A neural-radiance-field-based cloud-top height estimation method

Abstract

The application discloses a cloud top height estimation method based on a neural radiation field, integrates multi-angle and multi-spectral observations of multiple stationary meteorological satellites into a neural radiation field framework, continuously models cloud structure and radiation characteristics in three-dimensional space, overcomes the limitation that a cloud top is simplified as a thin layer target in a traditional double-satellite parallax method, and can depict a real three-dimensional cloud field. Through joint modeling of satellite positions, ground positions and cloud density fields in a unified three-dimensional coordinate system, and natural consideration of cloud distribution along a line-of-sight direction in a volume rendering process of the neural radiation field, the application significantly reduces projection errors under single-satellite observation and cloud height deviation caused by parallax measurement uncertainty in the double-satellite parallax method.

Description

Technical Field

[0001] This invention relates to the field of cloud top height estimation technology in meteorology, and in particular to a cloud top height estimation method based on Neural Radiance Fields (NeRF). Background Technology

[0002] Clouds are one of the most important components of the atmosphere, and their spatial distribution and vertical structure directly affect the energy balance, convection development, and precipitation formation processes of the land-atmosphere system. Cloud top height is an important physical quantity for characterizing the vertical structure of clouds and distinguishing cloud types. It is also a key input parameter in severe convection identification, heavy rainfall monitoring, aviation meteorological services, numerical weather prediction, and climate models. Improving the accuracy of cloud top height retrieval is of great significance for enhancing the ability to forecast severe weather, improving climate change assessment, and ensuring the safety of transportation and the operation of major engineering projects.

[0003] Existing methods for obtaining cloud top height mainly fall into two categories: ground-based observation and satellite remote sensing. Ground-based observation typically relies on active detection equipment such as lidar, ceilometers, and millimeter-wave cloud radar, which can acquire high-precision cloud top height data for single stations. However, its spatial coverage is limited, and it is mostly used for local monitoring and fixed-point verification. Satellite remote sensing, on the other hand, uses passive detection instruments mounted on polar-orbiting or geostationary meteorological satellites to achieve large-scale cloud top height inversion using visible light, infrared, or microwave observations. Among these, geostationary meteorological satellites have the characteristics of high temporal resolution and wide coverage, enabling continuous observation of large-scale cloud systems, and are the main data source for operational cloud top height products.

[0004] Currently, cloud top height retrieval methods based on geostationary meteorological satellites mainly fall into two categories: one is a physical retrieval method based on infrared brightness temperature and atmospheric temperature profiles. This method estimates cloud top temperature through infrared window channels or carbon dioxide absorption channels, and then calculates cloud top height based on the corresponding atmospheric temperature profile. This type of method has clear physical meaning and is easy to implement operationally, but it is highly sensitive to cloud optical thickness, cloud phase, and the quality of atmospheric temperature profiles, resulting in larger retrieval errors for thin clouds, multi-layered clouds, and cloud edge regions. The other category is a stereo geometric method based on dual-satellite parallax. This method utilizes different observation angles of the same cloud body from two geostationary meteorological satellites and solves for cloud top height through parallax geometry, which can reduce dependence on atmospheric profiles to some extent. However, this type of method is highly sensitive to image matching quality and geometric accuracy, and is easily affected by factors such as cloud motion, weak cloud texture, and poor imaging angle conditions, making it difficult to stably obtain a high-precision cloud top height field.

[0005] With the development of artificial intelligence technology, cloud top height retrieval methods based on deep learning have gradually emerged. These methods typically utilize multispectral observations and cloud parameter products from geostationary meteorological satellites, using cloud top height products as supervised labels to construct models such as convolutional neural networks and residual networks, achieving a direct mapping from multi-channel pixel features to cloud top height values. Compared to traditional lookup tables and empirical statistical methods, deep learning methods can better fit complex nonlinear relationships and exhibit better accuracy. However, most existing deep learning methods rely on data from a single satellite perspective, lacking a unified modeling of the overall three-dimensional structure of clouds and their radiative transmission processes, and have limited processing capabilities for multi-layered clouds and obscuring clouds.

[0006] In the field of 3D reconstruction, novel 3D scene representation methods, such as neural radiation fields, have developed rapidly in recent years. Neural radiation fields map 3D spatial coordinates and the observation direction to volume density and radiation intensity, and perform volume rendering integration along the line of sight, achieving an integrated representation of the geometry and appearance of a 3D scene. Existing research has shown that applying neural radiation fields to multi-view satellite optical imagery can achieve high-precision 3D reconstruction of surface buildings and terrain, as well as the generation of digital surface models (DSMs).

