A method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making
By coupling fuzzy C-mean clustering and Bayesian network model, integrating snow cover frequency and terrain factors, the problem of reconstruction of snow information under clouds in high-resolution images is solved, and high-precision snow monitoring is achieved. It provides important technical support for climate change and water resource management.
Patent Information
- Application Number
- CN202411139844.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-20
AI Technical Summary
The prior art is difficult to reconstruct the snow information under the cloud with high precision in high spatial resolution Landsat and Sentinel-2 images, especially in the cloud coverage area. The traditional method cannot effectively utilize the high-temporal resolution MODIS data, resulting in the accumulation of the cloud. Snow monitoring has great challenges.
Using Landsat image snow cover information reconstruction method based on multifactor Bayesian decision-making, the snow cover frequency, terrain elevation and south are integrated by coupling fuzzy C-mean clustering (FCM) and Bayesian network (SBN) models. calculating the membership degree and posterior probability of pixels to achieve accurate reconstruction of snow information in the cloud coverage area.
It significantly improves the reconstruction accuracy and reliability of snow accumulation information under the cloud, overcomes the problem of cloud coverage in remote sensing images, and provides a new technical approach for continuous monitoring of snow accumulation in space and space, and has important application value in climate change research and water resources. Management and disaster warning areas.
Smart Images

Figure CN118657849B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of high-resolution satellite image snow reconstruction under the cloud, and in particular relates to a method for reconstructing snow cover information under the cloud of Landsat images based on multi-factor Bayesian decision making. Background Art
[0002] Snow is one of the key elements in the cryosphere. Due to its high albedo, low thermal conductivity and snowmelt water effect, snow plays a vital role in global energy balance, ecological and hydrological modeling, and climate change research. However, most of the snow is distributed in areas with harsh natural conditions and difficult to access, making it difficult for traditional ground observations to capture the large-scale spatiotemporal variation characteristics of snow. In contrast, satellite remote sensing has become an ideal tool for monitoring snow cover with its wide coverage, short observation period and strong spatiotemporal continuity.
[0003] Microwave remote sensing can effectively detect various properties of snow (such as snow depth, snow water equivalent, melting state and density), but optical remote sensing has shown more advantages in snow identification and snow cover inversion. Since the 1960s, a large number of studies have explored the use of optical images to identify snow. However, optical remote sensing images are easily affected by cloud cover and cannot obtain snow information obscured by clouds. Studies have shown that nearly 50% of optical satellite remote sensing data are affected by clouds. Therefore, how to restore snow information under clouds is the key to continuous monitoring of snow cover.
[0004] Using optical images acquired by MODIS sensors on Terra and Aqua satellites, NASA developed the MOD / MYD10A1 snow cover product, which is widely used for dynamic monitoring of snow cover due to its high temporal resolution (two images are provided every day during the day, with an interval of about 3 hours). However, the spatial resolution of this product is only 500 meters, which is not enough to depict snow cover in mountainous areas with high spatial heterogeneity. Since 2008, 30-meter resolution images acquired by Landsat satellites have been open to users around the world free of charge. The European Space Agency launched the Sentinel-2A / B satellites in 2015 and 2017, which have higher spatial resolution (10 meters, 20 meters and 60 meters) and temporal resolution (revisit period can reach 3-4 days in mid-latitude regions). These high spatial resolution remote sensing image sequences have brought unprecedented opportunities for snow cover monitoring in mountainous areas.
[0005] Scholars have developed a variety of methods for reconstructing snow cover information under clouds for MODIS products, which can be roughly divided into three categories: multi-sensor fusion methods, temporal filtering methods, and spatial filtering methods. Multi-sensor fusion methods require the fusion of multi-sensor images, usually combining optical images and passive microwave data to comprehensively analyze snow cover. However, due to the limited availability of high-resolution image data sets and the huge difference in resolution between passive microwave images (spatial resolution of about 25 km) and high-resolution optical images, this method cannot meet the requirements for reconstructing snow cover under clouds with high spatial resolution images. The temporal filtering method uses images with high temporal resolution to reconstruct snow cover information under clouds through temporal interpolation. However, due to the trade-off between sensor time and spatial resolution, the temporal resolution of Landsat and Sentinel-2 satellite images is low, and this method is not suitable for rapidly changing snow. The spatial aggregation method reduces the spatial resolution and uses high-resolution effective observations in low-resolution pixels to infer the snow cover status of the entire low-resolution pixel. This method comes at the expense of reducing spatial resolution and cannot reconstruct snow cover information under clouds for pixels that are completely covered by clouds. The Snow Line Cloud Removal (SNOWL) algorithm reclassifies cloud pixels as snow-covered or snow-free pixels based on their position relative to the regional snowline elevation. It is more suitable for reconstruction of snow cover under clouds from single-period Landsat and Sentinel-2 images with lower temporal resolution. However, this method is based on the assumption that regional snow cover is mainly affected by elevation. In fact, in addition to elevation, factors such as slope aspect and southness that affect heat distribution, slope that affects snow redistribution, wind speed, and meteorological conditions can also affect local snow cover patterns. Therefore, the results of directly using the SNOWL method to reconstruct snow cover information under clouds in high-resolution images may be highly uncertain.
[0006] Unlike MODIS data with high temporal resolution, which can be comprehensively applied to the above methods, reconstruction of snow cover under clouds from Landsat and Sentinel-2 images with high spatial resolution but low temporal resolution faces great challenges. Therefore, it is urgent to propose a high-precision reconstruction method for snow cover under clouds suitable for high-resolution images, which can provide an effective technical approach for all-weather continuous observation of snow cover. Summary of the invention
[0007] The purpose of the present invention is to provide a method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making. The coupled FCM and SBN snow cover information reconstruction model proposed by us achieves accurate reconstruction of snow cover information in cloud-covered areas by comprehensively considering key factors such as snow cover frequency, terrain elevation, and southness, and shows significant advantages in terms of accuracy and reliability. This research result provides new ideas and methods for solving the cloud cover problem in remote sensing images, which is of great significance for carrying out continuous snow monitoring in time and space, and is expected to play an important role in climate change research, water resources management, and disaster warning.
