WO2020015326A1

WO2020015326A1 - Remote sensing image cloud shadow detection method supported by earth surface type data

Info

Publication number: WO2020015326A1
Application number: PCT/CN2018/124230
Authority: WO
Inventors: 吴昊; 周雪莹; 于会泳
Original assignee: 山东科技大学
Priority date: 2018-07-19
Filing date: 2018-12-27
Publication date: 2020-01-23
Also published as: CN109101894A; CN109101894B

Abstract

The present invention relates to the field of detection techniques, and provides a remote sensing image cloud shadow detection method supported by earth surface type data. The method comprises the following steps: constructing an earth surface clear sky pixel library and a shadow pixel library; selecting an optimal waveband; and generating a cloud shadow probability map. On the basis of the combination of the characteristic information of a pixel spectrum library and an earth surface type, the present invention relates to performing cloud shadow detection on a Landsat 8 image, resolves a problem caused by different cloud shadow spectrum characteristics under different earth surface types by using the support of the earth surface type, and resolves the problem of cloud shadow threshold setting in the manner of increasing a certain reflectivity value in step (a difference boundary between cloud shadow and a ground object spectrum characteristic is fuzzy, and therefore, it is not easy to determine a threshold). Therefore, by means of the method, the present invention has a good cloud shadow detection effect, reaches high cloud shadow detection precision, and brings a very beneficial cloud shadow detection effect.

Description

Cloud shadow detection method of remote sensing image supported by surface type data

[0001] The present invention belongs to the field of detection technology, and particularly relates to a method for detecting cloud shadow of a remote sensing image supported by ground surface type data.

Background technique

[0002] Cloud shadow detection has greater challenges than cloud detection. At present, the main cloud shadow detection methods are divided into threshold method, classification method and observation geometry-based method. Among them, threshold cloud shadow detection is the most widely used cloud shadow detection method, just like cloud detection.

[0003] The basic idea of the empirical threshold method is to perform cloud shadow detection based on spectral difference analysis using one or more bands or derived metrics. Shu SP et al. Proposed a shadow region recognition and brightness compensation method for the shadow problem of aerial images in the early days; Simpson JJ et al. Proposed a multispectral algorithm for AVHRR, which combined geometric and optical constraints from pixel-by-pixel geometry and Detection of cloud shadows in image spectral analysis; Martinuzzi S et al. Developed a simple semi-automatic method to detect and repair clouds and cloud shadows present in ETM + data; Luo Y et al. Developed an improved MODIS cloud and shadow mask The threshold method resamples the required band to a consistent 250m resolution and achieves high accuracy. Many domestic scholars have also made in-depth research on cloud shadow recognition. Chen Fen et al. Proposed a semi-automatic cloud shadow removal algorithm based on human-computer interaction for high-resolution images. This algorithm divided the remote sensing image into four parts: cloud, shadow, clear features, and transition area in a semi-automatic way of human-computer interaction. Then, the shadow area is compensated by establishing a histogram map of the shadow area and the clear area, and finally the boundary is processed; Mi Xueting et al. Proposed a cloud shadow detection method based on multitemporal remote sensing data. The main idea is to use the same As a support, clear-sky images of similar periods are supported. When the reflectance of the image to be detected is lower than a certain threshold, it is judged to be a cloud shadow. Yun Ya et al. Proposed a cloud and cloud shadow detection algorithm for GF-1. This method uses a combination of shadow spectral features and cloud-to-shadow geographic location relationships to detect shadows. Sun Lin et al. Proposed an improved cloud shadow detection method based on the Landsat-8 Fmask algorithm. There is a delineation of the cloud shadow area, which is then based on cloud and shadow using an improved cloud height calculation method Geometric relationships identify cloud shadows. Summary of invention

technical problem

Problem solution

Technical solutions

[0004] In view of the above technical problems in the prior art, the present invention proposes a method for cloud shadow detection of remote sensing images supported by surface type data, which is reasonably designed, overcomes the shortcomings of the prior art, and has good effects.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for detecting cloud shadow of a remote sensing image supported by surface type data includes the following steps:

[0007] Step 1: Construction of surface clear sky image element library and shade image element library;