[0007] For example, publication number CN117765168A discloses a method, apparatus, and device for three-dimensional reconstruction of satellite remote sensing images. This patent proposes a method for three-dimensional reconstruction of ground features using satellite remote sensing images. The basic idea is as follows: First, acquire multi-view satellite remote sensing images of the target area, and read the rational polynomial coefficients and metadata of each image. Using the multi-view images and rational polynomial coefficients, optimize the rational polynomial coefficients through bundle adjustment, and obtain the 3D coordinates of key points to obtain sparse 3D geometric priors. Based on this, construct a neural radiation field model of the target area. Encode the 3D coordinates and the direction of sunlight incidence into the network, and output parameters such as volume density, albedo, sunlight visibility, and sky color. Reconstruct pixel colors under each viewpoint using volume rendering formulas, and perform joint training using the depth-supervised loss of sparse 3D points and the transient loss established by pixel uncertainty to simultaneously fit static features and remove transient objects and shadows. For any new viewpoint, establish camera rays using the corresponding imaging geometry, sample the rays and input them into the neural radiation field model for volume rendering, and output the new viewpoint remote sensing image and digital surface model to achieve high-quality 3D reconstruction of the surface area. This method utilizes neural radiation field technology to achieve 3D reconstruction of areas captured by remote sensing satellites, improving the utilization rate of conventional satellite-captured remote sensing images. While the aforementioned 3D reconstruction method based on neural radiation fields from satellite remote sensing images can output new perspective images and digital surface models, its observation data comes from high-resolution optical remote sensing satellites, and the relevant observation geometry is not applicable to the geostationary meteorological satellite observation system commonly used in cloud top height estimation. Furthermore, geostationary meteorological satellites are much farther from Earth than high-resolution optical remote sensing satellites, requiring more targeted cloud sampling strategies to reduce training costs.

[0008] Publication No. CN117805940A proposes a method for measuring cloud top height based on a dual-satellite imager. The method includes: selecting two geostationary satellites with different longitudes and acquiring satellite remote sensing images recorded by both satellites at the same time. Due to the different longitudes of the two satellites, there is parallax in the image positions, which can be used for subsequent stereo ranging; selecting the target area for which the cloud top height needs to be measured in the image acquired by the first satellite, and recording the image's abscissa and geographic latitude and longitude coordinates of this point in the first image; determining the pixel coordinates of the target area using the latitude and longitude coordinates in the image acquired by the second satellite, and recording the image's abscissa of this point in the second satellite image; calculating the distance between the cloud top and the satellite plane based on the difference in the abscissa of the target area in the two satellite images and the imager parameters, and further determining the cloud top distance. This method utilizes the observation parallax of two geostationary meteorological satellites to achieve cloud top height measurement based on dual-satellite geometric parallax. However, the aforementioned cloud top height inversion based on dual-satellite parallax essentially simplifies the cloud top to a thin-layer target of a single height, making it difficult to accurately depict the true three-dimensional cloud field structure. On the one hand, this method establishes a matching relationship based on the latitude and longitude of the target area and pixel parallax, ignoring the projection error of cloud height under satellite imaging geometry, which leads to the error being amplified in geometric calculation and accumulated in the cloud top height result; on the other hand, this method is highly dependent on image resolution, and its cloud height measurement accuracy is limited by imaging resolution and matching accuracy, making it difficult to obtain a continuous and stable cloud top height field in complex cloud fields. Summary of the Invention

[0009] This invention addresses the problems of existing dual-satellite parallax methods, such as simplification of cloud field structure, high sensitivity to image matching accuracy, and susceptibility to projection errors and resolution limitations. It introduces neural radiation fields into the cloud top height estimation problem, utilizes multi-view data from multiple geostationary satellites, designs specialized sampling and modeling strategies for cloud characteristics, and proposes a cloud top height estimation method based on neural radiation fields. This provides a new technical approach for the reconstruction of cloud 3D structure and the inversion of cloud top height.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] Firstly, a method for estimating cloud top height based on neural radiation fields includes the following steps:

[0012] S1. Image data preprocessing: Acquire multispectral observation data from multiple geostationary meteorological satellites, and select visible and infrared bands to construct pseudo-color images that enhance the radiation contrast of sparse high clouds;

[0013] S2. Geostationary satellite observation three-dimensional spatial modeling: Based on the Earth ellipsoid model, the geographical location and satellite position corresponding to each pixel in the pseudo-color image are uniformly transformed to the geocentric and Earth-fixed coordinate system to construct the three-dimensional observation geometry;

[0014] S3. Geometric Prior Point Cloud Construction: Feature matching is performed on multi-view pseudo-color images at the same time. The corresponding observation rays are constructed using the matching points. The three-dimensional coordinates of the matching feature points are solved by spatial intersection to form a sparse three-dimensional point cloud as the geometric prior of the cloud field.

[0015] S4. Construction of observation rays and sampling points: In the geocentric coordinate system, a finite-length observation ray is constructed for each pixel of each image, pointing from the satellite position to the corresponding Earth surface position, and sampling is performed in the three-dimensional space covered by the ray. At the same time, the solar incidence direction corresponding to each pixel is calculated.

