Multi-scale feature fusion cloud detection method based on low-rank decomposition

Abstract

The invention discloses a multi-scale feature fusion cloud detection method based on low-rank decomposition. The method comprises the following steps: 1, obtaining an optical image needing to be detected; 2, inputting the optical image into a trained multi-scale feature fusion cloud detection model based on low-rank decomposition to obtain a detection result of the optical image; wherein the trained multi-scale feature fusion cloud detection model based on low-rank decomposition is trained by inputting a historical optical image into the multi-scale feature fusion cloud detection model based on low-rank decomposition; the multi-scale feature fusion cloud detection model based on low-rank decomposition is obtained by inserting a weight low-rank decomposition module into the multi-scale feature fusion cloud detection model; and designing a total loss function to restrain the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition. According to the method, the optimal detection precision is obtained in terms of objective evaluation indexes, and the image detection effect can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a multi-scale feature fusion cloud detection method based on low-rank decomposition. Background Art

[0002] With the development of remote sensing technology, especially the widespread application of deep learning technology, model-based cloud detection and removal methods have made significant progress. Large deep learning models have excellent performance in cloud detection and removal tasks due to their powerful feature extraction capabilities, but they also bring problems such as high computing resource consumption and difficulty in model deployment. Therefore, the development of lightweight models has become a current research hotspot. Lightweight models aim to reduce the number of model parameters and computational complexity so that they can run quickly and efficiently in resource-constrained environments, while maintaining the accuracy of cloud detection and cloud removal as much as possible to meet the needs of practical applications. Researching and developing efficient, accurate and lightweight cloud detection and removal technologies is of great significance to improving the application value of remote sensing images and promoting the application of remote sensing data in a wider range of fields. It can not only improve the accuracy and timeliness of surface observations, but also provide a more reliable basis for the subsequent processing and analysis of remote sensing data.

[0003] In the early stage of remote sensing cloud detection, graphics technology was mainly used to analyze pixel information in images to identify and distinguish between cloud and non-cloud areas. With the rapid development of deep learning technology, remote sensing cloud detection technology has been combined with deep learning methods to produce many cloud detection models. The current cloud detection network still has some shortcomings in performance, especially in terms of accuracy and real-time performance, which has not reached the ideal level. In addition, these networks often have a large number of parameters, resulting in huge computational complexity, which not only increases the difficulty of training and deployment, but also puts higher requirements on computing resources. These challenges limit the scope of application of cloud detection technology, especially in resource-constrained environments, affecting its potential for widespread deployment and practical application. Therefore, optimizing the network structure to reduce the amount of computation while improving the image detection effect has become a key issue that needs to be addressed in current cloud detection research. Summary of the invention

[0004] The present invention provides a multi-scale feature fusion image detection method based on low-rank decomposition, comprising the following steps:

[0005] Step 1: Obtain the optical image to be detected;

[0006] Step 2: inputting the optical image into the trained multi-scale feature fusion cloud detection model based on low-rank decomposition to obtain the detection result of the optical image;

[0007] Among them, the trained multi-scale feature fusion cloud detection model based on low rank decomposition is trained by inputting historical optical images into the multi-scale feature fusion cloud detection model based on low rank decomposition; the multi-scale feature fusion cloud detection model based on low rank decomposition is obtained by inserting a weighted low rank decomposition module into the multi-scale feature fusion cloud detection model; and the total loss function is designed to constrain the performance of the multi-scale feature fusion cloud detection model based on low rank decomposition.

[0008] Preferably, the multi-scale feature fusion cloud detection model based on low-rank decomposition includes: a multi-scale cloud mask network based on low-rank decomposition and a cloud difference fusion network, and the multi-scale cloud mask network based on low-rank decomposition and the cloud difference fusion network have encoders and decoders inside.

