A comprehensive cloud and cloud shadow under the snow information reconstruction method

Abstract

The application discloses a comprehensive snow information reconstruction method, tests Sentinel-2 and Landsat 8 images, identifies thick, thin and broken clouds through cloud spectrum testing and cloud probability testing, divides the ground cover type through NDVI by using an improved cloud shadow detection method, and identifies cloud shadows in the NIR band and the SWIR1 band; clear sky snow pixels in the cloud-free image and the multi-cloud shadow image of Landsat 8 and Sentinel-2 are extracted by using a Sen2Cor snow detection method; the improved SNOWL method extracts snow under clouds and cloud shadows, removes unstable snow area pixels from the preliminary snow reconstruction result, and obtains snow distribution. The reconstruction method improves the cloud shadow detection precision and the snow reconstruction precision under clouds and cloud shadows, applies Sentinel-2 and Landsat 8 images to monitor snow, enhances the description of the snow detail features in a complex mountainous area, and meets the high-precision snow detection requirement in the mountainous area.

Description

Technical Field

[0001] This invention belongs to the field of Earth and Space Science and Technology, and relates to a comprehensive method for restoring cloud and snow cover information under cloud shadows (combining Sen2Cor, Fmask4.0 and SNOWL algorithms) to reconstruct cloud and snow cover information under cloud shadows in Landsat 8 and Sentinel-2 images. Background Technology

[0002] Optical satellite remote sensing provides an important data source for snow cover monitoring and research. Currently, polar-orbiting and geostationary satellites have released various snow cover products, such as AVHRR, MODIS, Meteosat / MSG, and FY. However, mountain snow cover is characterized by high albedo, strong spatial heterogeneity, and rapid changes, making it difficult for these low- to medium-spatial-resolution snow cover products to accurately describe the spatiotemporal distribution characteristics of small-scale mountain snow cover, and they often exhibit strong saturation for snow cover. The successful launch of Landsat 8 and Sentinel-2A / B satellites has brought new opportunities for remote sensing snow cover identification and monitoring in mountainous areas. Compared with the Landsat 1-7 series satellites (8 bits), Landsat 8 has a higher radiometric resolution (12 bits); while the Sentinel-2 satellites not only provide higher radiometric resolution (12 bits), spectral resolution (13 bands), and spatial resolution imagery (10 m), but its A / B satellite network can also provide higher temporal resolution imagery of 2–5 days. Therefore, the combined use of these two types of high-resolution satellite remote sensing data with low revisit cycles is highly beneficial for studying rapid changes and redistribution processes of snow cover, and greatly improves the ability to monitor temporal changes in snow cover in mountainous areas.

[0003] However, optical remote sensing imagery, such as Landsat 8 and Sentinel-2 images, is severely affected by clouds and cloud shadows. Clouds not only alter the energy radiation transfer between the sun, the land surface, and the sensor, preventing accurate snow cover information from reaching the sensor, but also, cloud shadows share similar spectral characteristics with wetlands, water bodies, and other land features, leading to low accuracy in identifying these features. Therefore, accurately obtaining snow cover information under clouds and cloud shadows in Landsat 8 and Sentinel-2 A / B images is the primary task in snow cover time-series analysis. Cloud removal methods based on snowline elevation are more suitable for reconstructing snow cover information in mountainous areas under clouds and cloud shadows in Landsat 8 and Sentinel-2 images with low temporal resolution. By comparing cloud pixels with regional snowline elevations, cloud pixels are reclassified as either snow or land pixels. Therefore, snow cover information reconstruction under clouds and cloud shadows in Landsat 8 and Sentinel-2 images based on the SNOWL method mainly includes cloud and cloud shadow detection and snow cover information reconstruction under clouds and cloud shadows.

[0004] Mainstream cloud and cloud shadow detection methods include ACCA, Fmask, MCM, LaSRC, MAJA, Sen2cor, and ATSA. ACCA is primarily used for Landsat 4 / 5 and Landsat 7 imagery, but it suffers from high cloud omission and misclassification errors. MCM detects clouds and cloud shadows by utilizing the reflectance difference between clear-sky pixels and cloud and cloud image pixels, but it can only detect thick clouds. LaSRC generates a cloud mask for Landsat 8 imagery during atmospheric correction and extracts cloud shadows using band thresholds, but its accuracy is low for thin clouds. The Sen2Cor tool provided by the European Space Agency (ESA) can effectively distinguish Sentinel-2 clouds, cloud shadows, and snow cover. MAJA is a spectral-temporal method for cloud detection and atmospheric correction, widely used in the Landsat series and Sentinel-2 satellites, but it overestimates clouds and cloud shadows. ATSA is used in Landsat 4 / 8 and Sentinel-2 imagery, but it has poor separation of clouds and snow. The Fmask method has high accuracy and fast computation speed for cloud and cloud shadow detection, but cloud shadow detection is determined by the geometric positional relationship between clouds and cloud shadows, which is very easy to mix with dark surface pixels (water bodies, terrain shadows, etc.), resulting in cloud shadow omissions.