[0008] To achieve the above object, the present invention provides a method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making, comprising:
[0009] Extracting snow cover data and constructing snow cover proxy indicators; wherein the snow cover data is snow cover information in cloudless areas extracted using Landsat image data, and the snow cover proxy indicators include: monthly snow cover frequency data, south degree data, and DEM data;
[0010] Constructing a snow accumulation reconstruction model under clouds, wherein the snow accumulation reconstruction model under clouds includes: an FCM sub-model and an SBN sub-model;
[0011] The snow cover proxy index is input into the FCM sub-model to calculate the pixel I in the preset image. ij For the kth signal class ω k The membership degree of ij represents the pixel located at the i-th row and j-th column; wherein the preset image is a superimposed synthetic image of the snow cover proxy indicator;
[0012] The snow cover information of the cloud-free area and the membership degree are input into the SBN sub-model to estimate the pixel I ij The posterior probability of belonging to a certain snow cover state;
[0013] Based on the posterior probability, the maximum synthesis method is used to reconstruct the snow cover information under the clouds of Landsat images.
[0014] Optionally, before inputting the snow cover proxy indicator into the FCM sub-model, the method further includes:
[0015] Performing projection transformation on the snow cover information of the cloud-free area and the snow proxy index to convert them into a consistent projection coordinate system;
[0016] The converted snow cover information of the cloud-free area is cropped so that the size of the snow cover data is consistent with the size of the study area.
[0017] Optionally, the method for extracting snow from the snow cover data is:
[0018]
[0019] Among them, Green and SWIR represent the surface reflectance of the green band and shortwave infrared band respectively, and NDSI is the normalized difference snow index.
[0020] Optionally, calculate the pixel I in the preset image ij For the kth signal class ω k The membership includes:
[0021] Assume the preset image C={I ij |1≤i≤N, 1≤j≤M}, the preset image is composed of N rows and M columns of pixels, and the preset image is fuzzy divided into n signal classes, u k (i,j)(1≤k≤n) represents pixel I in the preset image ij For the kth signal class ω k The degree of membership;
[0022] Among them, the membership set U={u k (i,j)} satisfies the following constraints:
[0023]
[0024] Wherein, N represents the number of rows of the preset image, M represents the number of columns of the preset image, i represents the row number of the preset image, and j represents the column number of the preset image;
[0025] The membership set U is iteratively updated to achieve the optimal clustering that minimizes the objective function J, thereby obtaining the final membership.
[0026] Optionally, the objective function J is:
[0027]
[0028] Among them, Ψ={Ψ1,…,Ψ k ,…,Ψ n} is the signal class center point set, Ψ k Is a signal type k The center of the cluster, q is the fuzzy parameter that controls the fuzziness of the cluster. represents the norm of a vector, Represents the pixel value at row number i and column number j.
[0029] Optionally, the posterior probability is:
[0030]
[0031] Among them, P(S v |I ij ) is pixel I ij The posterior probability of being covered by snow, P(S v ) is the snow state S v The prior probability of v is 0 or 1, P(S0) represents the prior probability of no snow, P(S1) represents the prior probability of snow, and the snow and non-snow sample set T extracted from the cloud-free area v Get; P(ω k |I ij ) represents pixel I ij Belongs to the signal classω k The probability of , calculated by the FCM sub-model, is the pixel I in the preset image ij For the kth signal class ω k The membership degree of P(ω k |S v ) is the snow state S v Medium signal classω k The probability of occurrence, P(ω k ) is the signal type ω k probability.
[0032] Optionally, the snow state S v Medium signal classω k The probability of occurrence P(ω k |S v ) is obtained by: extracting snow and non-snow training sample sets T from cloud-free areas v To calculate the probability P(ω k |S v );
[0033] The training sample set T of snow and non-snow extracted from cloud-free areas v To calculate the probability P(ω k |S v )include:
[0034] Define the snow state S v Medium signal classω k The frequency formula is:
[0035]
[0036] in, Indicates the snow state S v Medium signal classω k The frequency and represents the training sample set of snow and non-snow extracted from the cloud-free area, u k (x,y) represents pixel I in the preset image xy For the kth signal class ω kThe degree of membership;
[0037] For the training sample set T v All pixel pairs in the signal class ω k The snow state S is calculated by summing up the membership v Medium signal classω k frequency;
[0038] After calculating all signal classes ω in the training set l After the frequency of (l=1,…,n), the probability P(ω k |S v ) is approximated as:
[0039]
[0040] in, Indicates the snow state S v Medium signal classω k The frequency and Indicates the snow state S v Medium signal classω l The frequency and n is the number of signal classes, and l takes values between 1 and n.
[0041] Optionally, the signal class ω k The probability P(ω k ) is used to calculate the posterior probability P(S v |I ij ) is calculated using the total probability formula:
[0042]
[0043] Among them, P(ω k |S v ) is the snow state S v Medium signal classω k The probability of occurrence, P(S v ) is the snow state S v The prior probability of v is 0 or 1, P(S0) represents the prior probability of no snow, P(S1) represents the prior probability of snow, and the training sample set T of snow and non-snow extracted from the cloud-free area is v get.
[0044] In summary, the pixel I can be calculated ij The posterior probability vector of the snow cover state ρ = (P(S0|I ij ), P(S1|I ij )) to realize the snow state estimation based on the posterior probability vector, and the posterior probability vector satisfies the condition that the probability sum is 1, that is:
[0045] .
