Snowfall area cloud identification method based on multi-dimensional information fusion

Abstract

The invention discloses a snowfall area cloud identification method based on multi-dimensional information fusion. The method comprises the following steps: acquiring a remote sensing image of a snowfall area and selecting a wave band; preprocessing the remote sensing image of the selected wave band to obtain a preprocessed image; carrying out filtering processing on the pre-processed image through a hybrid tuned filtering technology to obtain a grayscale image; separating the grayscale image according to normal distribution characteristics by using a false anomaly screening separation model to obtain cloud body information and non-cloud body information; spatial statistical analysis is carried out through an Anisjen local Moran index to eliminate a small part of residual false anomalies of the false anomaly screening separation model. According to the invention, a hybrid tuned filtering technology is introduced, so that the identification capability of cloud body information is enhanced, and spectrum interference generated by a complex earth surface background is effectively inhibited; and then, through a false anomaly screening separation model, cloud layer features are locked, and through spectral information and spatial statistical information, multi-dimensional information fusion is carried out, so that the cloud identification precision is further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image cloud recognition, and in particular relates to a snowfall area cloud recognition method based on multi-dimensional information fusion. Background Art

[0002] As an important form of precipitation in nature, the distribution, intensity and duration of snowfall have a profound impact on water resource management, agricultural production, transportation and ecosystem balance. However, in remote sensing images, both clouds and snowfall show high reflectivity, and the spectral characteristics of clouds and snowfall in the visible and infrared bands are very similar. This spectral similarity makes it particularly difficult to distinguish between clouds and snow in remote sensing images, and snowfall areas are often misjudged as clouds.

[0003] In addition, the texture characteristics of clouds and snow are complex and varied. The shape, thickness, and density of clouds will affect their texture performance in the image, while the snow coverage and terrain undulations will also affect the texture characteristics of the snow layer. This complexity of texture increases the difficulty of distinguishing between clouds and snow. The confusion between clouds and snowfall areas not only reduces the accuracy of weather forecasts, but also affects the reliability of environmental monitoring and climate change research.

[0004] At present, most of the existing cloud recognition algorithms are based on thresholds, texture features or deep learning methods, but these methods have different applicability in different regions, different times and different weather conditions, and have low universality, and are not effective in snowy areas. At the same time, most cloud recognition algorithms do not fully utilize the advantages of spatial information technology and cannot provide rich surface information, which directly affects the accuracy and effect of cloud-snow distinction. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a cloud recognition method for snowfall areas based on multi-dimensional information fusion to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a cloud recognition method for snowfall areas based on multi-dimensional information fusion, comprising:

[0007] Obtain remote sensing images of snowfall areas and select bands;

[0008] Preprocessing the remote sensing image of the selected band to obtain a preprocessed image;

[0009] Filtering the spectral information of the pre-processed image by hybrid tuned filtering technology to obtain a grayscale image;

[0010] Separating the grayscale image according to normal distribution characteristics using a false anomaly screening and separation model to obtain cloud body information and non-cloud body information;

[0011] The cloud information is spatially statistically analyzed using the Anselm Local Moran Index to quantify the spatial autocorrelation at a specific location, eliminate false anomalies in snowfall areas, and obtain cloud identification results.

[0012] Preferably, the remote sensing image of the selected band is 9 bands of Landsat 8OLI data, wherein the 9 bands are specifically Band 1 to Band 9.

[0013] Preferably, the process of preprocessing the remote sensing image of the selected band includes:

[0014] The remote sensing images of the selected bands were resampled using cubic convolution interpolation, and the resolution of Band 8 was resampled from 15 meters to 30 meters to match the resolution of other multispectral bands.

[0015] Preferably, the process of filtering the pre-processed image by using a hybrid tuned filtering technique comprises:

[0016] The endmember spectral information of interest to the user is extracted through a partial unmixing strategy to generate a hybrid tuned filter image.

[0017] Based on the mixed tuned filter image, a target area with low spectral contrast between the target and the background and inside the target is detected;

[0018] The feasibility score is used to adjust the abundance image of the target area to obtain a grayscale image.

[0019] Preferably, the process of separating the grayscale image according to the normal distribution characteristics by using the false anomaly screening and separation model includes:

[0020] According to the normal distribution characteristics, the grayscale image is separated by using a false anomaly screening and separation model to obtain cloud body information and non-cloud body information of a bimodal normal distribution;

[0021] The separability between the probability distributions of the cloud body information and the non-cloud body information is evaluated by using the Bhattacharyya distance.

[0022] Preferably, the snowfall area cloud identification method further includes a process of improving the false anomaly screening and separation model:

[0023] According to the inherent difference between the target area and its background, a reasonable threshold range is calculated, and based on the threshold range, it is used to distinguish the information area to be extracted and the background area;

[0024] A local subdivision strategy is introduced to perform multiple subdivisions on the local data according to the specific parameters of each sub-distribution in the mixed distribution to extract the interesting ground object information.

[0025] Separation degree evaluation is introduced to determine the degree of separation between the object to be extracted and the background information according to the size of the separation degree.

