Cloud particle spectrum optimization method and system based on mask autoencoder, medium and device

Abstract

The application discloses a cloud particle spectrum optimization method and system based on a mask autoencoder, a medium and equipment, and the method comprises the following steps: obtaining a cloud particle shadow image to be completed, performing pretreatment, and obtaining a pretreated image; based on a trained cloud particle image completion model, a reconstruction patch is performed to obtain a completed image; the completed image is evaluated and screened, the cloud particle geometric scale statistics and particle size grading of the completed image passing the evaluation are performed, the flight sweeping distance and the effective sampling width corresponding to each particle size grading are combined, the grading sampling volume and the grading concentration are calculated, and finally the cloud particle spectrum is obtained. The overall process of the method does not depend on cloud particle level artificial labeling, and only the original image is needed to complete the training and application, thereby reducing the data preparation cost, being suitable for popularization and application in large-scale airborne observation data, and increasing the number of available large cloud particle samples entering the cloud particle spectrum statistical link without increasing the hardware complexity.

Description

Technical Field

[0001] This invention relates to a cloud particle spectrum optimization method, system, medium, and device based on a mask autoencoder, belonging to the field of cloud microphysics observation and image processing technology. Background Technology

[0002] Airborne in-situ imaging instruments such as optical array detectors (OAPs) can record the projection of cloud particles onto a photodiode array as they pass through a laser beam during flight, thus obtaining a two-dimensional image of the cloud particles. This image can then be used to deduce cloud particle size distribution, number concentration, and volumetric parameters such as ice water content and extinction coefficient. However, due to limitations in the probe's field of view and depth of field, in actual observations, a large number of cloud particles are recorded only in a truncated form, with only a portion entering the field of view. If truncated cloud particles are discarded directly or if the truncated size is used in the statistics, it will lead to insufficient samples of large cloud particles, an underestimation of the effective sampling volume, and a systematic bias in the cloud particle spectrum at the large particle size end.

[0003] To address the aforementioned issues, existing methods reconstruct the contours of cloud particles that cross the field of view boundary through geometric extrapolation, and define sampling criteria such as complete entry, center entry, and reconstruction based on this to expand the effective sampling volume of large cloud particles. Although such methods improve the statistical representativeness of large cloud particles to some extent, they usually rely on idealized geometric assumptions and empirical thresholds, have limited applicability to dendrites, aggregates, and complexly oriented ice crystals, and require repeated calibration on different platforms and cloud environments.

[0004] In recent years, self-supervised visual models such as masked autoencoders (MAEs) have learned the global structure and local texture of images without manual annotation by occluding large proportions of the input image and reconstructing the missing regions. OAP cloud particle images have characteristics such as sparseness, strong contrast, and local connectivity, which are highly compatible with the "occlusion-reconstruction" task.

[0005] Therefore, there is an urgent need for a technical solution that can complete OAP truncated cloud particle images using a data-driven approach without increasing hardware complexity, in order to improve the reliability of cloud particle geometric scale and cloud particle spectrum estimation. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of existing OAP cloud particle image reconstruction methods, which rely on geometric assumptions, are sensitive to thresholds, and are not adaptable to complex ice crystal morphologies. This invention provides a cloud particle spectrum optimization method, system, medium, and device based on a mask autoencoder, which is used to effectively expand the effective sampling volume of large cloud particles and improve the reliability of cloud particle spectrum and related volume parameter estimation.

[0007] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0008] In a first aspect, the present invention provides a cloud particle spectrum optimization method based on a mask autoencoder, comprising:

[0009] Obtain the cloud particle shadow image to be completed, perform preprocessing, and obtain the preprocessed image;

[0010] Based on the preprocessed image, a reconstruction patch is performed using a trained cloud particle image completion model to obtain the completed image.

[0011] The completed images are evaluated and screened. The completed images that pass the evaluation are statistically analyzed for cloud particle geometric scale and classified into particle size groups. The sampling volume and concentration of each particle size group are calculated by combining the flight sweep distance and the effective sampling width corresponding to each particle size group, and the final cloud particle spectrum is obtained.

[0012] Furthermore, the training process of the cloud particle image completion model includes:

[0013] Obtain complete original images of cloud particle shadows, generate corresponding missing images by simulating truncation or missing scenarios in actual observations, and construct training sample pairs;

[0014] The network architecture is trained using a mask autoencoder that adapts to single-channel grayscale input and has enhanced structure. The model takes the missing image as input and performs completion learning through an encoder-decoder mechanism.

