A deep learning cloud detection method based on HY-1C satellite COCTS data
By using the STIGNet network model in the HY-1C satellite COCTS data, combining spectral and texture information, the threshold setting difficulties in cloud detection and inaccurate background field information are solved, and the accuracy and adaptability of cloud edge detection are significantly improved. Efficient ocean monitoring is achieved.
Patent Information
- Application Number
- CN202411076345.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-08-07
AI Technical Summary
The existing cloud detection methods have problems with threshold setting, inaccurate background field information, and texture characteristics and cloud edge errors, especially in marine satellite data processing.
Using deep learning cloud detection method based on HY-1C satellite COCTS data, by building a STIGNet network model, combining spectral and texture information, multi-level feature extraction and edge feature fusion are used to reduce cloud edges Detect errors and optimize the model by focusing loss function and edge loss function.
Improve the accuracy and adaptability of cloud detection, significantly reduce cloud edge detection errors, improve data processing efficiency, and support automated and semi-automated marine monitoring.
Smart Images

Figure CN118587610B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing technology, and in particular to a deep learning cloud detection method based on HY-1C satellite COCTS data. Background Art
[0002] Currently existing cloud detection methods can be roughly divided into: threshold method, which uses the physical properties, texture characteristics and spatial characteristics of clouds to establish a specific threshold combination to determine whether clouds exist; Bayesian method, which combines the background atmospheric state and surface parameters to calculate the probability of cloud existence; deep learning method, which builds a large number of data-driven models to learn the correspondence between spectral data characteristics and clouds.
[0003] The rise of artificial intelligence technology has led researchers to discover that machine learning can help find the best threshold. Therefore, cloud detection methods that combine thresholds with machine learning have also developed. This method essentially still uses thresholds to determine cloud and clear sky pixels. At present, the continuous development of deep learning has expanded people's attention to the research of using deep learning to solve remote sensing tasks. Deep learning is based on a large number of labeled samples to find the optimal structure and coefficients. The deep learning model can accurately learn the complex nonlinear relationship between environmental parameters and capture the potential connection between feature variables. At the same time, deep learning also has obvious advantages in obtaining multi-level and multi-scale features in remote sensing images and combining them from low levels to high levels. It has broad application prospects in image classification and processing problems. Deep learning has great potential in mining information from marine remote sensing data, and marine satellite big data can provide data support for deep learning. Deep learning can combine spectral information and spatial information, and has a strong feature extraction capability, which is very suitable for cloud detection tasks with a huge data base.
[0004] However, the existing technical solutions still have the following technical problems:
[0005] The threshold method for cloud detection starts from the spectral physical characteristics of clouds, combines the image texture characteristics and regional spatial characteristics, sets a specific threshold, and arranges different threshold combinations to determine whether a pixel is a cloud. Once the processing flow of this method is determined, it is not difficult to implement. However, the determination of the threshold is often related to environmental factors such as sensor characteristics, variable attributes, surface type, and regional scope. Therefore, a fixed threshold is often difficult to accurately detect clouds under all environmental factors.
[0006] The Bayesian cloud detection algorithm adds background field prior knowledge to the cloud detection process, improving the drawback of setting thresholds for each variable parameter. However, background field prior information needs to be obtained through additional resources such as numerical weather forecasts and atmospheric radiation transfer models, and the spatial and temporal resolution of numerical weather forecast information is usually much lower than the spatial and temporal resolution of satellite sensors, which leads to errors in the background field. In addition, for the domestic Ocean No. 1 satellite, there is no applicable fast radiation transfer model, and the background field information can only be obtained through complex and slow calculations of models such as MODTRAN, which increases the time cost.
[0007] Deep learning is based on a large number of labeled samples to find the optimal structure and coefficients. Deep learning models can accurately learn the complex nonlinear relationships between environmental parameters and capture the potential connections between feature variables. Most existing deep learning cloud detection methods only use spectral data as input, ignoring many cloud-related texture features. In addition, since the cloud edge has complex characteristics compared to other areas, the error at the cloud edge can be very significant.
[0008] In summary, existing technical methods have problems in cloud detection, such as difficulty in setting thresholds, inaccurate background field information, and errors in texture features and cloud edges. To address these problems, this paper proposes a deep neural network model (STIGNet) guided by spectral and texture information to solve the technical problems of traditional methods and achieve automatic, fast and accurate cloud detection. Summary of the invention
[0009] In view of the above-mentioned prior art, the present invention provides a deep learning cloud detection method based on HY-1C satellite COCTS data, which mainly solves the technical problems existing in the above-mentioned background technology.
