the total loss function.
[0043] S300 uses the total loss function to train the blind spot network until the total loss function is minimized, thus generating the target blind spot network.
[0044] like Figure 4 This application uses the Unet architecture as the blind spot network. The network first processes the input through two 3×3 convolutional layers (64 channels), followed by downsampling using 2×2 max pooling. The encoder contains four resolution levels with channel numbers of 64, 128, 256, and 512, respectively, each containing two 3×3 convolutional layers and a ReLU activation function. The bottom of the network contains a 512-channel convolutional layer for extracting the feature representation of the deepest layer. The decoder progressively restores the spatial resolution through 2×2 upconvolutions with channel numbers of 256, 128, and 64, respectively. In each upsampling step, the feature maps of the corresponding layers of the encoder and decoder are concatenated through skip connections to preserve multi-scale spatial information. The concatenated feature maps are then fused using two 3×3 convolutional layers.
[0045] In one embodiment, training a blind spot network using random blind spot samples includes: inputting local block data containing blind spots into the blind spot network; using the blind spot network to predict blind spot pixels based on the local block data containing blind spots to generate pixel prediction values; constructing a mean squared error loss function for the mask based on the pixel prediction values and the actual pixel values; and training the blind spot network using the mean squared error loss function until the loss function is minimized.
[0046] The formula for the mean squared error loss function is:
[0047] in, This is the network's predicted value; The actual value; mask i The mask is marked with 0 or 1. 1 indicates that the i-th position is the area where the loss needs to be calculated, i.e., the blind spot position, and 0 indicates that the area where the loss does not need to be calculated, i.e., the non-blind spot position.
[0048] In one embodiment, a blind spot network is used to predict blind spot pixels based on local block data containing blind spots. This includes: the blind spot network takes pixels other than the blind spot to be predicted within the local block as input features, learns local statistical properties, and then achieves accurate prediction of the blind spot pixel value.
[0049] In one embodiment, training the blind spot network using random blind spot samples further includes: introducing a vertical smoothing prior constraint term on the basis of the mean squared error loss function, improving the loss function to a total loss function with the vertical smoothing prior constraint, and training the blind spot network using the total loss function until the total loss function is minimized.
[0050] The formula for the total loss function is as follows:
[0051] in, and These are the predicted values for adjacent height pairs at the same observation time; The parameter is to be adjusted, and its value range is [0,1].
[0052] The vertical smoothing prior constraint forces the network output to maintain reasonable continuity in the height dimension, avoiding artifacts such as edge distortion that may be caused by blind spot networks, and enhancing the model's ability to preserve the vertical gradient structure.
[0053] Atmospheric echo signals exhibit clear physical continuity in the vertical direction: the vertical gradients of molecular scattering, aerosols, and clouds typically change smoothly with altitude, with only a limited number of abrupt change points at the cloud base, cloud top, or strong aerosol layer boundaries. Furthermore, lidar structural noise exhibits significant horizontal uniformity, with its gradient showing weak variation in the altitude direction and periodic repetition in the temporal direction, classifying it as typical directional structural noise. In contrast, real atmospheric structures (such as cloud boundaries and aerosol layers) have complex gradient distributions in the vertical direction, structurally exhibiting a statistically orthogonal relationship with stripe noise.
[0054] This application introduces a transfer learning strategy: First, the blind spot network is pre-trained using LiDAR data without structured noise, allowing the network to fully learn the physical continuity, layered structure distribution, and typical gradient features of the signal in the vertical direction, thus forming a "vertical structure priority" representation capability. Subsequently, fine-tuning is performed on data with strong stripe noise. Because the network has learned the vertical structure priority, it can automatically ignore horizontal noise inconsistent with the vertical structure during the transfer phase, achieving significant suppression of stripe noise.
[0055] By combining the aforementioned blind spot mechanism, loss function optimization, vertical smoothing prior, and transfer learning, a LiDAR denoising method that does not require ground truth samples is achieved. This method can simultaneously suppress random noise and structural noise, significantly improving the quality of LiDAR signals.
[0056] S400 inputs the LiDAR echo data to be processed into the target blind spot network, and uses the target blind spot network to perform noise reduction processing on the LiDAR echo data to be processed.
[0057] To comprehensively evaluate the performance of denoising algorithms, this paper uses signal-to-noise ratio (SNR) and root-mean-square error (RMSE). SNR is the power ratio of the original signal to the noise, measuring the noise content in the signal; a higher SNR value indicates a smaller noise proportion and better denoising performance. RMSE is the square root of the variance of the signal before and after denoising, reflecting the high-frequency details of the signal; a lower RMSE value indicates a closer resemblance between the denoised and original signals. Combining these two metrics provides a more comprehensive assessment of the denoising effect.
