Polar region cloud layer detection method and device based on day and night adaptive multi-mode fusion

Abstract

The invention discloses a polar region cloud layer detection method and device based on day and night adaptive multi-mode fusion. The method comprises the following steps: determining a polar region time period mode according to a solar altitude and environment illumination intensity of a polar region observation point; in the polar day mode, collecting a multi-angle polarization thermal image sequence and a visible light image, generating a cloud probability distribution map, and generating a polar day optical thickness map in combination with a physically guided thickness inversion model; in an extreme night mode, quantum radar echoes and terahertz radiation spectrums are collected, a cloud layer vertical distribution probability graph is generated, and an extreme night three-dimensional cloud mask is generated in combination with radar distance-Doppler constraint; and generating a polar region cloud layer full-time detection result based on the polar day optical thickness map and the polar night three-dimensional cloud mask. According to the invention, high-precision and all-weather detection of the polar region cloud is realized by introducing sensing technologies such as quantum radar and terahertz radiation spectrum and combining deep learning and quantum information processing methods.

Description

Technical Field

[0001] This application relates to the field of polar environment monitoring technology, and in particular to a polar cloud detection method and device based on day-night adaptive multimodal fusion. Background Technology

[0002] As a crucial component of the Earth's climate system, the cloud characteristics of polar regions significantly impact global energy balance and climate change. However, the unique geographical environment and extreme climatic conditions of the polar regions present enormous challenges to cloud detection. Existing cloud detection technologies are typically single-mode detection methods, which are ill-suited to the extreme diurnal observation conditions in the polar regions. Consequently, continuous monitoring throughout the entire period is impossible, resulting in substantial gaps and errors in polar cloud data, making it difficult to meet the demands of high-precision climate research. Summary of the Invention

[0003] The main purpose of this application is to provide a polar cloud detection method and device based on day-night adaptive multimodal fusion, which aims to solve the technical problems of existing technologies being unable to adapt to the special polar environment, having poor detection continuity, and insufficient accuracy.

[0004] To achieve the above objectives, this application proposes a polar cloud detection method based on day-night adaptive multimodal fusion. The polar cloud detection method based on day-night adaptive multimodal fusion includes: The polar time period pattern is determined based on the solar altitude angle and ambient light intensity at the polar observation points; When the polar time mode is polar day mode, multi-angle polarized thermal image sequences and visible light images are collected, and cloud probability distribution maps are generated based on the multi-angle polarized thermal image sequences and visible light images. When the polar period mode is polar night mode, quantum radar echo and terahertz radiation spectrum are collected, and cloud vertical distribution probability map is generated based on quantum radar echo and terahertz radiation spectrum. Based on the probability distribution map of cloud types, combined with a physically guided thickness inversion model, an optical thickness map of polar day is generated. Based on the vertical distribution probability map of clouds and combined with radar range-Doppler constraints, a three-dimensional cloud mask for polar night is generated. The full-time detection results of polar clouds were generated based on polar day optical thickness maps and polar night 3D cloud masks.

[0005] Preferably, when the polar time mode is polar day mode, multi-angle polarized thermal image sequences and visible light images are acquired, and a cloud probability distribution map is generated based on the multi-angle polarized thermal image sequences and visible light images, including: When the polar period mode is polar day mode, multi-angle polarized thermal image sequences output by the polarization thermal imaging camera and visible light images output by the all-sky imager are acquired. Radiometric calibration and nonlinear response correction are performed on multi-angle polarized thermal image sequences to generate a corrected set of polarized thermal images. Thermal radiation intensity images acquired under multiple different polarization directions were extracted from the corrected polarization thermal image set to obtain multiple registered polarization direction intensity maps; Stokes vectors are calculated pixel-by-pixel based on polarization direction intensity maps of multiple registrations to generate Stokes parametric maps; Based on the Stokes parametric map, the polarization degree, polarization angle and linear polarization difference characteristics of each pixel are calculated to generate a three-channel polarization feature map. The Stokes parametric map and the three-channel polarization feature map are stitched together along the channel dimension to form a multi-channel polarization feature set. Channel normalization and spatial alignment are performed on the multi-channel polarization feature set to obtain the polarization thermal feature tensor. The polarization thermal feature tensor and the visible light image are input into a multimodal fusion classification network for cloud class identification, resulting in a cloud class probability distribution map.

[0006] Preferably, when the polar period mode is polar night mode, quantum radar echoes and terahertz radiation spectra are collected, and a cloud vertical distribution probability map is generated based on the quantum radar echoes and terahertz radiation spectra, including: The quantum radar sends entangled microwave pulses of a preset frequency into the sky, receives the reflected quantum radar echo, and simultaneously collects the terahertz radiation spectrum of the preset frequency band of the terahertz passive radiometer. Quantum state tomography is performed on the quantum radar echo to reconstruct the density matrix of each resolution cell, and a quantum dot cloud containing phase, amplitude, and entanglement is generated based on the density matrix. A quantum topology graph is constructed using quantum dot clouds as nodes and entanglement degree as edge weights, and a radiation spectrum graph is constructed using terahertz radiation spectrum as nodes and spectral slope as edge weights. The quantum topology graph and radiation spectrum graph are input into the event branch and radiation branch of the graph neural network, respectively, to extract quantum topology features and radiation spectrum features; By using a cross-modal graph attention mechanism with learnable edge weights, the quantum topological graph and the radiation spectrum are fused into a joint quantum-radiative feature of the cloud. Obtain the polar ionospheric scintillation index and map the polar ionospheric scintillation index into a graph reparameterized vector; Dynamic edge weight correction is performed on the joint quantum-radiation features of clouds based on graph reparameterized vectors to generate perturbation adaptive features. Cloud vertical distribution probability is predicted based on perturbation adaptive features, resulting in a cloud vertical distribution probability map.

[0007] Preferably, a quantum topology graph is constructed using quantum dot clouds as nodes and entanglement degrees as edge weights, including: The quantum dot cloud is projected into a four-dimensional feature space composed of spatial coordinates and quantum entropy; In the four-dimensional feature space, the K-nearest neighbor algorithm is used to find a preset number of neighboring nodes for each node to construct an initial undirected graph; Calculate the entanglement similarity between two nodes on each edge in the initial undirected graph to generate an initial set of edge weights; The initial set of edge weights is compared with a preset entanglement threshold, and edges in the initial set of edge weights with entanglement similarity lower than the preset entanglement threshold are deleted to form a sparse quantum topological graph. A graph attention network is used to aggregate and update the node features in a sparse quantum topology graph to generate enhanced node features. By leveraging enhanced node features, the correlation strength between nodes is recalculated, edge weights are dynamically updated, and a dynamically evolving quantum topology graph is output.

[0008] Preferably, based on the cloud probability distribution map and combined with a physically guided thickness inversion model, a polar day optical thickness map is generated, including: A simulation dataset of cloud optical thickness was generated based on a physical radiative transfer model. The physical-guided thickness inversion neural network is trained based on the cloud optical thickness simulation dataset to obtain the trained thickness inversion neural network. The polarization thermal feature tensor is input into the trained thickness inversion neural network for forward inference, and the initial optical thickness map is output. By performing Bayesian fusion of the cloud probability distribution map and the initial optical thickness map, the polar day optical thickness map is obtained.

[0009] Preferably, the cloud probability distribution map and the initial optical thickness map are Bayesianly fused to obtain the polar day optical thickness map, including: The initial optical thickness map is segmented into superpixels, and the optical thickness variance within each superpixel region is calculated to generate a local homogeneity map. By comparing the local homogeneity map with a preset homogeneity threshold, abnormal region masks are identified. Extract visible light texture gradient and polarization angle anisotropy within the mask of the abnormal region to generate a correction weight map; The correction weight map is multiplied pixel by pixel with the initial optical thickness map to obtain the abnormal correction thickness map; By performing Bayesian fusion of the cloud probability distribution map and the anomaly correction thickness map, the polar day optical thickness map is obtained.

[0010] Preferably, the cloud probability distribution map and the anomaly correction thickness map are Bayesianly fused to obtain the polar day optical thickness map, including: Based on historical observation data of different cloud types, a priori optical thickness distribution model is constructed for each cloud type. Using the anomaly-corrected thickness map as the observation value, and combined with the cloud class probability distribution map, the thickness likelihood function of each pixel belonging to each cloud class is calculated. Based on Bayes' theorem, the posterior optical thickness distribution of each pixel is determined using the prior optical thickness distribution model and the thickness likelihood function. Extract the maximum posterior probability from the posterior optical thickness distribution of each pixel as the optimal optical thickness estimate for the corresponding pixel, and generate a maximum posterior probability thickness map. A guided filter is applied to the maximum a posteriori probability thickness map to generate a polar day optical thickness map.

