Cloud detection method based on multi-feature fusion of jilin-1 high-resolution series satellite

Abstract

The application relates to a cloud detection method based on multi-feature fusion of Jilin No. 1 high-resolution series satellites, and belongs to the technical field of satellite remote sensing, and comprises the following steps: detecting clouds with high brightness and obvious cloud characteristics; identifying thin clouds covered above in ground objects classified into the same color tone system; and identifying accurate thick clouds and thin clouds. The application can universally process all Jilin No. 1 high-resolution series satellites with four spectral bands, does not need manual intervention, is accurate in effect, and is good in thin cloud detection effect for complex high-brightness underlying surfaces. The application does not need time-consuming manual labeling and does not need a complex running environment, especially when a new satellite of the same series and different models is added, the application does not need retraining, and has the advantages of easy adjustment and universality. The cloud detection method based on multi-feature fusion of Jilin No. 1 high-resolution series satellites can accurately detect thick clouds with obvious characteristics and thin clouds covered on the ground.

Description

Technical Field

[0001] The present invention relates to the field of satellite remote sensing technology, and in particular to a cloud detection method based on multi-feature fusion of Jilin-1 high-resolution series satellites. Background Art

[0002] Automatically classify and estimate total cloud cover for the Jilin-1 high-resolution satellite series. Cloud detection targets a multispectral remote sensing image in four bands: blue, green, red, and near-infrared. Commonly used methods include traditional feature recognition and deep learning. However, deep learning methods rely on complex training sets and annotations. Adding new satellites to the same series requires retraining, resulting in a complex process and limited applicability.

[0003] Traditional cloud detection and recognition using multispectral imagery is limited by the number of spectra in four bands, resulting in good detection of thick clouds but poor detection of thin clouds. This is because the DN and apparent reflectance values ​​of ground features captured by satellites at the same location but at different times and with different camera parameters vary significantly due to variations in solar altitude, radiation conditions, and satellite parameters in different remote sensing images. This makes it impossible to accurately and automatically process large quantities of imagery from a full range of satellites using only these four spectral bands. Summary of the Invention

[0004] The present invention aims to solve the technical problems in the prior art and provide a cloud detection method based on multi-feature fusion of Jilin-1 high-resolution satellite series.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] A cloud detection method based on multi-feature fusion of Jilin-1 high-resolution satellite series includes the following steps:

[0007] Step i: Detect clouds with high brightness and significant cloud features;

[0008] Step ii: Identify the thin clouds covering the ground objects classified as the same color system;

[0009] Step iii: Identify accurate thick and thin clouds.

[0010] In the above technical solution, step i specifically includes: identifying the basic framework of the image based on the color model.

[0011] In the above technical solution, step ii specifically includes: cloud recognition based on IS feature relationship; cloud recognition based on IH feature relationship; and white light increment recognition with H similarity.

[0012] In the technical solution, the step iii is specifically: fusing the three feature results obtained in the step ii to obtain the final cloud detection result.

[0013] In the technical solution, the image base framework recognition based on the color model is specifically:

[0014] Ideally, when there is a cloud on the underlying surface, the RGB value of the original ground object is increased by a cloud energy E cloud value, the thicker the cloud layer, the greater the E cloud value:

[0015] R addcloud = R + E cloud ;

[0016] G addcloud = G + E cloud ;

[0017] B addcloud = B + E cloud .

[0018] In the technical solution, the cloud recognition based on the I-S feature relationship is specifically:

[0019] According to the I-S feature relationship, a significant formula W1 of ground objects and clouds is established:

[0020] W1 = (I-S)

[0021] W1 = 1000, when W1 > 1000;

[0022] The W1 is normalized to obtain a significant relationship of clouds and ground objects based on the I-S feature;

[0023] Then, a minimum cross-entropy threshold segmentation method is used to identify the cloud.

[0024] In the technical solution, the cloud recognition based on the I-H feature relationship is specifically:

[0025] According to the I-H feature relationship, a significant formula W2 of ground objects and clouds is established:

[0026]

[0027] The W2 is normalized to obtain a significant relationship of clouds and ground objects based on the I-H feature, wherein p is 1, I thresh is 1500;

[0028] Then, a minimum cross-entropy threshold segmentation method is used to identify the cloud.