[0008] Samples of cultivated land, forests, grasslands, shrubs, wetlands, artificial surfaces, and bare land are selected globally to build a library of clear sky pixels

[0009] and Yunyin image metabase; at the same time, a certain number of virtual shadow samples must be selected to ensure the accuracy of virtual shadow recognition during threshold simulation and probability calculation;

[0010] Step 2: Selection of an optimal band;

[0011] The optimal wavebands for different feature recognition shadows are different. In order to select the optimal wavebands for each surface, according to the established prior pixel library, the correctness and misjudgment rate of the negative image elements corresponding to each threshold range are simulated. If a certain band satisfies a higher shadow accuracy rate and a lower false positive rate, then this band is selected as one of the detection bands;

[0012] Step 3: Generation of a cloud shadow probability map;

[0013] The cloud shadow probability map is generated by calculating the cloud shadow probability at each wavelength by using a priori surface database of various types of tables, and fitting the obtained result to the cloud shadow probability calculation formula by using the S function; the feature requires multiple bands for detection , The final cloud shadow probability result is weighted by using the standard error of the function fitted at each band as the weight, and the cloud shadow binary result is obtained at the same time.

[0014] Preferably, in step 1, when constructing the surface clear sky image element library and the negative image element library, the following conditions must be satisfied:

[0015] First of all, the pixels in the pixel library must ensure that the pixels are shadow or clear sky and the numbers are close. When there is a thin cloud over the shadow, it should not be selected as a pixel library pixel to ensure the pixel library. Correctness; secondly, The pixel library must ensure that there are enough representative samples; finally, the selection of samples needs to consider the type of shadow.

[0016] Preferably, in step 2, in order to find the best band selection of each feature, according to the established prior pixel library, the reflectivity is changed from 0 to 1 at intervals of 0.01, respectively, with the change of the threshold value. Calculate the correctness rate of overcast image elements and the false positive rate of clear sky pixels in the corresponding pixel library; the correct ratio of overcast image elements is the ratio of the correct number of overcast image elements to the total number of overcast image elements; The ratio of the number of clear-sky pixels to the total number of clear-sky pixels judged as shadows;

[0017] When there is a threshold value that meets the conditions of overcast image element correctness rate and clear sky pixel misjudgment rate that are higher than 0.95 and lower than 0.1, respectively, this band is included in one of the best bands of the surface type .

[0018] Preferably, in step 3, the following steps are specifically included:

[0019] Step 3.1: The maximum correct value of the negative image element and the false positive rate of the clear-sky image element are 0 when the maximum value of the threshold value is MINI, and the minimum value of the threshold value when it is 1 is MAXI. The expression is as follows, where SCR represents the correctness of the shadow. , FR is the clear-sky error rate, and T is the threshold:

Indicates that [MN _I -F (I- 1) x0 OOlMH -f ix OJOl]

The number of vulvar image elements,

Table

[0025] shows the total number of pixels in the interval, from which the ratio of shadows to all pixels in the interval is calculated, and the probability value of the corresponding cloud shadow at a certain threshold is obtained;

[0026] Step 3.2: According to formula (5), calculate the cloud shadow probability based on the surface type support;

[0027]

[0028] In this calculation, a threshold value corresponding to the largest value when the probability is 1 is MIN, and a minimum threshold value corresponding to the probability is 0 when the probability is 0. When the image pixel value is less than MIN, the probability of cloud shadow is 1, When the element value is greater than M AX, the cloud shadow probability is 0; when the threshold is between MIN and MAX, according to statistics, the cloud shadow probability corresponding to each threshold will show a certain trend. According to statistics, the function with the best curve fitting effect is Sigmoid function, which is defined by the following formula:

[0029]

[0030] According to the fitting statistics, based on the optimal band selected by the table type in each place, the above cloud shadow probability simulation statistics are separately obtained to obtain a single band S-shaped curve fitting coefficient, goodness of fit, and standard error;

[0031] A formula for calculating the total cloud shadow probability, as shown in formula (6);

[0032]

[0033] wherein a, b and

The fitting coefficients of the S-type function are the number of bands required for the surface type i,

I is the weight of j-band of surface type i;

P _x

For the apparent reflectance of the remote sensing image, the formula (5) is used to calculate the cloud shadow probability of the j-band on the surface type i. When the apparent reflectance is lower than MIN, the cloud shadow probability is 1, and the pixel is considered to be a negative image. When the apparent reflectance is higher than MAX and the cloud shadow probability is 0, the pixel is considered to be a clear-sky pixel;

[0034] Step 3.3: According to formula (6), calculate and obtain the weighted synthesis of each optimal band, and finally obtain the cloud shadow detection probability result.