[0011] Preferably, based on the cloud vertical distribution probability map and combined with radar range-Doppler constraints, a three-dimensional cloud mask for polar night is generated, including: The cloud vertical distribution probability map is input into a 3D U-shaped network for edge enhancement and noise suppression, and the optimized 3D probability field is output. Acquire range-Doppler cube data synchronously collected by quantum radar, and generate a velocity divergence constraint field based on the range-Doppler cube data; Construct a learnable constant diagonal structure projection matrix to map the optimized 3D probability field onto a voxel grid of preset resolution, generating an initial voxel probability volume; Using the velocity divergence constraint field as the physical consistency loss function, the initial voxel probability volume is iteratively optimized to obtain the optimized voxel probability volume. The optimized voxel probability volume is discretized and connected component analyzed based on stochastic gradient to generate a three-dimensional cloud mask for polar night.

[0012] Preferably, the full-time detection results of polar clouds are generated based on the polar day optical thickness map and the polar night three-dimensional cloud mask, including: The polar night 3D cloud mask is multiplied on a voxel-by-voxel basis with the optimized 3D probability field to obtain the polar night optical thickness volume. The optical thickness volume of polar night and the optical thickness map of polar day are uniformly resampled to a common grid to obtain a common thickness volume; Using the solar elevation angle as the fusion weight, an adaptive weighted fusion of the common thickness volume is performed to obtain the fused thickness volume; By performing a logical AND operation on the fused thickness volume, the 3D cloud mask for polar night, and the probability map of cloud types for polar day, a cloud mask for all time periods is generated. Based on the fused thickness volume, the thickness level of the cloud mask is classified for all time periods to obtain the detection results of polar cloud layers for all time periods.

[0013] Furthermore, to achieve the above objectives, this application also proposes a polar cloud detection device based on day-night adaptive multimodal fusion. The polar cloud detection device based on day-night adaptive multimodal fusion includes: The determination module is used to determine the polar time period pattern based on the solar altitude angle and ambient light intensity of the polar observation point; The acquisition module is used to acquire multi-angle polarized thermal image sequences and visible light images when the polar time mode is polar day mode, and generate a cloud probability distribution map based on the multi-angle polarized thermal image sequences and the visible light images; The acquisition module is also used to acquire quantum radar echoes and terahertz radiation spectra when the polar time mode is polar night mode, and generate a cloud vertical distribution probability map based on the quantum radar echoes and terahertz radiation spectra. The generation module is used to generate a polar day optical thickness map based on the cloud probability distribution map and combined with a physically guided thickness inversion model. The generation module is also used to generate a three-dimensional cloud mask for polar night based on the cloud vertical distribution probability map and combined with radar range-Doppler constraints. The generation module is also used to generate full-time detection results of polar clouds based on the polar day optical thickness map and the polar night three-dimensional cloud mask.

[0014] This application proposes one or more technical solutions to determine the polar time period mode based on the solar altitude angle and ambient light intensity of the polar observation point. When the polar time period mode is a polar day mode, multi-angle polarized thermal image sequences and visible light images are acquired, and a cloud probability distribution map is generated based on these images. When the polar time period mode is a polar night mode, quantum radar echoes and terahertz radiation spectra are acquired, and a cloud vertical distribution probability map is generated based on these images. Based on the cloud probability distribution map, a polar day optical thickness map is generated using a physics-guided thickness inversion model. Based on the cloud vertical distribution probability map, a polar night three-dimensional cloud mask is generated using radar range-Doppler constraints. Finally, a polar cloud all-weather detection result is generated based on the polar day optical thickness map and the polar night three-dimensional cloud mask. Through the above methods, by introducing sensing technologies such as quantum radar and terahertz radiation spectra, and combining them with deep learning and quantum information processing methods, high-precision, all-weather detection of polar clouds is achieved. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an embodiment of the polar cloud detection method based on day-night adaptive multimodal fusion provided in this application. Figure 2 This is a schematic diagram of the module structure of the polar cloud detection device based on day-night adaptive multimodal fusion, as described in an embodiment of this application.

[0018] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0020] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0021] This application provides a solution that achieves high-precision, all-weather detection of polar clouds by introducing sensing technologies such as quantum radar and terahertz radiation spectrum, and combining deep learning and quantum information processing methods.

[0022] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a polar cloud detection device based on day-night adaptive multimodal fusion. The following description uses a polar cloud detection device based on day-night adaptive multimodal fusion as an example to illustrate this embodiment and the subsequent embodiments.

[0023] Based on this, this application provides a polar cloud detection method based on day-night adaptive multimodal fusion, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the polar cloud detection method based on day-night adaptive multimodal fusion of this application.

[0024] In this embodiment, the polar cloud detection method based on day-night adaptive multimodal fusion includes steps S10~S60: Step S10: Determine the polar time period pattern based on the solar altitude angle and ambient light intensity of the polar observation point.

[0025] It should be noted that the polar observation station can be either an Arctic or Antarctic observation station; this implementation method does not impose specific limitations on this. This implementation method uses an Arctic observation station as an example for illustration. The observation station is equipped with necessary hardware, including a polarization thermal imaging camera, an all-sky imager, a quantum radar, and a terahertz radiometer.

[0026] The solar altitude angle, or the angle between sunlight and the horizon, directly reflects the sun's altitude in the sky. Ambient light intensity characterizes the illumination level of the environment surrounding the observation point. A digital sunmeter installed at the top of the polar observation pole outputs the solar altitude angle θ (°) and ambient light intensity L (lx).

[0027] Understandably, based on a pre-set threshold range, when the solar altitude angle is greater than a certain value and the ambient light intensity exceeds the corresponding threshold, it is determined to be in polar day mode; conversely, when the solar altitude angle is less than the certain value and the ambient light intensity is less than the corresponding threshold, it is determined to be in polar night mode. Different data acquisition paths are triggered by real-time determination of whether polar day / polar night mode has been entered. The formula for calculating the solar altitude angle is:

[0028] in, The solar altitude angle, The geographical latitude of the observation point. The solar declination, calculated from the Julian day. It is the hour angle, determined by the local true solar time.

[0029] In the specific implementation, a threshold of 4° corresponding to the solar altitude angle and 1000 lx corresponding to the ambient light intensity are used for explanation. When the solar altitude angle is greater than 4° and the illuminance is greater than 1000 lx, it is determined to be in polar day mode; when the solar altitude angle is less than -4° and the illuminance is less than 100 lx, it is determined to be in polar night mode. This determination process is repeated every preset time interval, such as 5 minutes, to achieve dynamic switching of modes.

[0030] By intelligently identifying polar day and polar night modes and adaptively switching to the optimal detection scheme, the problem of observation interruption caused by extreme day and night conditions in polar regions has been effectively solved, and 24-hour uninterrupted cloud monitoring has been achieved.

[0031] Step S20: When the polar time mode is polar day mode, acquire multi-angle polarized thermal image sequence and visible light image, and generate cloud probability distribution map based on the multi-angle polarized thermal image sequence and visible light image.

[0032] It should be noted that after determining the polar time mode, different data acquisition paths are triggered. In polar day mode, the polarization thermal imaging camera and the all-sky imager are activated to synchronously acquire data. The polarization thermal imaging camera continuously captures a sequence of polarization thermal images from multiple preset angles, such as 0°, 45°, 90°, and 135°. Each frame contains polarization information such as Stokes parameters (I, Q, U, V). Stokes parameters are four basic parameters describing the polarization state of light waves, where I represents the total light intensity, Q and U describe the intensity difference of the linear polarization component in different directions, and V is related to circular polarization. The all-sky imager simultaneously captures high-resolution visible light images.

[0033] Understandably, by solving Stokes parameters through multi-angle polarization thermal images, a high-dimensional polarization thermal feature tensor is constructed, which is then fused with visible light images to perform cloud type recognition and generate a cloud type probability distribution map.

