Deep learning cloud detection method fusing object-oriented dynamic threshold and feature enhancement

Abstract

The invention relates to a deep learning cloud detection method fusing object-oriented dynamic threshold and feature enhancement, and belongs to the field of remote sensing image cloud detection. The method comprises the steps that original satellite remote sensing images are acquired, and a dynamic threshold method based on a triangulation method is adopted for each image to generate a mask threshold; according to the mask threshold, generating a high-confidence cloud mask, a low-confidence cloud mask and a shadow mask in the remote sensing image in combination with an improved ACCA algorithm; according to the generated mask, matching a cloud object and a shadow object in the image to obtain a cloud mask image; for a wrong mask in the cloud mask image, eliminating false detection through mask replacement; constructing all finally generated cloud mask images into a remote sensing image training data set, and training a texture feature enhanced deep learning cloud detection network model through the data set; the trained model is used for predicting cloud in a satellite remote sensing image. According to the invention, the accuracy of cloud detection in complex scenes such as snow mountains, desert saline-alkali soil and urban highlight earth surfaces can be improved.

Description

Technical Field

[0001] The present invention belongs to the fields of remote sensing image cloud detection and remote sensing image cloud detection dataset construction, and relates to a deep learning cloud detection method that combines object-oriented dynamic threshold and feature enhancement. Background Art

[0002] With the rapid development of remote sensing technology, satellite remote sensing images play an important role in fields such as meteorological monitoring, environmental observation, and disaster warning. Cloud detection, as a key step in remote sensing image preprocessing, directly affects the accuracy of subsequent ground object analysis and data application. However, existing cloud detection technologies still have significant limitations, especially in terms of performance in complex surface scenes that urgently need to be improved.

[0003] In the field of traditional cloud detection, mainstream methods rely on the threshold segmentation principle of multi-spectral reflection bands. Such methods distinguish clouds from the ground by setting fixed or semi-fixed spectral thresholds (such as the reflectance ratio of visible and near-infrared bands). However, such methods have the following defects: First, fixed threshold parameters are difficult to adapt to the spectral feature differences of different surface cover types. For example, in high-brightness surface areas such as snow-capped mountains and desert saline-alkali lands, their reflectance characteristics are similar to those of cloud layers, resulting in the traditional threshold method being prone to misjudging the ground as clouds; while in areas with dense vegetation, thin or semi-transparent cloud layers may be missed because their reflectance is close to that of vegetation. Second, traditional methods lack the ability to dynamically adjust. In practical applications, changes in cloud shape, thickness, and lighting conditions will significantly affect the spectral response, but existing methods cannot adaptively adjust the threshold according to the local features of the image, resulting in insufficient robustness of the detection results. Third, the correlation between clouds and shadows has not been fully exploited. Shadows cast by clouds are often misjudged as independent ground objects, and traditional methods do not effectively use the geometric relationship of the cloud-shadow pair for correction, further reducing the detection accuracy.

[0004] With the rise of deep learning technology, cloud detection methods based on convolutional neural networks (CNNs) have gradually become a research hotspot. However, existing deep learning methods still face three core challenges: First, data dependence is strong. Deep learning models require a large amount of pixel-level annotated cloud mask data for training, but remote sensing image annotation is costly and inefficient. Manual annotation requires distinguishing between clouds and bright surfaces pixel by pixel, which is time-consuming and labor-intensive and prone to subjective errors. Second, the feature extraction capability is limited. Existing models mostly rely on conventional convolution and downsampling operations to extract features, but their ability to capture cloud texture, edge details and spatial distribution features is insufficient. For example, low-contrast areas between thin clouds and bright surfaces are easily ignored, and irregular boundaries of clouds are prone to lose details due to pooling operations, resulting in blurred boundaries of segmentation results. Third, poor adaptability to dynamic scenes. Existing models lack an effective fusion mechanism for multi-scale features and contextual information, and are difficult to cope with dynamic changes in different resolutions, lighting conditions and surface types. For example, the detection performance of the same model in thick cloud coverage areas and thin cloud mixed scenes is significantly different, and the false detection rate of shadow areas remains high.

[0005] In addition, there are few solutions in the existing technology that combine the advantages of traditional threshold segmentation with deep learning feature enhancement. Although traditional methods have the advantage of high computational efficiency in specific scenarios, their static threshold setting and the dynamic feature learning ability of deep learning have not yet formed a complementary relationship. How to construct a fusion method that can both use prior knowledge to reduce labeling dependence and improve the detection accuracy of complex scenes through deep learning is still a technical problem that needs to be solved urgently in this field. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a deep learning cloud detection method that integrates object-oriented dynamic threshold and feature enhancement. This method designs a cloud mask annotation method by using object-oriented principles and dynamic thresholds, thereby efficiently constructing a cloud detection dataset, solving the problem of low efficiency and high cost in constructing remote sensing image cloud detection datasets; and solving the problem of low accuracy of cloud detection on highlighted surfaces through a deep learning network based on a hybrid attention mechanism.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A deep learning cloud detection method integrating object-oriented dynamic threshold and feature enhancement, the method comprising:

[0009] S1. Obtaining an original satellite remote sensing image, and generating a mask threshold for each image using a dynamic threshold method based on triangulation, wherein the mask threshold includes a high-confidence cloud mask threshold, a low-confidence cloud mask threshold, and a shadow mask threshold;

[0010] S2. Generate a high-confidence cloud mask, a low-confidence cloud mask, and a shadow mask in the remote sensing image according to the mask threshold and in combination with the improved ACCA algorithm;

[0011] S3. Match the cloud objects and shadow objects in the image according to the generated high-confidence cloud mask, low-confidence cloud mask, and shadow mask to obtain a cloud mask image;

[0012] S4. For the incorrect masks in the cloud mask image, eliminate false detections through mask replacement to improve the accuracy of the cloud mask;

[0013] S5. Construct all the finally generated cloud mask images into a remote sensing image training data set, and train a deep learning cloud detection network model with enhanced texture features through this data set;

[0014] S6. After the model is trained, it is used for the prediction of clouds in satellite remote sensing images.

