Remote sensing image cloud and cloud shadow detection method based on context information fusion

Abstract

The application provides a remote sensing image cloud and cloud shadow detection method based on context information fusion, and relates to the fields of deep learning and remote sensing image processing. The remote sensing image cloud and cloud shadow detection method based on context information fusion designs a remote sensing image cloud and cloud shadow detection method based on context information fusion. The application designs a novel residual module (ResBlock-cloud) to capture global and local features and prevent network degradation. A global context fusion module (GCF) is designed to fuse global context information at different levels through dense skip connections and guide feature fusion with decoder path features. Finally, a multi-scale context fusion module (MCF) is designed to extract the multi-scale context relationship between clouds and cloud shadows. The application can effectively detect small-area cloud blocks and thin clouds and significantly improve the accuracy of cloud shadow detection.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and remote sensing image processing technology, specifically to a method for detecting clouds and cloud shadows in remote sensing images based on contextual information fusion. Background Technology

[0002] With the significant success of deep convolutional neural networks (DCNNs) in image classification and recognition, this technology has also been introduced into the field of cloud detection in remote sensing images. Early cloud detection methods based on DCNNs, such as those by Wohlfarth and Varshney, predicted the category of each pixel by testing surrounding image patches. These methods were not only inefficient but also noisy. To further improve spatial coherence, Xie et al. first performed superpixel segmentation on the pixels of the remote sensing image before performing cloud detection. However, the accuracy limitation of superpixel clustering algorithms is based on the performance of superpixel cloud detection methods. Zi et al. further optimized the cloud detection results using fully connected conditional random fields. However, this post-processing optimization is very time-consuming when processing large-size remote sensing images.

[0003] Inspired by the excellent performance of Fully Convolutional Networks (FCNs) in image semantic segmentation, cloud detection methods based on FCNs have achieved significant results. Long et al. used the FCN-8 method to detect clouds on Landsat 8 satellite imagery. However, the FCN-8 method cannot accurately detect the edge details of clouds, resulting in significant information loss. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a remote sensing image cloud and cloud shadow detection method based on contextual information fusion, achieving accurate detection of clouds and cloud shadows in Gaofen-1 WFV optical remote sensing images. A novel residual module (ResBlock-cloud) is designed to capture global and local features and prevent network degradation. A global context fusion module (GCF) is designed to fuse global contextual information from different levels through dense skip connections and guide it to feature fusion with decoder path features. Finally, a multi-scale context fusion module (MCF) is designed to extract multi-scale contextual relationships between clouds and cloud shadows. This invention can effectively detect small cloud patches and thin clouds and significantly improve the accuracy of cloud shadow detection.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: a remote sensing image cloud and cloud shadow detection method based on contextual information fusion, comprising:

[0008] The remote sensing image dataset is preprocessed to create labels corresponding to the remote sensing images, so that the labels form mask images corresponding to the remote sensing images. The dataset is divided into training set, validation set and test set according to the proportion. The training set and validation set are input into the supervised network for training and validation.

[0009] When the Epoch is 110-120, the loss function reaches a low value and remains slightly fluctuating. All indicators of the validation set reach their highest values. Save the target network model and stop training to obtain the cloud and cloud shadow detection model.

[0010] The remote sensing images from the test set are input into the saved target network model for cloud and cloud shadow detection.

[0011] Preferably, the target network model introduces a residual module (ResBlock-cloud) to capture global and local features and prevent network degradation; a global context fusion module (GCF) is designed to fuse global context information at different levels through dense skip connections and guide it to perform feature fusion with decoder path features; and a multi-scale context fusion module (MCF) is designed to extract multi-scale contextual relationships between clouds and cloud shadows.

[0012] Preferably, the experimental parameters for training the target network model include: a learning rate of α = 0.0003, decay rates of β1 = 0.9 and β2 = 0.999, and a batch size of 8 (an empirical value).

[0013] Preferably, the residual module (ResBlock-cloud) includes a left branch and a right branch. The left branch is responsible for extracting more features from the input remote sensing image and is called the feature extraction branch. The right branch is a skip connection branch that is responsible for saving context information and is called the context branch.