[0005] Clear-sky snow cover extraction is a prerequisite for determining the instantaneous snowline using the SNOWL method. Based on the high reflectivity of snow, and by comprehensively applying NDSI, visible light, and near-infrared thresholds, snow cover can be easily distinguished from other ground features. The Sen2Cor snow cover detection method can extract clear-sky snow cover pixels from Sentinel-2 imagery with high accuracy. However, previous SNOWL methods for reconstructing snow cover information under clouds and cloud shadows have some shortcomings, such as the inability to effectively detect clouds and cloud shadows, or the failure to consider the influence of unstable snow cover areas when estimating the instantaneous snowline, resulting in lower accuracy of the snow cover information reconstruction results.

[0006] The reconstruction method of this invention not only improves the accuracy of cloud shadow detection, but also improves the accuracy of reconstructing information on clouds and snow cover under cloud shadows. Furthermore, for mountainous areas with complex terrain, by simultaneously applying Sentinel-2 and Landsat 8 images to monitor snow cover, the description of detailed features of snow cover in mountainous areas is enhanced, and the requirements for high-precision snow cover detection in mountainous areas are met. Summary of the Invention

[0007] The purpose of this invention is to provide a comprehensive method for reconstructing snow cover information under clouds and cloud shadows. By reducing the impact of some clouds and cloud shadows on snow cover monitoring in optical remote sensing images, the method reconstructs snow cover under clouds and cloud shadows in Landsat 8 and Sentinel-2 images, thereby improving the accuracy of snow cover monitoring in mountainous areas.

[0008] The technical solution adopted in this invention is: a comprehensive method for reconstructing snow cover information under clouds and cloud shadows, which is carried out according to the following steps:

[0009] Step 1: Cloud and cloud shadow detection:

[0010] 1) Based on the physical characteristics of clouds, conduct basic tests, whiteness tests, ratio tests, haze-to-optimal conversion tests, and cirrus tests on clouds in Sentinel-2 and Landsat 8 images to distinguish them from other ground features with similar spectral characteristics to clouds.

[0011] In the basic test, NDSI, NDVI, and SWIR2 band thresholds were used to separate clouds from vegetation and snow cover areas, specifically:

[0012] LBasic = ρ SWIR2 >0.03 andT <27℃ andNDSI <0.8 and NDVI <0.8 (5)

[0013] SBasic = ρ SWIR2 >0.03 andNDSI <0.8 and NDVI <0.8 (6)

[0014] In the formula, LBasic This represents the basic tests for Landsat 8; SBasic Represents the basic tests of Sentinel-2; NDSI It is the normalized snow cover index; NDVI It is the normalized vegetation index; T It is the thermal infrared band of Landsat 8 data. ρ SWIR2 It is the surface reflectance of the shortwave infrared band 2;

[0015] 2) Detecting inaccurately identified fragmented clouds and cloud edges in cloud spectral testing using cloud probability, divided into two cases:

[0016] If there is no large body of water in the detection area, and only land cloud probability calculation is performed, then the haze-optimized conversion value is normalized to generate a new temperature probability from the Sentinel-2 image, replacing the temperature probability that cannot be calculated due to the lack of thermal infrared bands. Then, the temperature probabilities from Landsat 8 and Sentinel-2 are combined with the spectral difference probability to calculate the final cloud probability.

[0017] Spectral difference probability: lVar =1-max[ White , abs ( NDVI ), abs ( NDSI )] (7)

[0018] Land temperature probability:

[0019]

[0020] Cloud probability: SlCloudp = lVar × lHOT + Cirrus ×0.5

[0021] LlCloudp = lVar × lTemp+Cirrus ×0.3 (9)

[0022] in, lVar , lHOT , lTemp , SlCloudp , LlCloudp These represent the spectral difference probability of Sentinel-2 and Landsat 8, the land temperature probability of Sentinel-2, the land temperature probability of Landsat 8, the cloud probability of Sentinel-2, and the cloud probability of Landsat 8, respectively. HOT high The high value represents the haze-to-optimal conversion value of Sentinel-2; HOT low The low value represents the haze-to-optimal conversion value of Sentinel-2; T high This represents the high value of the temperature band in Landsat 8; T low This represents the low value of the temperature band in Landsat 8;

[0023] If a large body of water exists in the detection area, the probability of clouds over water is calculated. The probability of clouds over water is a combination of the probability of water temperature and the probability of brightness. Clear-sky water pixels are identified through "water testing" and the SWIR1 band reflectivity threshold, and the probability is estimated using the higher level of clear-sky water temperature.

[0024] Probability of surface water temperature:

[0025]

[0026] Brightness probability: wBri =min( ρ SWIR1 ,0.11) / 0.11 (11)

[0027] Probability of clouds over water:

[0028] SwCloudp = wBri + Cirrus ×0.5

[0029] LwCloudp = wTemp × wBri + Cirrus ×0.3 (12)

[0030] in, Water Represents water testing; wTemp , wBri These represent the probability of water surface temperature and the probability of light, respectively. SwCloudp , LwCloudp These represent the probabilities of clouds over water in Sentinel-2 and Landsat 8, respectively.