[0026] Preferably, the false anomaly screening and separation model is a mixed distribution model, and the expression is:

[0027]

[0028] Among them, the parameter set μ=(n,μ1,...,μ n , a1, ..., a n ) contains a series of key elements used to define multiple branches, n represents the total number of branches, f(x, μ i ) represents the probability density of the i-th branch, μ i is the corresponding parameter, a i represents the weight of the i-th branch, and

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] The present invention proposes a cloud recognition method for snowfall areas based on multi-dimensional information fusion. For accurate cloud recognition in snowfall areas, Landsat 8OLI data is first used as research data and bands are selected. Then, preprocessing technology is used to eliminate the resolution inconsistency between different bands to ensure that all bands have the same sampling interval in space. Secondly, hybrid tuned filtering technology is introduced to enhance the recognition ability of cloud body information and effectively suppress the spectral interference caused by complex surface background. Subsequently, the false anomaly screening and separation model is used to lock the cloud layer characteristics. Finally, the Anselm local Moran index is used to perform spatial statistical analysis to eliminate the small part of false anomalies remaining in the false anomaly screening and separation model to obtain cloud recognition results. The present invention performs multi-dimensional information fusion through spectral information and spatial statistical information, thereby further improving the cloud recognition accuracy in snowfall areas.

[0031] By continuously optimizing the cloud recognition algorithm and improving its recognition accuracy and stability in snow-containing areas, the present invention can provide more accurate and reliable data support for fields such as weather forecasting, environmental monitoring, and climate change research. At the same time, with the continuous development of remote sensing technology and the continuous improvement of data acquisition capabilities, it is expected that cloud recognition services with higher accuracy, higher resolution, and wider coverage will be achieved in the future, contributing more wisdom and strength to global climate change response and sustainable development goals. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0033] Figure 1 A remote sensing image of a representative area a in an embodiment of the present invention;

[0034] Figure 2 is a false color composite image of a representative area a of an embodiment of the present invention;

[0035] Figure 3 is a grayscale image after hybrid tuning filtering of a representative area a of an embodiment of the present invention;

[0036] Figure 4 is a remote sensing image of a representative area b in an embodiment of the present invention;

[0037] Figure 5 A false color composite image of a representative area b of an embodiment of the present invention;

[0038] Figure 6 A grayscale image after mixed tuning filtering of a representative region b in an embodiment of the present invention;

[0039] Figure 7 This is a diagram showing the results of the improved false anomaly screening and separation model for the representative region a of an embodiment of the present invention;

[0040] Figure 8 This is a diagram showing the results of the improved false anomaly screening and separation model for the representative region b of an embodiment of the present invention;

[0041] Fig. 9 This is a schematic diagram of cloud extraction results after the representative area a of an embodiment of the present invention is processed by the improved false anomaly screening and separation model;

[0042] Fig.10 This is a schematic diagram of cloud extraction results after the representative area b of an embodiment of the present invention is processed by the improved false anomaly screening and separation model;

[0043] Fig.11 This is a schematic diagram of the final cloud extraction result after the Anselm local Moran index is calculated for the representative area a of the embodiment of the present invention;

[0044] Fig.12 This is a schematic diagram of the final cloud extraction result after the Anselm local Moran index is calculated for the representative area b of the embodiment of the present invention;

[0045] Fig.13 A comparison diagram of the embodiment of the present invention, the artificial neural network and Fmsk;

[0046] Fig.14 This is a step diagram of a method for identifying clouds in a snowfall area according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0049] Embodiment 1

[0050] like Fig.14 As shown, in this embodiment, a cloud recognition method for snowfall areas based on multi-dimensional information fusion is provided, wherein the multi-dimensional information includes: spectral information and spatial statistical information, and the cloud recognition method steps include:

[0051] S1, obtain remote sensing images of snowfall areas and select bands;

[0052] Specifically, Landsat 8OLI data is used as research data, and a multi-dimensional information deep fusion strategy is used to achieve high-precision recognition of clouds in snowfall areas. In the process of band selection, this embodiment strictly follows the following four core considerations to ensure that the selected bands can fully and accurately serve the research objectives of this embodiment:

[0053] ① Comprehensiveness of information and feature retention: In order to ensure that the selected bands can retain the characteristic information of clouds and underlying surfaces to the maximum extent, this embodiment strives to ensure that the selected band combination can cover a wide and comprehensive spectral range, thereby avoiding the omission of any key information and laying a solid foundation for subsequent analysis.

[0054] ② Minimization of spectral information loss: In different bands, the spectral characteristics of various types of objects may vary. Therefore, when selecting bands, this embodiment pays special attention to avoiding the spectral information loss caused by these differences, ensuring that each type of object, including snowfall areas and clouds, can be truly and completely expressed in its most suitable band.

[0055] ③True presentation of the spectral information of the target area: In order to accurately reflect the true spectral characteristics of the area of ​​interest in subsequent research, this embodiment carefully selects bands that can clearly show the spectral characteristics of these areas. The selection of these bands not only helps this embodiment to more deeply understand the physical properties of the target area, but also provides reliable input data for the subsequent cloud recognition algorithm.

[0056] ④ Accurately capture the spectral differences between clouds and snow: Although easily confused underlying surfaces such as clouds and snow show the commonality of high reflectivity in the visible light band, their actual reflectance spectral curves in specific bands are significantly different. These differences, such as specific absorption peaks and reflection valleys, are important diagnostic features for distinguishing clouds from snow. Therefore, in the process of band selection, this embodiment pays special attention to selecting those bands that can highlight these subtle spectral differences to ensure that the cloud recognition method of this embodiment can achieve accurate recognition of clouds.