[0046] The present invention has the following beneficial effects:
[0047] The proposed model for reconstructing snow information under clouds by coupling FCM and SBN integrates auxiliary parameters such as snow cover frequency, terrain elevation, and southness, fully utilizes the spatial correlation of snow distribution and the influence of terrain on snow distribution, and realizes accurate reconstruction of snow information under cloud cover. The present invention provides a new idea and method for overcoming the cloud cover problem in remote sensing images, which is of great significance for the spatiotemporal continuous monitoring of snow, and is expected to play an important role in climate change research, water resources management, and disaster warning. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] 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:
[0049] Figure 1 Schematic diagram of data processing flow in a method for reconstructing snow cover information under clouds from Landsat images according to an embodiment of the present invention;
[0050] Figure 2 A schematic diagram of a simple Bayesian network according to an embodiment of the present invention;
[0051] Figure 3 A schematic diagram of weight parameter sensitivity according to an embodiment of the present invention;
[0052] Figure 4 It is a schematic diagram of the fuzzy parameter q (LC08_133034_20200130) of an embodiment of the present invention;
[0053] Figure 5 It is a schematic diagram of the fuzzy parameter q (LC08_133034_20220220) of an embodiment of the present invention;
[0054] Figure 6 It is a schematic diagram of the fuzzy parameter q (LC08_133034_20221018) of an embodiment of the present invention;
[0055] Figure 7 It is a schematic diagram of the fuzzy parameter q (LC08_133034_20221103) of an embodiment of the present invention;
[0056] Figure 8 It is a schematic diagram of the fuzzy parameter q (LC09_133034_20230303) of an embodiment of the present invention;
[0057] Fig. 9It is a schematic diagram of the fuzzy parameter q (LC08_133034_20230514) of an embodiment of the present invention;
[0058] Fig.10 It is a schematic diagram of the fuzzy clustering number (LC08_133034_20200130) of an embodiment of the present invention;
[0059] Fig.11 It is a schematic diagram of the fuzzy clustering number (LC08_133034_20220220) of an embodiment of the present invention;
[0060] Fig.12 It is a schematic diagram of the fuzzy clustering number (LC08_133034_20221018) of an embodiment of the present invention;
[0061] Fig.13 It is a schematic diagram of the fuzzy clustering number (LC08_133034_20221103) of an embodiment of the present invention;
[0062] Fig.14 It is a schematic diagram of the fuzzy clustering number (LC09_133034_20230303) of an embodiment of the present invention;
[0063] Fig.15 It is a schematic diagram of the fuzzy clustering number (LC08_133034_20230514) of an embodiment of the present invention;
[0064] Fig.16 This is a schematic diagram of the spatiotemporal changes of snow cover from September 2019 to September 2020 according to an embodiment of the present invention;
[0065] Fig.17 This is a schematic diagram of the spatiotemporal changes of snow cover from September 2020 to September 2021 according to an embodiment of the present invention;
[0066] Fig.18 This is a schematic diagram of the spatiotemporal variation of snow cover from September 2021 to September 2022 according to an embodiment of the present invention;
[0067] Fig.19 This is a schematic diagram of the spatiotemporal variation of snow cover from September 2022 to May 2023 according to an embodiment of the present invention;
[0068] Fig. 20 This is a schematic diagram of the spatiotemporal variation of snow cover from May 2023 to August 2023 according to an embodiment of the present invention;
[0069] Fig.21 Schematic diagram of the change characteristics of the snow cover rate in the upper reaches of the Heihe River according to an embodiment of the present invention;
[0070] Fig. 22The present invention is a flowchart of a method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making according to an embodiment of the present invention. DETAILED DESCRIPTION
[0071] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments 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.
[0072] 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.
[0073] like Figure 1 and Fig. 22 As shown, this embodiment proposes a method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making, including:
[0074] Extracting snow cover data and snow cover proxy indicators; wherein the snow cover data is snow cover information in cloud-free areas extracted using Landsat image data, and the snow cover proxy indicators refer to proxy indicators (such as elevation, slope aspect, etc.) that can indicate the spatial distribution of snow, including but not limited to: monthly snow cover frequency data, south degree data, and DEM data;
[0075] Constructing a snow accumulation reconstruction model under clouds, wherein the snow accumulation reconstruction model under clouds includes: an FCM sub-model and an SBN sub-model;
[0076] The snow cover proxy index is input into the FCM sub-model, and the fuzzy clustering method is used to calculate the pixel I in the preset image. ij For the kth signal class ω k The membership degree of ij represents a pixel located at the i-th row and the j-th column; wherein the preset image is a superimposed synthetic image of the proxy indicator data of the snow accumulation;
[0077] The snow cover information of the cloud-free area and the membership degree are input into the SBN sub-model to estimate the pixel I ij The posterior probability of belonging to a certain snow cover state (snow or no snow);
[0078] Based on the posterior probability, the maximum synthesis method is used to reconstruct the snow cover information under the clouds of Landsat images.
[0079] Furthermore, before extracting snow from the snow cover data, the method further includes:
[0080] Performing projection transformation on the snow cover data into a consistent projection coordinate system;
[0081] The converted snow cover data is cropped to make the size of the snow cover data consistent with the size of the study area.
[0082] Further, calculate the pixel I in the preset image ij For the kth signal class ω k The membership includes:
[0083] Assume the preset image C={I ij |1≤i≤N, 1≤j≤M}, the preset image is composed of N rows and M columns of pixels, and the preset image is fuzzy divided into n signal classes, u k (i,j)(1≤k≤n) represents pixel I in image C ij For the kth signal class ω k The degree of membership;
[0084] By iteratively updating the membership set U, the optimal clustering that minimizes the objective function J is achieved, and the final membership is obtained;
[0085] Based on the posterior probability, the maximum synthesis method is used to reconstruct the snow cover information under the clouds of Landsat images.
[0086] Specifically, the experimental area selected in this embodiment is located in the upper reaches of the Heihe River with Qilian County, Qinghai Province as the center, and its geographical location is between 37°53'N ~38°18'N, 100°7'E ~100°43'E. The area is located in the hinterland of the Qilian Mountains, with mountains in the north and south and valleys in the middle, and generally presents a terrain with high north and south and low middle. The study area presents typical cold-zone alpine landform characteristics and complex terrain. These features can better reflect the algorithm's reconstruction effect on snow cover information under complex terrain clouds. In this area, due to the high altitude and deep in the interior of the continent, far away from the ocean. The study area has obvious characteristics of continental climate and plateau climate. There are glaciers in the high-altitude areas in the north of the study area. The weather is clear and cloudless in winter, and the average annual temperature is only 1°C. The river recharge in the valley mainly comes from snowmelt and glacier meltwater from the mountains on both sides. In this study area, various hydrological elements such as precipitation, snowmelt, and evaporation interact with each other to form significant seasonal changes and nonlinear runoff characteristics.