[0014] Preferably, the Global Context Fusion (GCF) module establishes a new skip connection by fusing feature maps with feature maps from all higher stages; by inserting multiple GCF modules between the encoder and decoder, global context information from higher levels is guided to different levels, thereby enhancing the CIFNet network model's ability to guide and fuse global context information without introducing irrelevant clutter and semantic gaps.

[0015] Preferably, the multi-scale context fusion module (MCF) is used to effectively extract and fuse multi-scale context information. The MCF module is designed with four cascaded branches, including one pooling branch and three convolutional branches. Multiple skip connections are introduced between each convolutional branch to share feature information at different scales. Compared with a simple parallel structure, the cascaded structure can obtain a larger receptive field, maximizing the extraction of multi-scale context information.

[0016] Preferably, the detection method further includes quantitatively evaluating the detection performance of clouds and cloud shadows, using accuracy P. Precision Accuracy A Accuracy Recall rate R Recall F1 score, average intersection ratio M MIoU Five quantitative evaluation indicators for semantic segmentation.

[0017] Preferably, the accuracy includes:

[0018] The accuracy rate includes:

[0019] The recall rate includes:

[0020] The F1 value includes:

[0021] The average intersection-union ratio includes:

[0022] Where S represents the range of correctly detected cloud pixels; W represents the range of correctly detected non-cloud pixels; R represents the range of cloud pixels that are misdetected as non-cloud pixels; and E represents the range of non-cloud pixels that are misdetected as cloud pixels. When calculating the various evaluation metrics for cloud pixels, cloud shadow pixels are classified as underlying surface pixels; when calculating the various evaluation metrics for cloud shadow pixels, cloud pixels are classified as underlying surface pixels.

[0023] (III) Beneficial Effects

[0024] This invention presents a cloud and cloud shadow detection method based on contextual information fusion in remote sensing images, achieving accurate detection of clouds and cloud shadows in Gaofen-1 WFV optical remote sensing images. A novel residual module (ResBlock-cloud) is designed to capture global and local features and prevent network degradation. A global context fusion module (GCF) is designed to fuse global contextual information from different levels through dense skip connections and guide it to feature fusion with decoder path features. Finally, a multi-scale context fusion module (MCF) is designed to extract multi-scale contextual relationships between clouds and cloud shadows. This invention can effectively detect small cloud patches and thin clouds, and significantly improves the accuracy of cloud shadow detection. Attached Figure Description

[0025] Figure 1 This is a flowchart of the remote sensing image cloud and cloud shadow detection method based on context information fusion according to the present invention;

[0026] Figure 2 This is a schematic diagram of the residual module (ResBlock-cloud) in the embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the Global Context Fusion (GCF) module in an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of the multi-scale context fusion module (MCF) in an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram of the CIFNet cloud and cloud shadow detection network model in an embodiment of the present invention;

[0030] Figure 6 This is a line graph showing the overall accuracy of the ablation experiment according to an embodiment of the present invention.

[0031] Figure 7 This is a visual comparison diagram of the cloud detection results of Gaofen-1 WFV remote sensing images using four methods in this embodiment of the invention.

[0032] Figure 8 Visual comparison of cloud shadow detection results in Gaofen-1 WFV remote sensing images using the method of this invention and the MU-Net method. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] Example

[0035] refer to Figure 1 As shown, Figure 1 Here is a flowchart of a CIFNet remote sensing image cloud and cloud shadow detection method according to an embodiment, which includes the following steps:

[0036] Training phase: Step 1, Preprocess the remote sensing images;

[0037] The selected remote sensing images are from the Gaofen-1 WFV cloud and cloud shadow dataset. WFV is one of the most critical components of the Gaofen-1 satellite, comprising four integrated cameras with a spatial resolution of 16 meters and a temporal resolution of 4 days. Each camera has four multispectral bands, with an observation swath width of 800 kilometers. This dataset contains 108 remote sensing images, essentially covering different types of underlying surfaces.

[0038] Typically, a large number of parameters are needed during model training. To ensure these parameters function correctly, a large amount of data is required for training. However, in reality, remote sensing image datasets contain relatively little data and are large in size. Therefore, data augmentation processing such as rotation, brightness transformation, noise injection, saturation transformation, color transformation, and cropping is required for remote sensing images. The augmented remote sensing images are cropped into 15,552 images of size 256×256, with 80% used as the training set and 20% as the test set.