[0015] Using the complete original image as the reconstruction target, a region-weighted composite loss function is used to constrain the model output, optimize the model parameters, and make dynamic adjustments to obtain a trained cloud particle image completion model.

[0016] Furthermore, the preprocessing process involves quality control of the image and normalization of pixel grayscale values ​​to map pixel values ​​to a preset value range.

[0017] It also includes preserving the largest connected component in a single image and removing noise.

[0018] Furthermore, the generation of missing images includes cropping the original image on one side and centering the cloud particle subject.

[0019] The centering process includes locating the cloud particle shadow connected region or bounding box, translating the cropping window based on the geometric center or symmetrically padding the image, so that the main body of the cloud particle falls into the preset central area, which is used to reduce the dependence of training on absolute position and more closely resemble the natural truncation shape.

[0020] Furthermore, the proportion of the unilateral pruning is obtained by random sampling within a preset interval, which is set to 10%-60% to cover a set degree of natural truncation and improve the model's generalization ability to different missing proportions.

[0021] Furthermore, after cropping, the missing areas are filled with edges. Bright strips are used for equal-volume filling to ensure that the missing image and the original image maintain the same spatial dimensions, thus avoiding blurring, deformation, or scale shift caused by scaling interpolation.

[0022] In the method of this invention, a missing data simulation method that closely resembles the statistical characteristics of observation is adopted in the construction of training data. By performing unilateral strip missing in the shaded area and cooperating with equal amount backfilling, the image size is kept consistent and the blurring introduced by scaling interpolation is avoided. At the same time, the model's dependence on fixed-position priors is reduced, so that the training samples are closer to the real truncation situation in terms of missing morphology, thereby improving the model's generalization ability to actual observation data.

[0023] Furthermore, the adaptation of the mask autoencoder to single-channel input includes weighting the original three-channel input layer weights according to their brightness contribution to synthesize them into single-channel equivalent weights, and setting the number of input channels to one.

[0024] Furthermore, the structural enhancement includes adding one or more of the following to the decoding part: increasing the number of layers, the number of attention branches, and the hidden feature dimension, in order to improve the reconstruction capability of complex contours, fragmented branches and internal gaps, reduce completion artifacts and improve boundary integrity.

[0025] The method of this invention synthesizes the input layer weights of the training model into a single-channel equivalent form and expands the expressive power of the decoder. While taking advantage of the training structure, it can more effectively restore complex contour structures and improve the geometric consistency and boundary integrity of the completed image.

[0026] Furthermore, the region-weighted composite loss function is obtained by dividing the image into missing regions, visible regions, and transition regions and assigning weights to them respectively, with the weight of missing regions being higher than that of visible regions.

[0027] The composite loss function further includes one or more of the following: structural similarity loss, edge loss, dark area morphology consistency loss, and total variation regularization term, used to simultaneously constrain pixel consistency, structural consistency, and contour sharpness.

[0028] Furthermore, the dynamic adjustment of the model includes unified management of batch size, learning rate and weights of each loss term through external configuration, and automatic application of preset update boundaries during training.

[0029] The optimization process of model parameters adopts an adaptive moment estimation algorithm with weight decay, the learning rate adopts a scheduling strategy of cosine annealing and periodic hot restart, and the gradient norm is pruned to improve training stability.

[0030] The method of this invention assigns higher weights to missing regions and combines constraints such as structural consistency, edge consistency, and dark area topological consistency, so that model reconstruction no longer only pursues pixel approximation but also focuses more on shape and contour. At the same time, it manages the mask ratio, batch size, and various loss weights uniformly through external configuration and can be updated during training, thereby improving training stability and engineering usability.

[0031] Furthermore, the process of evaluating and filtering the completed image includes:

[0032] Determine the completion range: Apply different proportions of masking to cloud particles to simulate incomplete observations in actual detection, and only perform completion on cloud particles whose visible part, i.e. the detected part, occupies a proportion of the complete image area of ​​cloud particles that is not less than a set threshold.

[0033] Evaluation: Extract the shadow area of ​​cloud particles before completion and calculate the area ratio with the shadow area of ​​cloud particles after completion. If the area after completion is twice or more than that before completion, it is considered unreliable completion.

[0034] The roundness is calculated based on the shadow of the completed cloud particles. The roundness is defined as: cloud particle area / minimum circumscribed circle area. When the calculated roundness value is less than 0.1, it is discarded.

[0035] Filtering: When the area ratio or roundness exceeds the preset threshold range, the corresponding image completion result is judged as over-completion, under-completion, or morphological distortion, and is removed, marked, or its weight in cloud particle spectrum statistics is reduced.