[0087] The snow cover data collected in this embodiment include three categories: Landsat7 / 8 / 9 L1C product data, monthly snow cover frequency data (SCF), digital elevation model (DEM) data and southness data (SN). SCF, DEM and SN are proxy indicators of snow cover, which are used to describe the spatial distribution pattern of snow cover. These three indicators are used as inputs of FCM, and the probability of each pixel belonging to each signal class, i.e., the degree of membership, is obtained through clustering. The detailed information of the data is shown in Table 1.
[0088] Table 1 Data details
[0089] data quantity time use Landsat7 / 8 / 9L1C 176 2019.9-2023.9 Extraction of clear sky snow and snow under clouds SCF 12 - As algorithm input DEM 1 - As algorithm input SN 1 - As algorithm input
[0090] Landsat7 / 8 / 9 L1C product data: This example selects Landsat 7 / 8 / 9 images from September 2019 to September 2023, screens out images with cloud coverage less than 80%, and uses L1C product data from Landsat 7 ETM+, Landsat 8OLI, and Landsat 9 OLI. Snow in the upper reaches of the Heihe River Basin was extracted through these screened images.
[0091] Monthly snow cover frequency data: Monthly snow cover frequency data is processed and generated on the Google Earth Engine cloud computing platform, integrating all multispectral images acquired by Landsat5 / 7 / 8 / 9 satellites on the Qinghai-Tibet Plateau from 2000 to 2024. The snow extraction method adopts the optimized SNOMAP method.
[0092] Southness data: Southness is a key geomorphic attribute indicator and plays a core role in disciplines such as geography and ecology. It can quantitatively characterize the impact of terrain on the ability to capture solar radiation and is used to evaluate how terrain orientation characteristics affect the solar radiation energy received per unit area of the surface. Southness can be calculated using the following formula:
[0093] (1)
[0094] Where SN is the south degree, slope is the slope, and aspect is the slope direction.
[0095] DEM data: NASA DEM data was selected, which is generated based on the Shuttle Radar Altimetry Program (SRTM) data.
[0096] This embodiment preprocesses the snow cover data, including:
[0097] In ArcGIS, the SCF, DEM, and SN data were projected and transformed to make their coordinate systems consistent with the Landsat image data. The study area range file was clipped to make the range of 176 Landsat images, SCF, DEM, and SN data consistent with the study area range, which was convenient for subsequent experiments.
[0098] Specifically, in this embodiment, extracting snow from the snow cover data includes:
[0099] Snow cover extraction from optical remote sensing images is usually based on the Normalized Difference Snow Index (NDSI). As shown in formula (3), NDSI uses the difference in the reflectance spectrum of snow light in the visible light band and the short-wave infrared band. The snow cover extraction rule is shown in formula (2).
[0100] (2)
[0101] (3)
[0102] In the formula, Green , SWIR They represent the surface reflectance of the green band and the short-wave infrared band, respectively. Snow extraction is completed in the GEE cloud computing platform.
[0103] Specifically, in this embodiment, the reconstruction model of snow accumulation information under the cloud is constructed as follows:
[0104] In order to overcome the challenge of obtaining snow information from high spatial resolution remote sensing images under cloud cover conditions, this embodiment adopts the concept of snow spatial pattern to describe the spatial correlation between snow states at different locations. The snow spatial pattern is based on the following assumption: Affected by conditions such as terrain and sunlight, the distribution of snow in a specific area is a recurring process, which means that some pixels in the area always present similar snow states. Therefore, pixels that always present similar snow states are fuzzy clustered to form a signal class of snow spatial pattern.
[0105] This example proposes a cloud-under-snow reconstruction model (CUSRM) that uses the SBN method coupled with FCM. The model establishes a relationship between the signal class identified by FCM and the snow state S v (where v is 0 for the snow-free state S0 and 1 for the snow-covered state S1). The coupling between SBN and FCM in the CUSRM model is reflected in the following two aspects: (1) By using the fuzzy membership calculated by FCM to define the pixel I ij With signal class ω k (2) By using SBN to combine fuzzy membership and snow training sample data, the signal class ω is established. k With snow state S vThis method is designed to solve the complex many-to-many relationship between signal classes and snow cover states, providing a new way to accurately reconstruct snow cover information under clouds.
[0106] SCF, DEM and SN are three important indicators for reconstructing the model of snow cover information under clouds.
[0107] In order to measure the uncertainty within the category and avoid the error caused by hard clustering, this embodiment uses a fuzzy clustering method to associate pixels with different signal categories with different membership degrees to improve the accuracy of subsequent processing.
[0108] Assume the preset image C={I ij |1≤i≤N, 1≤j≤M}, which consists of N rows and M columns of pixels. The image blur is divided into n signal classes, u k (i,j)(1≤k≤n) represents pixel I in the preset image C ij For the kth signal class ω k The membership degree set U={u k (i,j)} satisfies the following constraints:
[0109] (4)
[0110] The FCM clustering algorithm iteratively updates the membership set U and the signal class center Ψ to achieve the optimal clustering that minimizes the objective function J, where the objective function J is defined as follows:
[0111] (5)
[0112] Among them, Ψ={Ψ1,…,Ψ k ,…,Ψ n} is the signal class center point set, Ψ k Is a signal type k , and q is the fuzzy parameter that controls the fuzziness of clustering.
[0113] P(ω) in SBN k |I ij ) and u in FCM k (i,j) both represent pixel I ij Belongs to the signal classω k From formula (4), we can see that both satisfy the same constraints, so we can use the u calculated by FCM k (i,j) to estimate P(ω k |I ij ) to realize the combination of SBN and FCM.
[0114] like Figure 2The SBN model designed in this embodiment is divided into three layers: pixel layer (I ij represents the pixel located at the i-th row and j-th column), signal class layer (ω k represents the signal class that represents the spatial association of snow cover status between different locations determined by the snow cover proxy indicator), and the snow cover status layer (S0 represents no snow cover, S1 represents snow cover). The goal of the model is to transform the problem of snow reconstruction under the cloud into calculating pixel I ij The posterior probability P(S v |I ij ).