[0039] Step 2: Constructing a cloud and cloud shadow detection network model. This invention proposes a convolutional neural network model based on contextual information fusion (CIFNet), which is mainly an improvement on the encoder-decoder architecture. It consists of four main parts: a residual module (ResBlock-cloud), a global context fusion module (GCF), a multi-scale context fusion module (MCF), and a decoder module (DecoderBlock). The multi-scale context fusion module is placed at the end of the encoder part to extract and fuse multi-scale contextual information features. Multiple global context fusion modules are inserted between the encoder and decoder to fuse global contextual information features at different levels and guide them to perform feature fusion with the decoder path features.

[0040] The residual module (ResBlock-cloud) can significantly reduce computational complexity in deeper convolutional neural networks and effectively solve the gradient vanishing problem. Therefore, the residual module is used as a feature extractor in the encoder path to extract global contextual information, and its structure is as follows: Figure 2 As shown.

[0041] Each residual module consists of two branches, left and right. The left branch is responsible for extracting more features from the input remote sensing image and is called the feature extraction branch. The right branch is a skip connection branch, which is responsible for preserving context information and is called the context branch. Extensive experimental results show that as the number of network layers increases, using skip connections can accelerate the convergence speed of the network model and suppress the performance degradation of the network model [72-74]. In the feature extraction branch, adding a 1×1 convolution before and after the 3×3 convolution can improve the efficiency of network computation. In the context branch, two 1×1 convolutions are used. Finally, the outputs of the two branches are summed. The features output from each convolutional layer are batch normalized and the LeakyReLU activation function is used to output the features. The output Pt of the residual module is...

[0042] Pt=F(Pt-1,S,N1,N2)

[0043] In the formula, t = 1, 2, 3, ..., T0, T0 represents the number of residual modules; Pt-1 represents the output of the previous residual module; F() represents the residual function of the residual module; S represents the stride size in the convolution; N1 and N2 represent the number of convolution kernels used in the feature extraction branch and the context branch, respectively.

[0044] The residual module brings two important advantages to the CIFNet network model. First, it facilitates backpropagation of errors, preventing the vanishing gradient problem in deep networks and thus preventing network degradation. Therefore, it enables faster fusion of deep network features. Second, the module can output the complete low-level semantic information from the original input, thereby extracting more global contextual information and avoiding the neglect of this low-level semantic information after consecutive convolutional operations. Therefore, the residual module helps the convolutional neural network preserve and transmit complete low-level semantic information.

[0045] The encoder path in a convolutional neural network can learn global contextual information from remote sensing imagery, including cloud and cloud shadow category features and surrounding environmental features. However, as the number of network layers increases, the global contextual information may gradually weaken. The original U-Net network model addresses this problem by introducing multiple skip connections. However, the introduction of skip connections introduces irrelevant clutter and semantic gaps. Therefore, this invention introduces a Global Context Fusion (GCF) module to solve these problems, with the structure as follows: Figure 3 As shown.

[0046] In the global context fusion module, a new skip connection is established by fusing the feature map of the current stage with the feature maps of all higher stages. Figure 3This paper demonstrates the specific structure of the first global context fusion module in the CIFNet network model. First, multiple 1×1 convolutions map the features from all stages to the same dimensional space, resulting in features F1, F2, F3, F4, and F5. Next, features F2, F3, F4, and F5 are upsampled to the same size as feature F1, and then fused. The fused features are then input into three parallel dilated convolution branches with dilation rates of 1, 2, and 4, respectively, to guide global context information. The dilated convolutions increase or decrease the receptive field by changing the dilation rate without altering the size of the convolution kernel or feature map. Finally, a 1×1 convolution is performed to obtain the final feature map. By inserting multiple global context fusion modules between the encoder and decoder, more global context information from higher layers can be guided to different levels, enhancing the CIFNet network model's ability to extract and guide global context information without introducing irrelevant clutter or semantic gaps.

