ICESat-2 signal photon extraction method, apparatus, and storage medium
By using convolutional neural networks to perform image processing and feature learning on ICESat-2 photon data, the problem of distinguishing between noise photons and signal photons was solved, and more efficient separation of signal photons and noise photons was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2023-11-02
- Publication Date
- 2026-06-05
AI Technical Summary
In existing ICESat-2 signal photon data, it is difficult to effectively distinguish between noise photons and signal photons. Traditional threshold segmentation algorithms fail to make full use of the position and shape features within the photon neighborhood, resulting in poor information compression and separation effects.
A convolutional neural network was used to image the ICESat-2 photon data. The GoogLeNet and CBAM modules were used to extract features in the neighborhood of the photon. The feature difference between signal photons and noise photons was increased through training and learning.
It achieves better separation of signal photons and noise photons, improving the signal-to-noise ratio and separation accuracy of the data.
Smart Images
Figure CN117392405B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing signal extraction, and in particular to a method, apparatus and storage medium for extracting photons from ICESat-2 signals. Background Technology
[0002] The ICESat-2 satellite is a new generation of land surveying satellite launched by NASA. It carries a new generation of Advanced Terrain Measurement Laser Altimeter System (ATLAS), which can calculate the distance between the satellite and the Earth's surface by measuring the propagation time of laser pulses, thereby determining changes in land topography and providing high-precision surface elevation information.
[0003] However, during data collection, ICESat-2 receives not only its own pulse signals but also atmospheric scattering noise, solar radiation noise, and instrument noise. This manifests in the data as a large amount of random noise distributed around the signal photons. These random noises must be removed using various methods before the data can be used for further processing.
[0004] For two-dimensional ICESat-2 photon data, since the local density of signal photons is much greater than that of noise photons, many researchers have proposed some algorithms to remove noise photons. Most of them are threshold segmentation algorithms based on local density or local distance.
[0005] Thresholding algorithms based on local density or local distance mostly utilize the single feature that the local density of signal photons in ICESat-2 photon data is much greater than that of noise photons. They describe all information within the photon neighborhood using a single density value, without taking into account the position and shape features of photons in the neighborhood of signal and noise photons. This leads to significant compression of information within the photon domain and prevents optimal application. In contrast, the position and shape features of photons in the neighborhood, which are abandoned by traditional methods, can be used effectively to distinguish between signal and noise photons. Furthermore, within the same data set, there are cases where all signal photons have significantly different local densities. The histogram of local point densities in such data is not a typical bimodal distribution, but may be trimodal or even multimodal. For such data, applying a thresholding method to all photons cannot effectively separate all signals from noise. A threshold that is too high will miss many signal photons with low local density, while a threshold that is too low may misclassify some noise photons close to the signal as signal photons. Summary of the Invention
[0006] To address the problem that existing methods cannot effectively distinguish between signals and noise, this invention proposes an ICESat-2 signal photon extraction method based on convolutional neural networks (CNNs). This method introduces CNNs to better utilize the position and shape features within the photon domain, achieving effective denoising of photon data under the aforementioned conditions. The method first images each photon point with a specific width and height, recording the feature distribution information within the neighborhood of each point. Then, the image data is fed into the CNN for training and learning. Utilizing the powerful feature learning capabilities of CNNs, it can learn the different features of signal photons with different local densities, as well as the different position and shape features between signal and noise photons. This continuously increases the feature difference between signal and noise photons, ultimately achieving better signal and noise photon separation.
[0007] Specifically, this invention provides a method, apparatus, and storage medium for extracting photons from ICESat-2 signals.
[0008] The method includes the following steps:
[0009] S1: Data Acquisition: Collecting ICESat-2 / ATL03 photon data from different regions;
[0010] S2: Photon data preprocessing: After the photon data undergoes image range determination, feature information statistics and feature matrix image processing, a photon image is obtained;
[0011] S3: Collect the photon images to form a dataset, and divide the dataset into a training set and a validation set;
[0012] S4: Construct a neural network model for photon extraction;
[0013] S5: Train the photon extraction neural network model using the training set to obtain the trained model;
[0014] S6: Validate the trained model using the validation set to obtain the final model;
[0015] S7: Use the final model to complete photon extraction.