[0115] According to the calculation principle of Bayesian network, P(S v |I ij ) can be calculated by the following formula:
[0116] (6)
[0117] According to the Bayesian formula, P(S v |ω k ) can be expressed as:
[0118] (7)
[0119] Further substituting formula (7) into formula (6) yields:
[0120] (8)
[0121] Among them, P(ω k |I ij ) is pixel I ij Belongs to the signal classω k The degree of can be calculated by FCM, P(ω k |S v ) is a certain snow state S v Medium signal classω k The probability of occurrence, P(S v ) is the prior probability of snow or no snow, P(ω k ) is the signal type ω k The probability, P(S v |I ij ) is pixel I ij The posterior probability of being covered by snow. To calculate the posterior probability of pixel I ij is the posterior probability of snow accumulation P(S1|I ij ), a training sample set T of snow and non-snow extracted from cloud-free areas is required v To calculate the conditional probability P(ω k |S v ). First, define the snow state Sv Medium signal classω k The frequency formula is:
[0122] (9)
[0123] Different from ordinary clustering algorithms, due to the training pixel I ij About Signal Class k The membership degree is not 0 or 1, but a fuzzy membership degree u between 0 and 1. k (i,j), so we need to train the sample set T v All pixel pairs in the signal class ω k The snow state S is calculated by summing up the membership v Medium signal classω k frequency.
[0124] After calculating all signal classes ω in the training set l After the frequency of (l=1,…,n), the conditional probability P(ω k |S v ) is approximated as:
[0125] (10)
[0126] In addition, in formula (8), the prior probability P(ω k ) is used to calculate the posterior probability P(S v |I ij ) is calculated using the total probability formula:
[0127] (11)
[0128] In the formula, P(ω k |S v ) can be estimated by formula (11).
[0129] In summary, the value reflecting pixel I can be calculated by equation (8) ij The posterior probability vector of the snow cover state ρ = (P(S0|I ij ), P(S1|I ij )) to realize the snow state estimation based on the posterior probability vector, and the posterior probability vector satisfies the condition that the probability sum is 1, that is:
[0130] (12)
[0131] Below, this embodiment also performs sensitivity analysis and accuracy verification of snow reconstruction parameters under clouds;
[0132] Parameter sensitivity analysis:
[0133] Based on the constructed SBN subcloud snow information reconstruction model coupled with FCM, this embodiment performs a sensitivity analysis of algorithm parameters. By performing a sensitivity analysis on the algorithm's parameters, we can gain an in-depth understanding of how the algorithm's performance is affected by different parameter settings. This helps to identify which parameters are critical to the results of the algorithm and the optimal range of values for these parameters. Performing a sensitivity analysis on the algorithm can also help find the best parameter combination for the algorithm's performance, thereby optimizing the algorithm's performance. This is crucial to improving the accuracy and efficiency of the SBN subcloud snow information reconstruction model coupled with FCM, and to solving problems such as overfitting and underfitting.
[0134] Weight sensitivity analysis:
[0135] In the algorithm, the corresponding monthly SCF data is selected as the input indicator according to the image date. In addition to the monthly snow cover frequency, DEM and SN are also selected as algorithm input indicators in view of the important influence of topography on snow distribution. Different weights are assigned to SCF, DEM and SN, and the Kappa coefficients under different weight combinations are calculated. The optimal weight combination is found by comparing the size of the Kappa coefficient. The Kappa values obtained by different weight combinations are selected to draw charts. Figure 4 It is a line graph of the Kappa values obtained when the weight parameters are combined in different ways. The weight values of SCF, DEM and SN are respectively increased from 0 to 1 with an increment of 0.1. The sum of the three weight values is 1 for combination, and there are 66 combinations in total. For the results of the reconstruction of the snow cover information under the clouds obtained for each combination, through the calculation of 8 precision evaluation indicators, it was found that: when the SCF weight is 0.7, the DEM weight is 0.2, and the SN weight is 0.1, the Kappa coefficient is 0.727, OA is 0.877, Specificity is 0.91, OE is 0.187, CE is 0.174, Precision is 0.826, Recall is 0.813, and F1 is 0.82. The 8 precision evaluation indicators of the reconstructed image are the best among the 66 weight combinations. From Figure 3 It can also be seen that the Kappa value is relatively the highest when SCF is 0.7, DEM is 0.2 and SN is 0.1, so this weight combination is considered to be the most accurate combination for reconstructing snow cover information under clouds.
[0136] Fuzzy parameter sensitivity analysis:
[0137] To test the fuzzy parameters q The impact on algorithm performance, this embodiment sets 8 different qThe values are 1.2, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, and 5.0. Two images are selected from each of the three periods: the snow accumulation period (October to December), the snow stability period (January to February), and the snow melting period (March to May) for sensitivity analysis of the fuzzy parameter q. Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Fig. 9 This is a bar chart calculated from 8 accuracy evaluation indicators. Figures 4 to 9 The values of OA, Kappa and F1 scores shown in the figure are all flat, with the difference being only around 0.02, and the changes of the other four positive indicators (SPECIFICITY, PRECISION, Recall, F1) are also between 0.1. Among the negative indicators, OE and CE also change only between 0.1 as the q value increases.
[0138] Therefore, it is concluded that the CUSRM algorithm is not sensitive to changes in the fuzzy parameter q value. This shows that the CUSRM algorithm has high stability, is highly robust to small changes in parameters, and is not very dependent on specific parameter values. It is easier to use and adjust in practical applications, because even if the parameter settings are not very precise, relatively stable and reliable results can be obtained. The above results show that the effective working range of the CUSRM algorithm (that is, the parameter range that can produce satisfactory results) is relatively wide, and there is greater freedom to choose parameters without worrying about significantly affecting the quality of the reconstructed image. This also reflects that the mechanism of processing data and generating results within the CUSRM algorithm is relatively mature, and can maintain good performance under different parameter configurations, indicating that wide applicability and stability have been considered in the design of the algorithm.