[0057] In summary, this patent finally selected 9 bands (band1 to band9) of Landsat 8OLI data. These bands not only meet all the above selection criteria, but also provide a new cloud identification method for snowfall areas based on multi-dimensional information fusion for this embodiment, providing strong support for research and application in related fields.

[0058] S2, preprocessing the remote sensing image of the selected band to obtain a preprocessed image;

[0059] Specifically, when processing Landsat 8 satellite image data, since the resolution of Band 8 of the satellite reaches 15 meters, while the resolution of the remaining multispectral bands is 30 meters, this difference brings inconvenience to the subsequent data processing and analysis. In order to eliminate this inconsistency and ensure that all bands have the same sampling interval in space, this embodiment adopts a resampling technical method. Specifically, this experiment selected cubic convolution interpolation as a means of resampling, and resampled the Band 8 band with an original resolution of 15 meters to 30 meters to match the resolution of other bands. This method is a resampling method with higher precision. It takes into account the information of multiple original pixels around the output pixel and uses cubic polynomial interpolation to calculate the grayscale value of the output pixel. This method can produce smoother and more accurate results while maintaining image details.

[0060] S3, filtering the spectral information of the pre-processed image by hybrid tuned filtering technology to obtain a grayscale image;

[0061] Specifically, the hybrid tuned filtering technology uses a partial unmixing strategy to extract the endmember information of interest to the user in a customized manner. In actual operation, the technology first focuses on a selected set of pixels of interest, and then treats the remaining unclassified pixels as background data. The average spectral characteristics of these regions of interest are calculated and used as the reference endmember information for the hybrid tuned filtering. In order to ensure data accuracy, this experiment deliberately collected training samples in relatively flat areas with stable lighting conditions, effectively avoiding the potential interference of complex terrain and lighting changes on the results.

[0062] Hybrid tuned filtering presents an efficient detection approach for specific target objects. Its advantage is that it does not require a comprehensive analysis of all pixel details of the entire satellite image. Instead, it accurately identifies areas that match the endmember spectrum based on the high degree of match with the preset sample spectrum or image endmember spectrum. This feature makes this method excellent in reducing confusion errors caused by high-reflectivity surface covers such as snow and deserts, thereby improving the accuracy of cloud recognition in complex terrain environments. The focus of this study is to extract endmembers and use them to generate hybrid tuned filter images, which is an advanced matched filtering method designed to reduce false anomalies. This method can also detect targets with low spectral contrast between the target and the background and within the target. Multivariate sample endmembers are denoted as "x".

[0063] Fi(x),i=(1,2,3,...,p) (1)

[0064] Any data set is a signal component that contains some background noise caused by various sensors and atmospheric factors. Therefore, the total signal has the following characteristics:

[0065] F(x)=A(x)+B(x) (2)

[0066] The signal and noise components are A(x) and B(x), respectively. Hybrid tuned filtering is used to evaluate the abundance of end members and uses a feasibility score between 0 and 1 to adjust the abundance image by reducing false positives. The abundance image is a distribution map of the relative intensity or significance of certain specific features or information in the image after filtering.

[0067] In this embodiment, the construction of end-member spectral information adopts an innovative strategy, that is, based on the spectral average of four pixels around each pixel in the training area. Compared with a single pixel spectrum, this method can better represent the comprehensive characteristics of the area and enhance the stability and reliability of the spectral information. At the same time, through averaging, the interference of spectral noise is effectively reduced, and the purity of the data is further improved. In addition, this method also significantly weakens the bright and dark area effect caused by changes in lighting conditions and solar azimuth in complex terrain areas, providing more accurate basic data for subsequent cloud identification. In order to fully cover the diversity of clouds, this study specifically selected two areas of interest, thick clouds and thin clouds containing all bands, as analysis objects. This choice is based on the difference in the physical properties of clouds: thin clouds are often mixed with surface information to form mixed pixels, while thick clouds may be misjudged as other natural covers such as snow or deserts due to spectral similarity. This step is efficiently completed on the MATLAB platform, ensuring the accuracy and convenience of data processing.

[0068] In order to further verify the practical application effect of this method in cloud identification, this embodiment selects two representative complex background areas for case analysis. These areas not only contain a variety of surface cover types, but also show a variety of cloud forms, thereby comprehensively evaluating the feasibility and effectiveness of this embodiment. For the application potential of this method on large-scale data sets and its accuracy evaluation, subsequent research will further explore and quantify its performance in actual operations.

[0069] Take the representative area a as an example, Figure 1 As shown in the figure, in the northern part of the image, the hills are covered with heavy snow, which shows high brightness and exists in patches, a feature that significantly increases the complexity of automatic cloud recognition. At the same time, the image is dotted with irregular frozen rivers, whose bright features and disordered distribution patterns further exacerbate the challenges in the cloud extraction process. It is particularly noteworthy that there are also tiny point clouds scattered in the image, especially in the south direction. These point clouds are very easy to be visually confused with the patchy snow in the northern region due to their small size, which further increases the difficulty of accurate cloud extraction. In summary, the complex characteristics of the image, including the patchy distribution of snow, the disordered highlights of the frozen rivers, and the subtle confusion of the point clouds, together constitute multiple obstacles in the cloud extraction task.

[0070] Next, this embodiment will use false color synthesis technology for image a. In this process, this embodiment selects bands 7, 6, and 3 in the Landsat 8OLI data set, which correspond to the red, green, and blue channels of the false color image for synthesis. Through this process, this embodiment obtains a brand new false color image, and its display effect is shown in the attached figure. Figure 2 shown.