[0010] To achieve the above object, the technical solution of the embodiment of the present invention is implemented as follows:
[0011] A deep learning cloud detection method based on HY-1C satellite COCTS data, the method comprising the following steps:
[0012] The method comprises the following steps:
[0013] Step S1: construct the HY-1C satellite COCTS global ocean cloud detection dataset;
[0014] Step S2: preprocessing the cloud detection data set;
[0015] Step S3: construct and train a STIGNet network model, wherein the STIGNet network model includes an encoder, an edge learning module, and a decoder, and classify the cloud detection data set through the trained STIGNet model to obtain a classification result;
[0016] The encoder includes four consecutive feature extraction blocks FEB, the edge learning module includes four edge learning blocks ELB and a feature fusion block FFB, and the decoder includes four feature recovery blocks FRB; the preprocessed cloud detection data set is subjected to multi-level feature extraction through the four consecutive feature extraction blocks FEB, the edge information of the features extracted by the feature extraction block FEB is extracted and enhanced through the four edge learning blocks ELB to obtain multiple edge features, the multiple edge features obtained by the edge learning block ELB are fused through the feature fusion block FFB to obtain multi-scale edge features, the multi-scale edge features and the edge truth value establish a loss function to enhance the learning of the edge features by the feature fusion block FFB, the features extracted by the feature extraction block FEB are restored through the four consecutive feature recovery blocks FRB, and the features are spliced with the features extracted by the corresponding feature extraction block FEB, so as to finally realize the cloud mask output or the clear sky mask output.
[0017] Specifically, the cloud detection data set in step S1 specifically includes spectral data, texture data, and geographic data of clouds.
[0018] Specifically, the preprocessing includes converting the radiance of 443nm, 490nm, 565nm, 670nm and 865nm bands in the cloud detection data set into reflectance, and converting the radiance of 11μm and 12μm bands into brightness temperature.
[0019] Specifically, the radiance of the 443nm, 490nm, 565nm, 670nm and 865nm bands in the cloud detection dataset is converted to reflectance by the following formula:
[0020]
[0021] in, is the wavelength, is pi, is the reflectivity of the band, is the radiance of the band, is the cosine function value of the solar zenith angle, is the solar irradiance at the top of the atmosphere.
[0022] Specifically, the brightness temperature is obtained by constructing a lookup table for converting the radiance in the 11 μm and 12 μm bands into brightness temperature.
[0023] Specifically, a lookup table for converting the radiance of the 11μm and 12μm bands to brightness temperature is constructed using the following formula. The specific process is as follows:
[0024] Preset brightness temperature , preset wavelengths within the band range , calculate the wavelength to be The blackbody radiance at :
[0025]
[0026] in, is Planck's constant, is the speed of light, is the Boltzmann constant, is the blackbody radiance, For a wavelength of , brightness temperature is The blackbody radiance at ;
[0027] Then calculate the brightness temperature according to the following formula The radiance corresponding to a wavelength range :
[0028]
[0029] in, is at wavelength The spectral response function of this band is defined above, Indicates wavelength The integral of
[0030] Get the radiance of different bands and brightness temperature The one-to-one correspondence between the radiances of the 11μm and 12μm bands in the cloud detection dataset is found in the lookup table. value.
[0031] Specifically, the preprocessing process further includes constructing classification labels in the cloud detection data set, where the classification labels include cloud and clear sky, wherein the classification label of clear sky is shown in the following formula:
[0032]
[0033] in, is the clear sky condition over the ocean, with the superscript Indicates "observation", superscript represents the “prior background”, represents the state vector of the prior background, represents the observation vector, represents the probability of clear sky over the ocean, represents the probability of clear sky given the prior background state and observation conditions, represents the probability of observations given the prior background state and clear sky conditions, represents the background state probability under clear sky conditions, is the probability of an observation given the prior background state, represents the probability of the prior background state;
[0034] The pixel labels with the calculated Bayesian clear sky probability greater than 0.9 are labeled as clear sky pixels, and this label is added to the cloud detection dataset.
[0035] Specifically, the feature extraction blocks FEB include a 3×3 convolution layer, a batch normalization layer, a rectified linear unit, and a maximum pooling layer;
[0036] The edge learning blocks ELB each include a 3×3 convolutional layer, a batch normalization layer, a rectified linear unit, and a 1×1 convolutional layer;
[0037] The feature recovery blocks FRB include a deconvolution layer, a 3×3 convolution layer, a batch normalization layer and a ReLU activation function.
[0038] Specifically, the loss function is the sum of a focus loss function and an edge loss function;
[0039] The focus loss function is defined as follows:
[0040]
[0041] in, Represents the number of categories, Indicates categories, Representative The true value of each category, represents the predicted value of the network model, represents the weight, Represents the logarithm of the network model prediction value;
[0042] The marginal loss function is defined as follows:
[0043]
[0044] in, represents the Laplace smoothing term, represents the marginal truth value, Representative The marginal truth value of each category, Representative The output value of the feature fusion block FFB of each category, Indicates The logarithm of the output value of the feature fusion block FFB for each category.