[0139] Sensitivity analysis of cluster number:
[0140] In order to test the impact of the fuzzy clustering number on the algorithm performance, this embodiment sets five different clustering numbers of 2, 5, 10, 30, and 50, and selects two images from each of the three periods: the snow accumulation period (October to December), the snow stability period (January to February), and the snow melting period (March to May) to perform clustering number sensitivity analysis. Fig.10 , Fig.11 , Fig.12 , Fig.13 , Fig.14 , Fig.15 This is a line chart calculated from 8 accuracy evaluation indicators. Figures 10 to 15It can be seen that the overall accuracy is the highest when the number of clusters is 10, achieving the best snow reconstruction effect. Among the negative indicators, OE and CE have lower values when the number of fuzzy clusters is 10. Overall, in this experiment, when the number of clusters is 10, most of the accuracy evaluation indicators reach the optimal level. With the increase of the number of clusters, the reconstruction accuracy of the snow does not improve. This may be due to the increase in information redundancy and randomness caused by too many clusters, which reduces the stability of the algorithm. Moreover, with the increase of the number of clusters, the running time of the algorithm is significantly improved. Nevertheless, for different fuzzy cluster numbers, its impact on the performance of the algorithm is not significant, and the accuracy evaluation indicators remain basically stable.
[0141] Verification of the accuracy of reconstruction of snow cover information under clouds:
[0142] Since the true value of the snow cover information under the clouds cannot be obtained directly, in order to verify the accuracy of the cloud removal algorithm, an accuracy test based on the "cloud hypothesis" was adopted.
[0143] Accuracy test method based on "cloud hypothesis": randomly select 20 clear sky images without cloud coverage from the three time periods of snow accumulation period (October-December), snow stability period (January-February) and snow melting period (March-May) from September 2019 to September 2023 for experiments. Since these images are cloud-free, the snow status of all pixels under clear sky conditions can be accurately obtained. Randomly select an area in the study area and assume that it is covered by clouds (that is, cloud masking is performed), assign the pixel value of the cloud mask area to 0, and then use the cloud removal algorithm to remove the cloud from the image after the "cloud hypothesis" to obtain the corresponding cloud-free snow map; finally, compare the "calculated value" with the "true value", count the values of each evaluation index, and perform the accuracy test of the cloud removal algorithm.
[0144] In this embodiment, the confusion matrix is used to analyze the image after the CUSRM algorithm reconstructs the snow cover information of the "cloud hypothesis" area and the original image. The selected experimental area has obvious characteristics, complex and diverse terrain, and the mountains have shady slopes and sunny slopes, which can better demonstrate the reconstruction effect of the CUSRM algorithm on the snow cover information under the cloud in complex terrain areas. When the number of clusters of the CUSRM algorithm is set to 10 and the fuzzy parameter q value is set to 1.5, the snow cover map of the "cloud hypothesis" area generated by the CUSRM algorithm shows good results. Table 2 is the eight index results of the accuracy evaluation of 20 images of the reconstruction results of the snow cover information under the cloud by the CUSRM algorithm. The results show that after the snow cover information under the cloud is reconstructed by the CUSRM algorithm, the evaluation indicators include OA, Kappa, Specificity, OE, CE, Precision and F1-Score, which jointly evaluate the algorithm's ability to reconstruct snow cover information under complex cloud occlusion: OA reaches 0.911, indicating that the CUSRM algorithm is highly accurate in distinguishing snow and non-snow areas on the whole; the Kappa coefficient is 0 .786, emphasizing that the algorithm is significantly superior to random classification, especially in scenes with uneven pixel category distribution; OE is as low as 0.051, which means that the CUSRM algorithm performs well in avoiding missing actual snow areas, which is very important for ensuring the integrity of snow monitoring; CE is 0.074, indicating that the CUSRM algorithm can better distinguish different surface features, reduce misclassification between categories, and improve the purity of the results; Precision reaches 0.925, highlighting that when the CUSRM algorithm predicts that a certain area is snowy, this judgment is extremely reliable. F1 is as high as 0.937, which is a comprehensive measure of Precision and Recall, indicating that the CUSRM algorithm achieves a good balance between high precision and high recall in the task of snow recognition.
[0145] In summary, the CUSRM algorithm shows good performance in the task of reconstructing snow cover information under clouds. Although there is room for improvement in the specificity of certain areas, overall, the algorithm provides a powerful technical means for monitoring snow cover under clouds.
[0146] Table 2 Accuracy evaluation of CUSRM snow cover information reconstruction results
[0147] date OA Kappa Specificity OE CE Precision Recall F1 20191026 0.837 0.668 0.81 0.173 0.112 0.888 0.827 0.856 20201028 0.863 0.679 0.918 0.160 0.016 0.984 0.740 0.845 20201231 0.832 0.656 0.800 0.115 0.274 0.726 0.885 0.798 20210201 0.835 0.682 0.835 0.164 0.337 0.663 0.836 0.739 20221018 0.914 0.695 0.905 0.031 0.396 0.604 0.969 0.744 20221103 0.915 0.691 0.913 0.061 0.286 0.514 0.939 0.665 20221119 0.894 0.734 0.938 0.132 0.152 0.848 0.769 0.816 20211121 0.885 0.724 0.938 0.232 0.152 0.848 0.768 0.806 20220220 0.881 0.752 0.895 0.140 0.158 0.842 0.860 0.851 20220324 0.884 0.691 0.861 0.145 0.123 0.877 0.855 0.868 20211210 0.860 0.691 0.869 0.169 0.342 0.658 0.831 0.735 average value 0.872 0.697 0.885 0.138 0.203 0.778 0.843 0.800 20200114 0.874 0.687 0.834 0.163 0.112 0.896 0.827 0.864 20200130 0.867 0.720 0.957 0.257 0.074 0.926 0.743 0.824 20220119 0.869 0.696 0.893 0.104 0.333 0.568 0.896 0.695 20230114 0.87 0.694 0.916 0.148 0.039 0.961 0.5522 0.701 20230215 0.865 0.683 0.869 0.136 0.026 0.974 0.864 0.916 average value 0.871 0.696 0.894 0.162 0.117 0.865 0.776 0.800 20210508 0.939 0.746 0.942 0.087 0.318 0.682 0.913 0.780 20230303 0.911 0.786 0.824 0.051 0.074 0.925 0.943 0.937 20230319 0.877 0.744 0.945 0.188 0.060 0.940 0.802 0.866 20230514 0.935 0.787 0.930 0.035 0.279 0.721 0.965 0.825 average value 0.916 0.765 0.910 0.127 0.183 0.817 0.906 0.852 Overall average 0.881 0.712 0.890 0.134 0.181 0.805 0.840 0.808
[0148] Furthermore, this embodiment uses the CUSRM algorithm to extract snow and non-snow from Landsat7 / 8 / 9 images to extract snow cover, and analyzes the changing characteristics of snow coverage and daily spatiotemporal distribution of snow in the upper reaches of the Heihe River from September 2019 to September 2023.