[0015] Further, in step S1, generating the mask threshold by using the dynamic threshold method based on the triangulation method includes:

[0016] S11. Arbitrarily select an image from the original satellite remote sensing image, calculate the grayscale histogram of this image, set the origin in the grayscale histogram as point a, set the point with the highest frequency in the range of grayscale values from 125 to 254 in the grayscale histogram as point b, and connect points a and b to form a line L ab ;

[0017] S12. Find the point with the maximum distance from L in the range of grayscale values from 125 to 254, record the grayscale value corresponding to this point with the maximum distance, and then generate a pixel reference threshold through the inverse operation of grayscale conversion; ab

[0018] S13. Generate a high-confidence cloud mask threshold, a low-confidence cloud mask threshold, and a shadow mask threshold according to the reference threshold.

[0019] Further, in step S2, for any pixel in the image, the following judgment process is performed:

[0020] S21. Obtain the R-channel value B3 i , G-channel value B2 i , and NIR-channel value B4 i of pixel i;

[0021] S22. If B4 i < T S , and B4 i / B3 i > 1.5, then determine that this pixel is a shadow and generate a shadow mask; where T S represents the shadow mask threshold;​

[0022] S23. If B3 i > T H , then judge B4 i / B3 i whether it is less than 2.35; if B4 i / B3 i ≥2.35, then pixel i is clear sky; if B4 i / B3 i <2.35, and B4 i / B2 i <2.16, then pixel i is high-confidence cloud, and generate a high-confidence cloud mask; if B4 i / B3 i <2.35, and B4 i / B2 i ≥2.16, then pixel i is clear sky; where, T H represents the high-confidence cloud mask threshold;

[0023] S24. If B3 i ≤T H , then judge B3 i whether it is greater than the low-confidence cloud mask threshold T L ; if B3 i ≤T L , then pixel i is clear sky; if B3 i >T L , then enter step S25;

[0024] S25. If B4 i / B3 i ≥2.35, then pixel i is clear sky; if B4 i / B3 i <2.35, and B4 i / B2 i <2.16, then pixel i is low-confidence cloud, and generate a low-confidence cloud mask; if B4 i / B3 i <2.35, and B4 i / B2 i ≥2.16, then pixel i is clear sky.

[0025] Furthermore, in step S3, the cloud object and the shadow object are matched according to the affine invariant relationship between the cloud and the shadow, including:

[0026] S31. Screen the cloud objects in the high-confidence cloud object set to ensure that the number of pixels of the high-confidence cloud object meets the requirements, and the proportion of the area occupied by the high-confidence cloud object in the 8-neighborhood is less than one-eighth;

[0027] S32. For each screened high-confidence cloud object, using its center as the center point, search for shadow objects in the four quadrants of 0° - 90°, 90° - 180°, 180° - 270°, and 270° - 360°;

[0028] S33. When the ratio of the horizontal length to the vertical length between the high-confidence cloud object and the shadow object is approximately equal to 1, and the ratio of the number of pixels between the high-confidence cloud object and the shadow object is between 0.8 and 1.25, the high-confidence cloud object and the shadow object are successfully matched;

[0029] S34. Calculate the movement offset between the successfully matched high-confidence cloud object and the shadow object, and based on this movement offset, filter out non-cloud objects in the low-confidence cloud objects, thereby generating a cloud mask image.

[0030] Further, step S4 includes that when there is a missed detection in the generated cloud mask image, using the low-confidence cloud mask to replace the cloud mask image with missed detection; when there is a false detection in the generated cloud mask image, using the cloud-free mask to replace the false detection part in the image.

[0031] Further, the texture feature enhanced deep learning cloud detection network model is used to process the input satellite remote sensing image and output a prediction probability map including clear sky categories and cloud categories; wherein, the texture feature enhanced deep learning cloud detection network model includes an encoder and a decoder;

[0032] The encoder includes a double convolution module and multiple cascaded texture feature enhanced downsampling modules, wherein the multiple cascaded texture feature enhanced downsampling modules are connected to the output end of the double convolution module; the decoder includes multiple cascaded upsampling modules;

[0033] The modules in the first N - 1 layers of the encoder are skip-connected to the upsampling modules in the corresponding layers of the decoder.

[0034] Among them, the texture feature enhanced downsampling module includes a first-level wavelet transform layer, a splicing layer, a first convolution layer, a second convolution layer, and a CBAM attention layer connected in sequence. At the output ends of the first convolution layer and the second convolution layer, batch normalization and activation function layers are both connected.

[0035] The beneficial effects of the present invention are as follows:

[0036] (1) Through the synergistic effect of dynamic threshold generation and the feature enhancement network, the present invention effectively solves the spectral confusion problem between highlighted ground surfaces (such as snow-capped mountains, desert saline-alkali lands) and clouds. The dynamic threshold method adaptively adjusts the mask threshold parameters of the image by pixel-by-pixel calculation according to different types of remote sensing images such as urban forests, snow-capped mountains, and desert saline-alkali lands, avoiding systematic errors caused by fixed thresholds; while the deep learning model enhances the texture and spatial feature extraction ability through a hybrid attention mechanism to accurately distinguish clouds from complex backgrounds. Experiments show that the detection results of the present invention in thin cloud, highlighted ground surface, and cloud-shadow mixed regions are all superior to traditional methods and conventional deep learning models.

[0037] (2) The present invention innovatively proposes a cloud mask automatic annotation method based on dynamic threshold and improved ACCA algorithm, corrects the mask error through the cloud-shadow affine relationship, and introduces an artificial review and mask replacement strategy to improve the quality of the dataset. Compared with traditional manual annotation, this method reduces the dependence on a large amount of manually annotated data, significantly shortens the dataset construction cycle, and at the same time ensures that the annotation results have high consistency and scalability, providing efficient and low-cost data support for deep learning model training.

[0038] (3) The texture feature enhancement network (TENet) designed by the present invention realizes the refined extraction of multi-scale features of clouds through multi-level feature fusion of wavelet transform and attention mechanism. The encoder-decoder structure combined with skip connections effectively retains the cloud edge details; the CBAM module strengthens the model's attention to key regions through channel and spatial attention weighting. This design enables TENet to adapt to the dynamic changes of different resolutions, lighting conditions, and ground surface types, and can improve the accuracy of cloud detection in complex scenes such as snow-capped mountains, snow fields, desert saline-alkali lands, and urban highlighted ground surfaces.

[0039] (4) The present invention takes into account both algorithm accuracy and computational efficiency. The proposed method can be directly applied to the real-time processing of domestic satellite remote sensing images, providing a reliable technical solution for engineering applications in fields such as meteorological monitoring and environmental assessment. In addition, its modular design is convenient for integration with other remote sensing analysis tasks (such as land surface classification, change detection), and has broad expansion potential.

