Accumulated product cloud removal method and system based on random forest machine learning

Abstract

The invention provides an accumulated snow product cloud removing method and system based on random forest machine learning, and relates to the technical field of image processing, MOD10A1 pixel data and MYD10A1 pixel data are synthesized into NDSI values of cloudless pixels of a to-be-monitored area through adjacent time synthesis, clouded pixels are marked, and the accumulated snow product cloud removing method and system based on random forest machine learning are obtained by using NDSI values of adjacent similar pixels of the clouded pixels. Carrying out NDSI value synthesis on cloud pixels, carrying out water body correction on a to-be-monitored area, assigning water body pixels which are wrongly divided outside a water body boundary as cloud pixels, obtaining an optimal threshold value by adopting a dynamic threshold value method based on a confusion matrix, carrying out accumulated snow binaryzation on the pixels of the to-be-monitored area by adopting the optimal threshold value, and carrying out accumulated snow binaryzation on the basis of random forest machine learning. According to the method, the snowless binary snow accumulation product is obtained through the snowless binary prediction of the clouded pixels, the problem that the snow accumulation coverage accuracy of the existing areas with complex and diversified landforms is low is solved, and the method is suitable for generating the cloud-removed snow accumulation map of the complex and diversified areas.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method and system for removing clouds from snow products based on random forest machine learning. Background Art

[0002] Optical satellite remote sensing has important advantages such as spatio-temporal continuity and wide coverage in snow monitoring, and has become an important data source in snow research. At present, both polar-orbiting satellites and geostationary satellites have released a variety of snow products, such as Landsat, SPOT, AVHRR, VEGETATION, MODIS, etc., which can be used for the estimation of optical snow cover and binary snow products; in addition, snow depth monitoring is carried out through microwave remote sensing data such as SMMR, SSM / I, and AMSR-E.

[0003] Clouds change the energy radiation transfer between the sun, the earth's surface, and the sensor, making it impossible for the snow information under the clouds to be accurately transmitted to the sensor. Moreover, the reflection characteristics of clouds and snow are similar, resulting in low recognition accuracy of clouds and snow, which severely limits the practical application of MODIS optical remote sensing snow products on a daily time scale.

[0004] To solve the limitations of cloud occlusion on the application of MODIS products and accurately restore the pixels under the clouds to generate a spatio-temporally continuous full-coverage MODIS snow product, the current mainstream MODIS snow cloud removal methods can be summarized as adjacent time synthesis, spatial filtering, elevation snow line fitting, multi-source data fusion, etc.

[0005] Adjacent time synthesis usually adopts the method of synthesizing one or more days of MODIS Terra and Aqua, which cannot restore the snow information of the pixels under the clouds for consecutive days, and multi-day synthesis cannot reflect the actual snow changes during the period of rapid snow changes.

[0006] Spatial filtering usually selects cloud-free pixels in the spatial neighborhood to infer and estimate the information of the pixels under the clouds, and it cannot obtain satisfactory results for large continuous cloud occlusion areas.

[0007] Elevation snow line fitting can achieve binary classification of cloud-occluded pixels (i.e., classified as snow or land) by comparing cloud pixels with the elevation of the regional snow line. However, this method still has some defects in complex terrain areas and low-altitude areas with intensive urbanization, such as not considering the influence of unstable snow areas when estimating the snow line, resulting in low accuracy of the restored results of the snow information under the clouds.

[0008] Multi-source data fusion generally uses different means such as optical remote sensing, microwave remote sensing, and ground observation to achieve information complementarity. However, the spatio-temporal resolution mismatch between different data sources usually causes a certain degree of uncertainty in the restoration of pixels under the clouds.

[0009] The terrain and landforms in some areas are complex and diverse, with undulating terrain. The spatial distribution of snow cover in this area is extremely uneven, characterized by strong spatial heterogeneity and rapid changes. At present, the existing mainstream remote sensing snow cloud removal methods cannot fully adapt to the snow cover distribution characteristics of this area, easily causing problems such as the inability to recover the pixels below the cloud or poor recovery accuracy. Summary of the Invention

[0010] The technical problem to be solved by the present invention: Provide a snow product cloud removal method and system based on random forest machine learning to solve the problem of low snow cover accuracy in areas with complex and diverse terrain and landforms.