[0036] After completing the actual truncated cloud particle incompleteness, the method of this invention uses the "shading area ratio before and after incompleteness" and "cloud particle roundness after incompleteness" for joint evaluation to achieve quantitative discrimination of overcompleteness, undercompleteness, and morphological distortion. Based on this, the incompleteness results are screened, marked, or removed to avoid unreliable incompleteness from entering the cloud particle spectrum statistical link and introducing secondary errors. At the same time, combined with experimental statistics, an acceptable upper limit for missing values ​​is given, which ensures the reliability of subsequent scale statistics from the process perspective.

[0037] Furthermore, during the training of the image completion model, the missing image is divided into several non-overlapping image patches and position information is superimposed. Random masking is performed on the patches, and only visible patches are input into the encoder to obtain features. The features and mask label vectors are then input into the decoder to reconstruct the masked patches, and the patches are stitched together according to their original spatial positions to obtain the completed image. The encoder is composed of multiple layers of stacked Transformer blocks, with the number of layers increased from the default value to 12 layers. Each layer contains a "layer normalization-multi-head self-attention-feedforward network" structure, and a convolutional residual refinement module is added after the output to ensure training stability. The decoder is also composed of multiple layers of stacked Transformer blocks.

[0038] Furthermore, the sampling volume is calculated using the effective sampling width. The effective sampling width is equal to the fixed sampling width plus the maximum diameter of each cloud particle. The formula for the sampling volume is: Sampling volume = DOF (depth of field) × EAW (effective width) × TAS (flight distance). The sampling volume before and after completion differs in calculation. The uncompleted sampling volume is calculated directly using the basic formula, while the completed sampling volume is calculated using a different method, but the effective width is still calculated using a fixed value.

[0039] The formula for effective sampling width is: Effective sampling width = baseline width + particle size correction term. The particle size correction term is a non-negative function that does not decrease with the particle size represented by the gradation. By expanding or correcting the effective sampling width of large-size cloud particles, the systematic underestimation of the concentration at the large-size end is reduced.

[0040] In a second aspect, the present invention provides a cloud particle spectrum optimization system based on a mask autoencoder, used to execute the cloud particle spectrum optimization method based on a mask autoencoder described in the first aspect, comprising:

[0041] The image acquisition module is used to acquire the cloud particle shadow image to be completed, perform preprocessing, and obtain the preprocessed image;

[0042] The image completion module is used to reconstruct and patch the image based on the preprocessed image and the trained cloud particle image completion model to obtain the completed image.

[0043] The cloud particle spectrum optimization module is used to evaluate and filter the completed image. It performs cloud particle geometric scale statistics and particle size classification on the completed image that passes the evaluation. Combining the flight sweep distance and the effective sampling width corresponding to each particle size classification, it calculates the classification sampling volume and classification concentration to obtain the final cloud particle spectrum.

[0044] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the cloud particle spectrum optimization method based on a mask autoencoder as described in any of the first aspects.

[0045] Fourthly, the present invention provides a computer device, comprising:

[0046] Memory, used to store computer programs / instructions;

[0047] A processor for executing the computer program / instructions to implement the steps of the cloud particle spectrum optimization method based on a mask autoencoder as described in any one of the first aspects.

[0048] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0049] 1. The cloud particle spectrum optimization method based on mask autoencoder provided by this invention preprocesses the cloud particle shadow image to be completed, and uses a trained cloud particle image completion model to reconstruct and patch the preprocessed image to obtain the completed image. This invention achieves the completion of cloud particles to be completed through a data-driven approach, reducing the systematic bias caused by directly discarding truncated cloud particles or statistically analyzing them by truncation size, and alleviating the problem of underestimation of the effective sampling volume at the large cloud particle end, which causes cloud particle spectrum shift. The completed image is evaluated and screened, and the cloud particle spectrum of the completed image that passes the evaluation is optimized. In the optimization process, this invention uses the effective sampling width related to the particle size to participate in the sampling volume calculation for different particle size categories, so that the sampling volume estimate is adaptively adjusted with the particle size, and more reasonably corrects the effective sampling differences of cloud particles of different particle sizes. Compared with the simplified processing using a fixed effective width, this further reduces the risk of systematic underestimation at the large particle size end.