[0040] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail and preferably in conjunction with the accompanying drawings, where:

[0042] Figure 1 Schematic diagram of the dynamic threshold method based on triangulation;

[0043] Figure 2 Process of generating cloud mask image;

[0044] Figure 3 Cloud mask image generated in Example 1, Figure 3 (a) RGB image, Figure 3 (b) HCC mask, Figure 3 (c) LCC mask, Figure 3 (d) Cloud mask image;

[0045] Figure 4 (a) Random point marking of the original remote sensing image of the forest area containing some thin cloud areas, Figure 4 (b) Random point marking of the modified mask, Figure 4 (c) Cloud mask image of the forest area containing some thin cloud areas before manual review;

[0046] Figure 4 (d) Original remote sensing image of saline-alkali land and snowfield, Figure 4 (e) Cloud mask image after manual review and replacement, Figure 4 (f) Cloud mask image of saline-alkali land and snowfield before manual review;

[0047] Figure 5 (a) Schematic diagram of randomly sampled points in the original remote sensing image generated by ArcMap, Figure 5 (b) Schematic diagram of random point marking of the mask;

[0048] Figure 6 Schematic diagram of the TEDM module structure;

[0049] Figure 7 Schematic diagram of the TENet model structure;

[0050] Figure 8 (a) Original remote sensing image of the desert surface, Figure 8 (b) Cloud detection result map generated by the TENet model, Figure 8 (c) Cloud detection result map generated by the U-Net model;

[0051] Figure 9 (a) Original remote sensing image, Figure 9 (b) Cloud detection result map after adding CAM to the U-Net model, Figure 9 (c) Cloud detection result map after adding CBAM to the U-Net model;

[0052] Figure 10 For the cloud detection results for the highlighted mountainous surface, Figure 10 (a) is the original image, Figure 10 (b) is the cloud detection result map generated by the TENet model, Figure 10 (c) is the cloud detection result map generated by the U-Net model with CAM added, Figure 10 (d) is the cloud detection result of the U-Net model with CBAM added;

[0053] Figure 11 For the cloud detection results for the snow mountain and cloud mixed scene, Figure 11 (a) is the original image, Figure 11 (b) is the cloud detection result generated by the TENet model, Figure 11 (c) is the cloud detection result generated by the U-Net model with CAM added, Figure 11 (d) is the cloud detection result of the U-Net model with CBAM added;

[0054] Figure 12 Schematic diagram of the actual prediction results of different models on the GF-1 ARD cloud detection dataset;

[0055] Figure 13 Schematic diagram of the process of the deep learning cloud detection method integrating object-oriented dynamic threshold and feature enhancement described in the present invention. Detailed implementation manners

[0056] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0057] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and cannot be understood as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which does not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0058] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the accompanying drawings. This is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and should not be construed as a limitation of the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0059] Embodiment 1

[0060] This embodiment provides a construction method for a remote sensing image cloud detection dataset. 1107 WFV scenes are obtained from GF-1 ARD remote sensing images. All data have undergone geometric correction and atmospheric correction, covering various land types, including forests, bare lands, gobi deserts, deserts, snowfields, snow-capped mountains, cultivated lands, and urban areas, etc. The GF-1 and GF-6 satellites are only equipped with blue, green, red, and near-infrared bands, lacking thermal infrared and short-wave infrared bands. The size of each image is 6863×6863 pixels, and the number of spectral bands is 4. These remote sensing images include desert saline-alkali lands in the northwest of China, snow-capped mountains in the southwest of China, forests and farmlands in the central part of China, grasslands and perennial snow-covered areas in the north of China, some ocean areas, the land-sea connected part of Hainan Island, and some urban areas.

[0061] The construction method provided in this embodiment includes:

[0062] 1. Design a dynamic threshold method based on the triangulation method. Specifically, for any remote sensing image x, calculate the grayscale histogram of image x, set the point (0, 0) as point a, and set the peak point within the grayscale value range of 125 - 254 as point b. Connect points a and b to form a line L ab ; calculate the distance between each grayscale value peak point within the grayscale value range of 125 - 254 and L ab , retain the peak point with the maximum distance, and record the pixel value corresponding to this point as Maxi. Through the inverse calculation of grayscale conversion, generate the reference threshold T, and generate the high-confidence cloud (HCC) mask threshold T H , low-confidence cloud (LCC) mask threshold T L and shadow mask threshold T S .

[0063] In this embodiment, according to experience, T H 、TL , T S are set to 120%, 80%, and 40% of T respectively.

[0064] 2. Apply the threshold to the improved ACCA algorithm as shown in Figure 1 to generate a high-confidence cloud mask, a low-confidence cloud mask, and a shadow mask. The main process is that after performing dynamic threshold calculation on each input GF-1 remote sensing image, use T H , T L , T S to perform initial classification of HCC, LCC, and shadow for each pixel in each remote sensing image, and output the final HCC mask, LCC mask, and shadow mask for subsequent cloud mask correction.

[0065] For each pixel in the image, first obtain the R-channel value B3, G-channel value B2, and near-infrared (NIR) channel value B4 of the pixel, and then make the following judgments:

[0066] 1) B3 (0.63μm - 0.69μm) is the red light band. Thick clouds, snow-capped mountains, saline-alkali land, etc. absorb less radiation in the red light band and show relatively high reflectivity on these bright surface scenes. Vegetation, water bodies, etc. can absorb a large amount of red light band radiation and generally show lower reflectivity on ordinary surface scenes.

[0067] Therefore, use B3 as the parameter for cloud area confidence. According to the threshold setting of the ACCA algorithm and the empirical judgment of the dynamic threshold, when the B3 of the pixel = T H , it is the high-confidence threshold, when B3 = T L , it is the low-confidence threshold. When B3 < T L , it is judged as a non-cloud pixel. When T L < B3 < T H , it is judged that it may be an LCC pixel. When B3 > T H , it is judged that it may be an HCC pixel.

[0068] 2) B4 (0.76μm - 0.90μm) is the near-infrared band. The radiation in this band is less absorbed by water bodies and vegetation and has a relatively high reflectivity. At the same time, B3 has a lower reflectivity in the vegetation surface scene, and the cloud absorption in the visible light band and the near-infrared band channels is similar.

[0069] Therefore, use B4 / B3 as the parameter to distinguish clouds from water bodies and vegetation. Referring to the threshold setting of the ACCA algorithm, when B4 / B3 < 2.35, it is determined that the pixel may be a cloud.