[0009] Preferably, the trained low-rank decomposition-based multi-scale feature fusion cloud detection model is trained by inputting historical optical images into the low-rank decomposition-based multi-scale feature fusion cloud detection model, including:

[0010] The acquired historical optical image is input into the multi-scale feature fusion cloud detection model based on low-rank decomposition for training, the multi-scale cloud mask network based on low-rank decomposition extracts multiple features from the input historical optical image through the encoder inside it, and transmits the multiple features to the decoder inside it after splicing; the decoder fuses the multiple features to form a multi-scale cloud mask;

[0011] The cloud difference fusion network based on low-rank decomposition has multiple inner layers. The cloud difference fusion network is based on a residual structure and captures subtle differences between multiple inner layer features through differential operations to form a difference cloud mask. Afterwards, the difference cloud mask is fused with the multi-scale cloud mask to form an overall cloud mask. The overall cloud mask is used as a multi-scale cloud mask network of the multi-scale feature fusion cloud detection model.

[0012] Afterwards, a weighted low-rank decomposition module is inserted into the encoder-decoder of the multi-scale feature fusion cloud detection model with the multi-scale cloud mask network as the overall cloud mask to obtain a trained multi-scale feature fusion cloud detection model based on low-rank decomposition.

[0013] Preferably, the multi-scale cloud mask based on low-rank decomposition is a multi-scale feature fusion cloud detection model in which each independent convolution block on the encoder-decoder in the overall cloud mask is inserted into a weighted low-rank decomposition module to obtain a trained multi-scale feature fusion cloud detection model based on low-rank decomposition, comprising the following steps:

[0014] Step 4.1: Assume that the convolution weight matrix of each independent convolution block in the new encoder and the new decoder in the multi-scale cloud mask network of the multi-scale feature fusion cloud detection model is W; the convolution weight matrix W is reduced to a two-dimensional matrix by rearranging the dimensions, and the two-dimensional matrix is ​​decomposed into a matrix W=USV by singular value decomposition, where W is the initial matrix, U is the left singular matrix, S is the singular value matrix, and V is the right singular matrix;

[0015] Step 4.2: Based on the singular value matrix, obtain multiple singular values, and form an approximate weight matrix with the singular values. The rank of the singular value matrix S is r. Slice from the first element of the matrix U, S, and V respectively. The matrix after U slicing is M, the matrix after S slicing is S', and the matrix after V slicing is V'. Then the approximate weight matrix W'=MS'V'=MN, m k is the kth column of M, n k is the kth row of N;

[0016] Step 4.3: Since W' is r rank-one components m k ×n k The sum of k The weight matrix W' is pruned using the formula:

[0017]

[0018] In the above formula, G = diag (z1, ···, zk) is the diagonal matrix of pruned variables, z k It is an element in the diagonal matrix G, which is used to control whether the corresponding component in the weight matrix participates in the calculation.

[0019] Preferably, the total loss function is designed to constrain the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition, including:

[0020] The total loss function is used to update the parameters of the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition until the preset loss function value is reached, and the model corresponding to the preset loss function value is output as the trained multi-scale feature fusion cloud detection model based on low-rank decomposition.

[0021] Preferably, the total loss function includes: a binary cross entropy loss function, an overlap loss function and a stable distribution loss function.

[0022] Preferably, when a binary cross entropy loss function is selected to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, for each pixel point of the input historical optical image, the multi-scale feature fusion cloud detection model based on low-rank decomposition will output a probability value between 0 and 1, indicating the probability that the pixel point is a cloud, and the difference between the cloud or non-cloud label of the historical optical image and the prediction result of the multi-scale feature fusion cloud detection model based on low-rank decomposition is used as the loss function value. The binary cross entropy loss function is expressed as follows:

[0023]

[0024] In the above formula, N is the total number of historical optical image samples involved in the calculation, and Y is the true label, 0 or 1; It is the predicted value of the multi-scale feature fusion cloud detection model based on low-rank decomposition.

[0025] Preferably, when the overlap loss function is used to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, the multi-scale feature fusion cloud detection model based on low-rank decomposition is optimized by maximizing the prediction output by the multi-scale feature fusion cloud detection model based on low-rank decomposition and the actual annotation of the historical optical image, and the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition is evaluated by using the overlap loss function, and the overlap loss function is expressed as follows:

[0026]

[0027] In the above formula, N is the total number of samples involved in the calculation, Y is the true label 0 or 1, is the predicted value of the model.