[0011] The technical solution adopted by the present invention to solve the above technical problem: A snow product cloud removal method based on random forest machine learning, including the following steps:

[0012] S1. Based on the MOD10A1 pixel data and MYD10A1 pixel data, use adjacent time synthesis to obtain the NDSI values of cloud-free pixels in the area to be monitored, and mark the cloudy pixels.

[0013] S2. Use the NDSI values of adjacent similar pixels of the cloudy pixels to synthesize the NDSI values of the cloudy pixels.

[0014] S3. Perform water body correction on the area to be monitored, and assign the misclassified water body pixels outside the water body boundary as cloudy pixels.

[0015] S4. Based on the confusion matrix, use the dynamic threshold method to obtain the optimal threshold, and use the optimal threshold to perform snow binarization on the pixels in the area to be monitored.

[0016] S5. Based on random forest machine learning, perform binary prediction of snow or no snow for the cloudy pixels to obtain a cloud-free binary snow product.

[0017] Further, in S1, the NDSI values of cloud-free pixels in the area to be monitored are synthesized at adjacent times, and cloud pixels are marked, including: for the MOD10A1 pixel data and MYD10A1 pixel data on the same day, if a certain pixel is a cloud or missing value in the MOD10A1 pixel data and is an NDSI value in the MYD10A1 pixel data, then this pixel is a cloud-free pixel on that day, and the NDSI value is the NDSI value of this pixel in the MYD10A1 pixel data; if a certain pixel is an NDSI value in the MOD10A1 pixel data and is a cloud or missing value in the MYD10A1 pixel data, then this pixel is a cloud-free pixel on that day, and the NDSI value is the NDSI value of this pixel in the MOD10A1 pixel data; if a certain pixel is an NDSI value in both the MOD10A1 pixel data and the MYD10A1 pixel data, then this pixel is a cloud-free pixel on that day, and the NDSI value is the average of the NDSI value of this pixel in the MOD10A1 pixel data and the NDSI value of this pixel in the MYD10A1 pixel data; if a certain pixel is a cloud or missing value in both the MOD10A1 pixel data and the MYD10A1 pixel data, then this pixel is a cloud pixel on that day.

[0018] Further, in S1, it also includes: if a cloud pixel is an NDSI value on the previous day or the next day, then replace this cloud pixel with a cloud-free pixel, and the NDSI value is the NDSI value of the previous day, the NDSI value of the next day, or the average value.

[0019] Further, in S2, the synthesis formula is: where M represents the synthesized NDSI value of the cloud pixel, M' represents the historical NDSI value of the cloud pixel within a preset time period, n represents the number of adjacent similar pixels, SM i represents the current NDSI value of the i-th adjacent similar pixel, SM' i represents the NDSI value of the i-th adjacent similar pixel at the time corresponding to the historical NDSI value, w i represents the weight of the i-th adjacent similar pixel.

[0020] Further, the weight is set according to the distance, and the weight formula is: where x i ,y i represent the coordinates of the i-th adjacent similar pixel, x 0 ,y 0 represent the coordinates of the cloud pixel, and L represents the window size.

[0021] Further, S4 specifically includes: taking the pixel where the station is located as the reference, using the overall accuracy in the confusion matrix as the accuracy judgment criterion, adopting the dynamic threshold method, with a step size of 1, counting the binary accuracy under different NDSI threshold conditions, and using the NDSI value corresponding to the highest overall accuracy as the snow binary NDSI threshold to perform snow / non-snow binary on the image.

[0022] Further, in S5, it includes: selecting meteorological elements, DEM, and terrain complexity as the characteristic variables of the random forest model, using the binary snow remote sensing images of the current day and the previous day, taking the snow / non-snow of the pixel as the label, and forming a matching data set with the corresponding date and characteristic variables, establishing multiple decision tree forests, constructing training samples, through the grid search algorithm, finding the best hyperparameter configuration, obtaining the random forest machine learning model, and using the random forest machine learning model to predict the snow / non-snow binary of the cloudy pixels to obtain a cloud-free binary snow product.