The beneficial effects of the invention

Beneficial effect

[0035] The beneficial technical effects brought by the present invention:

[0036] Aiming at the problem that the cloud shadow detection threshold is set without a ground surface reference and the detection accuracy is low due to the general threshold method, the present invention performs cloud shadow detection on Landsat 8 images based on the combination of the pixel spectral information and the surface type. The surface type support solves the problems caused by the different cloud shadow spectral characteristics under different surface types, and solves the problem of difficult cloud shadow threshold setting in a certain reflectance step size increase (cloud shadow and ground feature spectral characteristics have blurred boundaries , It is not easy to determine the threshold), so this method has better cloud shadow detection effect, achieves higher cloud shadow detection accuracy, and brings very useful cloud shadows Detection effect.

Brief description of the drawings

BRIEF DESCRIPTION OF THE DRAWINGS

[0037] FIG. 1 is a schematic diagram of apparent reflectance of cloud shadows under different underlying surfaces (ground surface), real shadows and virtual shadows; Figures (a), (b) and (c) are vegetation, artificial ground and bare ground, respectively Schematic diagrams of apparent reflectance curves under clear sky, virtual shadow coverage, and solid shadow coverage, respectively;

[0038] FIG. 2 is a schematic diagram of the simulation trend of the accuracy rate of the negative image elements and the false positive rate of the clear-sky image elements;

[0039] FIG. 3 is a schematic diagram showing the trend of the change in the probability of cloud shadows with reflectance values in different bands of different types of tables;

[0040] FIG. 4 is a regression analysis chart of the statistical results of the cloud shadow ratio by the LSCSD algorithm and visual interpretation;

[0041] FIG. 5 is a schematic diagram of Landsat 8 image data distribution in multiple scenes;

[0042] FIG. 6 is a schematic diagram of a vegetation surface detection result;

[0043] Among them, the underlying surface type (A)-cultivated land; (B)-forest; (C)-grassland; (D)-shrub; (E)-wetland;

[0044] FIG. 7 is a schematic diagram of the results of non-vegetation surface detection;

[0045] Among them, the underlying surface type (A)-artificial ground surface; (B)-bare ground.

Invention Examples

Embodiments of the invention

[0046] The present invention is further described in detail below with reference to the drawings and specific embodiments:

[0047] Cloud shadows are high-altitude clouds that block radiation from the sun, leaving less or no radiation on the ground. Cloud shadow detection is more challenging than cloud detection. The difference is that the apparent reflectance of cloud shadows has a great relationship with the surface type. Figure 1 shows the apparent reflectance of cloud shadows under different underlying surfaces, real shadows, and virtual shadows. ⑻, (b), and (c) are the apparent reflectance curves of vegetation, artificial ground, and bare ground, respectively, under clear sky, virtual shadow coverage, and solid shadow coverage. It can be seen from the figure that the shadow spectrum curves and The surface types are highly correlated, their trend distribution is consistent and overall is low. Cloud detection requires detection with a low reflectance band to ensure that there is sufficient difference from the high reflection of the cloud. Similarly, cloud shadows need to select a high reflectance band with a large band difference for detection, which can achieve better detection results. Through the simulation method, the optimal band is selected for detection using certain judgment conditions.

[0048] There is a large difference in reflectance between the cloud and the typical surface, and it is easy to determine with a certain threshold. The reflection of the cloud shadow changes with the underlying surface, and it is difficult to set a uniform threshold judgment. There are usually two threshold settings Methods: Image-based threshold determination method and empirical threshold determination method. Due to different image features and different reflectance distributions in these images, these two methods have certain spatial and temporal limitations. Aiming at the uncertainty of cloud shadow spectrum, it is a very effective method to use different table types to distinguish different thresholds for discrimination. Taking cultivated land, forest, grassland, shrubs, wetland, artificial surface and bare land as examples, cloud shadow detection based on surface type data is discussed.