[0070] 3) B2 (0.52μm - 0.60μm) is the green light band, which is easily absorbed by vegetation and has a low reflectivity on the surface of densely vegetated areas such as forests, grasslands, and cultivated lands.

[0071] Therefore, using B4 / B2 as the parameter to distinguish vegetation and cloud areas, referring to the threshold setting of the ACCA algorithm, when B4 / B2 < 2.16, it is determined that the pixel is a cloud.

[0072] 4) For the pixels in the shadow area, the reflectance of the B4 band is less than that of the B3 band.

[0073] Therefore, using B4 and B4 / B3 as the parameters to judge shadow pixels, according to the threshold setting of the improved ACCA algorithm and the empirical judgment of the dynamic threshold, let the pixels that simultaneously meet the two conditions of B4 < T S and B4 / B3 > 1.5 be shadow pixels.

[0074] 3. According to Figure 2 the shown process, using the affine invariant relationship between clouds and cloud shadows, matching cloud objects and shadow objects, calculating the offset factor, screening out the false detection results in the low-confidence cloud mask, and finally generating the cloud mask image as Figure 3 shown.

[0075] Given that the sun's rays can be approximated as a parallel light source, there is an affine transformation relationship between the cloud layer and its shadow. Within a limited imaging range, for example, at the same time in the same area, the geometric feature ratios of different cloud-shadow pairs are significantly consistent, and this relationship can be used to correct the image. Convert the mask image into an object set, and the information of each object includes the number of pixels, aspect ratio, etc. According to the solar azimuth angle and the quantization conditions of isolated cloud clusters, there are the following conditions:

[0076] Condition 1: Screen cloud objects in the HCC object set to ensure that the cloud clusters have sufficient reflectance to leave a significant shadow on the ground. Confirm that the number of pixels of the HCC object is between 2000 and 4000, and the proportion of the area occupied by the object in the 8-neighborhood is less than one-eighth, so as to screen out suitable small isolated cloud clusters.

[0077] Condition 2: According to the solar azimuth angle, confirm the matching shadow object. Taking the center of the HCC object that meets Condition 1 as the center point, and looking for the shadow object in the four quadrants of 0° - 90°, 90° - 180°, 180° - 270°, and 270° - 360° divided according to the size of the solar azimuth angle. Specifically, when the sun is in the due south direction, the solar azimuth angle is 0°; when the sun is in the due east direction, the solar azimuth angle is 90°; when the sun is in the due north direction, the solar azimuth angle is 180°; when the sun is in the due west direction, the solar azimuth angle is 270°.

[0078] Condition 3: The cloud shadow objects matched in the appropriate relative positions also need to meet the requirements that the area, length, and width are in proportion. It is necessary to find HCC objects and cloud shadow objects such that the ratio of the horizontal coordinate length to the vertical coordinate length is in the range of about 1, and the ratio of the number of pixels of the cloud object to the number of pixels of the cloud shadow object is between 0.8 and 1.25.

[0079] The matching that simultaneously meets the above three quantitative conditions is used to calculate the offset for removing non-cloud objects in the low-confidence cloud objects, and then the final image is corrected according to the affine transformation relationship between the cloud layer and its shadow.

[0080] 4. Use the strategy of replacing the mask to improve the accuracy of the dataset.

[0081] Manually review the thumbnails. For the results with obvious errors, use the cloud-free marked mask or the LCC mask to replace. For example Figure 4 (a) is a schematic diagram of random point markings of the original image of a forest area containing some thin cloud areas, and it can be clearly observed that Figure 4 (c) There are large omissions in the cloud mask image obtained by the cloud detection dataset generation module. During the manual review process, this was quickly discovered based on the thumbnails of the remote sensing image and the mask image. Therefore, use Figure 4 the LCC mask in (b) to replace the final cloud mask image that failed to match successfully. Figure 4 (d) is a scene of desert saline-alkali land and some snow-capped mountains and snow fields. Similarly, use Figure 4 the cloud-free mask image in (e) to replace Figure 4 (f) There are some misdetection results. These two replacement strategies can improve the quality of the cloud detection dataset and the accuracy of the cloud mask.

[0082] After constructing the dataset in the above manner, select the data with an accuracy rate of 90% after manual evaluation for the entire scene image, and use the cloud-free images of the entire scene manually added in the highlighted complex surface scene as negative samples. In some images with a very large proportion of thin clouds, replace the final cloud mask with a more accurate LCC mask. Generally speaking, the final manually reviewed dataset can be used in the deep learning model. This cloud detection dataset contains 475 scenes of GF-1 ARD remote sensing images and pixel-level cloud mask labels with cloud pixels and clear sky pixels. Multiple negative samples are added in the highlighted complex ground object scenes such as snow-capped mountains and snow fields, desert saline-alkali land. In the dataset constructed by the method described in this embodiment, there are 173 images containing these scenes, accounting for 36.42% of the total data volume.

[0083] The main work of validating the dataset is to evaluate the accuracy of the cloud mask. Specifically, it is to compare the final cloud mask image with the real remote sensing image and give an accuracy coefficient reflecting the matching degree between them to determine the accuracy of the cloud detection process. Given the huge amount of remote sensing data, the number of pixels in a single scene of the image can reach 10 6 orders of magnitude, and it is impossible to perform pixel-by-pixel detection on the detected images. Therefore, it is a good method to adopt the way of randomly extracting verification points for the dataset images for accuracy detection and verification.

[0084] ArcGIS is a powerful geographic information system software, which is widely used in the collection, storage, analysis, and visualization of geospatial data. It helps users manage and analyze geospatial information in the form of maps and 3D scenes, supports from simple map making to complex geospatial analysis and modeling. At the same time, ArcGIS provides rich tools for data editing, spatial analysis, cartographic output, and geospatial information sharing. In this embodiment, the ArcMap desktop component in the ArcGIS platform is used to randomly sample points on the image to reduce the influence of human factors on the verification results. As Figure 5 shown is a case of random sampling points.

[0085] Create a new feature layer in ArcMap, create a rectangular planar feature according to the size of the image, and in the built-in Python code area of the software, randomly select points within a feature to generate a new random point feature layer, and this layer covers both Figure 5 the original image shown in (a) and Figure 5 the mask image shown in (b). Divide an image into small regions and conduct manual visual verification for the random points within each region.