[0149] Snow Cover Ratio (SCR) refers to the percentage of snow-covered area in a specific area. In order to reflect the characteristics of the change of snow cover ratio in the upper reaches of the Heihe River, the period from September 1, 2019 to September 1, 2023 was divided into a hydrological year (September 1 to August 30 of the following year), and the change trend of snow cover ratio in the upper reaches of the Heihe River in the past five years was analyzed. October to December of each year is defined as the accumulation period of snow, January to February is defined as the stable period of snow, March to May is the melting period of snow, and June to September is the non-snow season. The snow cover ratio is calculated by formula (13):
[0150] (13)
[0151] Where SCR is the snow cover ratio, which is used to describe the spatiotemporal variation characteristics of snow cover in the upper reaches of the Heihe River. is the snow-covered area in the basin, It refers to the entire area of the upper reaches of the Heihe River. Figures 16 to 20 The spatiotemporal variation of snow cover in the upper reaches of the Heihe River from September 2019 to September 2023 obtained by the CUSRM algorithm is described. At the same time, in order to understand the interannual variation characteristics of the snow cover rate in the upper reaches of the Heihe River in the past five years, the snow cover rate of each image is calculated according to formula (13), as follows: Fig.16 shown.
[0152] Fig.16 This is a schematic diagram of the spatiotemporal changes in snow cover from September 2019 to September 2020. Due to the limitation of the temporal resolution (16 days) of Landsat7 / 8 / 9 satellites, images with excessive cloud cover are removed. By combining the data from Landsat7 / 8, there are 3 images per month, and a total of 39 images in this hydrological year can obtain snow cover information. Fig.16 It can be seen that the snow-covered area began to increase gradually from October to December 2019, and the snow-covered area fluctuated slightly from January to February, when the snow entered a stable period. From March to May, when the snow entered a melting period, the snow-covered area gradually decreased. After entering summer, the snow in the basin is only distributed in areas with higher altitudes.
[0153] Fig.17 This is a schematic diagram of the spatiotemporal changes in snow cover from September 2020 to September 2021. As the number of Landsat9 data images available from 2021 increases, images with better accuracy can be selected and are more conducive to time series analysis. A total of 40 images were obtained in this hydrological year. Fig.17It can be seen from the data that the snow cover area began to increase in October in this hydrological year. The largest snow cover area of the year appeared on November 29, accounting for about 72% of the total basin area. The snow cover area tended to be flat in January and February, and gradually decreased from March to enter the snow melting period. The snow cover area in summer only accounts for about 10% of the total basin on average.
[0154] Fig.18 This is a schematic diagram of the spatiotemporal changes in snow cover from September 2021 to September 2022. There are 38 images in this hydrological year. The changes in the snow cover area are similar to the previous year. The snow cover area is smaller than that of 2020. The snow cover reached its maximum extent on January 27, accounting for about 70% of the total basin. From March on, although snowfall still occurred, the snow cover area showed an overall downward trend. In summer, snow only exists in high-altitude areas within the basin. Combined with Fig.18 In this hydrological year, the snowfall is mainly concentrated in winter and spring, and the snowfall in spring is heavier.
[0155] Figure 19-20 This is a schematic diagram of the spatiotemporal changes in snow cover from September 2022 to August 2023. The number of images obtained in this hydrological year increased compared with previous years, with a total of 52 images. The snow-covered area began to increase in October and reached the maximum snow-covered area on November 11, accounting for about 87% of the basin area. However, the snow-covered area fluctuated greatly from October to January, and the snow-covered area tended to stabilize in February. Snow only exists in the high-altitude areas of the basin in summer.
[0156] Fig.21 As shown in the figure, the annual variation characteristics of snow cover rate in the upper reaches of the Heihe River in the past five years show a multi-peak pattern. There are significant differences in the snow cover area in each hydrological year, and the maximum snow cover area gradually increases. In addition, the concentrated snowfall seasons are also different in different hydrological years.
[0157] In the current field of earth observation, remote sensing technology, as an important means of obtaining surface information, faces a huge challenge in its application in snow monitoring: the lack of snow information under the cloud in low temporal resolution and high spatial resolution remote sensing images caused by cloud shielding. This problem greatly limits the development of relevant research such as accurate assessment of snow cover and management of ice and snow water resources. To this end, this embodiment innovatively proposes a snow information reconstruction model under the cloud that couples fuzzy C-means clustering and SBN, aiming to effectively solve the above problems.
[0158] Firstly, the FCM algorithm is used to classify the snow cover in cloudless areas to extract snow cover features. Then, the SBN model is constructed to predict the snow cover in cloud-covered areas by combining the advantages of Bayesian networks. By integrating auxiliary parameters such as snow cover frequency, terrain elevation, and southness in the spatiotemporal neighborhood, the model can make full use of spatial correlation and terrain influence laws to accurately reconstruct snow cover information under cloud cover. This method not only takes into account the spatial heterogeneity of snow distribution, but also enhances the model's ability to adapt to snow changes under complex terrain conditions.