[0071] From the false color composite image, the present embodiment can clearly observe that the entire image is widely distributed with intricate frozen rivers, which meander and stretch in an irregular state. At the same time, there are many tiny point clouds scattered in the image, which may be difficult to detect in the original true color image due to their tiny size, but become more conspicuous under false color synthesis. Such image features not only enrich the present embodiment's understanding of the study area, but also provide a more intuitive reference for subsequent cloud extraction work.

[0072] After hybrid tuned filtering, the resulting grayscale image clearly shows the filtering effect, as shown below Figure 3As shown in the figure. In this image, the areas with high hybrid tuned filter values ​​are mainly concentrated on the cloud layer, indicating that this method effectively highlights the cloud information. At the same time, the frozen rivers that were originally distributed irregularly over a large area almost completely disappeared after filtering, which fully demonstrated the excellent ability of hybrid tuned filtering in suppressing the spectral response of the reflectivity of frozen rivers and significantly reduced the complexity of cloud information extraction. In addition, the filtering process also enhanced the spectral response of some smaller point clouds, making these originally difficult to identify cloud features more obvious, further reducing the difficulty of cloud information extraction.

[0073] However, despite the above positive effects, some patchy high-value pixel residues can still be seen in the alpine snow areas in the northern region. These patchy high-value areas may be visually confused with the smaller point clouds in the image, which poses certain challenges to the accurate extraction of cloud information. Therefore, in subsequent processing, additional measures need to be taken to distinguish these patchy snow from point clouds to ensure the accuracy and reliability of cloud information extraction.

[0074] In order to evaluate the influence of misty thin clouds on image filtering, this embodiment selects a representative area b as the research object, such as Figure 4 As shown in the figure. The notable feature of this image is that its northeastern part shows a large area of ​​snow landscape, which is spread vastly on the tops of the endless mountains. Due to the difference in the slope direction of the mountain where it is located, the snow shows completely different characteristics under the action of light: on the side facing the sun, the reflectivity of the snow is significantly enhanced, and the unique lighting conditions in high-latitude areas make these areas abnormally bright; on the contrary, on the side facing away from the sun, the snow appears dim because the light is blocked by the mountain, forming a sharp contrast - the sunny slope is bright and dazzling, while the shady slope is dim. This natural phenomenon causes the entire alpine snow area to present mottled light and dark patches, which greatly complicates the automatic extraction process of the cloud layer. In addition, the shape of the cloud in the selected area is not very clear. Especially in the southwest of the image, the cloud layer presents a typical foggy feature with blurred edges and difficult to define, which increases the difficulty of accurate extraction of the cloud layer.

[0075] In order to more intuitively display the distribution of clouds and snow, this embodiment uses false color synthesis technology to process the true color image of the first representative area, and presents the synthesized image, such as Figure 5 From this false color composite image, we can clearly see that the clouds in the southwestern region are in a typical foggy state, with blurred boundaries and unclear features, which undoubtedly brings great difficulties to the accurate extraction of clouds.

[0076] The grayscale image obtained after processing by the hybrid tuned filtering method is as follows: Figure 6As shown in . This grayscale image is generated based on the hybrid tuned filter value, where the filter value follows a normal distribution and the mean is set to zero. In this grayscale image, lower pixel values ​​represent the background area (i.e., the part that does not contain the target component), while higher pixel values ​​indicate the area containing a higher proportion of the target component, which are the key objects extracted in this embodiment. It can be clearly found that the high hybrid tuned filter values ​​are mainly concentrated on the cloud body, especially the heavy cloud layer in the southeast, showing the high sensitivity of the method to the cloud layer. At the same time, the surface information in the overall highlighted background is effectively suppressed, which significantly reduces the complexity of cloud extraction. It is particularly noteworthy that the spectral characteristics of the misty cloud body in the southwest are significantly enhanced after filtering, which greatly simplifies the extraction process of such cloud body information. However, in the snow-covered area in the northeast, although most of them are patchy and highlighted, there are still a small number of pixel values ​​that are high. This once again shows that the snow in the high mountain area has complex and diverse brightness changes in remote sensing images due to its complex lighting conditions and slope differences, which brings additional challenges to cloud recognition.

[0077] In summary, the hybrid tuned filtering technology shows its unique advantages: it can effectively suppress the spectral response of the bright background area, while significantly enhancing the cloud information, especially those misty thin clouds and tiny point clouds that are difficult to capture in traditional methods. Given that the hybrid tuned filtering result values ​​show the characteristics of normal distribution, this experiment further introduced a false anomaly screening and separation model, aiming to extract more refined cloud information in complex terrain areas. This strategy not only improves the accuracy of cloud recognition, but also enhances the adaptability of the model in complex environments.

[0078] S4, using a false anomaly screening and separation model to separate the grayscale image according to normal distribution characteristics to obtain cloud body information and non-cloud body information.

[0079] In the analysis of remote sensing images, the existence of various surface covers constitutes a complex multi-modal data set, and these modes are essentially a mixture of different ground object information. From a mathematical perspective, this mixed distribution phenomenon of multi-class ground object information can be accurately described by the concept of false anomaly screening and separation model. Among them, the normal distribution is one of the key mathematical tools for distinguishing different ground object characteristics, and its importance is self-evident. It provides a powerful basis for the identification of ground objects in remote sensing images.