[0050] 2. The cloud particle spectrum optimization method based on mask autoencoder provided by this invention uses a mask autoencoder as the network framework during model training. By reconstructing the complete image from the image to be completed using a random mask, the model can infer the complete outline of the cloud particle image based on the remaining parts after truncation. This invention does not rely on manual annotation at the cloud particle level, and only requires the original image to complete training and application, reducing data preparation costs and making it suitable for widespread application in large-scale airborne observation data. In addition, this invention increases the number of available large cloud particle samples entering the cloud particle spectrum statistical link without increasing hardware complexity. Attached Figure Description

[0051] Figure 1 This is an overall flowchart of the cloud particle spectrum optimization method based on a mask autoencoder provided according to an embodiment of the present invention;

[0052] Figure 2 This is a flowchart illustrating the model training process of the cloud particle spectrum optimization method based on a mask autoencoder according to an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram illustrating the completion of different types of cloud particles with different truncation ratios in the cloud particle spectrum optimization method based on a mask autoencoder provided according to an embodiment of the present invention.

[0054] Figure 4 The image shows the loss statistics of different types of cloud particles after completion in the cloud particle spectrum optimization method based on mask autoencoder provided in the embodiment of the present invention.

[0055] Figure 5 The figure shows the experimental comparison results of the cloud particle spectrum optimization method based on mask autoencoder provided in the embodiments of the present invention. Detailed Implementation

[0056] It should be noted that:

[0057] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0058] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0059] Example 1

[0060] This embodiment introduces a cloud particle spectrum optimization method based on a mask autoencoder, including:

[0061] Obtain the cloud particle shadow image to be completed, perform preprocessing, and obtain the preprocessed image;

[0062] Based on the preprocessed image, a reconstruction patch is performed using a trained cloud particle image completion model to obtain the completed image.

[0063] The completed images are evaluated and screened. The completed images that pass the evaluation are statistically analyzed for cloud particle geometric scale and classified into particle size groups. The sampling volume and concentration of each particle size group are calculated by combining the flight sweep distance and the effective sampling width corresponding to each particle size group, and the final cloud particle spectrum is obtained.

[0064] like Figure 1 As shown, the training of the cloud particle image completion model is based on the aforementioned cloud particle spectrum optimization method. The specific training process includes:

[0065] Obtain the original image of cloud particle shadows, generate corresponding missing images by simulating truncation or missing situations in actual observations, and construct training sample pairs;

[0066] The network architecture is trained using a mask autoencoder that adapts to single-channel grayscale input and has enhanced structure. The model takes the missing image as input and performs completion learning through an encoder-decoder mechanism.

[0067] Using the original image as the supervision target, a region-weighted composite loss function is used to constrain the model output, optimize the model parameters, and make dynamic adjustments to obtain a trained cloud particle image completion model.

[0068] Furthermore, the training / evaluation data used in this embodiment originated from an airborne in-situ observation activity in the Hulunbuir region of Inner Mongolia, which conducted three research flights on May 31, 2021. The observed cloud system was a stratiform mixed-phase cloud formed during the passage of a warm front. The cloud observations were mainly concentrated in the altitude range of 2km–4km (AMSL), with an ambient temperature of approximately 2°C to -8°C. The cloud sampling durations for the three flights were 7388s, 4154s, and 10167s, respectively. The typical cruise speed of the aircraft was approximately 80m / s, and the time resolution of the observation products was 1Hz.

[0069] The key sensors on the observation platform include: a hot-wire probe (LWC) for measuring liquid water content, a cloud imaging probe (CIP) and a precipitation imaging probe (PIP), used for cloud particle and precipitation cloud particle imaging and cloud particle spectrum inversion, respectively; and an AIMMS-20 for acquiring parameters such as temperature, altitude, wind field, and location. The CIP has a channel resolution of approximately 15 μm and a measurement range of 15–1920 μm; the PIP has a resolution of approximately 150 μm. To reduce uncertainty at the small particle size end, this embodiment removes the first few particle size bins of the CIP / PIP, retaining only cloud particles with a diameter greater than a preset threshold (e.g., >90 μm) for analysis; simultaneously, to reduce the influence of probe fragmentation artifacts, a modified probe front end is used, and post-correction processing is performed in conjunction with the cloud particle arrival time interval.

[0070] Furthermore, such as Figure 2 As shown, this embodiment uses the MAE encoder-decoder structure to perform completion training on cloud particle shadow images that have truncated or missing elements:

[0071] A single-channel cloud particle shadow image is acquired in OAP, the cloud particle shadow image is subjected to quality control, and the pixel grayscale is normalized so that the pixel value is mapped to a preset value range, thereby obtaining a single-channel standardized image; in this embodiment, quality control is performed by removing empty frames and abnormal frames, and the preset value range is set to [-1,1].

[0072] Furthermore, the training process of the cloud particle image completion model is as follows: the single-channel image is divided into several non-overlapping image patches, and the patches are embedded into vector representations. The two-dimensional image is further transformed into a one-dimensional patch sequence, and position encoding is superimposed.