[0028] Preferably, when the stable distribution loss function is used to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, the weight parameters of the multi-scale feature fusion cloud detection model based on low-rank decomposition are optimized by approximate L0 norm regularization, and the stable distribution loss function is expressed as follows:

[0029]

[0030] In the above formula, Hard-Sigmoid is the activation function, α j is a value associated with the weight parameter used to calculate the input to the function, β is a constant used to calculate the input to the Hard-Sigmoid function, and γ and ξ are parameters used to stretch the distribution of the continuous random variable.

[0031] Beneficial effects of the present invention:

[0032] A multi-scale feature fusion image detection method based on low-rank decomposition of the present invention includes the following steps: Step 1: Acquire an optical image to be detected; Step 2: Input the optical image into a trained multi-scale feature fusion cloud detection model based on low-rank decomposition to obtain the detection result of the optical image; wherein the trained multi-scale feature fusion cloud detection model based on low-rank decomposition is trained by inputting historical optical images into a multi-scale feature fusion cloud detection model based on low-rank decomposition; the multi-scale feature fusion cloud detection model based on low-rank decomposition is obtained by inserting a weighted low-rank decomposition module into the multi-scale feature fusion cloud detection model; and a total loss function is designed to constrain the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition. The method of the present invention obtains the best detection accuracy in terms of objective evaluation indicators, and can effectively improve the image detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall algorithm framework diagram of the present invention;

[0034] Figure 2 It is a multi-scale fusion module architecture diagram of the present invention;

[0035] Figure 3 It is a graph of the 38-Cloud dataset;

[0036] Figure 4 This is the Landsat8 data set comparison experiment result diagram;

[0037] Figure 5 This is the comparative experimental result of WHUS2-CD dataset;

[0038] Figure 6 This is the comparative experimental results of the GF1-WHU dataset. DETAILED DESCRIPTION

[0039] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. For the purpose of clear description, many physical details will be described together in the following description. However, it should be understood that these physical details should not be used to limit the present invention. That is, in some embodiments of the present invention, these physical details are not necessary. In addition, for the purpose of simplifying the illustration, some conventional structures and components will be depicted in a simple schematic manner in the illustration.

[0040] like Figure 1-6 As shown, a multi-scale feature fusion image detection method based on low-rank decomposition includes the following steps:

[0041] Step 1: Obtain the optical image to be detected;

[0042] Step 2: inputting the optical image into the trained multi-scale feature fusion cloud detection model based on low-rank decomposition to obtain the detection result of the optical image;

[0043] Among them, the trained multi-scale feature fusion cloud detection model based on low rank decomposition is trained by inputting historical optical images into the multi-scale feature fusion cloud detection model based on low rank decomposition; the multi-scale feature fusion cloud detection model based on low rank decomposition is obtained by inserting a weighted low rank decomposition module into the multi-scale feature fusion cloud detection model; and the total loss function is designed to constrain the performance of the multi-scale feature fusion cloud detection model based on low rank decomposition.

[0044] Furthermore, the multi-scale feature fusion cloud detection model based on low-rank decomposition includes: a multi-scale cloud mask network based on low-rank decomposition and a cloud difference fusion network, and the multi-scale cloud mask network based on low-rank decomposition and the cloud difference fusion network have encoders and decoders inside.

[0045] Furthermore, the trained low-rank decomposition-based multi-scale feature fusion cloud detection model is trained by inputting historical optical images into the low-rank decomposition-based multi-scale feature fusion cloud detection model, including:

[0046] The acquired historical optical image is input into the multi-scale feature fusion cloud detection model based on low-rank decomposition for training, the multi-scale cloud mask network based on low-rank decomposition extracts multiple features from the input historical optical image through the encoder inside it, and transmits the multiple features to the decoder inside it after splicing; the decoder fuses the multiple features to form a multi-scale cloud mask;

[0047] The cloud difference fusion network based on low-rank decomposition has multiple inner layers. The cloud difference fusion network is based on a residual structure and captures subtle differences between multiple inner layer features through differential operations to form a difference cloud mask. Afterwards, the difference cloud mask is fused with the multi-scale cloud mask to form an overall cloud mask. The overall cloud mask is used as a multi-scale cloud mask network of the multi-scale feature fusion cloud detection model.