[0031] 3) In areas with dense vegetation cover, the land surface is divided into vegetation cover and bare ground cover, and cloud shadows are identified in the NIR and SWIR1 bands:

[0032] NDWI <0.1 (13)

[0033]

[0034] ρ NIR It is the surface reflectance in the near-infrared band; ρ SWIR1 It is the surface reflectance of the shortwave infrared band 1; NDWI It is the normalized water quality index; NDVI It is the normalized vegetation index;

[0035] Step 2: Snow Extraction

[0036] 1) Extraction of snow accumulation in clear skies

[0037] Simultaneously extract clear-sky snow pixels from Landsat 8 and Sentinel-2 cloudless and cloudy images using the Sen2Cor snow cover detection method:

[0038]

[0039] 2) Extraction of snow accumulation under clouds and their shadows:

[0040] Step 1: Calculate the average elevation of snow-covered pixels and the elevation of clouds and cloud image pixels respectively. After comparison, reclassify cloud pixels. The reclassified cloud image is the result of snow information reconstruction.

[0041] Step 2: Calculate the initial snow cover distribution for the number of days with snow cover over many years using clear-sky snow cover images identified by the Sen2Cor method:

[0042]

[0043] In the formula, SCD Indicates the number of days the snow cover lasted; N Indicates the number of years with snow cover; S i Indicates the first year of a snow-covered period i Images of snow-covered skies on a clear day;

[0044] Then, unstable snow cover distribution areas are obtained with a 22-day threshold. Pixels of unstable snow cover areas are removed from the preliminary snow cover information reconstruction results to obtain the final snow cover distribution.

[0045] The snow cover information reconstruction method of the present invention is a method for reconstructing snow cover information under clouds and cloud shadows by integrating Fmask cloud and cloud shadow detection, Sen2Cor snow cover detection, and SNOWL method. Using Sentinel-2 and Landsat 8 imagery as input data, the spectral characteristics of clouds were analyzed. For these two remote sensing images, cloud detection was performed on both Sentinel-2 and Landsat 8 images based on the spectral characteristics of clouds and the Fmask4.0 cloud detection method. The cloud shadow detection method was improved based on the spectral characteristics of cloud shadows in the near-infrared and short-wave infrared bands and the spectral characteristics under different land cover conditions to detect cloud shadows in both images. Clear-sky snow cover pixels were extracted from clear-sky and cloud-covered images using the Sen2Cor method. Based on cloud and cloud shadow detection, the snow cover information under clouds and cloud shadows in Sentinel-2 and Landsat 8 images was reconstructed using the improved SNOWL method, which introduces unstable snow cover regions. The accuracy of the clear-sky snow cover extraction results from Sentinel-2 and Landsat 8 images was verified using GF-2 imagery, and further verification was performed using GF-2 imagery from adjacent time periods and Landsat 8 imagery from other remote sensing images. The accuracy of the snow cover information reconstruction results of the eight nearby Sentinel-2 images was verified.

[0046] The reconstruction method of this invention is based on the GEE (Google Earth Engine) platform, which aggregates petabyte-scale massive geographic information spatial data, including satellite / aerial remote sensing, basic geographic information and its application products, and is updated daily.

[0047] Because clouds and cloud shadows severely pollute optical remote sensing, they significantly affect the accurate extraction of snow cover information. This invention's snow cover information reconstruction method uses Sentinel-2 and Landsat 8 as input data, considering the influence of clouds and cloud shadows, the strong spatial heterogeneity of snow cover in mountainous areas, and the band differences between different images. It reconstructs the snow cover under clouds and cloud shadows in Sentinel-2 and Landsat 8 images, not only improving the accuracy of snow cover information extraction results but also enhancing the ability to monitor temporal changes in snow cover in mountainous areas. Attached Figure Description

[0048] Figure 1 This is a comparison image of the improved cloud shadow detection algorithm and the original algorithm in the snow cover information reconstruction method of this invention for identifying cloud shadows in Landsat 8 imagery and Sentinel-2 imagery.

[0049] Figure 2 It is a spectral curve of clouds and snow.

[0050] Figure 3 This is a clear-sky snow map extracted from Landsat8 and Sentinel-2 cloudless images using the Sen2Cor snow detection method.

[0051] Figure 4 This is a flowchart of the algorithms for reconstructing snow cover information under clouds and cloud shadows using Sentinel-2 and Landsat 8.

[0052] Figure 5 This is a comparison of snow cover images from GF-2 and snow cover images under Sentinel-2 cloud and cloud shadow coverage reconstructed using the improved and original SNOWL algorithms.