[0023] Further, the meteorological elements include surface temperature, total surface net solar radiation, 2-meter temperature, and total precipitation.

[0024] The present invention also provides a MODIS snow product cloud removal system based on random forest machine learning to implement the above-mentioned snow product cloud removal method based on random forest machine learning. The system includes a neighboring time synthesis module, a neighboring similar pixel synthesis module, a water body correction module, a binary module, and a random forest machine learning model; the neighboring time synthesis module is used to synthesize the NDSI values of cloud-free pixels in the area to be monitored from MOD10A1 pixel data and MYD10A1 pixel data, and mark the cloudy pixels; the neighboring similar pixel synthesis module is used to synthesize the NDSI values of the cloudy pixels; the water body correction module is used to perform water body correction on the area to be monitored, and assign the misclassified water body pixels outside the water body boundary as cloudy pixels; the binary module is used to obtain the best threshold by using the dynamic threshold method based on the confusion matrix, and use the best threshold to perform snow binary on the pixels in the area to be monitored; the random forest machine learning model is used to predict the snow / non-snow binary of the cloudy pixels to obtain a cloud-free binary snow product.

[0025] Advantages of the present invention: The present invention provides a method and system for removing clouds from snow cover products based on random forest machine learning. By adjacent time synthesis, the NDSI values of cloud-free pixels in the MOD10A1 pixel data and MYD10A1 pixel data are synthesized for the area to be monitored, and the cloudy pixels are marked. Using the NDSI values of adjacent similar pixels of the cloudy pixels, the NDSI values of the cloudy pixels are synthesized, and the water body correction is performed on the area to be monitored. The misclassified water body pixels outside the water body boundary are assigned as cloudy pixels. Based on the confusion matrix, the optimal threshold is obtained by using the dynamic threshold method. The optimal threshold is used to perform snow cover binarization on the pixels in the area to be monitored. Based on random forest machine learning, the binary prediction of snow or no snow for the cloudy pixels is performed to obtain a cloud-free binary snow cover product, which solves the problem of low accuracy of snow cover in areas with complex and diverse terrain and landforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 FIG. is a flowchart of a method for removing clouds from snow cover products based on random forest machine learning provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] In view of the problem of low accuracy of snow cover in areas with complex and diverse terrain and landforms in the prior art, the present invention provides a method and system for removing clouds from snow cover products based on random forest machine learning. By adjacent time synthesis, the NDSI values of cloud-free pixels in the MOD10A1 pixel data and MYD10A1 pixel data are synthesized for the area to be monitored, and the cloudy pixels are marked, thereby completing the fusion of two satellite data and being able to remove some cloudy pixels. Then, using the NDSI values of adjacent similar pixels of the cloudy pixels, the NDSI values of the cloudy pixels are synthesized to remove another part of the cloudy pixels. For the remaining cloudy pixels, the misclassified water body pixels outside the water body boundary are supplemented to obtain all inaccurate pixels. Based on the confusion matrix, the optimal threshold is obtained by using the dynamic threshold method. The optimal threshold is used to perform snow cover binarization on the pixels in the area to be monitored to obtain a binary snow cover product with clouds, that is, an incompletely accurate binary snow cover product. Based on random forest machine learning, the binary prediction of snow or no snow for the cloudy pixels is performed to obtain a cloud-free binary snow cover product, thereby completing the binary prediction of inaccurate pixels and obtaining an accurate cloud-free binary snow cover product.

[0028] As Figure 1 shown, a method for removing clouds from snow cover products based on random forest machine learning provided by the present invention includes the following steps:

[0029] S1. Based on the MOD10A1 pixel data and MYD10A1 pixel data, the NDSI values of cloud-free pixels in the area to be monitored are synthesized by adjacent time synthesis, and the cloudy pixels are marked.