[0073] During the model training phase, a random mask is applied to the patch sequence, and only the visible patches are input into the encoder for feature extraction. The mask autoencoder is composed of multiple layers of Transformer modules stacked together. Each layer contains a structure of "layer normalization-multi-head self-attention-feedforward network" and promotes flow through residual connections to ensure the stability of training.

[0074] The visible patch features output by the encoder are recombined with the vector representing the mask mark according to the original spatial position and input into the decoder to reconstruct the missing patch; the decoder is also composed of multiple layers of Transformer blocks stacked together.

[0075] Furthermore, the cloud particle image to be completed is input into the trained cloud particle image completion model for reconstruction and patching, and then stitched back together to obtain the completed image.

[0076] The generated completed image is used for subsequent loss function calculation and cloud particle geometric scale statistics.

[0077] Furthermore, the generation of missing images includes cropping the original image on one side and centering the cloud particle subject.

[0078] The centering process includes locating the connected components or bounding boxes of the dark areas of cloud particles, translating the cropping window based on the geometric center, or symmetrically padding the image so that the main body of the cloud particles falls into the preset central area. This reduces the dependence of training on absolute position and makes it closer to the natural truncation shape.

[0079] Furthermore, the proportion of the unilateral pruning is obtained by random sampling within a preset range. In this embodiment, the preset range is set to 10%-60% to cover the natural truncation situation from mild to severe and to improve the model's generalization ability to different missing proportions.

[0080] Furthermore, after cropping, the missing areas are filled with edges. Bright strips are used for equal-volume filling to ensure that the missing image and the original image maintain the same spatial dimensions, thus avoiding blurring, deformation, or scale shift caused by scaling interpolation.

[0081] Furthermore, the region-weighted composite loss function is obtained by dividing the image into missing regions, visible regions, and transition regions and assigning weights to them respectively, with the weight of missing regions being higher than that of visible regions.

[0082] The composite loss function further includes one or more of the following: structural similarity loss, edge loss, dark area morphology consistency loss, and total variation regularization term, to simultaneously constrain pixel consistency, structural consistency, and contour sharpness.

[0083] Furthermore, the dynamic adjustment of the model includes unified management of batch size, learning rate and weights of each loss term through external configuration, and automatic application of preset update boundaries during training.

[0084] The optimization process of model parameters adopts an adaptive moment estimation algorithm with weight decay, the learning rate adopts a scheduling strategy of cosine annealing and periodic hot restart, and the gradient norm is pruned to improve training stability.

[0085] Furthermore, the process of evaluating and filtering the completed image is as follows:

[0086] Determine the completion range: Apply different proportions of masking to cloud particles to simulate incomplete observations in actual detection, and only perform completion on cloud particles whose visible part, i.e. the detected part, occupies a proportion of the complete image area of ​​cloud particles that is not less than a set threshold.

[0087] Evaluation: Extract the area of ​​the dark region of cloud particles before completion and calculate the area ratio with the area of ​​the dark region of cloud particles after completion. If the area after completion is twice or more than that before completion, it is considered unreliable completion.

[0088] The roundness is calculated based on the shadow of the completed cloud particles. The roundness is defined as: cloud particle area / minimum circumscribed circle area. When the calculated roundness value is less than 0.1, it is discarded.

[0089] Filtering: When the area ratio or roundness exceeds the preset threshold range, the corresponding image completion result is judged as over-completion, under-completion, or morphological distortion, and is removed, marked, or its weight in cloud particle spectrum statistics is reduced.

[0090] Furthermore, the basic formula for the effective sampling width is: Effective sampling width = baseline width + particle size correction term. The particle size correction term is a non-negative function that does not decrease with the particle size represented by the gradation. By expanding or correcting the effective sampling width of large-size cloud particles, the systematic underestimation of the concentration at the large-size end is reduced.

[0091] Furthermore, such as Figure 3 As shown, different cutoff ratios for different types of cloud particles are evaluated. In this embodiment, ice cloud particle images classified by a 2D-S probe are selected as representative of typical ice cloud particle shapes and their statistical characteristics. The images cover various morphologies from small-scale near-circular cloud particles to large-scale irregular cloud particles, used to evaluate the applicability and stability of the completion model under different structural complexities. The eight cloud particle types selected in this embodiment are as follows:

[0092] Class 1: Aggregates; Class 2: Dendritic; Class 3: Ring-shaped; Class 4: Graupel (soft hail); Class 5: Irregular; Class 6: Linear; Class 7: Flaky; and Class 8: Spherical.