[0034] In one feasible implementation, step S20 may include: when the polar time mode is polar day mode, acquiring a multi-angle polarized thermal image sequence output by a polarization thermal imaging camera and a visible light image output by an all-sky imager; performing radiometric calibration and nonlinear response correction on the multi-angle polarized thermal image sequence to generate a corrected polarized thermal image set; extracting thermal radiation intensity images acquired under multiple different polarization directions from the corrected polarized thermal image set to obtain multiple registered polarization direction intensity maps; calculating Stokes vectors pixel by pixel based on the multiple registered polarization direction intensity maps to generate a Stokes parameter map; calculating the degree of polarization, polarization angle, and linear polarization difference features of each pixel based on the Stokes parameter map to generate a three-channel polarization feature map; concatenating the Stokes parameter map and the three-channel polarization feature map in the channel dimension to form a multi-channel polarization feature set; performing channel normalization and spatial alignment on the multi-channel polarization feature set to obtain a polarization thermal feature tensor; inputting the polarization thermal feature tensor and the visible light image into a multimodal fusion classification network for cloud class recognition to obtain a cloud class probability distribution map.

[0035] It should be noted that a polarization thermal imaging camera was used to acquire thermal images at polarization angles of 0°, 45°, 90°, and 135°, and non-uniformity correction was performed on each frame. Using blackbody radiation source calibration data, the radiance of the images in the multi-angle polarization thermal image sequence was radiometrically calibrated to eliminate errors caused by sensor response nonlinearity. Simultaneously, a polynomial fitting method was used to correct the nonlinear response of the images, generating a corrected polarization thermal image set, as shown in the following equation:

[0036] in, For the corrected radiance, denoted as the original radiance, and a, b, and c as camera calibration coefficients.

[0037] From the corrected polarization thermal image set, thermal radiation intensity images acquired under multiple different polarization directions are extracted according to preset polarization direction parameters. For example, thermal radiation intensity images corresponding to polarization directions of 0°, 45°, 90°, and 135° are extracted respectively. After image registration processing, multiple registered polarization direction intensity maps (I0°, I...) are obtained. 45 °,I 90 °,I 135 Following the Stokes vector calculation method, the Stokes vector (I,Q,U) is calculated pixel by pixel to generate the Stokes parametric map, as shown in the following formula: I = I0° + I 90 ° Q=I0°-I 90 ° U=I 45 °-I 135 ° Among them, I α Let α be the thermal radiation intensity at a polarization angle, I be the total radiation intensity, and Q and U be the linear polarization components.

[0038] Based on the Stokes parametric map, the degree of polarization, polarization angle, and linear polarization difference characteristics of each pixel are calculated as follows:

[0039]

[0040]

[0041] Where DoP is the degree of polarization, representing the proportion of the polarized portion of the light wave; AoP is the polarization angle, describing the angle between the vibration direction of linearly polarized light and the reference direction; D... lin This is a characteristic of linear polarization difference, reflecting the difference in linear polarization components in different directions.

[0042] The generated three-channel polarization feature map corresponds to the degree of polarization, polarization angle, and linear polarization difference features, respectively.

[0043] The Stokes parametric map and the three-channel polarization feature map are concatenated along the channel dimension to form a multi-channel polarization feature set. This step aims to fuse information from different polarization characteristics to provide a more comprehensive description of cloud polarization features. Subsequently, channel normalization is performed on the multi-channel polarization feature set to eliminate differences in numerical ranges between different channels, ensuring that each channel feature has equal importance in subsequent processing. Simultaneously, spatial alignment is performed to ensure the consistency of the polarization thermal feature tensor with the visible light image in spatial location.

[0044] The preprocessed polarization thermal feature tensor and high-resolution visible light image are input into a multimodal fusion classification network. This network learns the complex relationship between polarization features and visible light features to achieve accurate cloud classification. The network outputs the probability that each pixel belongs to "clear sky", "cirrus", "cumulus", or "stratus", which is the cloud classification probability distribution map.

[0045] Step S30: When the polar time mode is polar night mode, collect quantum radar echo and terahertz radiation spectrum, and generate cloud vertical distribution probability map based on quantum radar echo and terahertz radiation spectrum.

[0046] It should be noted that while cloud identification cannot rely on visible light images in polar night mode, this implementation method achieves effective cloud detection by introducing advanced sensing technologies such as quantum radar and terahertz radiometers. Quantum radar utilizes quantum properties such as quantum entanglement to improve detection sensitivity and resolution, enabling the capture of weak echo signals from clouds under extremely low light conditions. Terahertz radiometers, by detecting terahertz wave radiation emitted by clouds, obtain key parameters such as cloud temperature and humidity, providing strong support for comprehensive analysis of cloud characteristics.

[0047] Understandably, after entering the polar night mode, entangled microwave pulses at a frequency of 35 GHz are emitted into the sky via quantum radar, and low-noise radiation spectra in the 0.3–3 THz band of a terahertz passive radiometer are collected simultaneously to obtain quantum radar echoes and terahertz radiation spectra. The density matrix is ​​reconstructed using quantum radar echoes, the degree of entanglement is extracted to construct a quantum topology map, and a radiation spectrum is constructed by combining the slope of the terahertz spectrum. Perturbation adaptive features are generated by fusing through graph neural networks, and then the vertical distribution probability map of clouds is predicted based on the perturbation adaptive features. The vertical distribution probability map of clouds is a visual representation of the probability distribution of clouds at different heights in the vertical direction. It can clearly reflect the structural characteristics and distribution patterns of clouds in vertical space.

[0048] By deeply fusing multimodal data, the complementarity of data from various sensors is fully utilized. In polar day mode, polarization thermal imaging and visible light data are combined, while in polar night mode, quantum radar and terahertz radiation data are fused. Feature fusion is achieved through advanced graph neural networks and attention mechanisms, significantly improving the accuracy and robustness of cloud detection.

[0049] In one feasible implementation, step S30 may include: transmitting entangled microwave pulses of a preset frequency into the sky via quantum radar, receiving the reflected quantum radar echoes, and simultaneously acquiring the terahertz radiation spectrum of a preset frequency band from a terahertz passive radiometer; performing quantum state tomography processing on the quantum radar echoes to reconstruct the density matrix of each resolution unit, and generating a quantum point cloud containing phase, amplitude, and entanglement degree based on the density matrix; constructing a quantum topology graph with the quantum point cloud as nodes and the entanglement degree as edge weights, and constructing a radiation spectrum graph with the terahertz radiation spectrum as nodes and the spectral slope as edge weights; and combining the quantum topology graph with the... The radiation spectrum is input into the event branch and radiation branch of the graph neural network, respectively, to extract quantum topological features and radiation spectrum features. Through a cross-modal graph attention mechanism with learnable edge weights, the quantum topological features and radiation spectrum features are fused into cloud quantum-radiation joint features. The polar ionospheric scintillation index is obtained and mapped to a graph reparameterization vector. Based on the graph reparameterization vector, the edge weights of the cloud quantum-radiation joint features are dynamically corrected to generate perturbation adaptive features. Based on the perturbation adaptive features, the vertical distribution probability of clouds is predicted to obtain a cloud vertical distribution probability map.

[0050] It should be noted that in this embodiment, the preset frequency is 35 GHz, and the preset frequency band is 0.3–3 THz. That is, the quantum radar transmits entangled microwave pulses at a frequency of 35 GHz, the receiver receives the quantum radar echo through a quantum mixer, and the terahertz radiometer simultaneously scans the brightness temperature spectrum T in the 0.3–3 THz frequency band. B (ν), which is the terahertz radiation spectrum.

[0051] For each range gate, i.e., the resolution cell, quantum state tomography is performed on the quantum radar echo to reconstruct its density matrix ρ. By calculating the eigenvalues ​​of ρ, the quantum entropy of the cell and its entanglement with neighboring cells are obtained, as follows:

[0052]

[0053] in, denoted as quantum entropy, which characterizes the degree of uncertainty of a quantum state, and C as entanglement, which reflects the degree of quantum correlation between the unit and its neighboring units.

[0054] Quantum dot clouds are generated based on the density matrix ρ, where each dot represents a resolution unit, and its position is determined by radar coordinates. The attributes include phase, amplitude, and entanglement degree C.

[0055] A quantum topology graph is constructed using quantum dot clouds as nodes and entanglement degree C as edge weights. This graph is a graph structure that intuitively reflects the quantum correlation characteristics between resolvable units in a quantum radar echo. The connection strength between nodes is quantified by the entanglement degree C; a higher entanglement degree indicates a stronger quantum correlation between the corresponding resolvable units. In the quantum topology graph, each node represents a resolvable unit, and edges are connected through K-nearest neighbors, where K=8, and the edge weight is the entanglement degree C between the two nodes.