[0048] Afterwards, a weighted low-rank decomposition module is inserted into the encoder-decoder of the multi-scale feature fusion cloud detection model with the multi-scale cloud mask network as the overall cloud mask to obtain a trained multi-scale feature fusion cloud detection model based on low-rank decomposition.

[0049] Furthermore, the multi-scale cloud mask based on low-rank decomposition is a multi-scale feature fusion cloud detection model in which each independent convolution block on the encoder-decoder in the overall cloud mask is inserted with a weighted low-rank decomposition module to obtain a trained multi-scale feature fusion cloud detection model based on low-rank decomposition, comprising the following steps:

[0050] Step 4.1: Assume that the convolution weight matrix of each independent convolution block in the new encoder and the new decoder in the multi-scale cloud mask network of the multi-scale feature fusion cloud detection model is W; the convolution weight matrix W is reduced to a two-dimensional matrix by rearranging the dimensions, and the two-dimensional matrix is ​​decomposed into a matrix W=USV by singular value decomposition, where W is the initial matrix, U is the left singular matrix, S is the singular value matrix, and V is the right singular matrix;

[0051] Step 4.2: Based on the singular value matrix, obtain multiple singular values, and form an approximate weight matrix with the singular values. The rank of the singular value matrix S is r. Slice from the first element of the matrix U, S, and V respectively. The matrix after U slicing is M, the matrix after S slicing is S', and the matrix after V slicing is V'. Then the approximate weight matrix W'=MS'V'=MN, m k is the kth column of M, n k is the kth row of N;

[0052] Step 4.3: Since W' is r rank-one components m k ×n k The sum of k The weight matrix W' is pruned using the formula:

[0053]

[0054] In the above formula, G = diag (z1, ···, zk) is the diagonal matrix of pruned variables, z k It is an element in the diagonal matrix G, which is used to control whether the corresponding component in the weight matrix participates in the calculation.

[0055] Furthermore, the designed total loss function constrains the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition, including:

[0056] The total loss function is used to update the parameters of the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition until the preset loss function value is reached, and the model corresponding to the preset loss function value is output as the trained multi-scale feature fusion cloud detection model based on low-rank decomposition.

[0057] Among them, the calculation formula of the total loss function is:

[0058]

[0059] The total loss function combines the binary cross entropy loss function, the intersection-over-union loss function, and the robust distribution loss function. By adjusting the weight coefficients λ1, λ2, and λ3, the model classification accuracy (measured by the binary cross entropy loss function) and the bounding box positioning accuracy (measured by the intersection-over-union loss function) can be guaranteed while achieving the lightweight of the model (weight sparsification by the robust distribution loss function), so that the cloud detection model after inserting the weight low-rank decomposition module can achieve a good balance between performance and resource utilization efficiency in the cloud detection task. During the model training process, by minimizing the total loss function, the model can continuously adjust the weight parameters and optimize its own prediction ability and structure to meet the needs of the cloud detection task.

[0060] Furthermore, the total loss function includes: a binary cross entropy loss function (BCE-Loss), an overlap loss function (IoU-Loss) and a robust distribution loss function (HC-Loss);

[0061] Further, when the binary cross entropy loss function is selected to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, for each pixel point of the input historical optical image, the multi-scale feature fusion cloud detection model based on low-rank decomposition will output a probability value between 0 and 1, indicating the probability that the pixel point is a cloud, and the difference between the cloud or non-cloud label of the historical optical image and the prediction result of the multi-scale feature fusion cloud detection model based on low-rank decomposition is used as the loss function value. The binary cross entropy loss function is expressed as follows:

[0062]

[0063] In the above formula, N is the total number of historical optical image samples involved in the calculation, and Y is the true label, 0 or 1; Prediction value of multi-scale feature fusion cloud detection model based on low-rank decomposition.