[0053] Figure 6 This is a comparison of snow cover images from Sentinel-2 and snow cover images under Landsat 8 cloud and cloud shadow coverage reconstructed using the improved and original SNOWL algorithms. Detailed Implementation

[0054] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0055] Optical remote sensing images are severely contaminated by clouds; therefore, the primary prerequisite for snow cover information reconstruction is the detection of clouds and their shadows. Cloud detection is a method based on the physical properties of clouds (whiteness, brightness, coldness, etc.) and uses methods such as band thresholding and Normalized Difference Vegetation Index (NDVI) and Normalized Difference Snow Index (NDSI) to separate cloud pixels from surface objects. It involves two steps: cloud spectral testing and cloud probability calculation. Because solar radiation is blocked by clouds, atmospheric scattering in cloud shadows results in strong energy at shorter wavelengths (such as the visible light band) and relatively weaker scattered radiation energy at longer wavelengths (such as the near-infrared / shortwave infrared band), making shadow pixels darker than their surroundings. Therefore, cloud shadows are easier to detect at longer wavelengths. However, in the SWIR2 band, the reflectivity of cloud shadows is similar to that of water, making them difficult to distinguish. In the NIR and SWIR1 bands, the reflectivity of cloud shadows is higher than that of water; therefore, the NIR and SWIR1 bands are typically used for cloud shadow detection.

[0056] Therefore, the snow cover information reconstruction method of this invention first uses Sentinel-2 and Landsat 8 imagery as input data to perform cloud and cloud shadow detection. Based on the cloud and cloud shadow detection, snow cover detection is performed, including clear-sky snow cover detection and snow cover information reconstruction under clouds and cloud shadows. Furthermore, improvements are made to address the issue of missed cloud image pixels, thereby increasing the accuracy of snow cover detection. The snow cover information reconstruction method of this invention is carried out according to the following steps:

[0057] Step 1: Cloud and Cloud Shadow Detection

[0058] 1) Cloud detection mainly includes two parts: cloud spectral testing and cloud probability testing. Cloud probability testing is further divided into cloud probability over water and cloud probability over land.

[0059] Cloud spectral testing detects cloud layers based on their physical characteristics, including brightness, whiteness, coolness, altitude, and the increase in temperature with altitude. This includes basic testing, whiteness testing, haze-to-brightness conversion detection, ratio detection, and cirrus cloud testing, thereby distinguishing other ground features with similar spectral characteristics to clouds. Specifically, it detects potential cloud layers in Sentinel-2 and Landsat 8 imagery.

[0060] Whiteness test:

[0061]

[0062] Haze-to-good conversion detection: HOT =( ρ Blue -0.5× ρ Red -0.08)>0 (2)

[0063] Ratio test: Rock = ρ NIR / ρ SWIR1 >0.75 (3)

[0064] Cirrus cloud detection: Cirrus = ρ Cirrus / 0.04>0.01(4)

[0065] In the formula, White , MeanVis , HOT , Rock , Cirrus These represent whiteness test, visible light band mean, haze-to-excellence conversion detection, ratio detection, and cirrus cloud detection, respectively. ρ Blue , ρ Green , ρ Red , ρNIR , ρ Cirrus , ρ SWIR1 These represent the surface reflectance of blue, green, red, near-infrared, cirrus, and shortwave infrared band 1, respectively.

[0066] The thermal infrared band (T) can effectively capture the "cold" characteristics of clouds. Based on the presence or absence of the thermal infrared band, the basic tests for cloud detection in Landsat 8 and Sentinel-2 images are distinguished. Other test modules (whiteness test, haze-to-excellence conversion detection, ratio detection, and cirrus cloud test) are used in the same way in Landsat 8 and Sentinel-2. Due to the "bright" and "cold" characteristics of clouds, the reflectance of the cloud in the SWIR2 band should be greater than 0.03, and T should be less than 27°C. The "white" characteristics of clouds usually result in NDSI and NDVI values ​​near zero. The NDVI and NDSI values ​​of thin clouds in vegetated areas are not higher than 0.8. Since Sentinel-2 lacks a thermal infrared band, it can only use NDSI, NDVI, and SWIR2 band thresholds to separate clouds from vegetation and snow cover areas, as shown in formulas (5) and (6):

[0067] LBasic = ρ SWIR2 >0.03 andT <27℃ andNDSI <0.8 and NDVI <0.8 (5)

[0068] SBasic = ρ SWIR2 >0.03 andNDSI <0.8 and NDVI <0.8 (6)

[0069] In equations (5) and (6), LBasic This represents the basic tests for Landsat 8; SBasic Represents the basic tests of Sentinel-2; NDSI It is the normalized snow cover index; NDVI It is the normalized vegetation index; T It is the thermal infrared band of Landsat 8 data. It is the surface reflectance of the shortwave infrared band 2.

[0070] 2) While cloud spectral analysis can detect potentially thick cirrus cloud pixels, it is inaccurate for identifying some fragmented thin clouds and cloud edges. Therefore, cloud probability is introduced for further detection of these pixels, divided into two cases:

[0071] If there is no large body of water in the detection area, and only the probability of land clouds is calculated, the Haze-Optimism Transformation (HOT) value is normalized to generate a new temperature probability from the Sentinel-2 image, replacing the temperature probability that cannot be calculated due to the lack of thermal infrared bands. Then, the temperature probabilities from Landsat 8 and Sentinel-2 are combined with the spectral difference probability to calculate the final cloud probability, as shown in formulas (7), (8), and (9).