[0080] The statistical characteristics of remote sensing images are usually manifested as a multi-modal distribution pattern formed by the superposition of multiple independent populations. In most cases, the ground object information in the image closely follows the normal distribution law. Even those few ground object information that deviate from or are obviously inconsistent with the normal distribution can be effectively converted into data forms that obey the normal distribution through in-depth statistical analysis, combined with logarithmic transformation or other types of mathematical transformation methods.

[0081] In view of this, this embodiment can regard cloud information as an example of normal distribution, and use a mathematical model for false anomaly screening and separation to accurately characterize it. In specific applications, this embodiment often assumes that the sample data in the remote sensing image follows a mixed distribution model, which integrates the distribution characteristics of multiple ground object information, laying a solid theoretical foundation for subsequent cloud information extraction and analysis.

[0082]

[0083] Among them, the parameter set μ=(n,μ1,...,μ n , a1, ..., a n ) contains a series of key elements used to define multiple branches. Specifically, n represents the total number of branches, f(x, μ i ) represents the probability density of the i-th branch, μ i is the corresponding parameter, a i represents the weight of the i-th branch, and

[0084] The model can achieve more accurate extraction of specific ground object information. At the same time, this embodiment improves the model performance with the separation parameter between sub-distributions, based on an in-depth analysis of representative training data of each target category, thereby quantifying the degree of distinguishability between different target categories. In statistics, especially under the assumption that the data follows a normal probability distribution, the Bhattacharyya distance (L) has been proven to be an effective and applicable metric for evaluating the separability between two probability distributions.

[0085] The Bhattacharyya distance not only reflects the probability of misclassification in the classification task, but is also closely related to the optimal classification performance under the Bayesian decision theory, so it has a solid theoretical basis as a separability measure. The calculation formula of the Bhattacharyya distance is specifically expressed as follows, which can be directly applied to the model of this embodiment to quantify and optimize the separation effect between each sub-distribution, thereby enhancing the accuracy and reliability of the model in extracting information about objects of interest.

[0086]

[0087] Among them, α1 and α2 represent the means of two distribution characteristics, and Represent the variance of the two distribution characteristics respectively. If the means of the two characteristic distributions are exactly the same, the first term in the formula will automatically return to zero, because the mean difference is the basis for calculating this term, and zero difference naturally causes this term to be zero. On the other hand, if the variances of the two characteristic distributions are exactly the same, the second term in the formula will also disappear to zero, because the variance difference is the key factor in evaluating this part, and equal variance means no difference, which makes this term zero. That is, the equality of the means eliminates the influence of the first term, and the equivalence of the variances makes the second term no longer contribute to the overall separability measure. Such characteristics help this embodiment to more clearly understand how different statistical characteristics independently affect the degree of separability between the two distributions.

[0088] When dealing with complex classification scenes containing confusing objects, the introduction of SD distance as a means of describing the degree of separation shows more significant advantages compared to the previously mentioned L degree of separation calculation method. SD distance not only provides a clear and limited dynamic range interval, this feature makes the comparison between different feature separation results more intuitive and effective. Through the SD distance, the present embodiment can more accurately identify those features with good separability, which is crucial to improving classification accuracy. The SD distance varies within its domain (a scale range of 0 to 2), and this range provides a standardized perspective for the present embodiment to evaluate the separability between two categories. Specifically, the L parameter is used as part of the SD distance calculation or directly as its metric to quantify the degree of separability between categories, thereby helping the present embodiment to make more scientific and reasonable decisions in practical applications.

[0089] SD=2(1-e -L ) (5)

[0090] Ideally, if the features of the two sub-distributions show perfect separability, their SD value will be close to 2, which means that the two categories can be completely distinguished on a specific feature dimension, and theoretically there will be no misjudgment when using this feature for classification. On the contrary, as the SD value decreases, the separability gradually decreases, which means that the confusion between the sub-distributions increases, and the possibility of errors in classification also increases. In extreme cases, SD = 0 indicates that the two categories are almost indistinguishable on this feature, showing a high degree of overlap.

[0091] In view of the above analysis, this embodiment optimizes the false anomaly screening and separation model and implements it on the MATLAB platform. The data processed by the model comes from the hybrid tuning value generated by the hybrid tuning filtering technology. After the representative area a is screened and separated for false anomalies, the results show that ( Figure 7)The overall data presents the characteristics of a bimodal normal distribution. At the same time, the hybrid tuned filtering effectively enhances the performance of cloud information in the grayscale image and weakens the interference of other non-cloud information, so that the entire image is naturally divided into two categories: cloud and non-cloud. Figure 7 The two clusters correspond to two significant normal distribution forms: the more concentrated part on the left is defined as normal distribution I, while the relatively dispersed part on the right is marked as normal distribution II. This once again proves the inherent structural complexity of the data and provides an important basis for subsequent data analysis and interpretation.

[0092] After the false anomaly screening and separation model processed the data in the representative area, the key parameter values ​​of each sub-distribution were obtained, as shown in Table 1. Specifically, the mean of the first normal distribution reached 87098.37, while the mean of the second normal distribution was significantly higher, at 272010.42. The weights of the two in the data set are equivalent, 0.48 and 0.52 respectively, indicating that the background information and cloud information are close in total, and the background information has a slight advantage. To distinguish the two sub-distributions, the algorithm automatically sets the corresponding threshold range and divides the data set into two parts: from the minimum value to 182059.38 as one part, and from 182059.38 to the maximum value as the other part. This division strategy is based on the natural distribution characteristics of the data.