[0016] A storage medium storing instructions and data for implementing an ICESat-2 signal photon extraction method.
[0017] An ICESat-2 signal photon extraction device includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement an ICESat-2 signal photon extraction method.
[0018] The beneficial effects provided by this invention are: after converting photon data into images, the feature extraction capability of convolutional neural networks can be used to learn the different features of signal photons with different local densities, as well as the different position and shape features between signal photons and noise photons, thereby continuously increasing the feature difference between signal photons and noise photons, and ultimately achieving a better separation effect between signal photons and noise photons. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0020] Figure 2 This is a schematic diagram of the photon preprocessing process;
[0021] Figure 3 This is a schematic diagram of the network structure of the present invention;
[0022] Figure 4 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0024] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the method flow of the present invention;
[0025] This invention provides a method for extracting photons from ICESat-2 signals, comprising the following steps:
[0026] S1: Data Acquisition: Collecting ICESat-2 / ATL03 photon data from different regions;
[0027] It should be noted that this invention collects ICESat-2 / ATL03 data from different regions and extracts information such as the orbital distance, elevation, and incident angle of each photon point from the original data set file to obtain a point cloud distribution map.
[0028] S2: Photon data preprocessing: After the photon data undergoes image range determination, feature information statistics and feature matrix image processing, a photon image is obtained;
[0029] It should be noted that the present invention mainly goes through three steps: image range determination, feature information statistics, and feature matrix visualization, and finally converts each photon point into an image.
[0030] As one example, please refer to Figure 2 , Figure 2 This is a schematic diagram of the photon preprocessing process;
[0031] To adapt to the input of a convolutional neural network and better utilize the position and shape features of all photons within a photon's neighborhood, this invention performs image-based transformation on the photon data, summarizing the neighborhood information of individual photons into a single image. This method involves the following steps.
[0032] The first step is to determine the imaging range of each photon, that is, to design a suitable grid neighborhood range for imaging all photons.
[0033] Because the actual distribution of signal photon data is a long, narrow strip along the track distance, the elevation fluctuation range is not large. For a 1-kilometer span along the track, the vertical span may be less than 100 meters, and in shallow sea areas, it may be only a few dozen meters. Therefore, in determining the image range, this invention proposes using a rectangular grid neighborhood, where the length of each cell along the track distance is much greater than its width along the elevation. This allows for accurate statistical analysis of the position and shape distribution of photons within each photon neighborhood along the track distance, and also effectively separates signal and noise photons along the elevation, resulting in purer photons within each cell and ultimately greater discriminability between image samples. The image pixel range and number of pixels are determined empirically and can be optimized and adjusted according to specific circumstances.
[0034] The second step is to summarize the feature information of all photons within the gridded neighborhood after determining the gridded neighborhood range, thereby realizing the transformation from discrete points to feature matrix.
[0035] This invention statistically analyzes three types of feature values within the image area: the number of photon points (Num), the difference between the average elevation of the photon points and the elevation of the center point (DMh), and the distance D of the photon points from the center point. Their calculation formulas are as follows:
[0036]
[0037] Where p1, p2, p3, ..., p k It is the number of k photon points contained in a single cell, and Count is the count function; It is the elevation value of the photon point. It is the elevation value of the photon point p0 at the center of the grid; L and H are the distance and elevation along the track at the center of the grid, respectively. and These are the distance and elevation values along the track for the photon point p0 at the center of the grid.
[0038] Then, each value in these three feature matrices is used as a single pixel value, thereby summarizing the feature information within the photon neighborhood into a three-band image.