[0086] Randomly select 47 images from 475 images in the cloud detection dataset after manual review. Among them, there are 40 images with automatically generated masks and 7 scenes with manually replaced masks. For the images containing cloud masks, use the above random sampling point strategy for manual visual verification and count the verification accuracy of the random sampling point strategy (accuracy refers to the percentage value of the number of correctly sampled points verified manually to the total number of randomly sampled points), as shown in the following table:

[0087]

[0088] Generally speaking, the accuracy is higher in most areas with green vegetation and areas with a large proportion of thick clouds. In areas with a large proportion of thin clouds and areas with less cloud cover that contain complex surface scenes with highlights, it is often necessary to manually replace the mask with an LCC mask image or a cloud-free mask image. After replacement, the accuracy of a single image without clouds can often reach over 99%, which also provides high-quality negative samples for subsequent deep learning. Among the 47 images verified, the overall accuracy of the automatically generated cloud mask scenes reached 90.29%, and the accuracy after manually replacing the mask reached 97.91%. In the actual dataset, the ratio of these two is also roughly consistent with the sampling ratio, and the overall accuracy of the total sampling of 91.43% also proves the credibility of the cloud detection dataset.

[0089] Example 2

[0090] This embodiment provides a remote sensing image cloud detection method based on deep learning. This method proposes a texture feature-enhanced deep learning cloud detection network (Texture-Enhanced Network for Cloud Detection, TENet) based on the texture spectral features, spatial distribution features, texture features, and encoding-decoding structure of clouds, and realizes the detection of clouds in remote sensing images through TENet.

[0091] This method includes:

[0092] 1. Design a texture feature-enhanced downsampling module (Texture-Enhanced Downsampling Module, TEDM). As Figure 6 shown, the processing process of this module is as follows:

[0093] 1) Perform a first-level wavelet transform operation on the input feature map F in and split it into four parts, including the low-frequency approximation component F LL , the horizontal detail component F LH , the vertical detail component F HL , and the diagonal detail component F HH , as shown in the following formula:

[0094] F LL , F LH , F HL , F HH = HaarWT(F in )

[0095] 2) Concatenate these four components:

[0096] F wt = Concat(F LL , F LH , F HL,F HH )

[0097] 3)F wt Feature extraction is performed through two convolutions. After obtaining the preliminary features in each convolution, they need to go through the BN layer for normalization to reduce internal covariate shift and accelerate the convergence of the model. Then, the ReLU activation function is used to perform a non-linear transformation on the normalized feature map to enhance the expressive power of the model. Among them, the feature maps obtained by the two convolutions are denoted as F C 、F DC :

[0098] F C =ReLU(BN(Conv 1×1 (F wt )))

[0099] F DC =ReLU(BN(Conv 3×3 (F C )))

[0100] 4) The CBAM attention module can be split into two modules, namely the channel attention mechanism (CAM) and the spatial attention mechanism (SAM), which enhance the channel features and spatial features respectively.

[0101] Let F represent the feature map input to the CBAM attention module. AvgPool(F) and MaxPool(F) represent global average pooling and max pooling operations respectively. MLP represents a multi-layer perceptron, and σ represents the Sigmoid activation function, which is used to output the channel attention weight. f 7×7 represents a convolution operation with a kernel size of 7×7. This process can be shown by the following formula:

[0102] Mc(F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F)))

[0103] Ms(F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0104] Input F DC into this CBAM attention module. The feature map Mc(F DC ) generated by the channel attention mechanism is multiplied by F DC to generate F out ′. The feature map Ms(F out ′) generated by the spatial attention mechanism is multiplied by F out ′ to generate the final feature map F out 。

[0105] 2. Based on the encoding-decoding structure, design the TENet model, as Figure 7 shown. The encoder part of this model includes a dual convolutional module and four cascaded TEDM modules, and the decoder structure includes four cascaded upsampling modules. Among them, the dual convolutional module and the first three TEDM modules are connected to the four upsampling modules in one-to-one correspondence.

[0106] The processing process of this TENet model is as follows:

[0107] 1) The input image to TENet is a GF-1 ARD remote sensing image of 320×320×4. In the encoder part, the input image first passes through a dual convolutional module to generate a feature map of 320×320×64. Among them, the dual convolutional module includes two convolutional layers with a convolutional kernel of 3×3, and there is a BN layer and a ReLU activation function after each convolutional layer.

[0108] 2) In the encoder part, the feature map output by the dual convolutional module passes through four TEDM modules in sequence.

[0109] In each TEDM, the feature map undergoes wavelet transform splicing and two convolutional layers with convolutional kernels of 1×1 and 3×3. Similarly, there is a BN layer and a ReLU activation function after each convolutional layer, and then the feature map passes through the CBAM module to enhance the channel and spatial features. In the first three TEDM channels, the output feature map is twice that of the input feature map, and the size of the image becomes half; the output feature map of the fourth TEDM only changes in size. The sizes of the feature maps output by the four TEDMs in the encoder part are 160×160×128, 80×80×256, 40×40×512, and 20×20×512 respectively.

[0110] 3) In the decoder part, with the help of the feature fusion mechanism of skip connection and transposed convolution, the low-level features in the encoder are fused with the high-level features in the decoder.

[0111] Specifically, it is processed through four upsampling modules in sequence. In each upsampling module, the input image is restored to the size of the input feature map of the same layer TEDM through a transposed convolution with a convolutional kernel of 2×2, and then spliced with the input feature map of the TEDM with skip connection, and the number of channels of the feature map is adjusted through a dual convolutional layer. Among them, the number of channels of the feature maps of the first three upsampling modules becomes half of that of the input of the same layer TEDM, and the number of channels of the output feature map of the last upsampling module remains 64. Specifically, the sizes of the feature maps output by the upsampling modules from bottom to top are 40×40×256, 80×80×128, 160×160×64, and 320×320×64 respectively.

[0112] 4) The feature map output by the decoder passes through a 1×1 convolutional layer and is normalized to 0-1 through sigmoid to obtain a predicted probability map of 320×320×2 containing clear sky classes and cloud classes.