[0093] Further analysis showed that there was a difference in the separation degree of the two sub-distributions. The separation degree of the first sub-distribution was 1.07, which was higher than the 0.79 of the second sub-distribution, indicating that the first sub-distribution was more independent and easier to distinguish in the data set. In contrast, the separation of the second sub-distribution was slightly more difficult.

[0094] In addition, the overlap rate between the two normal distributions was calculated, which was 33.57% and 30.23% respectively. This indicator reveals the proportion of the overlap between the two in the data set, which is of great significance for understanding the complexity of the data structure and the subsequent data processing strategy. The parameter values ​​of each sub-distribution of the false anomaly screening and separation model are shown in Table 1.

[0095] Table 1

[0096]

[0097] For the representative area b, the optimized false anomaly screening and separation model is used for analysis, such as Figure 8As shown, this embodiment obtains detailed sub-distribution parameter values, which accurately characterize the characteristics of each independent component in the mixed distribution. Here, this embodiment defines the normal distribution on the left as "Normal Distribution I", and the normal distribution on the right is marked as "Normal Distribution II". Through the parameter values ​​calculated by this model, this embodiment can deeply understand the specific information of each sub-distribution and provide a solid foundation for further refinement of data processing. The parameter values ​​of each sub-distribution of the false anomaly screening and separation model are shown in Table 2.

[0098] Table 2

[0099]

[0100] The average values ​​listed in the table refer to the central tendency of the internal data set of each normal distribution, which is specifically manifested as the average values ​​of the two sub-distributions reaching 2783.79 and 14790.59 respectively. The weight values ​​of the two sub-distributions are 0.38 and 0.62 respectively, which directly reflects that the background information occupies a large proportion of the whole, while the proportion of cloud information is relatively small. The false anomaly screening and separation model, with its intelligent characteristics, can automatically define the appropriate threshold range, thereby realizing the accurate extraction of feature bodies. For these two normal distributions, the threshold ranges set by the model are from the minimum value to 5305.43, and from 5305.43 to the maximum value, respectively. The separation degree of the two sub-distributions is different. The separation degree of the second sub-distribution reaches 1.09, which is higher than 0.73 of the first sub-distribution, indicating that the second sub-distribution is easier to distinguish in the overall data. In addition, the overlap rate reveals the degree of data overlap between sub-distributions, which is crucial to understanding the complexity of the data structure. The overlapping data of the two sub-distributions account for 47.27% and 35.01% of the overall data respectively. This information provides a valuable reference for subsequent data processing and analysis.

[0101] In order to intuitively understand the classification effect of the improved false anomaly screening and separation model, this embodiment displays the cloud body information automatically extracted by the improved false anomaly screening and separation model in the visualization software ArcGIS 10.8. Fig. 9 The cloud extraction result of the representative area a, where yellow represents the extracted cloud information and black represents the non-cloud underlying surface area. Figure 2 Compared with the cloud extraction results, it can be seen that the thick cloud area is completely extracted, and the small point cloud is also not missed, which proves the effectiveness of the algorithm for cloud information that is small and difficult to extract. However, the snow on the hills in the north of the image has some false anomaly information due to its patchy characteristics, and some ice and snow pixels and high reflectivity underlying surface pixels such as irregularly distributed frozen rivers in the background area are misjudged as clouds.

[0102] Fig.10 This is the cloud automatic extraction result of the second representative area b. The yellow area represents the cloud obtained by automatic cloud recognition, and the black area is the non-cloud area. Image synthesized by false color Figure 5 From the cloud extraction results, we can see that thick clouds are completely extracted. Figure 5 The misty thin clouds in the lower left corner that are almost invisible to the naked eye are also detected relatively completely, thus proving the practical feasibility of the algorithm. However, due to the difficulty of extracting the scene image, confusing false abnormal information is inevitably generated. In the ice and snow area in the north of the image, part of the ice and snow is misjudged as clouds, and the same situation also exists in the underlying surface area with high reflectivity.

[0103] Based on the above analysis, this embodiment finds that the false anomaly screening and separation model shows a high degree of accuracy in extracting cloud information, especially when faced with a series of complex or challenging image scenes. These scenes include but are not limited to: (1) extremely complex underlying surface backgrounds, such as mottled cloud areas formed by snow-covered areas in high mountains, or areas covered with irregular high-reflectivity frozen rivers, which are often difficult to identify with traditional cloud recognition methods; (2) the cloud and fog are fuzzy in shape and the contour features are not significant, resulting in foggy cloud information that is difficult to capture with traditional recognition methods; (3) the cloud itself is tiny in shape and almost imperceptible. In these cases, the false anomaly screening and separation model, with its unique algorithmic advantages, can more effectively extract cloud information, thereby showing significant application advantages.