[0029] In the technical solution, the white light increment recognition with similar H is specifically:

[0030] First, according to the range H area_i Carry out rough classification of different types of features; each range step is 160, and the interval is 480:

[0031] H area_i = [H i , H i + 3*160], H i = 0, 160, 320,...3040;

[0032] In the same tone H area_i , according to different S sub-order to obtain the intensity value of the dark domain pixel in the region as the threshold of cloud and underlying surface; the step length of S sub-order is 100, and the interval is 100:

[0033] S area_j = [S j , S j + 100], S j = 0, 100, 200,...4000;

[0034] Then, all pixel values in the image that satisfy H area_i and S area_j are normalized to [0-255], and statistics are carried out according to the histogram, and the minimum value that is not 0 is selected as the underlying surface of the feature;

[0035] When the S and I values thereof satisfy the following conditions, the pixel is judged to be a cloud:

[0036] I base_thresh = min[I Histogram ] I Histogram ∈ [1, 255]

[0037]

[0038]

[0039]

[0040] In the above technical solution, when I<1000, it is judged to be a non-cloud.

[0041] In the above technical solution, when the pixel intensity value I x,y satisfies the following conditions, the pixel Pixel x,y at the image position (x, y) is not a cloud:

[0042]

[0043] The present application has the following beneficial effects:

[0044] The cloud detection method based on the multi-feature fusion of the Jilin No.1 high-resolution satellite series of the application can accurately detect thick clouds with obvious features and thin clouds covering the ground surface.

[0045] The cloud detection method based on the multi-feature fusion of the Jilin No.1 high-resolution satellite series of the application can detect cloud detection results in batches and automatically produce processing based on the remote sensing image of the Jilin No.1 high-resolution satellite series.

[0046] Compared with the traditional method, the application can universally process all Jilin No.1 high-resolution satellite series with four spectral bands without manual intervention, and the effect is accurate, and the thin cloud detection effect of the complex high-brightness underlying surface is good.

[0047] Compared with the machine learning method, the application does not need time-consuming manual annotation and does not need a complex running environment, especially when a new satellite of the same series and different models is added, without retraining, and has the advantages of easy adjustment and universality. BRIEF DESCRIPTION OF DRAWINGS

[0048] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0049] Figure 1 Fig. 1 is a flowchart of the cloud detection method based on the multi-feature fusion of the Jilin No.1 high-resolution satellite series of the application. I-S Fig. 1(a) is a DN value image; Fig. 1(b) is an apparent reflectance image; Fig. 1(c) is an HSI image; and Fig. 1(d) is a Cloud I-S image based on I-S features.

[0050] Figure 2 Fig. 2 is a flowchart of the cloud detection method based on the multi-feature fusion of the Jilin No.1 high-resolution satellite series of the application.

[0051] Figure 3 Fig. 3 is a flowchart of the cloud detection method based on the multi-feature fusion of the Jilin No.1 high-resolution satellite series of the application. IS-H Fig. 3(a) is a DN value image; Fig. 3(b) is a Cloud I-H image based on I-H features; Fig. 3(c) is a Cloud image based on H-similar white light increments; Fig. 3(d) is a final cloud image based on multi-feature fusion; and Fig. 3(e) is a cloud mask image labeled on the underlying surface.

[0052] Figure 4 Fig. 4 is a flowchart of the cloud detection method based on the multi-feature fusion of the Jilin No.1 high-resolution satellite series of the application. DETAILED DESCRIPTION

[0053] The application idea of the application is as follows:

[0054] The cloud detection method based on the multi-feature fusion of the Jilin No.1 high-resolution satellite series of the application fuses multiple feature relationships and classifies different situations of clouds and underlying surfaces.

[0055] First, the cloud with high brightness and cloud significant characteristics is detected, then the thin cloud covered above is identified in the ground object classified into the same color system, finally the accurate thick cloud and thin cloud are identified by the method.

[0056] The application will be described in detail below with reference to the drawings.

[0057] As shown in the figure, the application provides a cloud detection method based on multi-feature fusion of "Jilin No. 1" high-resolution series remote sensing satellite images. Figures 1-4 The method mainly establishes the relationship among I-S, I-H and IS-H based on the HSI color conversion model in the multi-spectral remote sensing image, and fuses multi-features to effectively and accurately classify and detect the cloud, thin cloud and ground object.

[0058] As shown in the figure, the cloud detection method based on multi-feature fusion of "Jilin No. 1" high-resolution series satellite of the application comprises the following steps: Figure 4 Step 1, image basic framework identification based on color model:

[0059] Since the DN value of each pixel point of the multi-spectral image obtained by the high-resolution series remote sensing satellite has no actual physical meaning, the point value represents the size of the received spectral energy at the position, therefore, the value is first converted into an apparent reflectance image according to the absolute radiation correction coefficient. Secondly, the RGB (red, green and blue) value in the multi-spectral image is converted into an HSI color model.