[0016] S5. Neural Radiation Field Modeling and Training:

[0017] S5.1 Construct a neural radiation field network. The input of the network is the three-dimensional coordinates of the sampling point, the observation direction, and the solar incident direction. The output is the volume density of the sampling point, the view-dependent self-radiation component, the environmental radiation component, and the radiation determinants of each band.

[0018] S5.2. Based on the preset radiation mixing model, and combining the self-radiation component, the ambient radiation component, and the radiation determining variables, calculate the local radiation color of the sampling point;

[0019] S5.3 Perform volume rendering integration along the observation ray on the local radiation color and volume density to obtain the reconstructed pseudo-color image and the reconstructed depth field;

[0020] S5.4 Construct a joint loss function, including: calculating the radiometric reconstruction loss between the reconstructed pseudo-color image and the real observed pseudo-color image, and the geometric prior loss between the reconstruction depth and the geometric prior depth based on the sparse three-dimensional point cloud computing.

[0021] S5.5. Use multi-view satellite data and the joint loss function to train the neural radiation field network until convergence, and obtain a neural radiation field model that characterizes the three-dimensional structure and radiation properties of the cloud field.

[0022] S6. Cloud Top Height 3D Inversion: Based on the trained neural radiation field model, a vertical line of sight is constructed for geographic grid points. Cloud top depth is obtained by sampling along the line of sight and using volume rendering integral. Then, the cloud top height distribution of the target area is calculated through geometric relationships.

[0023] Furthermore, in step S1, constructing the pseudo-color image specifically involves:

[0024] For the FY-4A satellite, the 0.65μm, 0.83μm, and 10.7μm bands were selected;

[0025] For the FY-4B satellite, the 0.65μm, 0.83μm, and 10.8μm bands were selected;

[0026] For the Himawari-8 satellite, the 0.64μm, 0.86μm, and 10.4μm bands were selected;

[0027] The selected three band data are mapped to the blue, green, and red channels of a pseudo-color image, respectively.

[0028] Furthermore, in step S3, feature matching is implemented using traditional local feature operators or deep learning-based feature matching networks. The traditional local feature operators include SIFT or SURF, and the deep learning-based feature matching networks include LoFTR or MatchAnything.

[0029] Furthermore, in step S4, the length of the finite-length observation ray is determined by a preset cloud field height range, with its upper endpoint corresponding to the maximum possible cloud top height and its lower endpoint corresponding to the location of the land surface or sea level; the three-dimensional space covered by the observation ray is normalized and mapped to a standard cubic space.

[0030] Further, in step S5.1, the neural radiation field network includes:

[0031] A backbone network is used to encode the 3D coordinates of the input sampling points and output the volume density and intermediate features;

[0032] A view-dependent radiation branch network is used to input the intermediate features and observation direction codes, and output the self-radiation components;

[0033] An environmental radiation branch network is used to input the solar incidence direction code and output the environmental radiation component;

[0034] A radiation-determining branch network is used to take the intermediate features and solar incidence direction encoding as inputs and output the radiation-determining variables.

[0035] Further, in step S5.2, the radiation mixing model is:

[0036] c(x,ω view ,ω sun ) = c view (x,ω view )⊙[(η(x,ω sun )+(1-η(x,ω sun ))⊙c a (ω sun )]

[0037] Where c represents the local radiation color, c view For the self-radiation component, c aLet η be the environmental radiation component, ⊙ denote the channel-by-channel multiplication for each radiation band, η be the radiation-determining variable, x be the coordinates of the sampling point, and ω be the irradiance. view ω represents the direction of observation. sun This is the direction of the sun's incidence.

[0038] Further, in step S5.3, the volume rendering integral formula is:

[0039]

[0040] Where, C(ω) view ,ω sun D(ω) represents the reconstructed pseudo-color radiance value of a pixel in multispectral space. view ) represents the volume rendering weighted depth of a pixel along the ray direction, i is the index of the sampling point along the ray, and w i For the volume rendering weight of the i-th sampling point, c i Let z be the local radiative color of the i-th sampling point. i Let be the depth of the i-th sampling point.

[0041] Further, in step S5.4, the joint loss function L is:

[0042] L = L rad +λ depth L depth

[0043] Among them, L rad To calculate the radiometric reconstruction loss, the mean square error between the pixels of the reconstructed pseudo-color image and the actual observed pseudo-color image is calculated; L depth For the geometric prior loss, calculate the mean square error between the reconstruction depth and the geometric prior depth based on the sparse 3D point cloud; λ depth The weighting coefficients for the geometric prior loss are denoted as .

[0044] In a second aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the steps of the method described in any of the preceding claims.