[0104] The false anomaly screening and separation model has achieved the following key improvements and enhancements in terms of cloud feature screening and separation: (1) Automatic threshold calculation: The model can intelligently calculate a reasonable threshold range based on the inherent differences between the target area and its background, which is used to accurately distinguish whether each pixel in the satellite image belongs to the information area to be extracted or the background area. This innovation not only significantly reduces the human intervention and subjectivity in the threshold judgment process, but also greatly improves the efficiency and accuracy of target object extraction, ensuring the objectivity and reliability of the results. (2) Local segmentation optimization: In order to address the problem that the internal structure of the data may be obscured and the classification effect is poor due to global threshold classification, the false anomaly screening and separation model introduces a local segmentation strategy. It performs secondary or even multiple segmentation processing on the local data based on the specific parameters of each sub-distribution in the mixed distribution. This refined classification method effectively avoids the limitations brought by the global threshold, allowing the object information of interest to be extracted more accurately, improving the classification level and accuracy. (3) Introduction of separation evaluation: In order to more intuitively reflect the degree of distinction between the object to be extracted and the background information, the model introduces the separation index. The degree of separation directly reflects the separability between the two: a large degree of separation means that the boundary between the object to be extracted and the background information is clear and easy to separate; conversely, a small degree of separation indicates that there is a large confusion between the two and it is difficult to effectively distinguish them. The introduction of this indicator provides a quantitative basis for evaluating the extraction effect, which helps users to more scientifically evaluate model performance and optimize parameter settings.

[0105] S5, performing spatial statistical analysis on the cloud information by using the Anselm Local Moran Index to quantify the spatial autocorrelation at a specific location.

[0106] Specifically, in order to eliminate the small part of false anomalies remaining in the false anomaly screening and separation model, this embodiment adds the Anselm Local Moran Index to the precise cloud identification model in the snowfall area. As a spatial statistical tool, it is specifically used to quantify the spatial autocorrelation at a specific location. This feature is particularly important in geographic data analysis, environmental science and social science research. The Anselm Local Moran Index can identify high-value or low-value feature areas that are clustered in space, the so-called spatial clusters. In addition, it also has the ability to identify those spatial data points that deviate significantly from the overall pattern in spatial distribution and appear as abnormal. The statistical calculation of the Anselm Local Moran Index is as follows:

[0107]

[0108] Among them, x i is the attribute of feature i, is the mean of the corresponding attribute, w i,j is the spatial weight between features i and j.

[0109]

[0110] Among them, n represents the total number of characteristic elements.

[0111] In this embodiment, the Anselm Local Moran Index is used to eliminate potential false anomalies. Through its sophisticated calculation mechanism, this embodiment can quantify the spatial aggregation and dispersion of clouds and potential false anomalies. This index is calculated independently for each detection point, aiming to evaluate its spatial distribution pattern, whether it is presented as a tight cluster, loose dispersion, or a random dispersion state. The calculated I value, as the core output of the Anselm Local Moran Index, ranges from +1 (indicating extremely strong positive spatial autocorrelation) to -1 (indicating extremely strong negative spatial autocorrelation) in a continuous interval. When the I value is 0, it means that the spatial distribution of the area is random and there is no significant autocorrelation. The positive and negative and size of the I value directly map the aggregation or dispersion state of spatial entities in different geographical areas, providing a quantitative basis for understanding spatial structure.

[0112] Therefore, in this example, the areas with positive I values ​​are interpreted as spatial clustered distributions of clouds or related features. This finding is of great significance for further spatial analysis and decision making.

[0113] The visualization of the representative area a after Anselm local Moran index analysis is shown in the attached figure. Fig.11 As shown in the figure, the red area represents the cloud body, while the yellow area is the potential false anomaly area. This process effectively quantifies the clustering and dispersion characteristics of the entire data set in space, providing an intuitive and scientific evaluation of the spatial distribution pattern of the data.

[0114] It is worth noting that the I value of the snow layer is usually much lower than that of the cloud layer. Even if the cloud layer is relatively thin, its I value always remains above 0, which indicates that the cloud layer tends to form a compact clustered structure in space. In contrast, due to the terrain segmentation caused by the ridge line or the significant temperature difference between the sunny slope and the shady slope, the snow layer is often scattered in a patchy form, which leads to its lower I value. By applying the Anselm Local Moran Index, this embodiment reclassifies part of the false anomaly information in the image as noise false anomalies, and effectively distinguishes the salt and pepper noise generated by the snow-covered area and the high-brightness complex underlying surface. It can be clearly observed from the result figure that these noise points that were originally difficult to distinguish have now been effectively separated from the cloud information, further improving the accuracy and reliability of cloud identification. This process not only demonstrates the powerful ability of the Anselm Local Moran Index to distinguish spatial structural differences in complex environments, but also provides a solid foundation for subsequent spatial data processing and analysis.

[0115] The results of the Anselm local Moran index analysis in the representative area b are as follows Fig.12As shown in the figure, this example found that the false snow anomaly in the northern hilly area, which was originally misjudged due to its spectral characteristics similar to those of clouds, was effectively eliminated after this analysis. This achievement not only improves the accuracy of data processing, but also further verifies the ability of the Anselm Local Moran Index to distinguish between real clouds and false anomaly information in complex environments.

[0116] This embodiment thoroughly analyzes the core mechanism and implementation process of the cloud-snow separation algorithm based on multi-dimensional information fusion. For accurate cloud identification in snowfall areas, this embodiment introduces a hybrid tuned filtering technology to enhance the recognition ability of cloud information while effectively suppressing the spectral interference caused by complex surface background. Subsequently, the false anomaly screening separation model is used to lock the cloud characteristics and further improve the recognition accuracy. Finally, this embodiment uses the Anselm local Moran index to deeply explore the spatial correlation between clouds and potential noise false anomalies, and accurately eliminates the misjudgment information caused by complex terrain. This step not only enhances the robustness of the algorithm, but also ensures the purity of the cloud-snow separation results. The results show that the algorithm exhibits excellent cloud-snow separation performance when dealing with complex scenes such as complex terrain, diverse underlying surface background information, and blurred cloud contours and foggy distribution, and the separation effect is satisfactory. This discovery not only broadens the application scope of cloud identification technology, but also provides strong technical support for subsequent remote sensing data processing and analysis.