[0113] In the TENet model, the loss function used is Focal Loss, as shown in the following formula:

[0114]

[0115] In the formula, y represents the cloud mask label, which is the binary prediction value output by the model and is used to indicate whether the sample belongs to the positive class. It is a continuous value between 0 and 1, representing the confidence of the model in predicting that the pixel is a cloud. The role of the focusing parameter γ is to adjust the attention degree of the model to the samples in the difficult-to-separate thin cloud region and the easy-to-separate thick cloud region. By increasing the γ value, the model can reduce the attention to the easy-to-separate thick cloud region and focus more on the samples in the difficult-to-separate thin cloud region, thereby encouraging the model to improve the classification accuracy of these samples. The class balance parameter α is used to handle the data volume imbalance problem of cloud classes. By assigning different weights to different classes, it enhances the attention of the model to the minority classes. Especially in binary classification problems, an increase in the α value will cause the model to be more severely punished when making classification errors in the minority classes, thereby improving the classification performance of the model on the minority classes.

[0116] Based on the remote sensing image cloud detection method provided in this embodiment, the model performance is compared using the public dataset and the dataset constructed and verified in Example 1 respectively.

[0117] Among them, the public dataset is the GF-1 WFV cloud and cloud shadow coverage dataset, which includes 108 WFV level-2A scenes obtained from the GF-1 remote sensing satellite, covering various land types such as forests, bare lands, ice / snow, water, wetlands, and urban areas. Among them, there are 14 complex surface scenes containing highlights such as snow-capped mountains and desert saline-alkali lands, accounting for 12.96% of the total data volume, which is similar to the proportion of these surfaces in natural scenes. These images are all generated after radiometric correction and geometric correction. It should be noted that the radiance values of the images are compressed into 10-bit remote sensing image pixel brightness values for convenient storage and training. Each band only records the gray-scale information of the ground objects and has no direct physical meaning itself. The size of each image is approximately 15,700×16,200 pixels, and the number of bands is 4. In addition, the reference masks for quantitative evaluation are obtained manually by experienced experts and include clouds, cloud shadows, clear skies, and no values. According to the practice of the comparison model, images with a ratio of 40:40:28 are randomly selected from this public dataset for training, validation, and testing, and cloud shadow pixels are also redefined as clear sky pixels. To facilitate training, the remote sensing images and mask images are cropped into 320×320 pixel blocks from left to right and top to bottom to meet the limited computing power resources.

[0118] The dataset constructed in Example 1 includes 475 GF-1 ARD remote sensing image scenes. The dataset covers various land types, including forests, bare lands, gobi deserts, sandy deserts, snow-covered areas, snow-capped mountains, cultivated lands, and urban areas. These images are all generated after radiometric correction and geometric correction. The size of each image is 6,863×6,863 pixels, and the number of bands is 4. Images with a ratio of 8:8:3 are randomly selected for training, validation, and testing, and the images are also cropped into 320×320 pixel blocks.

[0119] In addition, all deep learning experimental models for performance comparison are implemented using the PyTorch framework. All experiments are conducted on a server with the CentOS 7 operating system and equipped with two NVIDIA Tesla-P100 12G GPUs. All input images are 320×320 pixels in size. The Adam optimizer is used for training, the batch size is set to 8, the weight decay is 1e-4, and the learning rate is set to 1e-4. The "poly" learning rate adjustment strategy with a factor of 0.9 is adopted. It is trained for 100 epochs on the GF-1 WFV cloud and cloud shadow coverage dataset and 80 epochs on the GF-1 ARD cloud detection dataset. The ablation experiments and the comparison with classical experimental methods try to use the same settings. The batch size of some methods is set to 4 due to the number of parameters and network structure. According to experience, α in the Focal loss is set to 0.5 and γ is set to 2. Only one GPU is used in the testing phase.

[0120] The evaluation metrics selected are Overall Accuracy (OA), F1-score (F1), Mean Intersection over Union (MIoU), and Cohen's Kappa Coefficient (Ka) to evaluate the performance of all models. All quantitative scores are calculated using the pixels provided in the dataset as a reference. These quantitative metrics are defined as follows:

[0121]

[0122]

[0123] Among them, TP represents that cloud pixels are correctly identified as clouds, TN represents that non-cloud pixels are correctly identified as non-clouds, FP represents that non-cloud pixels are misidentified as cloud pixels, FN represents that cloud pixels are misidentified as non-cloud pixels, and P and N represent the number of cloud and clear-sky pixels in the remote sensing image, respectively.

[0124] To evaluate the efficiency of the comparison method, the model parameters and the number of floating-point operations (FLOPs) of 320×320 pixel blocks are calculated using the model to evaluate the complexity and computational complexity of the model. It should be noted that in the GF-1 ARD cloud detection dataset, although the accuracy of the dataset has been improved by using manual review and replacing the mask image, there are still minor errors in some areas, and qualitative analysis needs to be combined with the actual image.

[0125] I. Ablation Experiment

[0126] In the ablation experiment, to avoid interference from other factors, other structures in the network are kept unchanged. The first independent variable of the experiment is set to the types of attention mechanisms added, including the CAM and CBAM modules, and the second independent variable is set to different downsampling methods, including max pooling and wavelet transform splicing. Among them, the positions where CAM and CBAM are added are at the end of the downsampling module in the encoder.

[0127] The comparison results of the evaluation metrics of each model are shown in the following table:

[0128] Table 2

[0129]

[0130] As can be seen from Table 2, the F1, mIoU, and Ka metrics of the Baseline model are significantly lower than those of the other three models used for comparison, and the overall accuracy gap is also 5%. This shows that adding CAM to the downsampling module in the encoder has a significant effect on improving the accuracy of the Baseline model. In observing the cloud characteristics of remote sensing images and the dataset generation module, the most easily noticeable feature is the spectral feature of the clouds. The significant improvement in accuracy using CAM with enhanced channel features in the experiment also demonstrates the effectiveness and necessity of the channel attention mechanism.

[0131] In Figure 8 (c), it can be clearly observed that the U-Net model has more false detection cases in the bare land or saline-alkali land scenes of the highlighted desert surface, while Figure 8 the detection of the TENet model shown in (b) has good results and can detect small cloud clusters near the lower edge and the middle part in difficult-to-classify scenes.

[0132] Considering the importance and necessity of the channel attention mechanism (CAM) for the cloud detection task, the spatial attention mechanism (SAM) is not analyzed separately, but the results of CAM and CBAM are analyzed to compare the effectiveness of SAM. From the evaluation metrics, the evaluation metrics of CBAM with the spatial attention mechanism slightly decrease compared to CAM, but from Figure 9 the actual prediction results, the model with SAM has a lower false detection rate on the highlighted complex surface.