[0045] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. The cloud top height estimation method based on neural radiation field proposed in this invention integrates multi-view and multi-spectral observations from multiple geostationary meteorological satellites into the neural radiation field framework, continuously models the cloud structure and radiation characteristics in three-dimensional space, overcomes the limitation of the traditional two-star parallax method which simplifies the cloud top to a thin-layer target, and can characterize the real three-dimensional cloud field.

[0048] 2. By jointly modeling the satellite position, the surface position, and the cloud volume density field in a unified three-dimensional coordinate system, and by naturally considering the cloud distribution along the line of sight during the volume rendering process using the neural radiation field, this invention significantly reduces the projection error under single-star observation and the cloud height deviation caused by the parallax measurement uncertainty in the two-star parallax method.

[0049] 3. By combining self-radiation with environmental radiation branches and achieving adaptive mixing of radiation determinants in different bands, the different radiation characteristics of clouds in the visible and infrared bands are fully considered. At the same time, a ray truncation and normalization strategy for the cloud height range is designed, which helps to reduce the redundant expression of the model. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0051] Figure 1 A flowchart of a cloud top height estimation method based on neural radiation field provided in an embodiment of the present invention.

[0052] Figure 2 The above is a comparison illustration of visible light images, pseudo-color images, and YunGao product images provided for embodiments of the present invention.

[0053] Figure 3 This is a schematic diagram of point cloud construction based on prior matching of cloud geometry provided in an embodiment of the present invention.

[0054] Figure 4 A schematic diagram of geostationary satellite observation geometry and sampling ray modeling provided for an embodiment of the present invention.

[0055] Figure 5 A diagram of the neural radiation field structure provided in an embodiment of the present invention.

[0056] Figure 6 This is a schematic diagram of three-dimensional inversion of cloud top height based on neural radiation field, provided for an embodiment of the present invention. Detailed Implementation

[0057] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0058] The cloud top height estimation method based on neural radiation field proposed in this invention has the following system architecture: Figure 1 As shown, it includes the following steps:

[0059] S1. Image data preprocessing: Acquire multispectral observation data from multiple geostationary meteorological satellites, and select visible and infrared bands to construct pseudo-color images that enhance the radiation contrast of sparse high clouds.

[0060] Specifically, this invention obtains L1b-level atmospheric cloud field observation data (L1b refers to the dataset acquired by the satellite imager, which is the raw data after outlier filtering without other processing; this data is in .nc format) from the official websites or data centers of the Fengyun series geostationary meteorological satellites (such as FY-4A and FY-4B) and the Himawari-8 satellite. The L1b data is spatially cropped according to the defined cloud height estimation observation area, and then projected and visualized using equidistant cylindrical projection to construct image data suitable for neural network training. Unlike constructing RGB color images using only the visible light band, this invention additionally introduces the infrared band, which characterizes cloud top brightness temperature and thin high clouds, to construct a pseudo-color image that is more sensitive to thin high clouds.

[0061] Specifically, for the FY-4A satellite, the 0.65μm, 0.83μm, and 10.7μm bands (channel numbers 2, 3, and 12) were selected; for the FY-4B satellite, the 0.65μm, 0.83μm, and 10.8μm bands (channel numbers 2, 3, and 13) were selected; and for the Himawari-8 satellite, the 0.64μm, 0.86μm, and 10.4μm bands (channel numbers 3, 4, and 13) were selected. The selected three band data were then mapped to the blue, green, and red channels of a pseudo-color image, respectively.

[0062] The aforementioned channels constitute a pseudo-color three-channel input, used to enhance the radiative contrast between cloud tops and sparse high clouds, providing richer spectral information for subsequent neural radiation field networks. Taking FY-4A satellite data as an example, the visible light image, pseudo-color image, and cloud height product image are compared as follows: Figure 2 As shown.

[0063] S2. Geostationary satellite observation three-dimensional spatial modeling: Based on the Earth ellipsoid model, the geographical location and satellite position corresponding to each pixel in the pseudo-color image are uniformly transformed to the geocentric Earth-fixed coordinate system to construct three-dimensional observation geometry.

[0064] Specifically, for each pseudo-color image after isometric cylindrical projection of the target area, since the latitude and longitude range of the target area is known, the latitude and longitude coordinates corresponding to the center of each pixel can be directly obtained by utilizing the approximately linear mapping relationship between pixel rows and columns and latitude and longitude under isometric cylindrical projection. Under the condition of preset surface elevation (e.g., uniformly set to 0 or introducing external terrain data), based on the Earth ellipsoid model, the latitude and longitude and surface elevation are converted into three-dimensional positions in the Earth-Centered Earth-Fixed (ECEF) coordinate system. The three-dimensional position of the geostationary meteorological satellite in the ECEF coordinate system can be obtained through the latitude and longitude of the nadir point and the orbital altitude. Thus, the three-dimensional positions of the surface points and the corresponding three-dimensional positions of the geostationary meteorological satellites are obtained simultaneously in the same coordinate system, providing a unified three-dimensional reference coordinate system for subsequent observation ray construction and volume rendering.