[0058]
[0059]
[0060] in, The original signal, This is the signal before noise reduction.
[0061] In actual experimental tests, Figure 5(a) in the table represents the actual collected data. It can be seen that the raw data in the high-altitude region (15-19km) is filled with a large number of random noise points above 15km. For example... Figure 5 As shown in (b), after denoising using the method of this application, high-altitude noise is significantly suppressed, and signal quality is significantly improved. Two distinct cloud structures exist in the mid-level cloud region (3-12km) between 3:00-6:00 and 18:00-21:00. The denoising results demonstrate that the method of this application effectively preserves the vertical structure and horizontal continuity of the clouds, with clear distinction between the cloud base and cloud top boundaries.
[0062] Figure 6 Robustness testing was conducted using real data. A noise-contaminated signal was constructed by artificially adding additive white Gaussian noise for robustness testing. Figure 6 (b) shows the structure after denoising using the method of this application. It can be seen that despite the extremely severe noise contamination of the input signal, the algorithm still exhibits remarkable denoising capabilities. Meanwhile, Figure 7 This involves selecting a single profile at midday, where the red line represents the denoised data and the blue line represents the undenoised data. Visually, the denoising effect is excellent; the first 100 elevation points, considered as the true effective signal structure, have an RMSE of 0.3038; the last 200 elevation points are almost entirely noise. A signal-to-noise ratio (SNR) test yielded a SNR of 18.72 dB.
[0063] Figure 8 In the comparative analysis of the proposed method with other denoising algorithms, it can be seen that traditional algorithms such as Gaussian filtering and wavelet denoising reduce noise to some extent, but still have high RMSE and limited signal-to-noise ratio improvement. In contrast, the denoising method of this application achieves the best denoising effect, with an RMSE of 0.36 and an SNR of 19.60 dB. The results show that the denoising method of this application can effectively suppress random noise while maintaining the continuity of atmospheric structure, and has a significant advantage in denoising low signal-to-noise ratio radar signals.
[0064] Figure 9 In the image (a), the denoising result is that of the traditional N2V algorithm, where the cloud boundary marked by the red circle shows significant distortion; while Figure 9 (b) shows the denoising result of the proposed denoising method. It can be seen that the cloud boundary structure is better preserved in this region, indicating that the improved method has advantages in suppressing boundary distortion.
[0065] Figure 10 The effectiveness of the proposed method in removing structural noise was tested. Figure 10 (b) shows the denoised signal, which clearly identifies the strong echo structure within the near-Earth boundary layer and the cloud-aerosol layer that evolves over time in the middle and lower layers, compared to the input signal. Figure 10In (a), the overall distribution and the spatiotemporal structure of the main cloud and aerosol layers are highly consistent, while the structural noise in the upper atmosphere and background region is reduced, indicating that our method can further suppress the structural noise that is difficult to remove by N2V to a certain extent.
[0066] Experimental results on real lidar measurement data show that the proposed method can effectively improve signal quality while maintaining atmospheric structure integrity: the root mean square error (RMSE) is 0.33, and the signal-to-noise ratio (SNR) can reach 19.60 dB. Compared with methods such as median filtering, wavelet denoising, empirical mode decomposition, and variational mode decomposition, the proposed method maintains stability and significantly outperforms traditional methods even in strong noise environments.
[0067] This invention boasts high real-time performance. The experiments described above utilized only a Ryzen 7 H255 processor for network training and fine-tuning, with input data consisting of automatically cropped 32×32 resolution 2D LiDAR image patches. During training, the Adam optimizer was used to minimize the loss function, with a learning rate of 0.0004 and a batch size of 64. The training process lasted for 30 epochs. Through this process, the network effectively learns key features from the LiDAR signal. Regarding computational efficiency, the algorithm's performance is affected by network structure complexity, blind spot mask generation method, and input data size. Training on the Ryzen 7 H255 processor took approximately 10 minutes for 30 epochs, demonstrating strong computational efficiency and real-time performance in practical applications. Once training is complete, the model can be directly applied to denoising the same batch of data without retraining, significantly improving processing speed and real-time performance. This characteristic gives this invention a significant advantage in complex LiDAR data processing, meeting the demands of real-time applications.
[0068] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0069] Based on the same inventive concept, this application also provides an atmospheric lidar data denoising system based on self-supervised learning. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the atmospheric lidar data denoising system based on self-supervised learning provided below can be found in the limitations of the atmospheric lidar data denoising method based on self-supervised learning described above, and will not be repeated here.