[0047] Clouds are typically surrounded by cloud shadows, which share similar boundaries and geometric shapes. Therefore, a strong contextual relationship exists between clouds and their shadows. Capturing more multi-scale contextual information can improve the resolution of remote sensing images, thereby increasing the detection accuracy of clouds and cloud shadows. Inspired by hollow spatial pyramid pooling and the PSPNet network model, this invention proposes a multi-scale context fusion module (MCF) to effectively extract and fuse multi-scale contextual information. Its structure is as follows: Figure 4 As shown.

[0048] This invention designs four cascaded branches in the multi-scale context fusion module, including one pooling branch and three convolutional branches. Multiple skip connections are introduced between each convolutional branch to share feature information at different scales. Compared to a simple parallel structure, the cascaded structure can obtain a larger receptive field, maximizing the extraction of multi-scale context information. Each convolutional branch consists of a 1×1 convolution and two 3×3 dilated convolutions, with dilation rates of (3, 5), (4, 6), and (5, 7), respectively. By gradually increasing the dilation rate, more comprehensive context information can be obtained. Since the multi-scale context fusion module uses feature fusion operations instead of feature addition operations, the channel dimension will increase rapidly. Therefore, 1×1 convolutions are used to reduce the channel dimension and decrease computational cost.

[0049] The input to each convolutional branch is the output of all previous convolutional branches. Each convolutional branch is fused with all previous convolutional branches along the channel dimension to improve feature reusability. Furthermore, to preserve the coarse-grained information of the initial input, global average pooling and 1×1 convolutions are used to extract global features, and bilinear interpolation is used to upsample the features to obtain the same size as the initial input. This is then fused with the outputs of the three convolutional branches to obtain a new feature map. Finally, a 1×1 convolution is performed to obtain the final output feature.

[0050] The structure of the context-information fusion-based convolutional neural network model (CIFNet) is as follows: Figure 5 As shown, the CIFNet network model is an improvement on the encoder-decoder architecture, mainly composed of four parts: the Resblock-cloud module, the Global Context Fusion (GCF) module, the Multi-Scale Context Fusion (MCF) module, and the DecoderBlock module. The MCF module is placed at the end of the encoder section to extract and fuse multi-scale contextual information features. Multiple GCF modules are inserted between the encoder and decoder to fuse global contextual information features from different levels and guide them to perform feature fusion with the decoder path features.

[0051] The original remote sensing image is input into the encoder path. Global context information is extracted using the residual module, and then multi-scale context fusion is used to extract and fuse this multi-scale context information. In the decoder path, features are recovered using the decoder module, which consists of one 1×1 convolution and two 3×3 convolutions. Bilinear interpolation is used to upsample the features to obtain detection results of the same size as the initial remote sensing image. Simultaneously, the global context fusion module guides the fusion of global context information with the decoder path features. Finally, a 1×1 convolution is applied to obtain the final detection result.

[0052] Since this invention aims to solve a three-class classification problem, the Softmax function is used to determine the class probability of each pixel in the feature, and the Loss value is calculated using the Softmax function. The Loss function L(9) is defined as follows:

[0053]

[0054] Among them, f u θ represents the category label, which has Q possible values. This method studies three-class semantic segmentation, so Q = 3; θ represents the model parameters. It is θ q transpose of; g uThe input image pixel observation vector G = {g1, g2, g3, ..., g...} represents the pixel observation vector of the input image. U The elements in}; U represents the number of pixels in each image; 1{g} is an explicit function.

[0055] Parameter optimization was performed using the Adam optimizer, with decay rates set to β1 = 0.9 and β2 = 0.999. The initial learning rate was set to α = 0.0003, and the batch size was set to 8. The loss function converged after 110-120 epochs. Three cloud and cloud shadow detection methods were used as comparison methods: the improved SegNet method (MSegNet), the DeepLabV3+ method, and the improved U-Net method (MU-Net). All experiments were conducted on a Windows 10 operating system, with an Intel Core i7 9700K CPU, 32GB of RAM, and an NVIDIA GeForce RTX 2080 Ti (11GB) GPU.