[0065] S3. Geometric Prior Point Cloud Construction: Feature matching is performed on multi-view pseudo-color images at the same time. Corresponding observation rays are constructed using the matching points. The three-dimensional coordinates of the matching feature points are solved by spatial intersection to form a sparse three-dimensional point cloud as the geometric prior of the cloud field.

[0066] Specifically, for multi-view pseudo-color satellite images taken at the same time, stereo vision methods or feature matching models are used to obtain sparse cloud feature matching points. This may include, but is not limited to: using traditional local feature operators such as SIFT and SURF for feature point detection and description, and screening reliable matching pairs through nearest neighbor matching and geometric consistency constraints; or using deep learning feature matching networks such as LoFTR and MatchAnything to extract highly robust matching points from pseudo-color images.

[0067] For each matched pixel, its pixel position and corresponding latitude and longitude coordinates in the images from various viewpoints are recorded. Then, based on the Earth ellipsoid model and a preset surface elevation, the latitude and longitude are converted to three-dimensional surface coordinates in the ECEF coordinate system. Next, based on the three-dimensional positions of the corresponding geostationary satellites in the ECEF coordinate system for each image, observation rays are constructed pointing from the satellite positions to the three-dimensional positions of the surface points. For multiple observation rays from different satellite viewpoints corresponding to the same cloud feature point, spatial intersection or nearest-point solving methods are used to estimate the actual three-dimensional coordinates of the feature point, thus forming a sparse three-dimensional point cloud as a geometric prior for the cloud field. A schematic diagram of the point cloud construction is shown below. Figure 3 As shown.

[0068] S4. Construction of observation rays and sampling points: In the geocentric coordinate system, a finite-length observation ray is constructed for each pixel of each image, pointing from the satellite position to the corresponding Earth surface position. Sampling is performed in the three-dimensional space covered by the ray, and the solar incidence direction corresponding to each pixel is calculated.

[0069] Specifically, for each pseudo-color image, based on the 3D spatial modeling in step S2, the ECEF coordinates of the surface point corresponding to each pixel and the ECEF coordinates of the corresponding geostationary satellite are obtained. Using the satellite's 3D position as the starting point and the direction connecting the satellite position and the pixel's corresponding surface position as the center direction, an observation ray for that pixel is constructed. According to a preset cloud field height range (e.g., from the surface / sea level to the tropopause or higher), the observation ray is truncated within this height range to obtain a finite-length sampling ray. The upper end of the truncated ray corresponds to the maximum possible cloud top height, and the lower end corresponds to the surface / sea level position. The 3D space covered by the aforementioned finite-length sampling ray is normalized and mapped to a standard cube (e.g., [-1,1)). 3 In this context, a numerically stable and scale-uniform network training space is provided.

[0070] Furthermore, based on the imaging time of each pseudo-color image and the geographical latitude and longitude corresponding to each pixel, the solar azimuth and solar altitude angles at that geographical location at the imaging time are calculated using an astronomical solar position calculation model. This yields the solar incidence direction in the local reference frame, which is then converted into a unit vector in the ECEF coordinate system. In this way, the solar incidence direction, the satellite observation ray direction, and the three-dimensional position of the cloud are all represented in the same coordinate system, which is beneficial for unified modeling and coding within the neural radiation field framework. A schematic diagram of geostationary satellite observation geometry and sampling ray modeling is shown below. Figure 4 As shown.

[0071] Along each normalized observation ray, the depth dimension is discretized uniformly or hierarchically in a standard cubic space, yielding a series of three-dimensional coordinates x of the sampling points. For each sampling point, its corresponding satellite observation direction vector ω is associated with it. view With respect to the direction vector of solar incidence ω sun The sampling point's 3D coordinates, observation direction, and solar direction are encoded separately, for example, using multi-frequency sine and cosine position encoding to improve the network's ability to represent high-frequency details and complex radiation distributions. The encoded 3D coordinates, viewing direction, and solar direction together constitute the sampling point's input features, providing training input for the subsequent neural radiation field network.

[0072] S5. Neural Radiation Field Modeling and Training:

[0073] S5.1 Construct a neural radiation field network. The input of the network is the three-dimensional coordinates of the sampling point, the observation direction, and the solar incidence direction. The output is the volume density of the sampling point, the view-dependent self-radiation component, the environmental radiation component, and the radiation determinants of each band.

[0074] Specifically, in step S5.1, the neural radiation field network is as follows: Figure 5 As shown, it includes:

[0075] A backbone network is used to encode the three-dimensional coordinates of the input sampling points and output the volume density σ(x) and intermediate features h(x). The volume density σ(x) describes the distribution of cloud medium at spatial location x. The larger the volume density, the higher the cloud particle concentration. The intermediate feature representation h(x) serves as the latent space representation for subsequent calculations of view-correlated radiation and environmental radiation, connecting geometric and radiation information.