[0072] Spectral difference probability: lVar =1-max[ White , abs ( NDVI ), abs ( NDSI )] (7)

[0073] Land temperature probability:

[0074]

[0075] Cloud probability: SlCloudp = lVar × lHOT + Cirrus ×0.5

[0076] LlCloudp = lVar × lTemp+Cirrus ×0.3 (9)

[0077] in, lVar , lHOT , lTemp , SlCloudp , LlCloudp These represent the spectral difference probability of Sentinel-2 and Landsat 8, the land temperature probability of Sentinel-2, the land temperature probability of Landsat 8, the cloud probability of Sentinel-2, and the cloud probability of Landsat 8, respectively. HOT high , HOT low , T high , T low These represent the high and low values ​​of the haze-to-optimal conversion value of Sentinel-2 and the high and low values ​​of the temperature band of Landsat 8, respectively.

[0078] If a large body of water exists in the detection area, the probability of clouds over water is calculated. The probability of clouds over water is a combination of the probability of water temperature and the probability of brightness. Clear-sky water pixels are identified by "water test" and SWIR1 band reflectivity threshold. The probability is estimated using a higher level of clear-sky water temperature (82.5%) to exclude other atmospheric influences that usually cool the water temperature. A constant of 4°C is used to rescale the temperature probability because if the T of a pixel is 4°C lower than the surface temperature, the probability of it being a cloud pixel is high. The probability of clouds over water is shown in formulas (10), (11), and (12):

[0079] Probability of surface water temperature:

[0080]

[0081] Brightness probability: wBri =min( ρ SWIR1 ,0.11) / 0.11 (11)

[0082] Probability of clouds over water:

[0083] SwCloudp = wBri + Cirrus ×0.5

[0084] LwCloudp = wTemp × wBri + Cirrus ×0.3 (12)

[0085] in, Water Represents water testing; wTemp , wBri Represents the probability of water temperature and the probability of water brightness; SwCloudp , LwCloudp This represents the probability of clouds over water in Sentinel-2 and Landsat 8.

[0086] 3) Improved cloud shadow detection: Cloud shadows have high reflectivity in the NIR and SWIR1 bands, and the reflectivity in the NIR band is higher than that of water bodies. In densely vegetated areas, the reflectivity of cloud shadows in the near-infrared and short-wave infrared bands is less than 0.25 and 0.11, respectively. Therefore, these two band thresholds are selected to identify cloud shadows. The identification accuracy of thick clouds is high, and some thin cloud shadow pixels are missed. Due to the high reflectivity of the surface in bright areas such as deserts and rocks, the reflectivity of cloud shadows, especially thin cloud shadows, increases significantly. In some areas, the reflectivity of thin cloud shadows is even higher than that of general ground objects (the reflectivity reaches 0.3 in the near-infrared band and 0.28 in the short-wave infrared band). In order not to ignore thin cloud shadows, the influence of the underlying surface needs to be considered. According to the surface vegetation cover, the surface of the study area is divided into two categories: vegetation cover and bare ground cover. The cloud shadow detection algorithm is improved to improve the accuracy of cloud shadow detection. The improved cloud shadow detection algorithm is shown in Equations (13) and (14):

[0087] NDWI <0.1 (13)

[0088]

[0089] The improved cloud shadow detection algorithm demonstrates better visual performance in identifying cloud shadows in Landsat 8 and Sentinel-2 imagery, such as... Figure 1 As shown. From Figure 1 It can be seen that the improved cloud shadow detection algorithm used in the snow cover information reconstruction method of this invention has a strong ability to identify thick clouds. The identified thick cloud edge details are well preserved, and the location corresponds to the cloud area, without identifying the snow in the original image as clouds. In contrast, the original algorithm only uses the NIR band and SWIR1 band thresholding method to detect cloud shadows, and compared with the algorithm of this invention, some cloud image pixels are missed. Therefore, it is clear that the improved cloud shadow detection algorithm in this invention has a stronger ability to detect cloud shadow details than the original algorithm. However, a limitation shared by both the improved and original algorithms is that some dark surface pixels (mountain shadows, water bodies) are misidentified as cloud shadows.

[0090] Step 2: Snow Extraction

[0091] 1) Snow Extraction in Clear Skies: Sentinel-2 satellite products exhibit good cloud and snow separation in the visible, near-infrared, and short-wave infrared bands. Spectral curves of clouds and snow cover are shown in the image. Figure 2The Sen2Cor software tool provided by the ESA official website has a snow cover detection function. It uses Sentinel-2 L2A level data as input data and can extract snow cover information well by comprehensively using NDSI, band spectral thresholds, and band ratio calculations. At the same time, the Landsat 8 image bands also meet the application conditions of the Sen2Cor snow cover detection method (having Blue, Green, Red, NIR, and SWIR1 bands). Therefore, the snow cover information reconstruction method of this invention uses the Sen2Cor snow cover detection method provided by ESA to simultaneously extract clear-sky snow cover pixels in Landsat 8 and Sentinel-2 cloudless and cloudy images, as shown in formula (15):