[0056] To quantitatively evaluate the detection performance of clouds and cloud shadows, the accuracy P was used. Precision Accuracy A Accuracy Recall rate R Recall F1 value, average intersection ratio (M) MIoU Five quantitative evaluation indicators for semantic segmentation.

[0057] Specifically, the accuracy includes:

[0058] The accuracy rate includes:

[0059] The recall rate includes:

[0060] The F1 value includes:

[0061] The average intersection-union ratio includes:

[0062] Where S represents the range of correctly detected cloud pixels; W represents the range of correctly detected non-cloud pixels; R represents the range of cloud pixels that are misdetected as non-cloud pixels; and E represents the range of non-cloud pixels that are misdetected as cloud pixels. When calculating the various evaluation metrics for cloud pixels, cloud shadow pixels are classified as underlying surface pixels; when calculating the various evaluation metrics for cloud shadow pixels, cloud pixels are classified as underlying surface pixels.

[0063] Ablation experiments were incorporated to investigate the impact of the global context fusion module and the multi-scale context fusion module on the performance of the CIFNet network model. The ablation experiments included removing one of the modules and removing all modules simultaneously. These experiments used the same dataset and set the same experimental parameters, selecting accuracy as the evaluation metric. The relationship between accuracy and the number of iterations (epochs) is shown below. Figure 6 As shown in the figure, both the global context fusion module and the multi-scale context fusion module in the CIFNet network model contribute to improving the accuracy of cloud detection.

[0064] Figure 7 This paper presents a visual comparison of cloud detection results of Gaofen-1 WFV remote sensing images under several typical conditions between our method and three comparative methods. Figure 7 The original image in (a) contains thick cloud patches and large cloud areas; Figure 7 The original images in (b) and (c) contain translucent thin clouds and barren underlying surfaces; Figure 7 The original images in (d), (e) and (f) contain ice and snow underlying surfaces, land-sea underlying surfaces, and ocean underlying surfaces, respectively; Figure 7 The original image in (g) contains small cloud patches and fragmented clouds.

[0065] Depend on Figure 7 It can be seen that the MSegNet method misses a lot of detailed information during cloud detection, especially in icy and snowy scenes. This method fails to correctly detect clouds from cloud-like objects and is prone to over-segmentation. The DeepLabV3+ method has better detection performance than MSegNet, but its cloud boundary detection capability is weaker than MU-Net and the method proposed in this invention. The MU-Net method utilizes high-level features of target orientation to obtain accurate pixel-level cloud detection results. However, due to its limited ability to fuse contextual information, the overall performance of the MU-Net method is slightly weaker than that of the CIFNet method. The CIFNet method proposed in this invention can better extract features from cloud pixels, produce better cloud boundaries, and preserve the overall shape of the cloud to the greatest extent, obtaining detection results closest to the cloud label.

[0066] Qualitative analysis of cloud detection results shows that the CIFNet method outperforms other comparative methods, extracting and fusing more contextual information. Even in complex environments such as snowy surfaces, fragmented clouds, and thin clouds, it still achieves good detection results. Experimental results also demonstrate the importance and effectiveness of the residual module, global context fusion module, and multi-scale context fusion module used in the CIFNet method.

[0067] Figure 8This paper presents a visual comparison of the cloud shadow detection results of our method and the MU-Net method on Gaofen-1 WFV remote sensing imagery. Figure 8 It can be seen that the MU-Net method performs poorly in cloud shadow detection, missing a large amount of cloud shadow details. The CIFNet method achieves better cloud shadow detection results, significantly outperforming the MU-Net method. It preserves more complete cloud shadow boundaries, performs better in detecting thin cloud shadows, and is closer to the cloud shadow label.

[0068] Table 1 shows the quantitative performance analysis of our method compared to three other methods, using precision (PPrecision), accuracy (AAccuracy), recall (RRecall), F1 score, and mean intersection-union ratio (MMIoU) as evaluation metrics. Table 4-1 shows the average values ​​of the quantitative evaluation metrics obtained by the four methods for cloud detection on Gaofen-1 WFV remote sensing images. As can be seen from Table 1, the CIFNet method exhibits the best overall performance, showing significant advantages over other methods and making it more suitable for cloud detection. Its precision, accuracy, recall, F1 score, and MMIoU reach 94.47%, 97.95%, 93.84%, 0.9415, and 0.8753, respectively. In the feature extraction process, the CIFNet method uses residual modules, multi-scale contextual information fusion technology, global contextual information fusion technology, and dense skip connections to obtain more semantic and spatial information.