[0133] From Figure 9 (b), it can be seen that there are obvious false detections in the middle part of the prediction results of the U-Net model with a single CAM added, and there are slight false detections above. These two surfaces are highlighted saline-alkali lands, and it is difficult for a deep learning network using only the channel attention mechanism to distinguish them from the clouds. In Figure 9 (c), due to the addition of SAM, no obvious false detections are observed in the prediction results, which is also related to the learning of spatial features. The same similar situation can also be observed in Figure 10 , within the area marked by the red circle, Figure 10 the prediction results of the model with CBAM added shown in (d) have significantly fewer false detections than Figure 10 the prediction results of the model with only CAM added shown in (c).

[0134] In checking some of the actual prediction results, it is found that the SAM module also affects the detection accuracy in the thin cloud area. For example, in Figure 10In the area marked by the red circle in , the recognition area of the thin cloud region with the prediction result of adding CBAM is larger than that with the prediction result of only adding CAM. In the GF-1 ARD cloud detection dataset used, even though the accuracy reaches 91.42%, there are still deviation problems of false detection in the results of a small part of the highlighted surface cloud mask, and there are also deviation problems of missed detection in the results of the cloud mask in a small part of the thin cloud region. The model with SAM can learn more spatial features, and in some thin cloud regions and highlighted complex surface scenes, these deviation problems of false detection and missed detection are corrected.

[0135] Although among these models, the evaluation index of TENet is the highest, the gap is not large compared with the model that only adds CAM. Since in the GF-1 ARD cloud detection dataset, the reference of the evaluation index is not necessarily completely accurate. For example, in the result analysis with CAM and CBAM as independent variables, the model with CBAM can learn more features, but the evaluation index is relatively slightly lower. Therefore, when considering verifying the effectiveness of different downsampling methods for the second independent variable, more qualitative analysis needs to be carried out from the details of the actual prediction results.

[0136] In Figure 10 As can be observed from the comparison of the prediction results of the sunny side of the highlighted mountainous surface shown in , the prediction result of TENet with the downsampling method changed to wavelet transform splicing has significantly fewer small false detections (as shown in Figure 10 (b)), while in the area marked by the red circle in the prediction results using max pooling and convolution as the downsampling methods, the sunny side of many mountains is falsely detected as small clouds (as shown in Figure 10 (c) and Figure 10 (d)). In the snow mountain and cloud mixed scene shown in Figure 11 (a), by comparing the details of the prediction results, it is not difficult to find that in the green box of the TENet prediction result in Figure 11 (b), there are basically no false detection pixel points for the highlighted mountainous surface, while in the green boxes of the prediction results shown in Figure 11 (c) and Figure 11 (d), there are different degrees of false detections on the highlighted mountainous surface at the lower boundary of the thick cloud. According to the evaluation index and the results of the actual qualitative analysis, it is consistent that the model with the downsampling module of wavelet transform splicing that retains more texture features can significantly improve the detection accuracy of the highlighted complex surface.

[0137] II. Comparative Experiments

[0138] First, compare the evaluation indexes of several methods / models such as MFC, DABNet, DCNet, and HCDNet with the TENet provided in this embodiment on the GF-1 WFV cloud and cloud shadow coverage dataset, as shown in the following table:​​​​​​​

[0139] Table 3

[0140] Methods OA(%) F1(％) mIoU(%) Ka(%) Params(MB) FLOPs(G) MFC 94.42 91.30 88.06 87.20 / / DABNet 96.60 94.39 92.30 91.95 4.12 16.58 DCNet 96.86 94.97 92.97 92.68 7.12 32.42 HDCNet 96.97 95.13 93.21 92.94 37.14 27.66 TENet 97.61 93.42 92.39 91.96 11.3 45.03

[0141] As can be seen from Table 3, the model proposed in this embodiment is the highest in the index OA, reaching 97.61%. The other evaluation indicators are slightly lower than those of DABNet, DCNet, and HCDNet designed based on edge details. However, the model proposed in this embodiment has a higher accuracy in the full scene and can adapt to the highlighted surface environment.

[0142] To verify the effectiveness of the cloud detection dataset and TENet constructed in Example 1. The GF-1 ARD remote sensing image data reduces a large amount of manual annotation cost and increases the dataset generation efficiency. However, there are still small misdetections or missed detections in the mask image details of complex surface scenes such as thin cloud regions, snow-capped mountains, and desert saline-alkali lands. Training on the GF-1 ARD cloud detection dataset and comparing the evaluation indicators and actual prediction results of FCN, SegNet, and U-Net are shown in the following table:

[0143] Table 4

[0144]

[0145] As can be seen from Table 4, the TENet model proposed in this embodiment shows excellent performance in the four selected indicators. It is worth noting that although FCN shows commendable accuracy in the weakly supervised cloud detection task, compared with the TENet model, for example, in Figure 12 FCN still results in a higher misdetection rate on difficult-to-classify surfaces. In addition, the parameter count of the FCN model is several orders of magnitude larger than that of the TENet model. In practical applications, the training time and prediction time for each image are also much higher than those of other comparison models.

[0146] From Figure 12 it can be seen that in forest and urban surface scenes, the prediction results of all four models are very clear. However, due to its characteristics, FCN produces more blurred result boundaries, and this scene is one of the easiest to classify in practice because the difference between the surface and the cloud is particularly significant. In the network trained with the GF-1 ARD cloud detection dataset, the prediction results are usually good in such easily classifiable surface scenes.

[0147] In the grassland and cultivated land prediction results, generally speaking, all four models show very high prediction accuracy. However, in the thin cloud areas marked by red circles, there are obvious differences in the prediction results. Since the FCN model has the largest number of model parameters and computational complexity, its prediction results perform the best. SegNet and U-Net seem to learn less from the features of the thin cloud areas, resulting in the thin clouds in this area being basically undetected. TENet focuses more on texture, channel, and spatial features, and its performance in detecting thin cloud areas is better than that of SegNet and U-Net, and slightly lower than that of FCN.

[0148] In the desert saline-alkali land prediction results, the differences between the four models are more obvious. In the upper left corner of the image, FCN shows a large misclassification error, misdetecting the high-brightness desert area as a thin cloud area. In the lower right corner of the image, in the highlighted saline-alkali land area, the first three models all show a certain degree of misclassification. Although there are small prediction errors in the lower right corner of the TENet results, TENet avoids the misdetection of a larger area of the highlighted complex surface.