[0064] Furthermore, when the overlap loss function is used to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, the multi-scale feature fusion cloud detection model based on low-rank decomposition is optimized by maximizing the prediction output by the multi-scale feature fusion cloud detection model based on low-rank decomposition and the actual annotation of the historical optical image. The performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition is evaluated by using the overlap loss function, and the overlap loss function is expressed as follows:

[0065]

[0066] In the above formula, N is the total number of samples involved in the calculation, Y is the true label 0 or 1, is the predicted value of the model (a probability value between 0 and 1).

[0067] Furthermore, when the stable distribution loss function is used to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, the weight parameters of the multi-scale feature fusion cloud detection model based on low-rank decomposition are optimized by approximate L0 norm regularization, so that unimportant parameters are structured pruned to achieve the purpose of parameter sparsification. This helps to reduce the computational requirements of the model and alleviate the overfitting problem, making the model more lightweight while ensuring certain performance. The stable distribution loss function is expressed as follows:

[0068]

[0069] In the above formula, Hard-Sigmoid is the activation function, α j is a value associated with the weight parameter used to calculate the input to the function, β is a constant used to calculate the input to the Hard-Sigmoid function, and γ and ξ are parameters used to stretch the distribution of the continuous random variable.

[0070] The method of the present invention will be further described below using the 38-Cloud dataset:

[0071] The experimental data comes from the 38-Cloud dataset, which contains 38 Landsat 8 scene images with manually extracted ground truth of pixel-level cloud detection. The entire image of these scenes is cropped into multiple 384×384 small blocks, each of which has 4 corresponding spectral channels, including blue channel (band 2), green channel (band 3), red channel (band 4) and near infrared channel (band 5), as shown in Figure 2. Figure 6 shown.

[0072] In the Landsat8 (Landsat 8) scene images, thin clouds and stratus clouds with a relatively simple background were selected as the experimental data set. For cloud detection in simple scenes, the difference between traditional methods and deep learning methods is not particularly obvious. In terms of subjective visual effects, OTSU (maximum between-class variance method) as a traditional threshold detection method has certain difficulties in thin cloud detection. In simple scenes, the Fmask (mask function) method showed good performance in thin cloud detection and successfully captured some thin clouds. GAN-CDM (generative adversarial network-cloud detection method) and WDCD (weighted distance cloud detection) are slightly lacking in thin cloud detection effects. In contrast, the method of the present invention showed very significant experimental results on the dataset of Landsat8 (Landsat 8) scene images, performed well in thin cloud detection, and presented a visual effect that was almost consistent with GroundTruth (ground truth). The experimental results are as follows Figure 4 shown.

[0073] In terms of objective evaluation indicators, Table 1 shows the image indicators obtained by three experimental data sets under five comparison methods: IoU (intersection over union), OA (overall accuracy) and F1_score (harmonic mean of precision and recall, F1 score) values. Since the background is relatively simple, the evaluation indicators of each method are not much different. Compared with other methods, the low-rank decomposition multi-scale feature fusion cloud detection method proposed in the present invention performs well in terms of IoU (intersection over union), OA (overall accuracy) and F1_score values. This means that the method of the present invention can detect thin cloud areas more accurately.

[0074] Table 1

[0075]

[0076] like Figure 5 As shown, in WHUS2-CD (Wuhan University thick cloud detection and thin cloud removal dataset), thin clouds, stratus clouds and cirrus clouds with slightly complex road backgrounds were selected as experimental datasets. In terms of subjective visual effects, for the first experimental dataset, since the road color is not consistent with most of the background, the OTSU (maximum between-class variance method), Fmask (mask function), Unet (U-type network) and GAN-CDM (generative adversarial network-cloud detection method) methods all have misjudgments, while the WDCD (weighted distance cloud detection) method and the method of the present invention have no misjudgments. For the stratus and cirrus experimental datasets, the experimental results of the method of the present invention are basically consistent with GroundTruth (ground truth), and are superior to other methods in the detection of thin clouds, stratus clouds and cirrus clouds. The experimental results are shown in the figure. Figure 5 shown.