[0159] The experimental results show that the coupled FCM-SBN model proposed in this embodiment shows excellent performance in the task of reconstructing snow information under clouds. Among multiple accuracy evaluation indicators, especially in overall accuracy and Kappa coefficient, the model shows high consistency and stability, which verifies the effectiveness of the model. More importantly, through the sensitivity analysis of algorithm parameters, it is found that the algorithm has relatively low requirements for parameter settings, and the change of the fuzzy parameter q value has little effect on the accuracy of the algorithm, which greatly reduces the risk of overfitting caused by parameter sensitivity, thereby ensuring the generalization ability and robustness of the model and improving its practical value. The optimal weight combination of the three indicators SCF, DEM, and SN was found, and it was found that when the number of clusters is 10, the accuracy and running time of the algorithm are better than other values. Through accuracy verification, it was found that the overall accuracy averaged 0.881, and the Kappa value averaged up to 0.712. The coupled FCM-SBN snow information reconstruction model proposed in this embodiment not only achieves accurate reconstruction of snow information in cloud-covered areas by comprehensively considering key factors such as snow cover frequency, terrain elevation, and southness, but also shows significant advantages in improving the accuracy and reliability of reconstruction results. This research result provides new ideas and methods for overcoming the cloud cover problem in remote sensing images, which is of great significance for improving the accuracy and timeliness of snow monitoring, and is expected to play an important role in climate change research, water resources management, and disaster warning.
[0160] 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 method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making, characterized in that: include: Extracting snow cover data and constructing snow cover proxy indicators; wherein the snow cover data is snow cover information in cloudless areas extracted using Landsat image data, and the snow cover proxy indicators include: monthly snow cover frequency data, south degree data, and DEM data; Constructing a snow accumulation reconstruction model under clouds, wherein the snow accumulation reconstruction model under clouds includes: an FCM sub-model and an SBN sub-model; The snow cover proxy index is input into the FCM sub-model to calculate the pixels in the preset image. I ij For k Signal Class ω k The membership degree of ; among them, I ij Indicates that it is located at i Line j The preset image is a superimposed composite image of the snow cover proxy indicator; Calculate the pixels in the preset image I ij For k Signal Class ω k The membership includes: Assume the preset image C={ I ij |1≤ i ≤ N , 1≤ j ≤ M }, the preset image is N OK M The preset image is blurred and divided into n Signal class, u k ( i , j ), 1≤ k ≤ n , indicating the pixels in the preset image I ij For k Signal Class ω k The degree of membership; Among them, the membership set U={ u k ( i , j )} satisfies the following constraints: in, N Indicates the number of rows of the preset image. M Indicates the number of columns of the preset image. i Indicates the row number of the preset image. j Indicates the column number of the preset image; Update the membership set by iteration U , in order to achieve the objective function J The smallest optimal clustering is used to obtain the final membership degree; The objective function J for: in, Ψ = { Ψ 1 ,…,Ψ k ,…,Ψ n } is the signal class center point set, Ψ k Is a signal type k The center of q is the fuzzy parameter that controls the fuzziness of clustering. represents the norm of a vector, Represents the pixel value at row number i and column number j; The snow cover information of the cloud-free area and the membership degree are input into the SBN sub-model to estimate the pixel I ij The posterior probability of belonging to a certain snow cover state; The posterior probability is: in, P ( S v | I ij ) is a pixel I ij The posterior probability of being covered by snow, P ( S v ) is snow state S v The prior probability of v The value is 0 or 1. P ( S 0) represents the prior probability of no snow, P ( S 1) Represents the prior probability of snow, a training sample set of snow and non-snow extracted from cloud-free areas T v get; P ( ω k | I ij ) represents pixels I ij Belongs to the signal class ω k The probability of is calculated by the FCM sub-model, that is, the pixel in the preset image I ij For k Signal Class ω k The degree of membership; P ( ω k | S v ) is snow state S v Medium signal class ω k The probability of occurrence, P ( ω k ) is a signal class ω k The probability of Snow status S v Medium signal class ω k Probability of occurrence P ( ω k | S v ) is obtained by: extracting snow and non-snow training sample sets from cloud-free areas T v To calculate the probability P ( ω k | S v ); Snow and non-snow training sample sets extracted from cloud-free areas T v To calculate the probability P ( ω k | S v )include: Defining snow status S v Medium signal class ω k The frequency formula is: in, Indicates snow conditions S v Medium signal class ω k The frequency and represents the training sample set of snow and non-snow extracted from cloud-free areas, u k ( x , y ) represents the pixels in the preset image I xy For k Signal Class ω k The degree of membership; Snow and non-snow training sample sets extracted from cloud-free areas T v All pixel pairs in the signal class ω k The membership degree of is summed to calculate the snow state S v Medium signal class ω k frequency; Snow and non-snow training sample sets extracted from the calculated cloud-free area T v All signal classes in ω l , l=1,…,n, After the frequency, the probability P ( ω k | S v ) is approximated as: in, Indicates snow conditions S v Medium signal class ω k The frequency and Indicates snow conditions S v Medium signal class ω l The frequency and n is the number of signal classes, l The value range is 1 to n between; Signal Class ω k Probability P ( ω k ) is to calculate the posterior probability P ( S v | I ij ) is calculated using the total probability formula: in, P ( ω k | S v ) is snow state S v Medium signal class ω k The probability of occurrence, P ( S v ) is snow state S v The prior probability of v The value is 0 or 1. P ( S 0) represents the prior probability of no snow, P ( S 1) Represents the prior probability of snow, a training sample set of snow and non-snow extracted from cloud-free areas T v get; Based on the posterior probability, the maximum synthesis method is used to reconstruct the snow cover information under the clouds of Landsat images.
2. The method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making according to claim 1, characterized in that: Before inputting the snow cover proxy index into the FCM sub-model, the following steps are also included: Performing projection transformation on the snow cover information of the cloud-free area and the snow proxy index to convert them into a consistent projection coordinate system; The converted snow cover information of the cloud-free area is cropped so that the size of the snow cover data is consistent with the size of the study area.
3. The method for reconstructing snow cover information under clouds from Landsat images based on multi-factor Bayesian decision making according to claim 1, characterized in that: The method to extract snow cover data is: in, Green , SWIR They represent the surface reflectance of the green band and shortwave infrared band respectively, and NDSI is the normalized difference snow index.