[0076] A view-dependent radiation branch network is used as input for the intermediate features and observation direction encoding, and outputs the self-radiation component c. view (x,ω view );c view (x,ω view As a component of particle self-radiation, it reflects the anisotropic radiation characteristics of clouds in different observation directions;

[0077] An environmental radiation branch network is used to input the solar incidence direction code and output the environmental radiation component c. a (ω sun );c a (ω sun This reflects the distribution of solar incident radiation within the cloud environment;

[0078] A radiation-determining branch network is used as input for the intermediate features and the solar incidence direction encoding, and outputs the radiation-determining variable η(x,ω). sun ), η(x,ω sun It is used to characterize the proportional relationship between the response of particles to self-radiation and ambient radiation at different wavelengths.

[0079] S5.2. Based on the preset radiation mixing model, and combining the self-radiation component, the ambient radiation component, and the radiation determinant variable, calculate the local radiation color of the sampling point.

[0080] Specifically, in step S5.2, based on particle self-radiation c view (x,ω view ), environmental radiation c a (ω sun ) and radiation determinant η(x,ω) sun A pre-defined radiation mixing model is constructed. For a sampling point at location x, in the observation direction ω... view ω with respect to the direction of solar incidence sun The radiation mixing model is as follows:

[0081] c(x,ω view ,ω sun ) = c view (x,ω view )⊙[(η(x,ω sun )+(1-η(x,ωsun ))⊙c a (ω sun )]

[0082] Where c represents the local radiation color, c view For the self-radiation component, c a Let η be the environmental radiation component, ⊙ denote the channel-by-channel multiplication for each radiation band, η be the radiation-determining variable, x be the coordinates of the sampling point, and ω be the irradiance. view ω represents the direction of observation. sun This is the direction of the sun's incidence.

[0083] η(x,ω sun ) Calculations are performed separately in different bands to adapt to the different radiation response characteristics of visible light and infrared channels, thereby achieving an adaptive mixing of self-radiation and environmental radiation.

[0084] Obtain the local mixed radiation color c(x,ω) view ,ω sun After obtaining the volume density σ(x), the sampling points are integrated and accumulated along each observation ray using the volume rendering formula to obtain the reconstructed pseudo-color radiometric value and weighted depth of the corresponding pixel under that viewpoint, thus forming the reconstructed pseudo-color image corresponding to each satellite viewpoint.

[0085] S5.3 Perform volume rendering integration along the observation ray on the local radiation color and volume density to obtain the reconstructed pseudo-color image and the reconstructed depth field.

[0086] Specifically, in step S5.3, the observation direction ω from the satellite position is... view The emitted rays are discretely sampled to obtain a series of spatial positions. and corresponding depth The volume density σ at each sampling point is obtained from the neural radiation field network. i =σ(x i ) and local radiation color c i =c(x,ω) view ,ω sun ).

[0087] Define the step size between adjacent sampling points as

[0088] δ i =z i+1 -z i i = 1, ..., N-1,

[0089] A constant step size δ is given for the last sampling point. N The step size sequence on the ray is obtained. Based on the principles of volume rendering, the opacity and volume rendering weight of each sampling point are calculated. The opacity of the i-th sampling point is...

[0090] α i =1-exp(-δ i σ i ), i = 1, ..., N,

[0091] Its cumulative transmittance before reaching the sampling point is

[0092]

[0093] This yields the volume rendering weight of the sampling point for the final pixel radiance.

[0094] w i =α i T i ,i=1,…,N.

[0095] By using the aforementioned weights to weight and accumulate the local radiant color and depth along the ray, the final pseudo-color radiant value and weighted depth of the pixel at this viewpoint are expressed as follows:

[0096]

[0097]

[0098] Where, C(ω) view ,ω sun D(ω) represents the reconstructed pseudo-color radiance value of a pixel in multispectral space. view ) represents the volume rendering weighted depth of a pixel along the ray direction, i is the index of the sampling point along the ray, and w i For the volume rendering weight of the i-th sampling point, c i Let z be the local radiative color of the i-th sampling point. i Let be the depth of the i-th sampling point. Subsequently, by combining satellite position and observation geometry, the depth value can be converted into three-dimensional coordinates of the cloud top, and then further converted into the cloud top height.

[0099] S5.4 Construct a joint loss function, including: calculating the radiometric reconstruction loss between the reconstructed pseudo-color image and the real observed pseudo-color image, and the geometric prior loss between the reconstruction depth and the geometric prior depth based on the sparse 3D point cloud computing.