[0049] The cloud shadow algorithm of the present invention is based on Landsat 8 data of cloud shadow and clear sky reflection spectrum difference. The cloud shadow reflectance and ground surface reflectance are relatively different in some bands. Therefore, it is not determined using the traditional threshold. The method determines the threshold, but based on the cloud shadow and clear sky pixel library of different surface types, taking all the bands of Lan dsat 8 into account, and using a threshold of a certain step size to simulate the clear sky correct rate and shadow correct rate of different thresholds for each band Then, based on the difference in single-band reflectivity, the band and threshold for cloud shadow detection are determined, and a cloud shadow probability map is generated. The threshold value obtained by this method is more accurate and reasonable, and has high applicability.

[0050] 1. Landsat8 Yin Image Metabase Establishment

[0051] The Landsat8 pixel library is composed of Yunyin image elements and clear sky pixels. When constructing the pixel library, 40 scene images were selected globally, and artificial visual sampling was performed on the cultivated land, forests, grasslands, shrubs, wetlands, artificial surfaces and bare land. First of all, the pixels in the cell library must ensure that the pixels in them are shadows or clear sky and the number of them is close. When there is a thin cloud over the shadow, it should not be selected as a cell library cell to ensure the correctness of the cell library. Secondly, the pixel library must ensure that there are enough representative samples, and 40 scene images are selected for sampling, which guarantees the number of samples, and at the same time, it is selected globally, and the samples are representative in different regions. Spatial limitations make the samples richer and more comprehensive. Finally, the selection of samples needs to consider the type of shadows. The establishment of the cell library is to more comprehensively statistics the characteristics of shadows and calculation thresholds in different surface types, so virtual shadows also need to be considered in Inside, the virtual shadow contains both the surface information and the shadow features. In order to make the simulation and threshold more accurate, when sampling the samples, pixels with virtual shadows are also needed to make the shadow features contained in the negative image meta library more complete. Subsequent shadow threshold simulations provide accuracy guarantees.

[0052] 2. Optimal Band Selection

[0053] Because the cloud shadow reflectivity trend of different surface types is related to the reflectivity of the ground surface itself, and the cloud shadow itself has a low and non-constant reflectance, it is not possible to find a suitable threshold to define the range of cloud shadows through the spectrum curve Around. Taking into account the small differences in visible light, near-infrared, and short-wave infrared of different ground features and cloud shadow spectra, a single-band detection method was used to select the cloud shadow detection band.

[0054] In order to find the optimal band selection of each feature, according to the established prior pixel library, the reflectivity is changed from 0 to 1 at intervals of 0.01, and the shadows in the corresponding pixel library are respectively calculated as the threshold changes. Pixel accuracy rate (ratio of the correct number of shadow image elements to the total number of shadow image elements) and clear-sky pixel false positive rate (the ratio of the number of clear sky pixels misjudged as shadows to the total number of clear sky pixels). Figure 2 is a schematic diagram of the simulated trend of the accuracy rate of cloudy image elements and the false positive rate of clear sky pixels. When the threshold is very low, both the accuracy rate of cloudy image elements and the false positive rate of clear sky pixels are 0. After that, the threshold begins to identify some shadows. The accuracy rate of overcast image elements has been significantly improved, and the discrimination of clear sky pixels is also more accurate. It has a lower misjudgment rate of clear sky pixels, until the difference between the accuracy rate of overcast image elements and that of clear sky pixels is the largest. As the threshold value increases, the accuracy rate of the negative image elements approaches 1, and more high reflectance clear sky is misjudged as shadows, and the false positive rate of clear sky pixels increases and approaches 1. In Figure 2, when there is a threshold value that meets the conditions for the correctness of the negative image element and the false positive rate of the clear-sky image element for a certain surface type, which are higher than 0.95 and lower than 0.1, respectively, this band is included in the best band of the surface type. One

[0055] 3. Support cloud shadow probability map generation method based on surface type

[0056] After the band selection, the optimal band for identifying the shadow of each surface type is selected, and the maximum value of the threshold when the correct rate of the negative image element and the false positive rate of the clear-sky image element in FIG. 2 are 0 is MINI, and The minimum value of the threshold at 1 is MAX1, and the expression is as follows, where SCR (Shadow Correct Rate) represents the shadow correct rate, FR (Fault Rate) represents the clear sky error rate, and T represents the threshold:

⑵;

[0059] When the threshold value Te [MIN1, MAXI], it sequentially traverses N intervals at intervals of 0.001 in this range, and statistically calculates the cloud shadow probability in each 0.001 range interval, and takes the right end of each interval as the value representing the interval. Threshold, so the corresponding cloud shadow probability at each threshold is obtained, and the cloud shadow probability at the ith interval is calculated as follows:

[0060]

[0061] wherein

Represents the total number of pixels in the interval. From this, the ratio of the shadow to all pixels in the interval is calculated, and the probability value of the corresponding cloud shadow at a certain threshold is obtained. In this calculation, the maximum threshold corresponding to MIN is 1 when the probability is 1, and the minimum threshold corresponding to MAX is 0 when the probability is 0. When the image pixel value is less than M IN, the probability of cloud shadow is 1, and the pixel value is When it is greater than MAX, the cloud shadow probability is 0. When the threshold is between MIN and MAX, according to statistics, the cloud shadow probability corresponding to each threshold will show a certain trend. According to statistics, the function with the best curve fitting effect is the Sigmoid function (S-shaped curve). The function is given by the following formula definition:

(4);

[0063] According to the fitting statistics, based on the selected optimal bands of each table type, the above cloud shadow probability simulation statistics are separately obtained to obtain the single-band S-shaped curve fitting coefficient, goodness of fit (R2), and standard error (Standard Error of Estimate (SEE) As shown in Figure 3 and Table 1, the fitting function of each band of various types of surface is superior, has a strong correlation with sample data, and has a small standard error.

Table 1 Optimal band fitting equation coefficients, goodness of fit, and standard error for each table type

[]

[0065] According to formula (5), calculate the cloud shadow probability based on the surface type support;

[0066]

l3 p _u > MMX

(5);

[0067] A formula for calculating the total cloud shadow probability, as shown in formula (6);

[0068] ₅₅

[

(6)

[0069] wherein a, b and

Are the fitting coefficients of the sigmoid function (Table 1),

m

Is the number of bands required for surface type i,

Is the weight of j-band of surface type i;

For the apparent reflectance of the remote sensing image, the formula (5) is used to calculate the cloud shadow probability of the j-band on the surface type i. When the apparent reflectance is lower than MIN, the cloud shadow probability is 1, and the pixel is considered to be a negative image. When the apparent reflectance is higher than MAX and the cloud shadow probability is 0, the pixel is considered to be a clear-sky pixel.

[0070] The results of cloud shadow probability obtained by different wavebands and different S-curves are different. If the wavebands are simply merged, accumulation of shadow misjudgment errors will be formed, resulting in inaccurate probability results. , The probability information will be lost, and the shadow miss rate will increase. [0071] According to formula (6), a weighted combination of each optimal band is obtained by calculation, and finally a cloud shadow detection probability result is obtained.

[0072] Therefore, based on the above-mentioned problems, according to the standard error of the fitted S function, different weights are assigned to the wavebands involved. The weights with large errors are low and the weights with small errors are high. This setting makes the probability results more accurate. Credible. Table 2 is the weight value corresponding to each band in each table. Formulas (5) and (6) are the final cloud probability calculation formulas. According to the cloud shadow probability map, the cloud shadow result can be easily obtained, with a limit of 75%, which is larger than the probability. Pixels are judged as Yunyin image pixels, otherwise they are clear sky pixels.

Table 2 Cloud Shadow Probability MIN, MAX and Weight Distribution in Each Band

[0074] It should be noted that, because of the low reflectivity of water bodies and the ocean, the visual discrimination of shadows is difficult, and the accuracy of sample selection cannot be guaranteed. Therefore, a constant threshold is used for shadow discrimination over the two surface types. It can be seen from the spectrum curve that the reflectance of water body decreases with increasing wavelength. The reflectance is the largest (greater than 0.1) in the dark blue band, and the sensitivity to the effect of reducing the reflectance of the shadow is large. Therefore, when the reflectance in this band is lower than At 0.1, it is judged as a pixel with shadow.