[0060] The application detects and identifies the cloud based on the white mixed light characteristics: under ideal conditions, when there is a cloud on the underlying surface, the RGB value of the original ground object is increased by a cloud energy E cloud value, the thicker the cloud layer, the larger the E cloud value:

[0061] R addcloud = R + E cloud ;

[0062] G addcloud = G + E cloud ;

[0063] B addcloud = B + E cloud ; (1)

[0064] Step 2, cloud identification based on I-S feature relationship:

[0065]

[0066] ​In the HSI color model, I is the brightness, which is the average of the RGB three values. The greater the I value, the stronger the image brightness. S is the saturation, which represents the depth of color. Based on step (1), when there is thick cloud, the I value of the cloud is greater than that of other similar ground objects in the image. At the same time, the presence of the cloud dilutes the color of the ground object, and the S value of the cloud is less than that of the similar ground object in the image. Therefore, the significant formula W1 of the ground object and the cloud is established according to the I-S characteristic relationship:

[0067] W1 = (I-S)

[0068] W1 = 1000, when W1 > 1000; (2)

[0069] The I-S characteristic-based significant relationship of the cloud and the ground object is obtained by normalizing it, and then the minimum cross-entropy threshold segmentation method is used to identify the cloud. Finally, to avoid misjudgment of the ground object, when I < 1000, it is judged as non-cloud. Finally, Cloud I-S .

[0070] Step 3, cloud recognition based on I-H characteristic relationship:

[0071] Since the DN value of the remote sensing image is proportional to the energy collected by the sensor, which is a quantity without physical meaning, it is often necessary to perform pre-processing to represent a visual image friendly to the human eye. The energy of thick cloud is much greater than that of other ground objects, so to display the effective ground object, some thick clouds may appear cloud exposure phenomenon. In the HSI color model, I is the brightness, and the brightness of the cloud is large. H is the hue, which represents the frequency of color. When the cloud shows pure white characteristics on the image, the RGB three values are the same, and H is 0. In the exposure case, the significant characteristics of the cloud are large brightness and small hue. The significant formula W2 of the ground object and the cloud is established according to the I-H characteristic relationship:

[0072]

[0073]

[0074] Since the increment E cloud of the cloud has uncertainty, and the value of H is related to the size of B and G values, H is modified to obtain H'. W2 is normalized to obtain the I-H characteristic-based significant relationship of the cloud and the ground object, where p is 1, I thresh is 1500. Then the minimum cross-entropy threshold segmentation method is used to identify the cloud. Finally, to avoid misjudgment of the ground object, a restriction condition is introduced. When the pixel intensity value I x,y of the image at the position (x, y) satisfies the following condition, the pixel Pixel x,y at the position (x, y) of the image is not cloud, and finally CloudI-H .

[0075]

[0076] Step 4, white light increment identification with H similar:

[0077] The energy of remote sensing image DN value is positively correlated with the entrance pupil radiance, but the image results obtained by the same object are different when the shooting time and satellite model are different. This is because in different images, due to the difference of solar elevation angle, radiation working condition and satellite parameters, the DN value and apparent reflectivity value of the same object at the same place in different time and different satellite shooting have difference. In batch production, it is characterized that a pixel point of a certain object in an image and a pixel point of a cloud in another image have the same apparent reflectivity value, and the values of RGB and near infrared 4 spectrums have consistent characteristics. Therefore, for the problem that the high-brightness object and the low-brightness cloud have the same value, the I value is limited in the above steps 2 and 3.

[0078] Through experimental comparison, the thick cloud detection effect of steps 2 and 3 is good, but the thin cloud detection is poor. In order to accurately detect the thin cloud, the thin cloud and the object in a single image are observed, when the mixed energy of white light is small, the energy value of RGB three spectrums is increased, then the I brightness is increased and the S saturation is reduced. The hue H is related to the difference of RGB, when the increment of cloud is consistent, the hue of the underlying surface with and without cloud is consistent, due to the error, so the classification can be carried out according to the system of the same hue of the same object.

[0079] The same image may exist different types of underlying surface including river, desert, building and the like, the present application firstly carries out rough classification of different types of objects according to the range H area_i of different hues.

[0080] H area_i =[H i , H i +3*160], H i =0, 160, 320,...3040; (5)

[0081] Compared with the object, the intensity I of the cloud is strong and the saturation S is low. In order to reduce the calculation amount, in the same hue H area_i , the intensity value of the dark domain pixel in the region is calculated according to different S stages, as the threshold value of cloud and underlying surface. The step length of S stage is 100 and the interval is 100:

[0082] S area_j =[S j , S j +100], S j= 0, 100, 200,..., 4000; (6)

[0083] Then all the pixel values in the image satisfying H area_i and S area_j are normalized to [0-255], and statistics are performed according to a histogram, and the minimum value that is not 0 is selected as the underlying surface of the ground object. In order to avoid errors, the area of the underlying surface needs to be large enough, that is, the number of pixels in the range of the statistical value needs to be greater than 10000. At the same time, in order to avoid the influence of shadows on the dark domain threshold, the relationship between the near-infrared value and the intensity I is used to filter the shadow of the cloud. Finally, all the pixel points satisfying the range of H area_i are subjected to threshold judgment, and when the S and I values satisfy the following conditions, the pixel is judged to be a cloud. The above steps are repeatedly performed, and finally the Cloud IS-H .