[0093] Among the eight cloud particle types used to evaluate the performance of the incomplete image model, types 1, 2, and 4 have relatively large cloud particle scales, more complex shapes, and more fragmented boundary contours. From the perspective of incomplete image results, these large-scale cloud particles not only likely have the most missing pixels, but also present greater incomplete difficulty due to their larger spatial scale, more irregular internal structure, and more irregular edge shapes compared to other types. Therefore, this embodiment selects these large cloud particles as key test samples to reflect, to a certain extent, the model's incomplete ability under the "most unfavorable situation," providing a relatively conservative and physically meaningful upper limit for evaluating the performance of the cloud particle image incomplete image model. Based on this, selecting large cloud particles as the acceptable upper limit for incomplete image completion under different occlusion conditions is reasonable and representative.

[0094] In this embodiment, the applicable range of the completion is set to 50%. From the image completion results, it can be seen that when the occlusion rate is 0.5, the overall shape of the cloud particles can still be clearly identified in the completed image. The main axis direction of the cloud particles is basically consistent with the large-scale outline, but the edge structure and local details deviate to some extent from the original image. In this embodiment, it can be seen that the thickness and bifurcation structure of the local dendrites are weakened. This invention mainly studies the calculation of the maximum diameter of cloud particles, that is, the geometric extension range on the main scale, rather than the fine texture features. Therefore, the shape error is still within an acceptable range and will not cause significant systematic deviations in the maximum diameter statistical results.

[0095] When the coverage rate is increased to 0.6, the overall size, shape and contour of cloud particles in the completion results show significant deviations, and the principal axis direction and contour closure shape of some cloud particle types are significantly changed. At the same time, it can be observed that the completion effect of small-scale cloud particles also decreases significantly, indicating that small-scale cloud particles can tolerate a smaller deformation range and are more sensitive to the completion process.

[0096] In summary, the cloud particle image completion model can provide images with reliable geometric information under the condition that the occlusion rate does not exceed 0.5, which can be used for subsequent cloud particle scale statistics and maximum diameter estimation. However, in practical applications, if the scale of the completed cloud particle image is more than twice that of the original estimate or the typical scale of neighboring spatiotemporal samples, the completion result will be judged as unreliable completion and will be removed or separately marked in subsequent analysis.

[0097] Furthermore, such as Figure 4 As shown, the various losses of different types of cloud particles after completion are statistically analyzed. In this embodiment, the statistical results of three different loss indices under different occlusion ratios are presented respectively:

[0098] For pixel-by-pixel reconstruction error, this embodiment uses L1 loss, such as... Figure 4As shown in (a), the absolute value of the difference between the "reconstructed value" and the "original value" is taken and then the average value is taken. Since the gray values ​​of the input image have been standardized to the range [-1,1], when L1=0, it means that the pixels are completely consistent. The larger the L1 value, the greater the deviation between the overall brightness and the detail recovery.

[0099] To better characterize image perceptual quality, this embodiment introduces the structural similarity index SSIM, such as... Figure 4 As shown in (b), considering the similarity of brightness, contrast and structure, this method is closer to the subjective evaluation of image quality by the human eye. In this embodiment, the 1-SSIM value is used in the statistical chart. The smaller the value, the higher the structural similarity and the better the reconstruction effect.

[0100] To highlight the difference in cloud particle contours and edge sharpness, this embodiment also employs an edge-based loss (Edge) to measure the recovery of contour position and edge sharpness in the completion result. Figure 4 As shown in (c), the index directly reflects whether the completed contour is accurate and sharp enough, which is of great significance for subsequent particle size estimation that depends on the geometric features of the contour.

[0101] Furthermore, combining the statistical results of the three types of loss varying with the occlusion ratio, it can be seen that for large cloud particles of types 1, 2, and 4, the overall levels of the three types of loss are significantly higher than those of other cloud particle types. This indicates that large-scale cloud particles are more difficult to accurately complete after being occluded; this phenomenon is consistent with... Figure 2 The conclusions remain consistent: due to their more irregular shapes and more complex internal structures, the reconstruction errors of large cloud particles are more significant in multiple dimensions such as brightness, structure, and edges. Therefore, in this embodiment, it is reasonable and necessary to take the first, second, and fourth types of large cloud particles as key reference objects when evaluating the applicability of the model under different occlusion conditions.