[0075] In addition, in the vegetation area, there may be errors such as the type of vegetation changed to the type of water body, there is a certain boundary error between the vegetation and the water body on the surface library and the image, and the water body is misjudged as a shadow. Therefore, after obtaining the shadow detection results, in the detection results of the vegetation surface (arable land, forest, grassland, shrubs and wetlands), The condition that the ND VI is less than 0 is used to remove the case where the water body is misjudged as a shadow.

[0076] The quantitative verification of cloud shadow detection results still uses the method of verification of previous cloud detection results, that is, 6 types of sample areas of 500 * 500 pixels in each type are randomly selected in the corresponding image for visual interpretation of cloud shadow areas. When clouds and cloud shadows exist simultaneously, this type of pixel is divided into cloud pixels. The visual interpretation result is then used as the true value, and compared with the cloud detection result of the LSCSD (Land Cover Cloud Shadow Detection, surface-type cloud shadow detection) algorithm. Use Cloud Shadow Proportion (CSP) to quantitatively evaluate the accuracy of the LSCSD algorithm.

[0077] First, perform a regression analysis on the cloud shadow ratio of the visual interpretation result of the sample area and the cloud shadow ratio of the LSCSD algorithm result, and calculate the RMSE (Root Mean Square Error, root mean square error) of the LSCSD cloud shadow detection result; Then use CRCS (Correct Rate of Cloud

Shadow (Correct Rate of Clear-Sky) and CRS (Correct Rate of Clear-Sky) two indexes are used to perform detailed statistical analysis and accuracy evaluation of cloud overcast image elements and clear sky image elements on cloud detection results. Finally, all samples are counted. The overall correct rate of points, that is, the correct rate (TCR) of the LSCSD algorithm for correct recognition of both cloud and clear sky pixels.

[0078] FIG. 4 is a regression analysis chart of LSCSD algorithm and visual interpretation of the statistical results of cloud shadow ratio, and Table 3 is the RMSE calculation result of table type cloud shadow ratio in each place. It can be seen that when comparing the table types in different places with the true value, the overall trend of the cloud shadow ratio of the LSCSD algorithm is consistent with the visual interpretation result, and its overall RMSE is 3.37%, reaching a high correlation. Analyzing the error, the accuracy is generally underestimated, but the overestimation is more obvious than the cloud shadow, which indicates that the overestimation of the shadow is more than that of the cloud. This is because the shadow and the spectral characteristics of the surface in clear sky are more similar. It can be seen from Figure 4 and Table 3 that the wetlands deviate more from the reference line, and the underestimation phenomenon is more obvious than other surface types, with an RMSE of 5.83%; secondly, the surface types with large errors are cultivated land and forest, and the RMSE score is 4.62%. 4.54%; The type of surface with the smallest error in the statistics of cloud shadow ratio is artificial surface, with RMSE of 1.42%.

Table 3 RMSE Results of Cloud Shadow Proportion for Various Table Types

[] [Table 1]

[0080] This statistic indicates that the statistical error of cloud shadows in wetlands has the largest statistical error. Due to the existence of dark grounds of various degrees in wetlands, when selecting clear-air pixel samples, some dark-surface clear skies are included. Some light shadows on the bright surface are also selected. Therefore, compared with other ground surfaces, the difference between the wetland clear-sky pixel library and the shaded image pixel library is relatively small. Therefore, the error of the cloud shadow ratio in the wetland of the LSCSD algorithm is relatively larger than other ground objects. Due to its high reflectivity, the artificial surface is distinguished from the cloud shadow spectral information, so the statistics of cloud shadow ratio are closest to the true value. Generally speaking, the error of the proportion of cloud shadows on all surfaces is generally low and the accuracy is good.

[0081] For Landsat OLI and TIRS (Thermal InfRared Scanner) data, the DN (Digital Number, pixel brightness value) value of the image itself needs to be converted into a table with clear physical meaning through formulas (7-9). Observe the brightness of the radiation, and then calculate the corresponding apparent reflectance or brightness temperature, and then further process the detected image. The conversion formula is:

[0082] _£ Gsfs- Q.,-JMs

(7)

[0085] wherein

Gain Is the gain value,

Bias

Is the offset value,

Expressed as the apparent reflectance of the top layer of the atmosphere, the average distance between the sun and the earth,

Is the solar spectral irradiance at the top of the atmosphere

ft

Zenith angle

T

Is the brightness temperature on the star, the unit is K, and the calibration constants are respectively. For Landsat 8 TIRS 10, the values are 774.885.