[0030] Specifically, the NDSI values of cloud-free pixels in the area to be monitored are synthesized at adjacent times, and cloud pixels are marked, including: for the MOD10A1 pixel data and MYD10A1 pixel data on the same day, if a certain pixel is a cloud pixel or a missing value in the MOD10A1 pixel data and is an NDSI value in the MYD10A1 pixel data, then this pixel is a cloud-free pixel on that day, and the NDSI value is the NDSI value of this pixel in the MYD10A1 pixel data; if a certain pixel is an NDSI value in the MOD10A1 pixel data and is a cloud pixel or a missing value in the MYD10A1 pixel data, then this pixel is a cloud-free pixel on that day, and the NDSI value is the NDSI value of this pixel in the MOD10A1 pixel data; if a certain pixel is an NDSI value in both the MOD10A1 pixel data and the MYD10A1 pixel data, then this pixel is a cloud-free pixel on that day, and the NDSI value is the average of the NDSI value of this pixel in the MOD10A1 pixel data and the NDSI value of this pixel in the MYD10A1 pixel data; if a certain pixel is a cloud pixel or a missing value in both the MOD10A1 pixel data and the MYD10A1 pixel data, then this pixel is a cloud pixel on that day. Through testing, it is found that the above steps can remove about 6% of cloud cover. Further, if a cloud pixel is an NDSI value both on the previous day and the next day, then this cloud pixel is replaced with a cloud-free pixel, and the NDSI value is the NDSI value of the previous day, the NDSI value of the next day, or the average value. This step can remove about 24% of cloud cover.

[0031] S2. Use the NDSI values of adjacent similar pixels of cloud pixels to synthesize the NDSI values of cloud pixels.

[0032] Specifically, the synthesis formula is: where M represents the synthesized NDSI value of cloud pixels, M′ represents the historical NDSI value of cloud pixels within a preset time period, n represents the number of adjacent similar pixels, SM i represents the current NDSI value of the i-th adjacent similar pixel, SM′ i represents the NDSI value of the i-th adjacent similar pixel at the time corresponding to the historical NDSI value, and w i represents the weight of the i-th adjacent similar pixel. The weight is set according to the distance, and the weight formula is: where x i and y i represent the coordinates of the i-th adjacent similar pixel, x 0 and y 0 represent the coordinates of the cloud pixel, and L represents the window size. Among them, all non-target pixels in the window are adjacent pixels, and calculate SM′ iThe difference from M′ is arranged in ascending order, and the first n adjacent pixels are taken as n adjacent similar pixels. Usually, the window size is 15×15 and n is 20. This step can remove about 15% of cloud cover.

[0033] S3. Perform water body correction on the area to be monitored, and assign the misclassified water body pixels outside the water body boundary as cloudy pixels.

[0034] Specifically, by comparing the land use type data, there is a problem of misjudging the shadow area as lake ice or lakes, which is likely to cause error accumulation in the subsequent NDSI for judging the snow cover range. To reduce this misjudgment error, and considering that the time series of constructing the snow cover dataset in this study area is relatively long, the IGBP classification scheme in the MODIS land use product MCD12Q1 data is used to correct the water body. Specifically: lakes or lake ice within the water body boundary are retained, and misclassified lakes and lake ice outside the boundary are reassigned as cloudy pixels.

[0035] S4. Based on the confusion matrix, use the dynamic threshold method to obtain the optimal threshold, and use the optimal threshold to perform snow cover binarization on the pixels in the area to be monitored.

[0036] Specifically, the selection of the NDSI threshold has an important impact on the accuracy of snow cover extraction. There are certain differences in the NDSI values corresponding to snow cover pixels in different regions. It is difficult for a single NDSI threshold to meet the accuracy requirements under actual natural conditions. Therefore, taking the pixel where the station is located as the benchmark, using the overall accuracy in the confusion matrix as the accuracy judgment standard, and using the dynamic threshold method with a step size of 1, the binarization accuracy under different NDSI thresholds is statistically calculated. The NDSI value corresponding to the highest overall accuracy is used as the optimal threshold to perform snow cover binarization on the pixels in the image of the area to be monitored, and a binarized snow cover image is obtained.