[0102] From the trend of the occlusion ratio, when the cropping rate increases from 0.2 to 0.5, the three loss indicators remain relatively stable, showing only a slow increase or slight fluctuation. This indicates that under moderate occlusion, the model can still recover the overall brightness distribution, structural shape, and edge contour of cloud particles relatively well. When the cropping rate further increases to 0.6, the L1, 1-SSIM, and Edge losses all show a significant jump. Especially for large cloud particle categories, the increase in reconstruction error is more prominent, indicating that the completion result can hardly maintain consistency with the original cloud particles in terms of geometric shape and structural information. Combined with the above, subjective visual inspection of the completed image shows that when the cropping rate is 0.6, the cloud particle contours show relatively severe deformation, and the local structure even becomes mismatched or fictitious. The edges and details of small cloud particles are also obviously blurred.

[0103] Based on the comprehensive quantitative loss statistics and qualitative image comparison analysis, when the clipping rate reaches 0.6, the model's completion performance is insufficient to meet the accuracy requirements of subsequent particle size estimation and statistical analysis. In contrast, when the clipping rate does not exceed 0.5, although there are some deviations in local details, the overall geometric shape and maximum diameter of cloud particles can still be reliably recovered, providing effective support for subsequent particle size analysis.

[0104] Furthermore, in this embodiment, the complete model is applied to the flight observation data to construct and compare cloud particle spectrum results under different processing calibers.

[0105] The cloud particle samples were divided into two processing methods according to their geometric relationship into the sampling area, and the corresponding cloud particle spectrum curves were plotted:

[0106] First curve: Only cloud particle samples "completely located within the sampling area" are used for particle size classification and cloud particle spectrum statistics, and the color is blue;

[0107] The second curve: For the "truncated cloud particles that are partially located in the sampling area", the completion model of the present invention is first used to generate a complete image. Then, based on the completed cloud particle image, geometric scale estimation and particle size classification statistics are performed. Finally, it is merged with the "cloud particles that are completely located in the sampling area" to construct a cloud particle spectrum, which is orange in color.

[0108] like Figure 5 As shown, compared to the first curve, the second curve shows a shift in spectral distribution towards larger particle sizes and an increase in cloud particle number concentration at the large particle size end. This indicates that if statistics are based solely on the visible portion of truncated cloud particles within the field of view, it is easy to systematically underestimate the number of large-particle cloud particles. After employing the completion processing of this invention, the accuracy of estimating the true geometric scale of truncated cloud particles can be improved, thereby improving the statistical results of cloud particle spectra at the large particle size end. Thus, it can be seen that the method of this invention can improve the representativeness and reliability of cloud particle spectra at the large particle size end by completing and restoring the image of incompletely detected cloud particles without changing the original observation hardware.

[0109] Example 2

[0110] Based on the cloud particle spectrum optimization method based on a mask autoencoder described in Example 1, this example introduces a cloud particle spectrum optimization system based on a mask autoencoder, used to execute the cloud particle spectrum optimization method based on a mask autoencoder described in Example 1, including:

[0111] The image acquisition module is used to acquire the cloud particle shadow image to be completed, perform preprocessing, and obtain the preprocessed image;

[0112] The image completion module is used to reconstruct and patch the image based on the preprocessed image and the trained cloud particle image completion model to obtain the completed image.

[0113] The cloud particle spectrum optimization module is used to evaluate and filter the completed image. It performs cloud particle geometric scale statistics and particle size classification on the completed image that passes the evaluation. Combining the flight sweep distance and the effective sampling width corresponding to each particle size classification, it calculates the classification sampling volume and classification concentration to obtain the final cloud particle spectrum.

[0114] Example 3

[0115] Based on the cloud particle spectrum optimization method based on mask autoencoder described in Embodiment 1, this embodiment introduces a computer-readable storage medium storing a computer program / instruction. When the computer program / instruction is executed by a processor, it implements the steps of the cloud particle spectrum optimization method based on mask autoencoder as described in any of Embodiment 1.

[0116] Example 4

[0117] Based on the cloud particle spectrum optimization method based on mask autoencoder described in Example 1, this example provides a computer device, including:

[0118] Memory, used to store computer programs / instructions;

[0119] A processor is configured to execute the computer program / instructions to implement the steps of the cloud particle spectrum optimization method based on a mask autoencoder as described in any one of Embodiments 1.