[0100] Specifically, in step S5.4, the reconstructed pseudo-color image C(r) obtained by rendering the neural radiation field is compared with the actual observed pseudo-color image C obtained by preprocessing in step (1). GT (r) Perform pixel-level comparisons and construct a radiation reconstruction loss function based on the radiation differences of each channel, such as using mean squared error loss:

[0101]

[0102] in, The set of observed rays corresponding to the pixels participating in the training. This refers to one of the observed rays.

[0103] For the sparse 3D point cloud obtained in step (3), the set of observation rays passing through the point cloud is obtained based on its corresponding geostationary satellite position and ray construction method. Volume rendering is performed on these rays using neural radiation fields to obtain the reconstructed depth D(r); simultaneously, the geometrically corresponding true depth D of the ray can be calculated from the 3D coordinates of the point cloud. GT (r). Construct the geometric prior loss in the form of mean squared error:

[0104]

[0105] By minimizing the reconstruction depth D(r) and the geometric prior depth D GT The difference in (r) explicitly introduces the cloud field geometric prior provided by sparse point clouds into the neural radiation field training process, reducing the ambiguity of the three-dimensional shape of the cloud and improving the stability and accuracy of cloud height inversion.

[0106] Combining the above-mentioned radiation reconstruction loss and geometric prior loss, the joint loss function L is constructed as follows:

[0107] L = L rad +λ depth L depth

[0108] Among them, L rad To calculate the radiometric reconstruction loss, the mean square error between the pixels of the reconstructed pseudo-color image and the actual observed pseudo-color image is calculated; L depth For the geometric prior loss, calculate the mean square error between the reconstruction depth and the geometric prior depth based on the sparse 3D point cloud; λ depth The weighting coefficients for the geometric prior loss are denoted as .

[0109] S5.5. The neural radiation field network is trained using multi-view satellite data and the joint loss function until convergence, resulting in a neural radiation field model that characterizes the three-dimensional structure and radiation properties of the cloud field.

[0110] Specifically, using pseudo-color images from multiple geostationary satellites at various perspectives as input samples, the corresponding observation rays, sampling point positions, observation viewpoint directions, solar incidence directions, and sparse geometric prior depths are extracted to form training data. Stochastic gradient descent or adaptive optimization algorithms are employed to iteratively update the neural radiation field network parameters. In each iteration, volume density and radiation distribution are calculated through forward propagation, volume rendering is performed to obtain the reconstructed image and reconstructed depth field, and the loss value is calculated based on the joint loss function. Backpropagation is then performed to update the network parameters. Training stops when the preset number of iterations is reached or the loss converges. The network parameters at this point are saved as a stable neural radiation field model representing the three-dimensional structure and radiation characteristics of the cloud field in the target region, for subsequent three-dimensional inversion of cloud top height.

[0111] S6. Cloud Top Height 3D Inversion: Based on the trained neural radiation field model, a vertical line of sight is constructed for geographic grid points. Cloud top depth is obtained by sampling along the line of sight and using volume rendering integral. Then, the cloud top height distribution of the target area is calculated through geometric relationships.

[0112] Specifically, a schematic diagram of the three-dimensional inversion of cloud top height based on the neural radiation field is shown below. Figure 6 As shown. After model training is complete, based on the spatial modeling method in step S4, an observation line of sight perpendicular to the land surface / sea level is constructed for each geographic grid point in the target area. Unlike the ray construction from the satellite direction in the training phase, here it extends upward from the land surface point along the local normal direction, covering the preset cloud top height range to form a vertical line of sight. In the trained neural radiation field, the volume density is sampled along each vertical line of sight, and the cloud volume distribution along the line of sight is integrated and judged using the volume rendering formula or the corresponding depth estimation strategy to obtain the cloud top depth value corresponding to each geographic location. Subsequently, the three-dimensional coordinates of the cloud tops are recovered in the ECEF coordinate system and then converted into latitude, longitude and altitude to form the cloud top height distribution data of the target area. Finally, the cloud top height can be output in the form of a regular grid, a two-dimensional raster product or a three-dimensional point cloud, generating two-dimensional / three-dimensional cloud top height products for operational applications such as weather analysis, cloud microphysics inversion and numerical forecasting.

[0113] In a second aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the steps of the method described in any of the preceding claims.

[0114] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0115] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for estimating cloud top height based on neural radiation fields, characterized in that, Includes the following steps: S1. Image data preprocessing: Acquire multispectral observation data from multiple geostationary meteorological satellites, and select visible and infrared bands to construct pseudo-color images that enhance the radiation contrast of sparse high clouds; S2. Geostationary satellite observation three-dimensional spatial modeling: Based on the Earth ellipsoid model, the geographical location and satellite position corresponding to each pixel in the pseudo-color image are uniformly transformed to the geocentric Earth-fixed coordinate system to construct the three-dimensional observation geometry; S3. Geometric Prior Point Cloud Construction: Feature matching is performed on multi-view pseudo-color images at the same time. The corresponding observation rays are constructed using the matching points. The three-dimensional coordinates of the matching feature points are solved by spatial intersection to form a sparse three-dimensional point cloud as the geometric prior of the cloud field. S4. Construction of observation rays and sampling points: In the geocentric coordinate system, a finite-length observation ray is constructed for each pixel of each image, pointing from the satellite position to the corresponding Earth surface position. Sampling is performed in the three-dimensional space covered by the ray, and the solar incidence direction corresponding to each pixel is calculated. S5. Neural Radiation Field Modeling and Training: S5.1 Construct a neural radiation field network. The input of the network is the three-dimensional coordinates of the sampling point, the observation direction, and the solar incidence direction. The output is the volume density of the sampling point, the view-dependent self-radiation component, the environmental radiation component, and the radiation determinants of each band. S5.