[0037] The calculation process of the overall accuracy is as follows: based on the K-day historical data of the station, when the NDSI value of the pixel where the station is located is greater than the set threshold and the snow depth at the station is greater than 1 cm, this pixel is classified as A; when the NDSI value of the pixel where the station is located is less than the set threshold but the snow depth at the station is greater than 1 cm, this pixel is classified as B; when the NDSI value of the pixel where the station is located is greater than the set threshold but the snow depth at the station is equal to 0 cm, this pixel is classified as C; when the NDSI value of the pixel where the station is located is less than the set threshold and the snow depth at the station is equal to 0 cm, this pixel is classified as D. The K-day historical data of the station are divided to obtain the number of pixels corresponding to each category, and the classification data of all stations in the area to be monitored every day are statistically calculated. The formula for the overall accuracy is: Where OA represents the overall accuracy, a represents the number of pixels corresponding to all stations classified as A, b represents the number of pixels corresponding to all stations classified as B, c represents the number of pixels corresponding to all stations classified as C, and d represents the number of pixels corresponding to all stations classified as D.

[0038] S5. Based on random forest machine learning, perform binary prediction of snow presence or absence for cloudy pixels to obtain a cloud-free binary snow cover product.

[0039] Specifically, after the above processing steps, there are still a small number of remaining cloudy pixels in the binary snow cover image. Considering that snow cover is not only related to elevation but also closely related to meteorological elements such as air temperature, precipitation, and surface temperature, relevant auxiliary variables are introduced, and cloud removal for the last pixels is performed based on machine learning. Specifically: Select meteorological elements, DEM (digital elevation model) data, and terrain complexity as the feature variables of the random forest model. Use the binary snow cover images of the current day and the previous day, with the snow presence or absence of pixels as labels, and form a matching data set with the corresponding dates and feature variables. Establish multiple decision tree forests, construct training samples, and through the grid search algorithm, find the best hyperparameter configuration to obtain a random forest machine learning model. Use the random forest machine learning model to perform binary prediction of snow presence or absence for cloudy pixels to obtain a cloud-free binary snow cover product. In this way, a daily cloud-free binary snow cover product can be obtained. The meteorological elements include surface temperature, total surface net solar radiation, 2-meter temperature, and total precipitation.

[0040] The present invention also provides a MODIS snow cover product cloud removal system based on random forest machine learning to implement the snow cover product cloud removal method based on random forest machine learning as described above. The system includes a neighboring time synthesis module, a neighboring similar pixel synthesis module, a water body correction module, a binary module, and a random forest machine learning model; the neighboring time synthesis module is used to synthesize the NDSI values of cloud-free pixels in the MOD10A1 pixel data and the MYD10A1 pixel data for the area to be monitored and mark the cloudy pixels; the neighboring similar pixel synthesis module is used to synthesize the NDSI values of cloudy pixels; the water body correction module is used to perform water body correction on the area to be monitored and assign the misclassified water body pixels outside the water body boundary as cloudy pixels; the binary module is used to obtain the best threshold by using the dynamic threshold method based on the confusion matrix and perform snow cover binary on the pixels in the area to be monitored using the best threshold; the random forest machine learning model is used to perform binary prediction of snow presence or absence for cloudy pixels to obtain a cloud-free binary snow cover product.

Claims

1. A snow product cloud removal method based on random forest machine learning, characterized in that: The following steps are involved: S1. Based on MOD10A1 pixel data and MYD10A1 pixel data, the NDSI values ​​of cloud-free pixels in the monitored area are synthesized using adjacent time, and cloud pixels are marked; S2, using the NDSI values ​​of similar pixels adjacent to the cloud pixel, synthesize the NDSI value of the cloud pixel; S3, perform water body correction on the monitored area, and assign the misclassified water pixels outside the water body boundary to cloud pixels; S4. Based on the confusion matrix, the dynamic threshold method is used to obtain the optimal threshold, and the optimal threshold is used to binarize the snow accumulation in the pixels of the monitored area; S5. Based on random forest machine learning, the presence or absence of snow in cloud pixels is binary predicted to obtain cloud-free binary snow accumulation products.