[0056] Simultaneously, a straight line T is fitted to the terahertz radiation spectrum TB(ν). B =aν+b, then the spectral slope s=a. A radiation spectrum is constructed using frequency points as nodes and slope k as edge weights. The radiation spectrum is a graph structure that intuitively reflects the variation characteristics of the terahertz radiation spectrum across different frequency points. The connection strength between nodes is quantified by the spectral slope s; the greater the change in spectral slope, the more significant the difference in radiation characteristics between corresponding frequency points. In the radiation spectrum, each node is a frequency sampling point, and edges connect adjacent frequency points, with the edge weight being the spectral slope s between the two nodes.

[0057] In this embodiment, the graph neural network comprises two core components: an event branch and a radiation branch. The event branch processes the quantum topology graph, aggregating node information layer by layer through graph convolution operations to capture the complex correlation characteristics between resolvable units in the quantum radar echo. This branch first initializes the features of each node in the quantum topology graph, encoding attributes such as phase, amplitude, and entanglement into high-dimensional vectors. Subsequently, through multiple layers of graph convolution operations, it dynamically adjusts the information transmission strength between nodes using edge weights, i.e., entanglement, gradually aggregating local features to form global quantum topology features. The radiation branch is designed for the radiation spectrum, employing a similar structure to extract frequency domain correlation features of the terahertz radiation spectrum. It generates radiation spectrum features by fusing node information guided by spectral slope edge weights. The two branches are processed in parallel through independent graph convolutional layers, ultimately outputting dimension-matched quantum topology feature vectors and radiation spectrum feature vectors.

[0058] The cross-modal graph attention mechanism dynamically generates attention weights by calculating the similarity matrix between quantum topological features and radiation spectrum features, thereby achieving adaptive fusion between features, as shown in the following equation:

[0059] in, Let w be the attention weight of node i in the quantum topological graph to node j in the radiation spectrum graph, and w be a learnable weight vector. and Let be the feature vectors of node i and node j, respectively. This represents a vector concatenation operation. Let i be the set of neighboring nodes of node i.

[0060] This mechanism dynamically adjusts the fusion ratio between different modal features by using attention weights, making the fusion process more targeted and flexible.

[0061] Based on the calculated attention weights, the quantum topological features and radiation spectrum features are weighted and summed to generate a fused cloud quantum-radiation joint feature. This feature integrates the quantum correlation characteristics of quantum radar echoes and the frequency domain variation characteristics of terahertz radiation spectra.

[0062] The polar ionospheric scintillation index, which reflects the degree of electromagnetic wave perturbation by the polar ionosphere, is obtained. This index is mapped to a graph reparameterized vector and encoded into a high-dimensional vector form compatible with graph structures through a nonlinear transformation. This vector serves as external perturbation information, used to dynamically correct the edge weights of the cloud quantum-radiative joint characteristics.

[0063] Specifically, the graph reparameterized vector is fused with the edge weights in the cloud quantum-radiation joint features. Based on the real-time changes in the polar ionospheric scintillation index, the weights of different edges in the feature transfer process are adjusted. For example, when the ionospheric scintillation index is large, indicating strong electromagnetic disturbance, the weights of the more affected edges are appropriately reduced to minimize interference with feature fusion; conversely, when the scintillation index is small, the edge weights are kept relatively stable. This dynamic correction method generates a disturbance-adaptive feature, which better adapts to the complex and variable electromagnetic environment of the polar region, providing a more accurate and reliable input for subsequent prediction of cloud vertical distribution probability.

[0064] The perturbation adaptive features are input into the pre-trained cloud vertical distribution prediction model, which outputs the probability of cloud presence at each altitude level. A cloud vertical distribution probability map is constructed, with the vertical axis representing altitude and the horizontal axis representing probability values. Color-coded maps display different probability intervals.

[0065] By introducing quantum topology map construction technology and ionospheric scintillation dynamic correction mechanism, the interference of ionospheric disturbances on the detection signal during the polar night is effectively suppressed, solving the technical problem of poor reliability of traditional microwave radar in polar environments.

[0066] In one feasible implementation, constructing a quantum topology graph using the quantum dot cloud as nodes and the entanglement degree as edge weights includes: projecting the quantum dot cloud into a four-dimensional feature space composed of spatial coordinates and quantum entropy; using the K-nearest neighbor algorithm to find a preset number of neighboring nodes for each node in the four-dimensional feature space to construct an initial undirected graph; calculating the entanglement degree similarity between two nodes on each edge in the initial undirected graph to generate an initial edge weight set; comparing the initial edge weight set with a preset entanglement degree threshold and deleting edges in the initial edge weight set whose entanglement degree similarity is lower than the preset entanglement degree threshold to form a sparse quantum topology graph; using a graph attention network to aggregate and update the node features in the sparse quantum topology graph to generate enhanced node features; recalculating the correlation strength between nodes using the enhanced node features, dynamically updating the edge weights, and outputting a dynamically evolving quantum topology graph.

[0067] It should be noted that this implementation maps the quantum dot cloud to a four-dimensional space, constructs a graph using K-nearest neighbors, prunes based on entanglement, and dynamically updates edge weights using graph attention, ultimately obtaining a dynamically evolving quantum topology graph.

[0068] Understandably, the construction of the four-dimensional feature space effectively integrates spatial location and quantum state information. Spatial coordinates ensure the accurate positioning of nodes in physical space, while quantum entropy reflects the quantum state complexity of each resolvable unit. The K-nearest neighbor algorithm selects eight nearest neighbors for each quantum point by calculating the Euclidean distance between nodes, constructing an initial undirected graph structure. The initial edge weight set is generated by calculating the entanglement similarity of the density matrices of adjacent nodes; the closer this value is to 1, the more similar the quantum states of the two nodes are. A preset entanglement threshold of 0.7 is set, retaining only edges with similarity higher than this value, ensuring that the quantum topology graph retains the most quantum-related connections.

[0069] The graph attention network employs a multi-head attention mechanism, where each attention head independently calculates the attention weights between nodes. Enhanced node features are obtained by concatenating the output features of each head and performing a linear transformation.

[0070] During the dynamic edge weight update process, the new edge weight is determined by the weighted sum of the cosine similarity of the enhanced node features and the initial entanglement degree, with weight coefficients set to 0.6 and 0.4 respectively. This allows the quantum topology graph to reflect both static quantum correlations and capture dynamic evolutionary characteristics. The final output dynamic quantum topology graph is updated every 10 minutes, with a node feature dimension of 128 and edge weights normalized to the [0,1] interval, enabling the graph structure to better adapt to complex scenarios with different cloud layers.

[0071] In one feasible implementation, a radiation spectrum is constructed using the terahertz radiation spectrum as nodes and the spectral slope as edge weights. This includes: sampling the terahertz radiation spectrum at specific frequencies, treating each sampling point as a node in the radiation spectrum, with the node position determined by the frequency index; calculating the spectral slope between adjacent frequency points, using this slope value as the edge weight connecting the two nodes, the spectral slope being obtained by linearly fitting the radiation spectrum data, and its value reflecting the rate of change of radiation intensity with frequency; after constructing the initial radiation spectrum, normalizing the edge weights to map the spectral slope values ​​to the [0,1] interval, eliminating the influence of radiation intensity differences in different frequency bands on the graph structure; and employing a threshold filtering strategy to delete weak connections with edge weights less than a preset threshold, such as 0.2, while retaining frequency point associations with significant differences in radiation characteristics, thus forming a sparse radiation spectrum.

[0072] Step S40: Based on the cloud probability distribution map and combined with the physical-guided thickness inversion model, generate a polar day optical thickness map.

[0073] It should be noted that the physically guided thickness inversion model is a physical model that takes polarization thermal characteristics as input and optical thickness as output. This model fully considers the radiative transfer characteristics of clouds under polar day conditions and the intrinsic relationship between polarization information and optical thickness.

[0074] Understandably, the initial optical thickness map output by the cloud probability distribution map and the physical-guided thickness inversion model is fused by Bass fusion. Through the Bass fusion algorithm, information from two different sources is effectively integrated to generate the polar day optical thickness map.

[0075] Step S50: Based on the cloud vertical distribution probability map and combined with radar range-Doppler constraints, generate a three-dimensional cloud mask for polar night.

[0076] It should be noted that radar range-Doppler constraint refers to generating a velocity divergence constraint field using range-Doppler cube data synchronously acquired by quantum radar. The velocity divergence constraint field is used as the physical loss to iteratively optimize the initial voxel probability volume obtained based on the cloud vertical distribution probability map. The optimized voxel probability volume is then binarized to generate a three-dimensional cloud mask for polar night.