[0092]

[0093] Figure 3 It is an image extracted from a clear sky with snow. Figure 3 a~ Figure 3 c represents the original images of Sentinel-2, GF-2, and Landsat 8, respectively; Figure 3 d~ Figure 3 f represents the clear-sky snow cover images extracted from Sentinel-2, GF-2, and Landsat 8 using Sen2Cor, respectively. The clear-sky snow cover extraction results from Landsat 8 and Sentinel-2 images have extremely similar spatial distribution characteristics to the GF-2 snow cover image. However, there are also some missed snow cover pixels. These misclassified pixels are mainly concentrated in the transition zone between snow cover and non-snow cover, where the snow cover is thin and mostly consists of mixed pixels, which are unstable and easily affected by external environmental factors such as lighting and atmospheric conditions.

[0094] 2) Improved SNOWL Snowline Extraction: The SNOWL snowline method, based on the zonal distribution characteristics of snow cover, reclassifies cloud pixels identified in snow-covered images into snow, non-snow, and patchy snow. In practical applications, although the SNOWL snowline method removes less cloud cover, it has the lowest error, making it an effective method for reconstructing snow cover information under clouds in mountainous areas using Landsat 8 and Sentinel-2 single-phase images. Due to the uncertainty of patchy snow, these pixels are not considered in the reconstruction of clouds and snow cover under cloud shadows. This invention's snow cover information reconstruction method uses an improved SNOWL snowline method with unstable snow cover areas to extract snow cover information under clouds and cloud shadows from Landsat 8 and Sentinel-2 cloud-covered images, such as... Figure 4 As shown, it includes two steps:

[0095] The first step is to calculate the average elevation of the snow cover pixels and the elevation of the cloud and cloud image pixels, and then convert the average elevation value ( H SM) and the elevation values ​​of cloud and cloud image metadata ( H CS In comparison, the rules for reclassifying cloud pixels are as follows:

[0096] like H SM ≤ H CS The cloud image pixels are then reclassified as snow, with the aim of preserving more snow pixels.

[0097] like H SM > H CS Then the cloud image elements will be reclassified as snowless land.

[0098] The reclassified cloud imagery provides preliminary results for snow cover reconstruction.

[0099] The second step is to use the clear-sky snow image identified by the Sen2Cor method to calculate the initial snow distribution of the number of snow days over many years, as shown in formula (16).

[0100]

[0101] in, SCD Indicates the number of days the snow cover lasted; N This indicates the number of years in a snow-covered year (a snow-covered year runs from September 1st to August 31st of the following year). S i Indicates the first snow cover of the year i Images of snow-covered skies on a clear day; if S i = 1 indicates the presence of snow. S i = 0 indicates no snow;

[0102] Then, using 22 days as a threshold, the distribution area of ​​unstable snow cells is obtained. Snow cells in unstable snow areas are deleted from the preliminary snow information reconstruction results to obtain the final snow distribution.

[0103] The acquisition of unstable snow cover areas includes all available Landsat8 and Sentinel-2 clear-sky snow cover information.

[0104] Figure 5 and Figure 6 These are reconstructed images of clouds and snow cover under cloud shadows from Sentinel-2 and Landsat 8 imagery, respectively. Figure 5 a and Figure 5 d are the original images of GF-2 and Sentinel-2, respectively; Figure 5 b is a cloudless snow cover map from GF-2. Figure 5e is a Sentinel-2 cloudless snow cover image extracted using the original SNOWL. Figure 5 g is a cloudless snow cover map of Sentinel-2 extracted by the improved SNOWL; Figure 5 c is a map of GF-2 clouds and snow cover under their shadows. Figure 5 f is the original SNOWL-extracted image of Sentinel-2 clouds and snow cover under cloud shadows. Figure 5 h is an improved SNOWL-extracted image of Sentinel-2 clouds and snow cover under cloud shadows. (Compared to...) Figure 5 resemblance, Figure 6 a and Figure 6 d are the original images of Sentinel-2 and Landsat 8, respectively; Figure 6 b is a Sentinel-2 image showing snow cover without clouds. Figure 6 e is a Landsat 8 cloudless snow map extracted using the original SNOWL image. Figure 6 g is a Landsat 8 cloudless snow map extracted using the improved SNOWL method; Figure 6 c is a diagram of Sentinel-2 clouds and snow cover under their shadows. Figure 6 f is a Landsat 8 cloud and snow cover under cloud shadows extracted from the original SNOWL image. Figure 6 h represents the Landsat 8 cloud and snow cover image under cloud shadows extracted using the improved SNOWL method. Compared to the original SNOWL method snow cover reconstruction results, the snow cover reconstructed by the improved SNOWL method in this invention has a shape that is closer to the verification image, significantly improving the missed and false detection of snow cover pixels, and providing richer details of snow cover in mountainous areas.