2. Based on the preset radiation mixing model, and combining the self-radiation component, the ambient radiation component, and the radiation determining variables, calculate the local radiation color of the sampling point; S5.3 Perform volume rendering integration along the observation ray on the local radiation color and volume density to obtain the reconstructed pseudo-color image and the reconstructed depth field; S5.4 Construct a joint loss function, including: calculating the radiometric reconstruction loss between the reconstructed pseudo-color image and the real observed pseudo-color image, and the geometric prior loss between the reconstruction depth and the geometric prior depth based on the sparse three-dimensional point cloud computing. S5.

5. Use multi-view satellite data and the joint loss function to train the neural radiation field network until convergence, and obtain a neural radiation field model that characterizes the three-dimensional structure and radiation properties of the cloud field. S6. Cloud Top Height 3D Inversion: Based on the trained neural radiation field model, a vertical line of sight is constructed for geographic grid points. Cloud top depth is obtained by sampling along the line of sight and using volume rendering integral. Then, the cloud top height distribution of the target area is calculated through geometric relationships.

2. The cloud top height estimation method based on neural radiation field according to claim 1, characterized in that, In step S1, constructing the pseudo-color image specifically involves: For the FY-4A satellite, the 0.65μm, 0.83μm, and 10.7μm bands were selected; For the FY-4B satellite, the 0.65μm, 0.83μm, and 10.8μm bands were selected; For the Himawari-8 satellite, the 0.64μm, 0.86μm, and 10.4μm bands were selected; The selected three band data are mapped to the blue, green, and red channels of a pseudo-color image, respectively.

3. The cloud top height estimation method based on neural radiation field according to claim 1, characterized in that, In step S3, feature matching is implemented using traditional local feature operators or deep learning-based feature matching networks. The traditional local feature operators include SIFT or SURF, and the deep learning-based feature matching networks include LoFTR or MatchAnything.

4. The cloud top height estimation method based on neural radiation field according to claim 1, characterized in that, In step S4, the length of the finite-length observation ray is determined by a preset cloud field height range, with its upper endpoint corresponding to the maximum possible cloud top height and its lower endpoint corresponding to the location of the land surface or sea level; the three-dimensional space covered by the observation ray is normalized and mapped to a standard cubic space.

5. The cloud top height estimation method based on neural radiation field according to claim 1, characterized in that, In step S5.1, the neural radiation field network includes: A backbone network is used to encode the 3D coordinates of the input sampling points and output the volume density and intermediate features; A view-dependent radiation branch network is used to input the intermediate features and observation direction codes, and output the self-radiation components; An environmental radiation branch network is used to input the solar incidence direction code and output the environmental radiation component; A radiation-determining branch network is used to take the intermediate features and solar incidence direction encoding as inputs and output the radiation-determining variables.

6. The cloud top height estimation method based on neural radiation field according to claim 1, characterized in that, In step S5.2, the radiation mixing model is: ， in, For localized radiation colors, This is the self-radiative component. For environmental radiation components, This indicates channel-by-channel multiplication over each radiation band. For radiation, These are the coordinates of the sampling point. For the direction of observation, This is the direction of the sun's incidence.

7. The cloud top height estimation method based on neural radiation field according to claim 1, characterized in that, In step S5.3, the volume rendering integral formula is: ， ， in, The reconstructed pseudo-color radiance value of the pixel in multispectral space. The volume rendering weighted depth of a pixel along the ray direction, where i is the index of the sampling point along the ray. Render the volume weights for the i-th sampling point. The local radiant color of the i-th sampling point. Let be the depth of the i-th sampling point.

8. The cloud top height estimation method based on neural radiation field according to claim 1, characterized in that, In step S5.4, the joint loss function L is: ， in, To account for the radiation reconstruction loss, the mean square error between the pixels of the reconstructed pseudo-color image and the actual observed pseudo-color image is calculated. As the geometric prior loss, the mean square error between the reconstruction depth and the geometric prior depth based on the sparse 3D point cloud is calculated. The weighting coefficients for the geometric prior loss are denoted as .

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 8.

10. 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 as described in any one of claims 1 to 8.

Images