[0077] In terms of objective evaluation indicators, due to the complex background of the experimental data set, the method of the present invention is higher than other methods in evaluation indicators, indicating that the method of this paper is robust, more accurate in cloud detection and less susceptible to background interference. The evaluation indicators are shown in Table 2 below.

[0078] Table 2

[0079]

[0080] In GF1-WHU (Gaofen-1 Satellite-Wuhan University Dataset), mountains, lakes and vegetation with more complex backgrounds were selected as experimental data sets. In terms of subjective visual effects, the OTSU and Fmask (mask function) methods cannot detect some thin clouds. For experimental data set 2, due to the green background interference near the stratus clouds, the OTSU (maximum inter-class variance method), Fmask (mask function), GAN-CDM (generative adversarial network-cloud detection method) and WDCD (weighted distance cloud detection) methods are all affected, and the method of the present invention has better visual effects. The comparative experimental results of the GF1-WHU (Gaofen-1 Satellite-Wuhan University Dataset) data set are shown in Figure 2. Figure 6 shown.

[0081] In terms of objective evaluation indicators, the method of the present invention is 12.5%, 12.3%, and 16.5% higher than the OTSU (maximum inter-class variance method) method in terms of IoU (intersection over union), OA (overall accuracy), and F1_score (F1 score) indicators; 8.1%, 9.9%, and 13.4% higher than the Fmask (mask function) method; 1.9%, 1.3%, and 1.4% higher than the Unet (U-type network) method; 3.9%, 1.7%, and 2.2% higher than the GAN-CDM (generative adversarial network-cloud detection method) method; 4.9%, 4%, and 2.9% higher than the WDCD (weighted distance cloud detection) method; Among all experimental methods, the method of the present invention obtains the best detection accuracy in objective evaluation indicators. The evaluation indicators are shown in Table 3.

[0082] Table 3

[0083]

[0084] The FLOPs (floating point operations) index of the method of the present invention and the comparative method is shown in Table 4, and the unpruned and pruned methods of the present invention are used as experimental comparisons. Through low-rank decomposition, the amount of computation after pruning is reduced by 52.6G, and the evaluation indexes of the method of the present invention are better than those of other methods.

[0085] Table 4

[0086]

[0087] The above description is only a specific embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A multi-scale feature fusion image detection method based on low-rank decomposition, characterized in that: Including the following steps: Step 1: Obtain the optical image to be detected; Step 2: inputting the optical image into the trained multi-scale feature fusion cloud detection model based on low-rank decomposition to obtain the detection result of the optical image; Among them, the trained multi-scale feature fusion cloud detection model based on low rank decomposition is trained by inputting historical optical images into the multi-scale feature fusion cloud detection model based on low rank decomposition; the multi-scale feature fusion cloud detection model based on low rank decomposition is obtained by inserting a weighted low rank decomposition module into the multi-scale feature fusion cloud detection model; and the total loss function is designed to constrain the performance of the multi-scale feature fusion cloud detection model based on low rank decomposition.

2. The method according to claim 1, characterized in that The multi-scale feature fusion cloud detection model based on low-rank decomposition includes: a multi-scale cloud mask network based on low-rank decomposition and a cloud difference fusion network, wherein the multi-scale cloud mask network based on low-rank decomposition and the cloud difference fusion network have encoders and decoders inside.

3. The method according to claim 2, characterized in that The trained low-rank decomposition-based multi-scale feature fusion cloud detection model is trained by inputting historical optical images into the low-rank decomposition-based multi-scale feature fusion cloud detection model, including: The acquired historical optical image is input into the multi-scale feature fusion cloud detection model based on low-rank decomposition for training, the multi-scale cloud mask network based on low-rank decomposition extracts multiple features from the input historical optical image through the encoder inside it, and transmits the multiple features to the decoder inside it after splicing; the decoder fuses the multiple features to form a multi-scale cloud mask; The cloud difference fusion network based on low-rank decomposition has multiple inner layers inside, and the cloud difference fusion network captures the subtle differences between multiple inner layer features through differential operations to form a difference cloud mask; then, the difference cloud mask is fused with the multi-scale cloud mask to form an overall cloud mask; the overall cloud mask is used as a multi-scale cloud mask network of the multi-scale feature fusion cloud detection model; Afterwards, a weighted low-rank decomposition module is inserted into the encoder-decoder of the multi-scale feature fusion cloud detection model with the multi-scale cloud mask network as the overall cloud mask to obtain a trained multi-scale feature fusion cloud detection model based on low-rank decomposition.