[0069] Table 1

[0070]

[0071] Table 2 shows the average values ​​of quantitative evaluation indicators obtained by the method of this invention and the MU-Net method for cloud shadow detection on Gaofen-1 WFV remote sensing images. As can be seen from Table 2, the CIFNet method outperforms the MU-Net method in all evaluation indicators, demonstrating excellent cloud shadow detection performance. Its precision, accuracy, recall, F1 score, and mean intersection-over-union ratio reach 86.31%, 96.46%, 82.74%, 0.8449, and 0.7335, respectively. Compared with the MU-Net method, the CIFNet method introduces a residual module, a global context fusion module, and a multi-scale context fusion module, effectively fusing global and multi-scale context information, resulting in significantly better cloud shadow detection performance than the MU-Net method.

[0072] Table 2

[0073]

[0074] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for detecting clouds and cloud shadows in remote sensing images based on contextual information fusion, characterized in that, The detection method includes: The remote sensing image dataset is preprocessed to create labels corresponding to the remote sensing images, so that the labels form mask images corresponding to the remote sensing images. The dataset is divided into training set, validation set and test set according to the proportion. The training set and validation set are input into the supervised network for training and validation. When the Epoch is 110-120, the loss function reaches a low value and remains slightly fluctuating. All indicators of the validation set reach their highest values. Save the target network model and stop training to obtain the cloud and cloud shadow detection model. The remote sensing images in the test set are input into the saved target network model for cloud and cloud shadow detection; The target network model introduces a residual module to capture global and local features and prevent network degradation; a global context fusion module is designed to fuse global context information at different levels through dense skip connections and guide it to perform feature fusion with decoder path features; a multi-scale context fusion module is designed to extract multi-scale context relationships between clouds and cloud shadows. The residual module includes a left branch and a right branch. The left branch is responsible for extracting more features from the input remote sensing image and is called the feature extraction branch. The right branch is a skip connection branch that is responsible for saving context information and is called the context branch. The global context fusion module establishes a new skip connection by fusing feature maps with feature maps from all higher stages. By inserting multiple global context fusion modules between the encoder and decoder, global context information from higher levels is guided to different levels, enhancing the network model's ability to guide and fuse global context information without introducing irrelevant clutter or semantic gaps. The multi-scale context fusion module is used to effectively extract and fuse multi-scale context information. The multi-scale context fusion module is designed with four cascaded branches, including one pooling branch and three convolutional branches. Multiple skip connections are introduced between each convolutional branch to share feature information at different scales.

2. The remote sensing image cloud and cloud shadow detection method based on contextual information fusion according to claim 1, characterized in that: The experimental parameters for training the target network model include: a learning rate of α = 0.0003, decay rates of β1 = 0.9 and β2 = 0.999, and a batch size of 8 (an empirical value).

3. The remote sensing image cloud and cloud shadow detection method based on contextual information fusion according to claim 1, characterized in that: The detection method also includes quantitatively evaluating the detection performance of clouds and cloud shadows, using accuracy P. Precision Accuracy A Accuracy Recall rate R Recall F1 score, average intersection ratio M MIoU Five quantitative evaluation indicators for semantic segmentation.

4. The remote sensing image cloud and cloud shadow detection method based on contextual information fusion according to claim 3, characterized in that: The accuracy includes: The accuracy rate includes: The recall rate includes: The F1 value includes: The average intersection-union ratio includes: Where S represents the range of correctly detected cloud pixels; W represents the range of correctly detected non-cloud pixels; R represents the range of cloud pixels that are misdetected as non-cloud pixels; E represents the range of non-cloud pixels that are misdetected as cloud pixels; when calculating the various evaluation indicators of cloud pixels, cloud shadow pixels are classified as underlying surface pixels; when calculating the various evaluation indicators of cloud shadow pixels, cloud pixels are classified as underlying surface pixels.

Images