[0077] In one feasible implementation, step S50 may include: inputting the cloud vertical distribution probability map into a three-dimensional U-shaped network for edge enhancement and noise suppression, and outputting an optimized three-dimensional probability field; acquiring range-Doppler cube data synchronously collected by quantum radar, and generating a velocity divergence constraint field based on the range-Doppler cube data; constructing a learnable constant diagonal structure projection matrix, mapping the optimized three-dimensional probability field onto a voxel grid of a preset resolution, and generating an initial voxel probability volume; using the velocity divergence constraint field as a physical consistency loss function, iteratively optimizing the initial voxel probability volume to obtain an optimized voxel probability volume; and performing stochastic gradient-based discretization and connected component analysis on the optimized voxel probability volume to generate a polar night three-dimensional cloud mask.

[0078] It should be noted that the 3D U-Net is a network that takes the vertical distribution probability map of clouds as input and performs edge enhancement and noise suppression through an encoder-decoder structure and skip connections. The output is an optimized 3D probability field, which has significant improvements in both spatial resolution and feature representation.

[0079] From the range-Doppler cube data acquired by quantum radar, the radial velocity field V_r(x,y,z) is extracted, and its velocity divergence field is then calculated. The specific calculation method of the velocity divergence field is to perform partial derivative operations on the radial velocity field in three spatial dimensions and then sum them. That is, by taking the partial derivatives of V_r(x,y,z) in the x, y, and z directions respectively, and then adding the partial derivatives in the three directions, the velocity divergence constraint field Div(V_r) is obtained. This constraint field can reflect the motion change trend of cloud particles in space and provides a key basis for the subsequent physical consistency loss function.

[0080] When constructing a learnable constant diagonal structure projection matrix, the dimension of the matrix is ​​determined according to the voxel grid of the preset resolution. The elements in the matrix are initialized through learnable parameters and continuously optimized and adjusted during the training process to achieve the goal of accurately mapping the optimized three-dimensional probability field to the voxel grid.

[0081] Using the velocity divergence constraint field as the physical consistency loss function, the initial voxel probability volume is iteratively optimized by continuously adjusting the probability values ​​within the voxel probability volume. This ensures that the optimized voxel probability volume, while satisfying physical laws, more accurately reflects the actual cloud distribution. During the iterative optimization process, optimization algorithms such as gradient descent are employed to progressively update the parameters of the voxel probability volume based on the gradient information of the physical consistency loss function, until the loss function reaches its minimum or a preset stopping condition is met.

[0082] When performing discretization and connected component analysis based on stochastic gradients, the voxel probability volume is first discretized using stochastic gradient information, converting continuous probability values ​​into discrete category labels, usually set to 0 or 1, representing no cloud and cloud, respectively. Then, through the connected component analysis algorithm, small noise points are removed from the discretized voxels, and the main cloud clusters are retained to generate the final three-dimensional cloud mask for polar night.

[0083] Step S60: Generate full-time detection results of polar cloud layers based on the polar day optical thickness map and the polar night three-dimensional cloud mask.

[0084] It should be noted that the polar day optical thickness map and the polar night 3D cloud mask are unified into a single grid and then weighted and fused according to the solar altitude angle to generate a full-time cloud product with thickness levels.

[0085] Understandably, the unified grid operation transforms the polar day optical thickness map and the polar night 3D cloud mask to the same spatial resolution and grid coordinate system, ensuring precise spatial correspondence between the two. When weighted by solar altitude angle, the solar altitude angle serves as a key factor. During the polar day, when the solar altitude angle is high, the polar day optical thickness map dominates the fusion result, providing accurate information on cloud optical thickness that reflects the absorption and scattering characteristics of solar radiation by the clouds. During the polar night, when the solar altitude angle is negative or zero, the polar night 3D cloud mask plays a major role in the fusion result, presenting reliable information on cloud spatial distribution and morphology for polar night cloud detection. This weighted fusion method fully combines the advantages of cloud detection during different periods of the polar day and polar night, generating a full-time cloud product with thickness levels.

[0086] In one feasible implementation, step S60 may include: multiplying the polar night three-dimensional cloud mask with the optimized three-dimensional probability field on a voxel-by-voxel basis to obtain the polar night optical thickness volume; and resampling the polar night optical thickness volume and the polar day optical thickness map to a common grid to obtain a common thickness volume. Using the solar altitude angle as the fusion weight, the common thickness volume is adaptively weighted and fused to obtain a fused thickness volume; the fused thickness volume and the polar night 3D cloud mask are logically ANDed to generate an all-time cloud mask; the all-time cloud mask is classified into thickness levels based on the fused thickness volume to obtain the all-time detection results of polar clouds.

[0087] It should be noted that multiplying the three-dimensional cloud mask in polar night with the optimized three-dimensional probability field voxel by voxel can utilize the probability information in the three-dimensional probability field to finely adjust the cloud mask, thereby obtaining a more accurate polar night optical thickness volume, which can more accurately reflect the optical characteristics of the cloud layer under polar night conditions.

[0088] The two-dimensional polar day optical thickness map is expanded into a three-dimensional thickness volume by vertical interpolation, assuming a cloud thickness of 1 km. Then, the three-dimensional thickness volume and the polar night optical thickness volume are resampled to the same latitude-longitude-height common grid, such as 0.1°×0.1°×0.1km, ensuring complete consistency in spatial resolution and grid coordinates, thus obtaining the common thickness volume. This unified resampling operation provides a data foundation of the same dimensions for subsequent weighted fusion, avoiding fusion errors caused by inconsistent spatial resolution.

[0089] When using the solar elevation angle as the fusion weight for adaptive weighting, the influence of solar radiation on cloud detection during different periods of polar day and polar night is fully considered. During the polar day, the solar elevation angle is larger, and the optical characteristics of clouds reflected in the polar day optical thickness map are more significant. Therefore, it is given a larger weight in the fusion process, so that the fusion result can highlight the absorption and scattering characteristics of clouds on solar radiation during the polar day. During the polar night, the solar elevation angle is smaller or even negative. At this time, the spatial distribution information of clouds presented by the polar night 3D cloud mask becomes crucial, so it is given a larger weight to ensure that the fusion result can accurately reflect the actual distribution of clouds during the polar night. Through this adaptive weighting fusion method, the proportion of the polar day optical thickness map and the polar night optical thickness volume in the fusion result can be dynamically adjusted according to the characteristics of solar radiation at different times, thereby generating a fused thickness volume that is more consistent with the actual situation.

[0090] A logical AND operation is performed between the fused thickness volume and the 3D cloud mask for polar night conditions. The rule of the logical AND operation is that the result is "cloudy" only if the voxels or pixels at corresponding positions in both operations meet certain conditions, such as both indicating a cloudy state; otherwise, it is "cloudless." Through this operation, the cloud information provided by both operations can be combined to further remove possible noise and misjudgments, generating a more accurate all-time cloud mask. This cloud mask can clearly identify the spatial distribution of clouds at different times.

[0091] Based on the fused thickness volume, the cloud mask is classified into different thickness levels for all time periods. According to the pre-set thickness threshold range, the cloud mask is divided into different levels based on the thickness value in the fused thickness volume, such as thin cloud, medium cloud, thick cloud, etc., and finally the detection results with thickness level labels are output.

[0092] This embodiment provides a polar cloud detection method based on day-night adaptive multimodal fusion. The polar time period mode is determined according to the solar altitude angle and ambient light intensity of the polar observation point. When the polar time period mode is polar day mode, multi-angle polarized thermal image sequences and visible light images are acquired, and a cloud probability distribution map is generated based on these images. When the polar time period mode is polar night mode, quantum radar echoes and terahertz radiation spectra are acquired, and a vertical cloud distribution probability map is generated based on these images. Based on the cloud probability distribution map, a polar day optical thickness map is generated using a physics-guided thickness inversion model. Based on the vertical cloud distribution probability map, a polar night three-dimensional cloud mask is generated using radar range-Doppler constraints. Finally, a full-time polar cloud detection result is generated based on the polar day optical thickness map and the polar night three-dimensional cloud mask. Through this method, by introducing sensing technologies such as quantum radar and terahertz radiation spectra, and combining them with deep learning and quantum information processing methods, high-precision, all-weather detection of polar clouds is achieved.