[0117] In order to verify the effectiveness of the method, this example compares the proposed method with artificial neural network and Fmask method, which have good effects in cloud extraction. Fig.13 Four regions of interest in the cloud extraction results are shown. Fig.13 The first row is the original image, covering various surface environments including bright background, mountains, ice and snow. The second row is a false color composite image, using bands 7, 4 and 2 to represent red, green and blue, respectively, which helps readers visually distinguish snow / ice from clouds. The 3rd to 5th rows are the results of cloud extraction using ANN, Fmask and the proposed method, respectively, where yellow represents the extracted cloud information and black represents the non-cloud background area.

[0118] In contrast, Fig.13 As shown in the figure, both the ANN and Fmask methods over-recognize cloud information, which also verifies the fact that images containing snow and ice in the background are easily misclassified as clouds. It can be seen from the comparison chart that the method proposed in this article is better in extracting cloud information and effectively reduces the problem of confusion between clouds and snow / ice. The following table is a quantitative display of the recognition results. This embodiment uses three different accuracy evaluation methods to evaluate the accuracy of the algorithm results. The producer accuracy, user accuracy, and overall accuracy of this embodiment are compared with those of the artificial neural network and Fmsk algorithm, as shown in Table 3.

[0119] Table 3

[0120]

[0121] The results of this embodiment are better than those of the artificial neural network algorithm and the Fmask method, especially in terms of user accuracy, which is significantly improved compared with the other two methods. This is because the method effectively improves the phenomenon of misjudging ice and snow as clouds and reduces the occurrence of confusion between clouds and snow. In areas where clouds are difficult to extract, the method achieves high-precision cloud recognition. The results show that the producer accuracy exceeds 95%, the user accuracy exceeds 83%, and the overall accuracy exceeds 90%. Therefore, this method can solve the problem of ice and snow in the background being misjudged as clouds, and the problem of missing point clouds with extremely small shapes, and also achieves better results in cloud-snow separation in complex terrain areas.

[0122] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cloud recognition method for snowfall areas based on multi-dimensional information fusion, characterized in that: The following steps are involved: Obtain remote sensing images of snowfall areas and select bands; Preprocessing the remote sensing image of the selected band to obtain a preprocessed image; Filtering the spectral information of the pre-processed image by hybrid tuned filtering technology to obtain a grayscale image; Separating the grayscale image according to normal distribution characteristics using a false anomaly screening and separation model to obtain cloud body information and non-cloud body information; The cloud information is spatially statistically analyzed using the Anselm Local Moran Index to quantify the spatial autocorrelation at a specific location, eliminate false anomalies in snowfall areas, and obtain cloud identification results.

2. The snowfall area cloud recognition method based on multi-dimensional information fusion according to claim 1 is characterized in that: The selected bands of remote sensing images are 9 bands of Landsat 8OLI data, wherein the 9 bands are specifically Band 1 to Band 9.

3. The snowfall area cloud recognition method based on multi-dimensional information fusion according to claim 1 is characterized in that: The process of preprocessing remote sensing images of selected bands includes: The remote sensing images of the selected bands were resampled using cubic convolution interpolation, and the resolution of Band 8 was resampled from 15 meters to 30 meters to match the resolution of other multispectral bands.

4. The snowfall area cloud recognition method based on multi-dimensional information fusion according to claim 1 is characterized in that: The process of filtering the pre-processed image by using the hybrid tuned filtering technology includes: The endmember spectral information of interest to the user is extracted through a partial unmixing strategy to generate a hybrid tuned filter image. Based on the mixed tuned filter image, a target area with low spectral contrast between the target and the background and inside the target is detected; The feasibility score is used to adjust the abundance image of the target area to obtain a grayscale image.

5. The snowfall area cloud recognition method based on multi-dimensional information fusion according to claim 1, characterized in that: The process of separating the grayscale image according to the normal distribution characteristics by using the false anomaly screening and separation model includes: According to the normal distribution characteristics, the grayscale image is separated by using a false anomaly screening and separation model to obtain cloud body information and non-cloud body information of a bimodal normal distribution; The separability between the probability distributions of the cloud body information and the non-cloud body information is evaluated by using the Bhattacharyya distance.

6. The snowfall area cloud recognition method based on multi-dimensional information fusion according to claim 1, characterized in that: It also includes the process of improving the false anomaly screening and separation model: According to the inherent difference between the target area and its background, a reasonable threshold range is calculated, and based on the threshold range, it is used to distinguish the information area to be extracted and the background area; A local subdivision strategy is introduced to perform multiple subdivisions on the local data according to the specific parameters of each sub-distribution in the mixed distribution to extract the interesting ground object information. Separation degree evaluation is introduced to determine the degree of separation between the object to be extracted and the background information according to the size of the separation degree.

7. The snowfall area cloud recognition method based on multi-dimensional information fusion according to claim 1, characterized in that: The false anomaly screening and separation model is a mixed distribution model, and the expression is: Among them, the parameter set μ=(n,μ1,...,μ n , a1, ..., a n ) contains a series of key elements used to define multiple branches, n represents the total number of branches, f(x, μ i ) represents the probability density of the i-th branch, μ i is the corresponding parameter, a i represents the weight of the i-th branch, and a i >0.

Images