4. The method according to claim 3, characterized in that The multi-scale cloud mask based on low-rank decomposition is used as the overall cloud mask, and each independent convolution block on the encoder-decoder in the multi-scale feature fusion cloud detection model is inserted with a weighted low-rank decomposition module to obtain a trained multi-scale feature fusion cloud detection model based on low-rank decomposition, including the following steps: Step 4.1: Assume that the convolution weight matrix of each independent convolution block in the new encoder and the new decoder in the multi-scale cloud mask network of the multi-scale feature fusion cloud detection model is W; the convolution weight matrix W is reduced to a two-dimensional matrix by rearranging the dimensions, and the two-dimensional matrix is ​​decomposed into a matrix W=USV by singular value decomposition, where W is the initial matrix, U is the left singular matrix, S is the singular value matrix, and V is the right singular matrix; Step 4.2: Based on the singular value matrix, obtain multiple singular values, and form an approximate weight matrix with the singular values. The rank of the singular value matrix S is r. Slice from the first element of the matrix U, S, and V respectively. The matrix after U slicing is M, the matrix after S slicing is S', and the matrix after V slicing is V'. Then the approximate weight matrix W'=MS'V'=MN, m k is the kth column of M, n k is the kth row of N; Step 4.3: Since W' is r rank-one components m k ×n k The sum of k The weight matrix W' is pruned using the formula: In the above formula, G = diag(z1,···,z k ) is the diagonal matrix of pruned variables, z k It is an element in the diagonal matrix G, which is used to control whether the corresponding component in the weight matrix participates in the calculation.

5. The method according to claim 1, characterized in that The designed total loss function constrains the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition, including: The total loss function is used to update the parameters of the performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition until the preset loss function value is reached, and the model corresponding to the preset loss function value is output as the trained multi-scale feature fusion cloud detection model based on low-rank decomposition.

6. The method according to claim 5, characterized in that The total loss function includes: a binary cross entropy loss function, an overlap loss function and a stable distribution loss function.

7. The method according to claim 5, characterized in that When the binary cross entropy loss function is selected to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, for each pixel point of the input historical optical image, the multi-scale feature fusion cloud detection model based on low-rank decomposition will output a probability value between 0 and 1, indicating the probability that the pixel point is a cloud. The difference between the cloud or non-cloud label of the historical optical image and the prediction result of the multi-scale feature fusion cloud detection model based on low-rank decomposition is used as the loss function value. The binary cross entropy loss function is expressed as follows: In the above formula, N is the total number of historical optical image samples involved in the calculation, and Y is the true label, 0 or 1; It is the predicted value of the multi-scale feature fusion cloud detection model based on low-rank decomposition.

8. The method according to claim 5, characterized in that When the overlap loss function is used to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, the multi-scale feature fusion cloud detection model based on low-rank decomposition is optimized by maximizing the prediction output by the multi-scale feature fusion cloud detection model based on low-rank decomposition and the actual annotation of the historical optical image. The performance of the multi-scale feature fusion cloud detection model based on low-rank decomposition is evaluated by using the overlap loss function, and the overlap loss function is expressed as follows: In the above formula, N is the total number of samples involved in the calculation, Y is the true label 0 or 1, is the predicted value of the model.

9. The method according to claim 5, characterized in that When the stable distribution loss function is used to train the multi-scale feature fusion cloud detection model based on low-rank decomposition, the weight parameters of the multi-scale feature fusion cloud detection model based on low-rank decomposition are optimized by approximate L0 norm regularization. The stable distribution loss function is expressed as follows: In the above formula, Hard-Sigmoid is the activation function, α j is a value associated with the weight parameter used to calculate the input to the function, β is a constant used to calculate the input to the Hard-Sigmoid function, and γ and ξ are parameters used to stretch the distribution of the continuous random variable.

Images