[0093] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, step S40 includes steps S401 to S404: Step S401: Generate a simulation dataset of cloud optical thickness based on the physical radiative transfer model.

[0094] It should be noted that the physical radiative transfer model is a mathematical model used to describe the process of light propagating in a medium. It takes into account the interaction between light and the medium, such as absorption and scattering.

[0095] In the process of generating cloud optical thickness simulation datasets, the physical radiation transfer model can simulate the propagation of light under different cloud conditions, such as cloud thickness, cloud droplet size distribution, and cloud water content, thereby generating simulation data that is similar to the actual optical characteristics of clouds.

[0096] Step S402: Train a physically guided thickness inversion neural network based on the cloud optical thickness simulation dataset to obtain the trained thickness inversion neural network.

[0097] It should be noted that the physics-guided thickness inversion neural network is a neural network structure that combines a physical model with deep learning techniques. During training, the network uses a simulated cloud optical thickness dataset as input and learns the mapping relationship from input data to cloud optical thickness through nonlinear transformations of multiple layers of neurons. Simultaneously, the physical model is integrated as guiding information into the neural network training process, ensuring that the network not only focuses on the fit of the data but also follows physical laws, thereby improving the accuracy and reliability of the inversion. Through training with a large amount of simulated data, the physics-guided thickness inversion neural network can learn the complex characteristics of cloud optical properties, providing strong support for subsequent cloud optical thickness inversion based on actual observation data.

[0098] In the specific implementation, a U-Net neural network is constructed. Using a simulated cloud optical thickness dataset as input and optical thickness as the supervisory signal, the network is trained. During training, prior knowledge provided by the physical radiative transfer model is used to constrain and optimize the parameters of each layer of the neural network. Fundamental laws regarding cloud optical properties in the physical model, such as the influence of different cloud droplet sizes on optical thickness and the relationship between cloud water content and optical thickness, are transformed into mathematical constraints and incorporated into the loss function of the neural network. Thus, while learning the data mapping relationship, the neural network must satisfy these physical constraints, ensuring that the inversion results not only fit the simulation data well but also conform to physical reality. Through continuous iterative training and adjustment of network parameters, the neural network's ability to invert cloud optical thickness gradually improves, ultimately resulting in a trained thickness inversion neural network capable of accurately and reliably inverting cloud optical thickness information based on input observation data.

[0099] Step S403: Input the polarization thermal feature tensor into the trained thickness inversion neural network for forward inference and output the initial optical thickness map.

[0100] It should be noted that the polarization thermal feature tensor contains key information about clouds in multi-angle polarization thermal image sequences, such as radiation intensity and polarization characteristics at different angles. This information reflects the physical structure and optical properties of clouds and is an important basis for inverting cloud optical thickness.

[0101] The acquired multi-angle polarization thermal image sequences were preprocessed to extract polarization thermal feature tensors, which were then input into a trained thickness inversion neural network. The neural network, through a forward inference process, utilizes the data mapping relationships and physical constraints learned during training to analyze and process the input polarization thermal feature tensors, ultimately outputting an initial optical thickness map. This initial optical thickness map preliminarily reflects the optical thickness distribution of clouds under polar day conditions.

[0102] By combining the physical radiative transfer model with a deep learning network, the thickness inversion results not only conform to physical laws but also possess the flexibility of data-driven approaches, overcoming the shortcomings of high complexity in pure physical models and poor physical consistency in pure data-driven models.

[0103] Step S404: Perform Bayesian fusion of the cloud probability distribution map and the initial optical thickness map to obtain the polar day optical thickness map.

[0104] It should be noted that Bayesian fusion is a fusion method based on probability and statistics theory. It can organically combine information from different sources to obtain more accurate and reliable fusion results.

[0105] In generating the polar day optical thickness map, the cloud class probability distribution map provides the probability information of clouds belonging to different classes at different locations, while the initial optical thickness map gives a preliminary estimate of the optical thickness of clouds at each location. By organically combining these two through Bayesian fusion, the probability of cloud class and the preliminary estimate of optical thickness can be comprehensively considered, thus obtaining a more accurate polar day optical thickness map.

[0106] Specifically, the Bayesian fusion process weights and adjusts the pixel value in the initial optical thickness map based on the probability information in the cloud class probability distribution map. For a given location, if the probability of it belonging to a certain cloud class is high, the fusion process will give more consideration to the optical thickness features corresponding to that cloud class, thus adjusting the initial optical thickness value for that location accordingly. This fusion method fully leverages the advantages of both the cloud class probability distribution map and the initial optical thickness map, effectively overcoming the limitations of a single data source and further improving the accuracy and reliability of the polar day optical thickness map.

[0107] The polar day optical thickness map obtained after Bayesian fusion processing can more accurately reflect the optical characteristics of clouds during the polar day period, providing a more solid data foundation for subsequent all-time polar cloud detection. This polar day optical thickness map not only includes cloud optical thickness information but also implicitly contains cloud category information because the influence of cloud category probability distribution maps is fully considered during the Bayesian fusion process. This fusion method significantly improves the spatial resolution and accuracy of the polar day optical thickness map, enabling a more detailed depiction of the complex structure and optical characteristics of clouds during the polar day period. For example, in the case of multi-layered or mixed clouds, the polar day optical thickness map can clearly distinguish the differences in optical thickness of different cloud layers and their superposition relationships, providing richer and more accurate information for subsequent cloud analysis and forecasting. Furthermore, this polar day optical thickness map also exhibits good stability and robustness, maintaining high accuracy and reliability under different observation conditions and environments, providing strong data support for all-time polar cloud detection.

[0108] In one feasible implementation, step S404 may include: performing superpixel segmentation on the initial optical thickness map and calculating the optical thickness variance within each superpixel region to generate a local homogeneity map; comparing the local homogeneity map with a preset homogeneity threshold to identify anomaly region masks; extracting visible light texture gradients and polarization angle anisotropy within the anomaly region masks to generate a correction weight map; multiplying the correction weight map pixel-by-pixel with the initial optical thickness map to obtain an anomaly correction thickness map; and performing Bayesian fusion of the cloud probability distribution map and the anomaly correction thickness map to obtain a polar day optical thickness map.

[0109] It should be noted that the SLIC algorithm is used to perform superpixel segmentation on the initial optical thickness map. For each superpixel region, the variance of its internal optical thickness is calculated. The smaller the variance value, the higher the homogeneity of the region, and vice versa. This generates a local homogeneity map. This map can intuitively present the homogeneity of different regions in the initial optical thickness map, providing a foundation for the subsequent identification of abnormal regions.

[0110] A homogeneity threshold is set, and all superpixels in the local homogeneity map are traversed. The corresponding optical thickness variance is compared with the preset homogeneity threshold. If the variance value of a certain region exceeds the threshold, the region is determined to be an abnormal region, and an abnormal region mask is generated. This mask clarifies the regions in the initial optical thickness map that need further correction.

[0111] Within the mask of the anomalous region, two key features—visible light texture gradient and polarization anisotropy—are extracted. The visible light texture gradient reflects the variations in cloud surface texture; different textures may correspond to different cloud structures and compositions. Polarization anisotropy reflects the differences in how clouds affect the polarization characteristics of light, and is closely related to the microscopic physical properties of clouds. Using specific algorithms and models, these two features are accurately extracted from the anomalous region, and a correction weight map is generated based on their influence on cloud optical thickness. Each pixel value in the correction weight map represents the degree of correction to the corresponding pixel value in the initial optical thickness map.

[0112] The corrected weight map is multiplied pixel-by-pixel with the initial optical thickness map. This process is equivalent to finely adjusting each pixel value in the initial optical thickness map based on the corrected weight map, making the optical thickness values ​​in abnormal areas more consistent with reality, thus obtaining an anomaly corrected thickness map. The corrected thickness map is more accurate in representing the optical thickness in abnormal areas, effectively correcting any errors that may exist in the initial optical thickness map.

[0113] Finally, the cloud class probability distribution map and the anomaly-corrected thickness map are fused using Bayesian fusion. During the fusion process, the cloud class probability distribution map and the corrected optical thickness value from the anomaly-corrected thickness map are fully considered. Using a Bayesian fusion algorithm, the optical thickness at each location is re-evaluated and calculated by integrating these two pieces of information, resulting in the final polar day optical thickness map. This fusion method maximizes the advantages of various data sources, effectively reduces errors and uncertainties, and achieves higher levels of accuracy and reliability in the generated polar day optical thickness map.