And 1321.08K.

[0086] In the accuracy verification, the multi-view Landsat 8 image data distribution of various types of typical surface areas and complex surface areas distributed globally is shown in FIG. 5, because North America and Asia are relatively large and the surface types are relatively diverse. Therefore, there are too many images selected in this area, and each type of selected image is guaranteed to be evenly distributed as much as possible. After the above pre-processing process and cloud detection, the accuracy verification analysis of the experimental results is performed. A total of 60 sample areas were selected, and cloud speckles were delineated based on the false-color composite image of the sample areas, and the LSCSD cloud detection results were quantitatively verified.

[0087] The reflectance of the water body itself is extremely low, no matter whether it is visual discrimination and comparison or the outline of the shadow area for quantitative analysis, it is impossible to obtain a more accurate judgment, which affects the accuracy evaluation. Therefore, the accuracy verification of cloud shadows does not consider the underlying surface as The case of water bodies. The spectral characteristics of cloud shadows are affected by the underlying spectral characteristics of the underlying surface. Therefore, in the analysis, the surface is directly divided into vegetation-type and non-vegetation-type surfaces. Figure 6 and Figure 7 are visual interpretations of the comparison results of vegetation and non-vegetation cloud shadow detection results, where the left side of the figure is a false color composite image, and the middle is the cloud shadow probability map (from black to white, cloud shadow The probability ranges from 0 to 1), and the right side is the binary result of the final cloud shadow detection. In order to more intuitively and completely show the results of cloud shadow detection, when selecting a region, real shadows and virtual shadows are considered comprehensively, and visual interpretation of different forms of shadows is also performed for comparison.

[0088] Vegetation surface

[0089] FIG. 6 is a schematic diagram of a detection result of a vegetation surface. The surface types include cultivated land, forest, grassland, shrub land, and wetland. The situation of cultivated land is more complicated, and the degree of surface type changes from vegetation to bare land. At the same time, there are differences in surface characteristics of different areas caused by sparseness and seasonal changes in grass and shrub land. The LSCS D algorithm can completely identify cloud shadows in the case of uneven bare land and vegetation coverage in the graph (A), and more bare land information in the graph (D), and has strong spatial continuity and adaptability. Sex. This is because in the selection of samples, not only the lush vegetation period, but also the sparse vegetation area or the area showing bare land characteristics are taken into account. The best bands obtained include near-infrared and short-wave infrared, especially In the near-infrared band, both vegetation and bare land are highly reflective and distinguishable from shadows. Therefore, discrimination errors caused by different types of ground surface can be minimized.

[0090] FIG. (B) shows the recognition result of shadows on the bare ground. Bare ground generally has a high reflectance, so it is easier to judge Don't shadow one of the figure types. It can be seen that for dense cloud shadows generated by thin clouds, including solid shadows and virtual shadows, most of them can be identified, but in areas where a small amount of clouds and shadows are mixed, they are affected by high reflectivity clouds. In general, the algorithm has higher detection accuracy over bare ground.

[0091] Based on the visual interpretation of the combined vegetation and non-vegetation surfaces, the LSCSD algorithm is applied to all types of surfaces, whether on the vegetation surfaces with different vegetation coverage or on the highlighted non-vegetation surfaces. Can show good recognition ability, especially for the recognition of various forms of virtual shadows. However, when some surface-type regions show special low reflectivity, the LSCSD algorithm may misjudge them as shadows.

[0092] Of course, the above description is not a limitation on the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention are also It should belong to the protection scope of the present invention.