[0084] I base_thresh = min[I Histogram ], I Histogram ∈ [1, 255]

[0085]

[0086]

[0087]

[0088] The above three feature results, including Cloud I-S , Cloud I-H and Cloud IS-H , are fused to obtain the final cloud detection result.

[0089] The cloud detection method based on the multi-feature fusion of the Jilin No. 1 high-resolution series satellite of the application can accurately detect thick clouds with obvious features and thin clouds covering the ground surface.

[0090] The cloud detection method based on the multi-feature fusion of the Jilin No. 1 high-resolution series satellite of the application can detect the cloud detection result in batches and automatically based on the remote sensing image of the Jilin No. 1 high-resolution series.

[0091] Compared with the traditional method, the application can universally process all Jilin No. 1 high-resolution series satellites with four spectral bands without manual intervention, and the effect is accurate, and the thin cloud detection effect of a complex high-brightness underlying surface is good. Compared with the machine learning method, the application does not need time-consuming manual labeling and does not need a complex running environment, especially when a new satellite of the same series is added, no retraining is needed, and the application has the advantages of easy adjustment and universality.

[0092] Obviously, the above embodiments are merely example for clearly illustrating but not limitation to the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments need not and can not be enumerated. The obvious changes or variations derived from the above description are still within the protection scope of the present application.

Claims

1. A cloud detection method based on multi-feature fusion of Jilin-1 high-resolution satellite series, characterized in that: The following steps are involved: Step i: Detect clouds with high brightness and significant cloud features; Step ii: Identify the thin clouds covering the ground objects classified as the same color system; Step iii: Identify accurate thick and thin clouds; Step ii specifically includes: cloud recognition based on IS feature relationship; cloud recognition based on IH feature relationship; white light increment recognition with H similarity; Step iii is specifically as follows: fusing the three feature results obtained in step ii to obtain the final cloud detection result; The cloud identification based on IS feature relationship is as follows: The significant formula W1 of ground objects and clouds is established based on the characteristic relationship of IS: W1=(IS) W1=1000, when W1>1000; Normalize W1 to obtain the significant relationship between clouds and ground objects based on IS features; Then, the minimum cross entropy threshold segmentation method is used to identify the clouds; Cloud identification based on IH feature relationships is specifically as follows: The significant formula W2 of ground objects and clouds is established based on the characteristic relationship of IH: Normalizing W2 yields the significant relationship between clouds and ground objects based on IH features, where p is 1 and I thresh is 1500; Then, the minimum cross entropy threshold segmentation method is used to identify the clouds; The specific identification of white light increments with H similarity is: First, according to the range of different tones H area_i Perform a rough classification of different types of land features; the step size of each range is 160 and the interval is 480: H area_i =[H i ,H i +3*160],H i =0,160,320,...3040; In the same color H area_i In the example, the intensity values ​​of the dark pixels in the region are obtained according to different S-orders as the thresholds for cloud and underlying surface; the step size of the S-order is 100 and the interval is 100: S area_j =[S j ,S j +100],S j =0,100,200,...4000; Then, for the image that satisfies H area_i and S area_j All pixel values ​​are normalized to [0-255], and statistics are performed according to the histogram, and the smallest non-zero value is selected as the underlying surface of the feature; When its S and I values ​​meet the following conditions, the pixel is judged to be a cloud; IN base_thresh =min[I Histogram ],IN Histogram ∈[1,255] 2. The cloud detection method based on multi-feature fusion of Jilin-1 high-resolution satellites according to claim 1 is characterized in that: Step i specifically includes: identifying the basic framework of the image based on the color model.

3. The cloud detection method based on multi-feature fusion of Jilin-1 high-resolution satellites according to claim 2 is characterized in that: The basic framework of image recognition based on color model is as follows: Ideally, when there are clouds on the underlying surface, the RGB values ​​of the original objects are all increased by the cloud energy E cloud The thicker the cloud, the higher the E cloud Larger values: R addcloud =R+E cloud ; G addcloud =G+E cloud ; B addcloud =B+E cloud 。 4. The cloud detection method based on multi-feature fusion of Jilin-1 high-resolution satellites according to claim 1 is characterized in that: When I<1000, it is judged as non-cloud.

5. The cloud detection method based on multi-feature fusion of Jilin-1 high-resolution satellites according to claim 1, characterized in that: When the pixel intensity value I x,y The pixel at the image position (x,y) is Pixel when the following conditions are met x,y Not for the Cloud:

Images