[0039] S3: Collect the photon images to form a dataset, and divide the dataset into a training set and a validation set;
[0040] S4: Construct a neural network model for photon extraction;
[0041] It should be noted that after the image-based step of the photon data, each individual photon point and its neighborhood information in the original ICESat-2 data can be transformed into an image, which can then be used for the learning and training of convolutional neural networks.
[0042] As one embodiment, the present invention selects GoogLeNet as the backbone network for deep learning.
[0043] GoogLeNet is a deep neural network model based on the Inception module, developed by the GoogLe team. It won the ImageNet competition in 2014. Prior to this, structures such as AlexNet and VGG achieved better training results by increasing the depth (number of layers) of the network. However, increasing the number of layers brings many negative effects, such as overfitting, vanishing gradients, and exploding gradients.
[0044] Inception offers another way to improve training results: by making more efficient use of computing resources and extracting more features with the same amount of computation, thereby improving training results.
[0045] In its implementation, the Inception module uses convolutional kernels of different sizes within a single convolutional layer. This increases the "width" of the network model and enhances its ability to extract details at different scales from the original feature information, resulting in a more comprehensive and deeper feature representation. Therefore, in this invention, since the shape and position features of signal photon images vary across different land cover scenarios and types, GoogLeNet, using the Inception module, can effectively extract features from signal photon images of different land cover types and allow the network to differentiate them, increasing the difference between the image and noisy photon images, thus achieving better classification results.
[0046] In addition, in order to enable the neural network to learn more effective features from the sample images, the present invention incorporates a convolutional attention mechanism (CBAM) module into the network structure.
[0047] The CBAM module primarily focuses on improving network performance through "attention." It helps the network learn which information to emphasize or suppress, thus effectively facilitating information flow within the network. The CBAM module consists of a Channel Attention Module (CAM) and a Spatial Attention Module (SAM). The CAM tells the network which channels in the input features are meaningful. Its implementation process is as follows: First, average pooling and max pooling operations are used to aggregate the spatial information of the input features. Then, these are fed into a multilayer perceptron with shared weights. The outputs are then summed pixel-by-pixel and passed through a sigmoid function to obtain the channel attention feature M. c (F):
[0048] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))
[0049] Where F is the input feature and σ is the sigmoid function.
[0050] Furthermore, SAM serves to inform the network which locations in the input features are more important. Its implementation is as follows: Unlike CAM, SAM first applies average pooling and max pooling operations along the channel dimension, then merges them along the channel dimension. The merged two-dimensional feature map is then passed through a convolutional layer to obtain the spatial attention feature M. s (F):
[0051] M s (F)=σ(f 7×7 ([AvgPool(F);MaxPool(F)]))
[0052] Where F represents the input features, σ is the sigmoid function, and f 7×7 It is a convolutional layer with a kernel size of 7×7.
[0053] The CBAM module combines the two modules mentioned above. The initial input features pass through the CAM module and the SAM module in sequence. During this process, the output features need to be multiplied with the input features pixel by pixel, thus obtaining the final CBAM weighted features.
[0054] As a specific example, please refer to Figure 3 , Figure 3 This is a schematic diagram of the network structure of the present invention;
[0055] This invention uses the main network architecture of GoogLeNet and incorporates a CBAM module. The main network can be divided into six parts. The first and second parts consist of convolutional layers and max pooling layers, with one and two convolutional layers respectively, followed by a CBAM module. The third, fourth, and fifth parts are all composed of Inception modules, with three, four, and five Inception modules respectively. Each Inception module is connected to a CBAM module and then to a max pooling layer. The sixth part is the output layer, consisting of an average pooling layer, a fully connected layer, and a softmax function, which is the final output of the network.
[0056] S5: Train the photon extraction neural network model using the training set to obtain the trained model;
[0057] S6: Validate the trained model using the validation set to obtain the final model;
[0058] S7: Use the final model to complete photon extraction.