Claims

Claim

[Claim i] A method for detecting cloud shadow of a remote sensing image supported by surface type data, which comprises the following steps:

Step 1: Construction of surface clear sky image library and shade image library;

Samples of cultivated land, forests, grasslands, shrubs, wetlands, artificial surfaces, and bare land are selected globally to build a clear sky pixel library and a cloud shadow image library; select virtual shadow samples to ensure the accuracy of virtual shadow recognition during threshold simulation and probability calculation ;

Step 2: Selection of the optimal band;

According to the established prior pixel library, the correctness and false positive rate of the negative image elements corresponding to each threshold range are simulated. If a certain band satisfies a higher shadow correct rate and has a lower false positive rate, then this band is selected. As one of the detection bands;

Step 3: Generation of cloud shadow probability map;

The cloud shadow probability map is generated by calculating the cloud shadow probability at each wavelength by using a prior surface surface library of each type of table, and fitting the obtained result to the cloud shadow probability calculation formula by using the S function. The feature requires multiple bands for detection. The standard error of the fitted function of each band is used as a weighted weight to synthesize the final cloud shadow probability result, and the cloud shadow binary result is obtained at the same time.

[Claim 2] The method for detecting cloud shadow of remote sensing image supported by the surface type data according to claim 1

, Which is characterized in that in step 1, when constructing the surface clear sky pixel library and the negative image pixel library, the following conditions must be met:

First of all, the pixels in the cell library must ensure that the pixels in them are shadows or clear sky and the number of them is close. When there is a thin cloud over the shadow, it should not be selected as a cell library cell to ensure the correctness of the cell library. Secondly, the pixel library must ensure that there are enough representative samples. Finally, the selection of the samples needs to consider the type of shadows.

[Claim 3] The method for detecting cloud shadow of remote sensing image supported by the surface type data according to claim 1

, Which is characterized in that, in step 2, in order to find the optimal band selection of each feature, according to the established prior pixel library, the reflectivity is changed from 0 to 1 at intervals of 0.01, and calculated separately with the change of the threshold Corresponds to the accuracy rate of the negative image elements and the false positive rate of the clear image elements in the pixel library; the correctness ratio of the negative image elements is the ratio of the correct number of negative image elements to the total number of negative image elements; the clear sky image The misjudgment rate is the ratio of the number of clear-sky pixels that are misjudged as shadows to the total number of clear-sky pixels. When a certain surface type exists, the correct rate of the overcast image elements and the clear-sky pixel misjudgment rates are higher than 0.95, Below the threshold of the 0.1 condition, this band is included in one of the best bands of the surface type.

[Claim 4] The method for cloud shadow detection of a remote sensing image supported by the surface type data according to claim 1, characterized in that: in step 3, the method specifically includes the following steps: Step 3.1: the accuracy rate of the negative image element and the clear-sky image element When the false positive rate is 0 at the same time, the maximum value of the threshold is MINI, and when the false value is 1, the minimum value of the threshold is MAXI.

When the threshold value Te [MINl, MAXI], it sequentially traverses N intervals at intervals of 0.001 in this range, and statistically calculates the cloud shadow probability in each interval of 0.001 range. The right end of each interval is the threshold value of the interval, so we get The cloud shadow probability at each threshold corresponds to the cloud shadow probability at the ith interval:

The number of vulvar image elements,

Represents the total number of pixels in the interval, from which the ratio of the shadow to all pixels in the interval is calculated, and the probability value of the corresponding cloud shadow at a certain threshold is obtained;

Step 3.2: Calculate the cloud shadow probability based on the surface type according to formula (5)

In this calculation, the maximum threshold corresponding to MIN is 1 and the minimum threshold is MAX to 0. When the image pixel value is less than MIN, the probability of cloud shadow is 1 and the pixel value is greater than At MAX, the cloud shadow probability is 0. When the threshold is between MIN and MAX, according to statistics, the cloud shadow probability corresponding to each threshold will show a certain trend. According to statistics, the function that has the best curve fitting effect is the Sigmoid function. The function is defined by the following formula:

According to the fitting statistics, based on the optimal band selected by the table type in each place, the simulation statistics of the above cloud shadow probabilities are respectively obtained, and the single-band S-shaped curve fitting coefficient, goodness of fit, and standard error are obtained;

The total cloud shadow probability calculation formula is shown in formula (6);

⑹;

Where a, b and

^X 0 is the fitting coefficient of s-type function,

Is the number of bands required for surface type i, where: is the weight of j-band for surface type i;

Step 3.3: According to formula (6), calculate the weighted composition of each optimal band, and finally obtain the cloud shadow detection probability result.