[0070] In one embodiment, an atmospheric lidar data denoising system based on self-supervised learning is provided, comprising: The training sample construction module acquires and preprocesses the original LiDAR echo data. After preprocessing, it randomly masks some pixels in the original LiDAR echo data as blind spots to generate random blind spot sample data. The preprocessing includes performing a logarithmic transformation on the original LiDAR echo data. The first model building module constructs a blind spot network based on the Unet architecture, inputs sample data into the blind spot network to generate pixel prediction values, constructs a mean squared error loss function based on the pixel prediction values and the pixel true values, and introduces a vertical smoothing prior constraint term to generate the total loss function. The second model building module trains the blind spot network using the total loss function until the total loss function is minimized, thereby generating the target blind spot network. The noise reduction module is used to input the LiDAR echo data to be processed into the target blind spot network, and then use the target blind spot network to perform noise reduction processing on the LiDAR echo data to be processed.
[0071] The modules in the aforementioned atmospheric lidar data denoising system based on self-supervised learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.
[0072] In one embodiment, a computer device is provided, which may be a server. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores relevant data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a self-supervised learning-based atmospheric lidar data denoising method.
[0073] In one embodiment, a computer device is provided, which may be a terminal. The computer device includes a processor, memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a self-supervised learning-based atmospheric lidar data denoising method.
[0074] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0075] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.
[0076] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.
[0077] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.
[0078] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0079] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0080] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A method for denoising atmospheric lidar data based on self-supervised learning, characterized in that, include: The raw lidar echo data is acquired and preprocessed. In the preprocessed raw lidar echo data, some pixels are randomly masked as blind spots to generate random blind spot sample data. The preprocessing includes performing a logarithmic transformation on the raw lidar echo data. A blind spot network is constructed based on the Unet architecture. Sample data is input into the blind spot network to generate pixel prediction values. A mean squared error loss function is constructed based on the pixel prediction values and the actual pixel values. A vertical smoothing prior constraint term is introduced to generate the total loss function. Train the blind spot network using the total loss function until the total loss function is minimized, and generate the target blind spot network. The lidar echo data to be processed is input into the target blind spot network, and the target blind spot network is used to perform noise reduction processing on the lidar echo data to be processed.
2. The method according to claim 1, characterized in that, The preprocessing of raw lidar echo data includes: After removing the background light from the original lidar echo data, range square correction is performed. The lidar echo data after range square correction is then logarithmically transformed and linearly normalized.
3. The method according to claim 1, characterized in that, The preprocessed raw lidar echo data is atmospheric lidar height-time two-dimensional echo intensity data. The row dimension of this data is the observation time series, and the column dimension is the detection height series. Each pixel corresponds to the lidar echo intensity value at a certain observation time and a certain detection height.
4. The method according to any one of claims 1 to 3, characterized in that, Randomly masking a portion of pixels in the preprocessed raw LiDAR echo data to generate random blind spot sample data includes: The preprocessed raw lidar echo data is segmented to generate raw local block data of multiple target sizes; A set number of non-overlapping pixels are randomly selected within the original local block as blind spots. The selected blind spot pixels are then masked to generate local block data containing blind spots. Using local block data containing blind spots as input samples, and the actual pixel values of the blind spot locations in the corresponding original local block data as supervision labels, random blind spot sample data is constructed.
5. The method according to claim 1, characterized in that, The formula for the total loss function is: in, This is the network's predicted value; The actual value; mask i The mask is marked with 0 or 1. 1 indicates that the i-th position is the area where the loss needs to be calculated, i.e., the blind spot position, and 0 indicates that the area where the loss does not need to be calculated, i.e., the non-blind spot position. and These are the predicted values for adjacent height pairs at the same observation time; The parameter is to be adjusted, and its value range is [0,1].
6. The method according to claim 1, characterized in that, The process of training the blind spot network also includes: introducing a transfer learning strategy, using unstructured noise LiDAR data to pre-train the blind spot network, so that the network can fully learn the physical continuity of the signal in the vertical direction, the layered structure distribution and typical gradient features, and then fine-tuning on data with strong stripe noise.
7. A self-supervised learning-based atmospheric lidar data denoising system, characterized in that, include: The training sample construction module acquires and preprocesses the original LiDAR echo data. After preprocessing, it randomly masks some pixels in the original LiDAR echo data as blind spots to generate random blind spot sample data. The preprocessing includes performing a logarithmic transformation on the original LiDAR echo data. The first model building module constructs a blind spot network based on the Unet architecture, inputs sample data into the blind spot network to generate pixel prediction values, constructs a mean squared error loss function based on the pixel prediction values and the pixel true values, and introduces a vertical smoothing prior constraint term to generate the total loss function. The second model building module trains the blind spot network using the total loss function until the total loss function is minimized, thereby generating the target blind spot network. The noise reduction module is used to input the LiDAR echo data to be processed into the target blind spot network, and then use the target blind spot network to perform noise reduction processing on the LiDAR echo data to be processed.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.