[0105] The method for reconstructing snow cover information in this invention was verified as follows:

[0106] The GF-2 snow cover verification data used for validation came from the State Key Laboratory of Remote Sensing Science. The processed GF-2 image had a spatial resolution of 0.8 m. Since there was a lack of GF-2 images contemporaneous with or near the Landsat 8 image, Sentinel-2 images from near the Landsat 8 image and meeting the research requirements were used for cross-validation of the snow cover reconstruction results. The accuracy verification of the snow cover extraction results included two parts: the accuracy verification of the snow cover extraction results under clear skies and the accuracy verification of the snow cover reconstruction results under clouds and cloud shadows. The GF-2 image used for the accuracy verification of the clear skies snow cover extraction results was contemporaneous with the Landsat 8 image, while the Sentinel-2 image was from a near-terminal period. The accuracy verification of the snow cover reconstruction results showed that the GF-2 image and the Sentinel-2 image were from near-terminal periods, and the Sentinel-2 image used for cross-validation was also from near-terminal periods with Landsat 8. Specific information is shown in Table 1.

[0107] Table 1 Basic Information of Accuracy Verification Data

[0108]

[0109] To quantitatively describe and evaluate the accuracy of snow extraction, the overall classification accuracy was selected. OA ), misclassification error ( CE ), omission error ( OE Snow cover classification accuracy () P )and Kappa The coefficient is used as an evaluation index to quantitatively analyze the snow extraction results of the reconstruction method of the present invention. The calculation formula is described in Table 2.

[0110] Table 2 Accuracy Verification Indicators

[0111]

[0112] Experimental results show that the snow cover information reconstruction method of this invention achieves high accuracy in snow cover information reconstruction, as shown in Table 3. The overall classification accuracy of Sentinel-2 snow cover information reconstruction increased from 77.27% to 80.74%, the correct snow cover classification accuracy was 86.26%, the misclassification error was 23.85%, and the omission error was relatively small at 13.74%. Kappa The coefficient is 0.616.

[0113] Table 3. Accuracy of Sentinel-2 cloud and cloud shadow snow cover reconstruction information.

[0114]

[0115] The accuracy of Landsat8 snow cover information reconstruction is shown in Table 4. Kappa With a coefficient of 0.735, the overall classification accuracy improved from 81.86% to 89.94%, and the snow classification accuracy improved from 68.83% to 74.41%. The misclassification error was 4.21%, and the omission error was relatively large at 25.59%, which is also the reason for the low accuracy of correct snow classification. This indicates that the improved algorithm still needs further improvement in reconstructing Landsat 8 clouds and snow under cloud shadows.

[0116] Table 4. Accuracy of Landsat 8 Snow Cover Information Reconstruction under Cloud Shadows and Clouds

[0117]

[0118] Since snow cover distribution varies significantly across different altitudes and slopes in mountainous areas, the impact of altitude and slope aspect on the accuracy of snow cover reconstruction is considered to further discuss the accuracy of the reconstruction results. The overall classification accuracy of Sentinel-2 and Landsat 8 at different elevation zones decreases with increasing elevation, but remains above 82% overall. The correct snow cover classification accuracy is higher in the mid-altitude zone (3600-4200 m), reaching over 80%, where snow cover is stable and widely distributed, accounting for 19.31% and 37.09% of the validation area, respectively. The snow cover reconstruction accuracy is lower in low and high altitude zones. Low-altitude areas are affected by climatic factors such as temperature and precipitation, resulting in more severe snowmelt; high-altitude areas have smaller snow cover areas. These factors contribute to the lower accuracy of snow cover reconstruction.

[0119] This invention's snow cover information reconstruction method uses Sentinel-2 and Landsat 8 data as input data, while considering the differences in different sensor bands and the impact of clouds and cloud shadows on remote sensing image quality. The results show that this invention's snow cover information reconstruction method under cloud and cloud shadow conditions improves in two aspects: firstly, it reduces the impact of clouds and cloud shadows on the quality of optical remote sensing images, improving the accuracy of cloud shadow detection; secondly, it improves the accuracy of snow cover information reconstruction, enhancing the ability to monitor snow cover in mountainous areas and providing a more detailed depiction of the characteristics of snow cover in mountainous regions.