[0149] In the snow-capped mountain and snow-covered ground scene, the misdetection of FCN is mainly concentrated in the high-brightness mountain areas, where some continuous high-brightness snow-covered areas are misdetected as thick cloud areas. SegNet and U-Net show larger misclassification areas in this region. The error rate of the model proposed in this embodiment is very low.

[0150] In summary, the model proposed in this embodiment performs well in the highlighted complex surface area, and the misdetection rate is particularly low. From the performance of the prediction results, the GF-1 ARD cloud detection dataset used in the experiment can be regarded as a weakly supervised dataset. Since TENet has a unique attention mechanism for texture features, channel features, and spatial features and can adapt to scenarios with a small amount of noise in the data, the overall accuracy is not much different from that under full supervision learning, and the actual prediction effect is of reliable quality.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A deep learning cloud detection method that integrates object-oriented dynamic thresholds and feature enhancement, characterized in that, The method includes: S1. Obtain the original satellite remote sensing image, and for each image, generate mask thresholds using a dynamic threshold method based on the triangulation method, where the mask thresholds include a high-confidence cloud mask threshold, a low-confidence cloud mask threshold, and a shadow mask threshold; S2. According to the mask thresholds, generate a high-confidence cloud mask, a low-confidence cloud mask, and a shadow mask in the remote sensing image by combining an improved ACCA algorithm; S3. According to the generated high-confidence cloud mask, low-confidence cloud mask, and shadow mask, match the cloud objects and shadow objects in the image to obtain a cloud mask image; S4. For the incorrect masks in the cloud mask image, eliminate false detections through mask replacement to improve the accuracy of the cloud mask; S5. Construct all the finally generated cloud mask images into a remote sensing image training data set, and train a deep learning cloud detection network model with enhanced texture features through this data set; S6. After the model is trained, it is used for the prediction of clouds in satellite remote sensing images.

2. The method according to claim 1, characterized in that, In step S1, generating the mask thresholds using the dynamic threshold method based on the triangulation method includes: S11. Arbitrarily select an image from the original satellite remote sensing images, calculate the grayscale histogram of this image, set the origin in the grayscale histogram as point a, set the point with the highest frequency in the range of grayscale values from 125 to 254 in the grayscale histogram as point b, and connect points a and b to form a line L ab ; S12. Find the point with the maximum distance from L within the range of gray values from 125 to 254, record the gray value corresponding to the point with the maximum distance, and then generate a pixel reference threshold through the inverse operation of graying; ab And record the gray value corresponding to the point with the maximum distance, and then generate a pixel reference threshold through the inverse operation of graying; S13. Generate a high-confidence cloud mask threshold, a low-confidence cloud mask threshold, and a shadow mask threshold according to the reference threshold.

3. The method according to claim 1, characterized in that, In step S2, for any pixel in the image, the following judgment process is performed: S21. Obtain the R-channel value B3 of pixel i i , the G-channel value B2 i and the NIR-channel value B4 i ; S22. If B4 i <T S , and B4 i / B3 i >1.5, then determine that the pixel is a shadow and generate a shadow mask; where T S represents the shadow mask threshold. S23. If B3 i > T H , then determine whether B4 i / B3 i is less than 2.35; if B4 i / B3 i ≥ 2.35, then pixel i is clear sky; if B4 i / B3 i < 2.35, and B4 i / B2 i < 2.16, then pixel i is high-confidence cloud, and generate a high-confidence cloud mask; if B4 i / B3 i < 2.35, and B4 i / B2 i ≥ 2.16, then pixel i is clear sky; where T H represents the high-confidence cloud mask threshold. S24. If B3 i ≤T H , then determine whether B3 i is greater than the low confidence cloud mask threshold T L ; if B3 i ≤T L , then pixel i is clear sky; if B3 i >T L , then go to step S25; S25. If B4 i / B3 i ≥ 2.35, then pixel i is clear sky; if B4 i / B3 i <2.35, and B4 i / B2 i <2.16, then pixel i is low-confidence cloud, and a low-confidence cloud mask is generated; if B4 i / B3 i <2.35, and B4 i / B2 i ≥ 2.16, then pixel i is clear sky.

4. The method according to claim 1, wherein In step S3, matching the cloud objects and shadow objects according to the affine invariant relationship between clouds and shadows includes: S31. Screen cloud objects in the high-confidence cloud object set to ensure that the number of pixels of the high-confidence cloud object meets the requirements, and the proportion of the area occupied by the high-confidence cloud object in the 8-neighborhood is less than one-eighth; S32. For each screened high-confidence cloud object, with its center as the center point, search for shadow objects in the four quadrants of 0° - 90°, 90° - 180°, 180° - 270°, and 270° - 360°; S33. When the ratio of the horizontal length to the vertical length between the high-confidence cloud object and the shadow object is approximately equal to 1, and the ratio of the number of pixels between the high-confidence cloud object and the shadow object is between 0.8 and 1.25, the high-confidence cloud object and the shadow object are successfully matched; S34. Calculate the movement offset between the successfully matched high-confidence cloud object and the shadow object, and based on this movement offset, screen out the non-cloud objects in the low-confidence cloud objects, thereby generating a cloud mask image.

5. The method according to claim 1, wherein Step S4 includes, when there is a missed detection in the generated cloud mask image, using the low-confidence cloud mask to replace the cloud mask image with the missed detection; When there is a false detection in the generated cloud mask image, use the cloud-free mask to replace the false detection part in the image.

6. The method according to claim 1, characterized in that The deep learning cloud detection network model with enhanced texture features is used to process the input satellite remote sensing image, and output a prediction probability map including clear sky categories and cloud categories; where the deep learning cloud detection network model with enhanced texture features includes an encoder and a decoder; The encoder includes a dual convolutional module and multiple cascaded texture feature enhanced downsampling modules, where the multiple cascaded texture feature enhanced downsampling modules are connected to the output end of the dual convolutional module; the decoder includes multiple cascaded upsampling modules; The modules of the first N-1 layers in the encoder are skip-connected to the upsampling modules of the corresponding layers in the decoder.

7. The method according to claim 6, wherein The texture feature enhancement downsampling module includes a first-level wavelet transform layer, a splicing layer, a first convolutional layer, a second convolutional layer, and a CBAM attention layer connected in sequence. At the output ends of the first convolutional layer and the second convolutional layer, a batch normalization and activation function layer are both connected.

Images