[0114] In one feasible implementation, the step of performing Bayesian fusion of the cloud class probability distribution map and the anomaly correction thickness map to obtain the polar day optical thickness map includes: constructing a prior optical thickness distribution model corresponding to each cloud class based on historical observation data of different cloud classes; using the anomaly correction thickness map as the observation value and combining it with the cloud class probability distribution map, calculating the thickness likelihood function of each pixel belonging to each cloud class; determining the posterior optical thickness distribution of each pixel based on Bayes' theorem according to the prior optical thickness distribution model and the thickness likelihood function; extracting the maximum posterior probability from the posterior optical thickness distribution of each pixel as the optimal optical thickness estimate of the corresponding pixel, generating a maximum posterior probability thickness map, and performing guided filtering on the maximum posterior probability thickness map to generate the polar day optical thickness map.

[0115] It should be noted that, based on historical databases, the prior distributions of optical thickness for different cloud types, such as cirrus, cumulus, and stratus, are statistically analyzed. Assuming they follow a log-normal distribution, parameter estimation methods, such as maximum likelihood estimation, are used to determine the parameters of the optical thickness distribution model corresponding to each cloud type, including the mean and variance, thereby constructing a prior optical thickness distribution model for each cloud type. These models can describe the statistical characteristics and distribution patterns of optical thickness for different cloud types, providing important prior information for subsequent Bayesian fusion. The formula for the optical thickness distribution model is:

[0116] in, This indicates that the cloud type is known to be... Under these conditions, optical thickness The prior probability density, For cloud optical thickness, These are cloud-type identifiers, including cirrus (Ci), cumulus (Cu), stratus (St), etc. For cloud The corresponding log-normal distribution location parameters, For cloud The corresponding log-normal distribution scaling parameter.

[0117] Using each pixel value in the anomaly-corrected thickness map as an observation, and assuming Gaussian noise in the observations, a corresponding thickness likelihood function can be constructed for each cloud class in the cloud class probability distribution map. This function describes the probability density distribution of a pixel belonging to a certain cloud class given the observations. The likelihood function combines the observation information provided by the anomaly-corrected thickness map with the assumed noise model, quantifying the degree of matching between each pixel and each cloud class through statistical and probabilistic methods. The thickness likelihood function is as follows:

[0118] in, Indicates that in the real cloud category Under the premise of observing abnormal correction thickness values The likelihood probability, To correct the observed optical thickness value of a pixel in the thickness map, For cloud Typical optical thickness average, The observation noise standard deviation is set for cloud type c, reflecting the inversion uncertainty.

[0119] Based on Bayes' theorem, a prior optical thickness distribution model is combined with a thickness likelihood function. Bayes' theorem provides a method for updating posterior probabilities given prior information and observational data. The formula for calculating the Bayesian posterior probability is as follows:

[0120] in, Indicates the observed thickness Under these conditions, the pixel belongs to the cloud category. The posterior probability, This pixel belongs to the cloud category. The prior class probabilities are derived from the cloud class probability distribution map. This is a collection of all cloud types.

[0121] For each pixel, select to make The largest cloud category The mean of the prior distribution corresponding to the cloud type is used as the optimal optical thickness estimate for the pixel, and a maximum posterior probability thickness map is generated.

[0122] Guided filtering is applied to the maximum a posteriori (MAP) thickness map. Guided filtering is an image filtering method based on a local linear model that can smooth images while preserving edges. Anomaly-corrected thickness maps are used as guide maps for filtering the MAP thickness map. Guided filtering utilizes the local structural information of the guide map to guide the filtering process, enabling the MAP thickness map to better preserve the edge and detail features of the cloud structure during smoothing.

[0123] Specifically, within each local window, guided filtering establishes a linear relationship between the output image and the guided map, adjusting the filtering weights based on the gradient information of the guided map. This smooths noise while avoiding excessive blurring of important structures such as cloud edges. After guided filtering, a final polar day optical thickness map is generated. This map not only has high accuracy and reliability but also clearly presents the complex structure and optical characteristics of clouds during the polar day period in terms of spatial resolution, providing crucial data support for subsequent all-time polar cloud detection, analysis, and forecasting.

[0124] In this embodiment, by combining the physical radiation transfer model with a neural network, the accuracy and reliability of optical thickness inversion are improved, and the quality of the polar day optical thickness map is further enhanced by combining Bayesian fusion.

[0125] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the polar cloud detection method based on day-night adaptive multimodal fusion. Any simple modifications based on this technical concept are within the protection scope of this application.

[0126] This application also provides a polar cloud detection device based on day-night adaptive multimodal fusion, please refer to... Figure 2 The polar cloud detection device based on day-night adaptive multimodal fusion includes: Module 10 is used to determine the polar time period mode based on the solar altitude angle and ambient light intensity of the polar observation point.

[0127] The acquisition module 29 is used to acquire multi-angle polarized thermal image sequences and visible light images when the polar time mode is polar day mode, and generate a cloud probability distribution map based on the multi-angle polarized thermal image sequences and the visible light images.

[0128] The acquisition module 20 is also used to acquire quantum radar echoes and terahertz radiation spectra when the polar time mode is polar night mode, and generate a cloud vertical distribution probability map based on the quantum radar echoes and terahertz radiation spectra.

[0129] The generation module 30 is used to generate a polar day optical thickness map based on the cloud probability distribution map and in combination with a physically guided thickness inversion model.

[0130] The generation module 30 is also used to generate a three-dimensional cloud mask for polar night based on the cloud vertical distribution probability map and combined with radar range-Doppler constraints.

[0131] The generation module 30 is also used to generate full-time detection results of polar clouds based on the polar day optical thickness map and the polar night three-dimensional cloud mask.

[0132] The polar cloud detection device based on day-night adaptive multimodal fusion provided in this application, employing the polar cloud detection method based on day-night adaptive multimodal fusion described in the above embodiments, can solve the technical problems of existing technologies being unable to adapt to the special polar environment, having poor detection continuity, and insufficient accuracy. Compared with the prior art, the beneficial effects of the polar cloud detection device based on day-night adaptive multimodal fusion provided in this application are the same as those of the polar cloud detection method based on day-night adaptive multimodal fusion provided in the above embodiments, and other technical features in the polar cloud detection device based on day-night adaptive multimodal fusion are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0133] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A polar cloud detection method based on diurnal self-adaptive multi-modal fusion, characterized in that, The method comprises: determining a polar time period mode according to a solar elevation angle of a polar observation point and an ambient light intensity; in a case where the polar time period mode is a polar day mode, collecting a multi-angle polarized thermal image sequence and a visible light image, and generating a cloud class probability distribution map based on the multi-angle polarized thermal image sequence and the visible light image; in a case where the polar time period mode is a polar night mode, collecting quantum radar echoes and terahertz radiation spectra, and generating a cloud layer vertical distribution probability map based on the quantum radar echoes and the terahertz radiation spectra; generating a polar day optical thickness map based on the cloud class probability distribution map and a physically guided thickness inversion model; generating a polar night three-dimensional cloud mask based on the cloud layer vertical distribution probability map and a radar range-Doppler constraint; generating a polar cloud layer full-time period detection result based on the polar day optical thickness map and the polar night three-dimensional cloud mask.

2. The method of claim 1, wherein, The method comprises: in a case where the polar time period mode is a polar day mode, collecting a multi-angle polarized thermal image sequence and a visible light image output by a polarized thermal imaging camera and a full-sky imager, respectively; performing radiation calibration and non-linear response correction on the multi-angle polarized thermal image sequence to generate a set of corrected polarized thermal images; extracting thermal radiation intensity images collected at a plurality of different polarization directions from the set of corrected polarized thermal images to obtain a plurality of registered polarization direction intensity images; calculating Stokes vectors pixel by pixel based on the plurality of registered polarization direction intensity images to generate a Stokes parameter map; calculating a polarization degree, a polarization angle and a linear polarization difference feature of each pixel based on the Stokes parameter map to generate a three-channel polarization feature map; splicing the Stokes parameter map and the three-channel polarization feature map in a channel dimension to form a multi-channel polarization feature set; performing channel normalization and spatial alignment on the multi-channel polarization feature set to obtain a polarized thermal feature tensor; inputting the polarized thermal feature tensor and the visible light image into a multi-modal fusion classification network for cloud class recognition to obtain a cloud class probability distribution map.