[0059] It should be noted that this invention ultimately uses the GoogLeNet network, which includes an attention mechanism, for training and classification. During the experiment, 80% of the sample dataset was used as training samples and 20% as validation samples. All of these samples were used in the network structure for training. After the training converged, the trained network structure was used for classification and denoising of other data.
[0060] As one embodiment, the verification part of the present invention includes two parts. The first part is a simulation experiment comparison: the present invention designs multiple sets of simulation data containing different noise levels, and then uses different methods to denoise the simulation data, and uses the denoising results to compare the denoising capabilities of different methods. The second part is absolute verification: for the signal points extracted by the method of the present invention, their elevation values are extracted, and then compared with the elevation of reference data at the same location to determine whether the extracted signal points are correct.
[0061] Please see Figure 4 , Figure 4 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: an ICESat-2 signal photon extraction device 401, a processor 402, and a storage medium 403.
[0062] An ICESat-2 signal photon extraction device 401: The ICESat-2 signal photon extraction device 401 implements the ICESat-2 signal photon extraction method.
[0063] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the ICESat-2 signal photon extraction method.
[0064] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the ICESat-2 signal photon extraction method.
[0065] The beneficial effects of this invention are: after converting photon data into images, the feature extraction capability of convolutional neural networks can be used to learn the different features of signal photons with different local densities, as well as the different position and shape features between signal photons and noise photons, thereby continuously increasing the feature difference between signal photons and noise photons, and ultimately achieving a better separation effect between signal photons and noise photons.
[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for extracting photons from an ICESat-2 signal, characterized in that: The method includes the following steps: S1: Data Acquisition: Collecting ICESat-2 / ATL03 photon data from different regions; S2: Photon data preprocessing: After the photon data undergoes image range determination, feature information statistics and feature matrix image processing, a photon image is obtained; S3: Collect the photon images to form a dataset, and divide the dataset into a training set and a validation set; S4: Construct a neural network model for photon extraction; S5: Train the photon extraction neural network model using the training set to obtain the trained model; S6: Validate the trained model using the validation set to obtain the final model; S7: Use the final model to complete photon extraction; The specific process of step S2 is as follows: S21. Determine a suitable grid neighborhood range for all photon data for visualization; S22. Summarize the feature information of all photons within the range in each grid to realize the transformation from discrete points to feature matrix; S23. Use the eigenvalues of the feature matrix as individual pixel values to summarize the feature information in the photon neighborhood into one image. The feature information in step S22 includes: the number of photon points. Num The difference between the average elevation of the photon point and the center elevation DMh and the distance of the photon point from the center point D .
2. The ICESat-2 signal photon extraction method as described in claim 1, characterized in that: Number of photons Num The difference between the average elevation of the photon point and the center elevation DMh and the distance of the photon point from the center point D The calculation formula is as follows: ： in, It is the numbering of the k photon points contained within a single cell. Count This is a count function; It is the elevation value of the photon point. It is the photon point at the center of the grid. Elevation value; L and H These represent the distance along the track from the center of the grid and the elevation value, respectively. and These are the photon points at the center of the grid. The distance and elevation along the track.
3. The ICESat-2 signal photon extraction method as described in claim 2, characterized in that: The photon extraction neural network model employs a GoogLeNet network with an added attention mechanism.
4. The ICESat-2 signal photon extraction method as described in claim 3, characterized in that: The structure of the photon extraction neural network model comprises six parts. The first and second parts are mainly composed of convolutional layers and max-pooling layers, with a CBAM module added after the convolutional layers. The third, fourth, and fifth parts are all mainly composed of Inception modules, each of which is connected to a CBAM module. Finally, a max-pooling layer is connected; the sixth part is the output layer.
5. A storage medium, characterized in that: The storage medium stores instructions and data to implement the ICESat-2 signal photon extraction method according to any one of claims 1 to 4.
6. An ICESat-2 signal photon extraction device, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the ICESat-2 signal photon extraction method according to any one of claims 1 to 4.