Claims

1. A comprehensive method for reconstructing snow cover information, characterized in that, The reconstruction method is performed according to the following steps: Step 1: Cloud and cloud shadow detection: 1) Based on the physical characteristics of clouds, conduct basic tests, whiteness tests, ratio tests, haze-to-optimal conversion tests, and cirrus tests on clouds in Sentinel-2 and Landsat 8 images to distinguish them from other ground features with similar spectral characteristics to clouds. In the basic test, NDSI, NDVI, and SWIR2 band thresholds were used to separate clouds from vegetation and snow cover areas, specifically: LBasic = ρ SWIR2 ＞0.03 andT ＜27℃ andNDSI ＜0.8 and NDVI ＜0.8 （5） SBasic = ρ SWIR2 ＞0.03 andNDSI ＜0.8 and NDVI ＜0.8 （6） In the formula, LBasic This represents the basic tests for Landsat 8; SBasic Represents the basic tests of Sentinel-2; NDSI It is the normalized snow cover index; NDVI It is the normalized vegetation index; T It is the thermal infrared band of Landsat 8 data. ρ SWIR2 It is the surface reflectance of the shortwave infrared band 2; 2) Detecting inaccurately identified fragmented clouds and cloud edges in cloud spectral testing using cloud probability, divided into two cases: If there is no large body of water in the detection area, and only land cloud probability calculation is performed, then the haze-optimized conversion value is normalized to generate a new temperature probability from the Sentinel-2 image, replacing the temperature probability that cannot be calculated due to the lack of thermal infrared bands. Then, the temperature probabilities from Landsat 8 and Sentinel-2 are combined with the spectral difference probability to calculate the final cloud probability. Spectral difference probability: lVar =1-max[ White , abs ( NDVI ), abs ( NDSI (7) Land temperature probability: Cloud probability: SlCloudp = lVar × lHOT + Cirrus ×0.5 LlCloudp = lVar × lTemp+Cirrus ×0.3 （9） in, lVar , lHOT , lTemp , SlCloudp , LlCloudp These represent the spectral difference probability of Sentinel-2 and Landsat 8, the land temperature probability of Sentinel-2, the land temperature probability of Landsat 8, the cloud probability of Sentinel-2, and the cloud probability of Landsat 8, respectively. HOT high The high value represents the haze-to-optimal conversion value of Sentinel-2; HOT low The low value represents the haze-to-optimal conversion value of Sentinel-2; T high This represents the high value of the temperature band in Landsat 8; T low This represents the low value of the temperature band in Landsat 8; If a large body of water exists in the detection area, the probability of clouds over water is calculated. The probability of clouds over water is a combination of the probability of water temperature and the probability of brightness. Clear-sky water pixels are identified through "water testing" and the SWIR1 band reflectivity threshold, and the probability is estimated using the higher level of clear-sky water temperature. Probability of surface water temperature: Brightness probability: wBri =min( ) / 0.11 (11) Probability of clouds over water: SwCloudp = wBri ＋ Cirrus ×0.5 LwCloudp = wTemp × wBri ＋ Cirrus ×0.3 (12) in, Water Represents water testing; wTemp , wBri These represent the probability of water surface temperature and the probability of light, respectively. SwCloudp , LwCloudp These represent the probabilities of clouds over water in Sentinel-2 and Landsat 8, respectively. T Water Represents water temperature under clear skies; 3) In areas with dense vegetation cover, the land surface is divided into vegetation cover and bare ground cover, and cloud shadows are identified in the NIR and SWIR1 bands: NDWI ＜0.1 （13） ρ NIR It is the surface reflectance in the near-infrared band; ρ SWIR1 It is the surface reflectance of the shortwave infrared band 1; NDWI It is the normalized water quality index; NDVI It is the normalized vegetation index; Whiteness test, ratio test, haze-to-optimal conversion test and cirrus cloud test in the detection of potential clouds in Sentinel-2 and Landsat 8 images; Whiteness test: Haze-to-good conversion detection: HOT =( ρ Blue -0.5× ρ Red -0.08)>0 (2) Ratio test: Rock = ρ NIR / ρ SWIR1 >0.75 (3) Cirrus cloud detection: Cirrus = ρ Cirrus / 0.04>0.01(4) In the formula, White , MeanVis , HOT , Rock , Cirrus These represent whiteness test, visible light band mean, haze-to-excellence conversion detection, ratio detection, and cirrus cloud detection, respectively. ρ Blue , ρ Green , ρ Red , ρ NIR , ρ Cirrus , ρ SWIR1 These represent the surface reflectance of blue, green, red, near-infrared, cirrus, and shortwave infrared band 1, respectively. Step 2: Snow Extraction 1) Extraction of snow accumulation in clear skies Simultaneously extract clear-sky snow pixels from Landsat 8 and Sentinel-2 cloudless and cloudy images using the Sen2Cor snow cover detection method: 2) Extraction of snow accumulation under clouds and their shadows: Step 1: Calculate the average elevation of snow-covered pixels and the elevation of clouds and cloud image pixels respectively. After comparison, reclassify cloud pixels. The reclassified cloud image is the result of snow information reconstruction. Step 2: Calculate the initial snow cover distribution for the number of days with snow cover over many years using clear-sky snow cover images identified by the Sen2Cor method: In the formula, SCD Indicates the number of days the snow cover lasted; N Indicates the number of years with snow cover; S i Indicates the first year of a snow-covered period i Images of snow-covered skies on a clear day; Then, unstable snow cover distribution areas are obtained with a 22-day threshold. Pixels of unstable snow cover areas are removed from the preliminary snow cover information reconstruction results to obtain the final snow cover distribution.

2. The comprehensive snow cover information reconstruction method as described in claim 1, characterized in that, The rules for reclassifying cloud pixels in step 2: like H SM ≤ H CS Then the cloud image element will be reclassified as snow; like H SM > H CS Then the cloud image elements will be reclassified as snowless land. H SM This represents the average elevation value of the pixels with snow cover in clear skies. H CS The elevation values ​​of the cloud and cloud image metadata.

Images