2. The method for removing cloud from snow products based on random forest machine learning according to claim 1, characterized in that: In S1, the NDSI values ​​of cloud-free pixels in the monitored area are synthesized using adjacent time periods, and cloud pixels are marked, including: for MOD10A1 pixel data and MYD10A1 pixel data within the same day, if a pixel has clouds or missing values ​​in the MOD10A1 pixel data, and has an NDSI value in the MYD10A1 pixel data, then the pixel is a cloud-free pixel within that day, and the NDSI value is the NDSI value of the pixel in the MYD10A1 pixel data; if a pixel has an NDSI value in the MOD10A1 pixel data, and has clouds in the MYD10A1 pixel data or a missing value, the pixel is a cloud-free pixel on that day, and the NDSI value is the NDSI value of the pixel in the MOD10A1 pixel data; if a pixel has an NDSI value in both the MOD10A1 pixel data and the MYD10A1 pixel data, then the pixel is a cloud-free pixel on that day, and the NDSI value is the average of the NDSI values ​​in the MOD10A1 pixel data and the NDSI values ​​in the MYD10A1 pixel data; if a pixel has clouds or a missing value in both the MOD10A1 pixel data and the MYD10A1 pixel data, then the pixel is a cloud pixel on that day.

3. The snow product cloud removal method based on random forest machine learning according to claim 2 is characterized in that: S1 also includes: if a cloud pixel has an NDSI value of the previous day or the next day, the cloud pixel is replaced by a cloud-free pixel, and the NDSI value is the NDSI value of the previous day, the NDSI value of the next day, or the average value.

4. The snow product cloud removal method based on random forest machine learning according to claim 1 is characterized in that: In S2, the synthesis formula is: Where M represents the synthetic NDSI value of the cloud pixel, M′ represents the historical NDSI value of the cloud pixel within the preset time period, n represents the number of adjacent similar pixels, and SM i Represents the current NDSI value of the i-th adjacent similar pixel, SM′ i represents the NDSI value of the i-th adjacent similar pixel at the time corresponding to the historical NDSI value, w i Represents the weight of the i-th neighboring similar pixel.

5. The method for removing cloud from snow products based on random forest machine learning according to claim 4, characterized in that: The weight is set according to the distance, and the weight formula is: in, x i ,y i represents the coordinates of the ith adjacent similar pixel, x0, y0 represent the coordinates of the cloud pixel, and L represents the window size.

6. The method for removing cloud from snow products based on random forest machine learning according to claim 1, characterized in that: S4 specifically includes: taking the pixel where the station is located as the benchmark, using the overall accuracy in the confusion matrix as the accuracy judgment standard, adopting the dynamic threshold method, with 1 as the step size, and counting the binarization accuracy under different NDSI thresholds. The NDSI value corresponding to the highest overall accuracy is used as the NDSI threshold for snow binarization, and the image is binarized with or without snow.

7. The method for removing cloud from snow products based on random forest machine learning according to claim 1, characterized in that: S5 includes: selecting meteorological elements, DEM and terrain complexity as characteristic variables of the random forest model, using the binary snow remote sensing images of the current day and the previous day, taking the presence or absence of snow in pixels as labels, forming a matching data set with the corresponding dates and characteristic variables, establishing multiple decision tree forests, constructing training samples, and finding the best hyperparameter configuration through a grid search algorithm to obtain a random forest machine learning model, and using the random forest machine learning model to make a binary prediction of the presence or absence of snow in cloud pixels to obtain a cloudless binary snow product.

8. The method for removing cloud from snow products based on random forest machine learning according to claim 7, characterized in that: The meteorological elements include surface temperature, total net solar radiation on the surface, 2-meter temperature and total precipitation.

9. MODIS snow product cloud removal system based on random forest machine learning, characterized by: The method for removing clouds from snow products based on random forest machine learning as described in claim 1 is implemented, and the system includes a neighboring time synthesis module, a neighboring similar pixel synthesis module, a water body correction module, a binarization module and a random forest machine learning model; the neighboring time synthesis module is used to synthesize the MOD10A1 pixel data and the MYD10A1 pixel data into the NDSI value of the cloud-free pixels in the monitored area, and mark the cloud pixels; the neighboring similar pixel synthesis module is used to synthesize the NDSI value of the cloud pixels; the water body correction module is used to perform water body correction on the monitored area, and assign the misclassified water pixels outside the water body boundary as cloud pixels; the binarization module is used to obtain the optimal threshold value by using the dynamic threshold method based on the confusion matrix, and use the optimal threshold value to binarize the snow accumulation of the pixels in the monitored area; the random forest machine learning model is used to binary predict the presence or absence of snow for the cloud pixels to obtain cloud-free binary snow products.

Images