3. The method of claim 1, wherein, The method comprises: sending entangled microwave pulses of a preset frequency to the sky by a quantum radar, receiving quantum radar echoes reflected back, and synchronously collecting terahertz radiation spectra of a preset frequency band by a terahertz passive radiometer; performing quantum state tomography processing on the quantum radar echoes to reconstruct a density matrix of each resolution unit, and generating quantum point clouds containing phases, amplitudes and entanglement degrees based on the density matrix; constructing a quantum topological graph with the quantum point clouds as nodes and the entanglement degrees as edge weights, and constructing a radiation spectrum graph with the terahertz radiation spectra as nodes and spectral slopes as edge weights; The quantum topology graph and the radiation spectrum graph are respectively input into an event branch and a radiation branch of a graph neural network to extract quantum topology features and radiation spectrum features; The quantum topology features and the radiation spectrum features are fused into cloud layer quantum-radiation joint features through a cross-modal graph attention mechanism of a learnable edge weight; A polar ionospheric scintillation index is obtained, and the polar ionospheric scintillation index is mapped into a graph reparameterization vector; Edge weight dynamic correction is performed on the cloud layer quantum-radiation joint features based on the graph reparameterization vector to generate a perturbation adaptive feature; Cloud layer vertical distribution probability prediction is performed based on the perturbation adaptive feature to obtain a cloud layer vertical distribution probability graph.

4. The method of claim 3, wherein, The quantum topology graph is constructed by taking the quantum dot cloud as a node and the entanglement degree as an edge weight, and includes the following steps: The quantum dot cloud is projected into a four-dimensional feature space composed of spatial coordinates and quantum entropy; In the four-dimensional feature space, a K-nearest neighbor algorithm is used to find a preset number of adjacent nodes for each node to construct an initial undirected graph; The entanglement degree similarity between two nodes on each edge in the initial undirected graph is calculated to generate an initial edge weight set; The initial edge weight set is compared with a preset entanglement degree threshold, and edges with an entanglement degree similarity lower than the preset entanglement degree threshold in the initial edge weight set are deleted to form a sparse quantum topology graph; A graph attention network is used to aggregate and update the node features in the sparse quantum topology graph to generate enhanced node features; The correlation strength between nodes is recalculated using the enhanced node features, and the edge weight is dynamically updated to output a dynamically evolved quantum topology graph.

5. The method of claim 1, wherein, The polar day optical thickness map is generated based on the cloud class probability distribution graph and a physically guided thickness inversion model, and includes the following steps: A cloud layer optical thickness simulation dataset is generated based on a physical radiation transfer model; A physically guided thickness inversion neural network is trained based on the cloud layer optical thickness simulation dataset to obtain a trained thickness inversion neural network; A polarized thermal feature tensor is input into the trained thickness inversion neural network for forward inference to output an initial optical thickness map; The cloud class probability distribution graph and the initial optical thickness map are Bayes fused to obtain a polar day optical thickness map.

6. The method of claim 5, wherein, The cloud class probability distribution graph and the initial optical thickness map are Bayes fused to obtain a polar day optical thickness map, and include the following steps: Superpixel segmentation is performed on the initial optical thickness map, and the optical thickness variance in each superpixel region is calculated to generate a local homogeneity map; The local homogeneity map is compared with a preset homogeneity threshold to identify an abnormal region mask; Visible light texture gradients and polarization angle anisotropy are extracted in the abnormal region mask to generate a correction weight map; The correction weight map and the initial optical thickness map are multiplied pixel by pixel to obtain an abnormal correction thickness map; The cloud class probability distribution graph and the abnormal correction thickness map are Bayes fused to obtain a polar day optical thickness map.

7. The method of claim 6, wherein, The cloud class probability distribution graph and the abnormal correction thickness map are Bayes fused to obtain a polar day optical thickness map, and include the following steps: construct a prior optical thickness distribution model corresponding to each cloud class based on historical observation data of different cloud classes; take the abnormal correction thickness map as an observation value, and combine the cloud class probability distribution map to calculate a thickness likelihood function of each pixel belonging to each cloud class; determine a posteriori optical thickness distribution of each pixel based on the prior optical thickness distribution model and the thickness likelihood function according to Bayes' theorem; extract the maximum a posteriori probability from the posteriori optical thickness distribution of each pixel as an optimal optical thickness estimation value of the corresponding pixel to generate a maximum a posteriori probability thickness map; perform guided filtering on the maximum a posteriori probability thickness map to generate a polar day optical thickness map.

8. The method of claim 1, wherein, The polar night three-dimensional cloud mask is generated based on the cloud layer vertical distribution probability map in combination with a radar range-Doppler constraint, including: inputting the cloud layer vertical distribution probability map into a three-dimensional U-shaped network for edge enhancement and noise suppression, and outputting an optimized three-dimensional probability field; acquiring range-Doppler cube data synchronously collected by a quantum radar, and generating a velocity divergence constraint field based on the range-Doppler cube data; constructing a learnable constant diagonal structure projection matrix, mapping the optimized three-dimensional probability field to a voxel grid of a preset resolution to generate an initial voxel probability volume; taking the velocity divergence constraint field as a physical consistency loss function, and iteratively optimizing the initial voxel probability volume to obtain an optimized voxel probability volume; performing random gradient-based discretization and connected domain analysis on the optimized voxel probability volume to generate a polar night three-dimensional cloud mask.

9. The method of claim 1, wherein, The polar night three-dimensional cloud mask is generated based on the cloud layer vertical distribution probability map in combination with a radar range-Doppler constraint, including: multiplying the polar night three-dimensional cloud mask and the optimized three-dimensional probability field voxel by voxel to obtain a polar night optical thickness volume; uniformly resampling the polar night optical thickness volume and the polar day optical thickness map to a common grid to obtain a common thickness volume; performing adaptive weighted fusion on the common thickness volume by taking the solar elevation angle as a fusion weight to obtain a fused thickness volume; performing a logical AND operation on the fused thickness volume and the polar night three-dimensional cloud mask to generate a full-time cloud mask; dividing the full-time cloud mask into thickness levels based on the fused thickness volume to obtain a polar cloud layer full-time detection result.

10. A polar cloud layer detection device based on diurnal adaptation multi-modal fusion, characterized in that, The device includes: a determination module configured to determine a polar time period mode according to a solar elevation angle and an ambient light intensity of a polar observation point; an acquisition module configured to, in a case where the polar time period mode is a polar day mode, acquire a multi-angle polarized thermal image sequence and a visible light image, and generate a cloud class probability distribution map based on the multi-angle polarized thermal image sequence and the visible light image; the acquisition module is further configured to, in a case where the polar time period mode is a polar night mode, acquire quantum radar echoes and terahertz radiation spectra, and generate a cloud layer vertical distribution probability map based on the quantum radar echoes and the terahertz radiation spectra; a generation module configured to generate a polar day optical thickness map based on the cloud class probability distribution map in combination with a physically guided thickness inversion model; and a determination module configured to determine a polar time period mode according to a solar elevation angle and an ambient light intensity of a polar observation point; an acquisition module configured to, in a case where the polar time period mode is a polar day mode, acquire a multi-angle polarized thermal image sequence and a visible light image, and generate a cloud class probability distribution map based on the multi-angle polarized thermal image sequence and the visible light image; the acquisition module is further configured to, in a case where the polar time period mode is a polar night mode, acquire quantum radar echoes and terahertz radiation spectra, and generate a cloud layer vertical distribution probability map based on the quantum radar echoes and the terahertz radiation spectra; a generation module configured to generate a polar day optical thickness map based on the cloud class probability distribution map in combination with a physically guided thickness inversion model; and a determination module configured to determine a polar time period mode according to a solar elevation angle and an ambient light intensity of a polar observation point; an acquisition module configured to, in a case where the polar time period mode is a polar day mode, acquire a multi-angle polarized thermal image sequence and a visible light image, and generate a cloud class probability distribution map based on the multi-angle polarized thermal image sequence and the visible light image; the acquisition module is further configured to, in a case where the polar time period mode is a polar night mode, acquire quantum radar echoes and terahertz radiation spectra, and generate a cloud layer vertical distribution probability map based on the quantum radar echoes and the terahertz radiation spectra; a generation module configured to generate a polar day optical thickness map based on the cloud class probability distribution map in combination with a physically guided thickness inversion model; and The generating module is further configured to generate a polar night three-dimensional cloud mask based on the cloud vertical distribution probability map and in combination with a radar range-Doppler constraint. The generating module is further configured to generate a polar region cloud layer all-time period detection result based on the polar day optical thickness map and the polar night three-dimensional cloud mask.

Images