[0120] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A cloud particle spectrum optimization method based on a mask autoencoder, characterized in that, include: Obtain the cloud particle shadow image to be completed, perform preprocessing, and obtain the preprocessed image; Based on the preprocessed image, a reconstruction patch is performed using a trained cloud particle image completion model to obtain the completed image. The completed images are evaluated and screened. The completed images that pass the evaluation are statistically analyzed for cloud particle geometric scale and classified into particle size groups. The sampling volume and concentration of each particle size group are calculated by combining the flight sweep distance and the effective sampling width corresponding to each particle size group to obtain the final cloud particle spectrum. The process of evaluating and filtering the completed image is as follows: Determine the completion range: Apply different proportions of masking to cloud particles to simulate incomplete observations in actual detection, and only perform completion on cloud particles whose visible part occupies a proportion of the complete cloud particle image area that is not less than a set threshold; Evaluation: Extract the shadow area of ​​cloud particles before completion and calculate the area ratio with the shadow area of ​​cloud particles after completion. If the area after completion is twice or more than that before completion, it is considered unreliable completion. Roundness is calculated based on the shadows of the completed cloud particles, which is used to evaluate the shape; Filtering: When the area ratio or roundness exceeds the preset threshold range, the corresponding image completion result is judged as over-completion, under-completion, or morphological distortion, and is removed, marked, or its weight in cloud particle spectrum statistics is reduced. The formula for the effective sampling width is: Effective sampling width = baseline width + particle size correction term. The particle size correction term is a non-negative function that does not decrease with the particle size of the segment, and is used to reduce the systematic underestimation of the concentration at the large size end.

2. The cloud particle spectrum optimization method based on mask autoencoder according to claim 1, characterized in that, The training process of the cloud particle image completion model includes: Obtain complete original images of cloud particle shadows, generate corresponding missing images by simulating truncation or missing scenarios in actual observations, and construct training sample pairs; The network architecture is trained using a mask autoencoder that adapts to single-channel grayscale input and has enhanced structure. The model takes the missing image as input and performs completion learning through an encoder-decoder mechanism. Using the complete original image as the reconstruction target, a region-weighted composite loss function is used to constrain the model output, optimize the model parameters, and make dynamic adjustments to obtain a trained cloud particle image completion model.

3. The cloud particle spectrum optimization method based on mask autoencoder according to claim 2, characterized in that, The process of generating the missing image includes cropping the original image on one side and centering the cloud particle body. The proportion of the unilateral cropping is determined by random sampling within a preset interval to cover the natural truncation situation; The centering process includes locating the cloud particle shadow connected region, translating the cropping window based on the geometric center, or symmetrically padding the image so that the main body of the cloud particle falls into the preset central region, which is used to reduce the dependence of training on absolute position and more closely resemble the natural truncation shape.

4. The cloud particle spectrum optimization method based on mask autoencoder according to claim 2, characterized in that, The adaptation to single-channel grayscale input refers to weighting the original three-channel input layer weights according to their brightness contribution and synthesizing them into a single-channel equivalent weight, and setting the number of input channels to one. The structural enhancement refers to adding one or more of the following to the decoding part: the number of layers, the number of attention branches, and the dimension of hidden features. This is used to improve the reconstruction capability of complex contours, fragmented branches, and internal gaps, reduce completion artifacts, and improve boundary integrity.

5. The cloud particle spectrum optimization method based on mask autoencoder according to claim 2, characterized in that, The region-weighted composite loss function is obtained by dividing the image into missing regions, visible regions, and transition regions and assigning weights to them respectively, with the weight of missing regions being higher than that of visible regions. The composite loss function also includes one or more of the following: structural similarity loss, edge loss, dark area morphology consistency loss, and total variation regularization term, which are used to simultaneously constrain pixel consistency, structural consistency, and contour sharpness. The dynamic adjustment refers to the unified management of batch size, learning rate, and weights of each loss term through external configuration, and the automatic application of the preset update boundary during training. The learning rate adopts a scheduling strategy of cosine annealing and periodic hot restart, and the gradient norm is pruned to improve training stability.

6. A cloud particle spectrum optimization system based on a mask autoencoder, characterized in that, The cloud particle spectrum optimization method based on a mask autoencoder as described in any one of claims 1 to 5 includes: The image acquisition module is used to acquire the cloud particle shadow image to be completed, perform preprocessing, and obtain the preprocessed image; The image completion module is used to reconstruct and patch the image based on the preprocessed image and the trained cloud particle image completion model to obtain the completed image. The cloud particle spectrum optimization module is used to evaluate and filter the completed image. It performs cloud particle geometric scale statistics and particle size classification on the completed image that passes the evaluation. Combining the flight sweep distance and the effective sampling width corresponding to each particle size classification, it calculates the classification sampling volume and classification concentration to obtain the final cloud particle spectrum.

7. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the cloud particle spectrum optimization method based on a mask autoencoder as described in any one of claims 1 to 5.

8. A computer device / equipment / system, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the cloud particle spectrum optimization method based on a mask autoencoder as described in any one of claims 1 to 5.

Images