[0045] The beneficial effects of the present invention are as follows: Compared with the prior art, the deep learning cloud detection method based on HY-1C satellite COCTS data provided by the present invention has the following beneficial effects:
[0046] 1. Improve the accuracy of cloud detection: By comprehensively considering the cloud's spectral data, brightness temperature differences, local statistical characteristics and geographic location information, the STIGNet network model can more accurately identify and distinguish between clouds and clear ocean surfaces. In particular, the edge learning module in the STIGNet network model is specifically optimized for the complex features of the cloud edge, significantly reducing the detection error in these areas; through the design of the edge learning module, the STIGNet network model strengthens the learning of cloud edge details. This optimization significantly improves the detection accuracy of the STIGNet network model in the cloud edge area;
[0047] 2. Strong adaptability: The STIGNet network model does not rely on a fixed threshold, but automatically learns the characteristics of the cloud through a deep learning model, so that it can adapt to different environmental conditions and sensor characteristics, and has good generalization ability;
[0048] 3. Improve data processing efficiency and support automated and semi-automated ocean monitoring: Since the STIGNet network model adopts an efficient deep learning architecture, it can quickly process large-scale remote sensing data sets and provide accurate cloud masks that can serve as the basis for subsequent ocean data analysis, thereby improving the efficiency and accuracy of the entire ocean monitoring process. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 A flow chart of the steps of the cloud detection method provided by the present invention.
[0050] Figure 2 This is a structural diagram of the STIGNet network model provided by the present invention.
[0051] Figure 3 This is an example diagram of the error image of cloud detection provided by the present invention. DETAILED DESCRIPTION
[0052] The technical solution of the present invention is further elaborated in detail below in conjunction with the drawings and specific embodiments of the specification. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. In the following description, the expression "some embodiments" is related to a subset of all possible embodiments, but it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.
[0053] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known in the art are not described.
[0054] It should be understood that the present invention can be implemented in different forms and should not be interpreted as being limited to the embodiments proposed herein. On the contrary, providing these embodiments will make the disclosure thorough and complete, and the scope of the present invention will be fully conveyed to those skilled in the art. And the purpose of the terms used herein is only to describe specific embodiments and is not intended to be a limitation of the present invention. When used herein, the singular forms of "one", "one" and "said / the" are also intended to include plural forms, unless the context clearly indicates another way. It should also be understood that the terms "compose" and / or "include" when used in this specification determine the presence of the features, integers, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, parts and / or groups. When used herein, the term "and / or" includes any and all combinations of the relevant listed items.
[0055] It should also be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "inside", "outside", "left", "right" and similar expressions used herein are for illustrative purposes only and are not intended to be the only implementation method.
[0056] In order to fully understand the present invention, a detailed structure will be proposed in the following description to illustrate the technical solution proposed by the present invention. The optional embodiments of the present invention are described in detail as follows, but in addition to these detailed descriptions, the present invention may also have other implementations.
[0057] Please refer to the attached Figure 1 and attached Figure 2 , the present application provides a deep learning cloud detection method based on HY-1C satellite COCTS data, the method comprising the following steps:
[0058] Step S1: construct the HY-1C satellite COCTS global ocean cloud detection dataset;
[0059] Step S2: preprocessing the cloud detection data set;
[0060] Step S3: Build and train the STIGNet network model, and use the STIGNet network model to preprocess
[0061] The STIGNet network model includes an encoder, an edge learning module and a decoder;
[0062] The encoder includes four consecutive feature extraction blocks FEB, the edge learning module includes four edge learning blocks ELB and one feature fusion block FFB, and the decoder includes four feature recovery blocks FRB;
[0063] For example, HY-1C satellite refers to Ocean No. 1C satellite, COCTS refers to Chinese Ocean Color and Temperature Scanner (COCTS), and the input data of the cloud detection dataset comes from the data of HY-1C satellite COCTS L1B product. The cloud detection dataset is preprocessed, and then the preprocessed cloud detection dataset is subjected to cloud detection through the STIGNet network model (the full name of the STIGNet network model is Spectral-and-Textural-Information-Guided deep neural Network, i.e., a deep neural network model guided by spectral and texture information). Since the shape and size of clouds are very irregular, the loss of edge information will be particularly obvious. Therefore, an edge learning module is designed in the STIGNet network model to enable the STIGNet network model to accurately detect and segment cloud areas in the global ocean area data of ocean satellites.
[0064] Multi-level feature extraction is performed on the preprocessed cloud detection data set through four consecutive feature extraction blocks FEB, edge information is extracted and enhanced on the features extracted by the feature extraction block FEB through four edge learning blocks ELB to obtain multiple edge features, and multiple edge features obtained by the edge learning block ELB are fused through the feature fusion block FFB to obtain multi-scale edge features. A loss function is established between the multi-scale edge features and the edge truth to enhance the learning of edge features by the feature fusion block FFB, and features extracted by the feature extraction block FEB are restored through four consecutive feature recovery blocks FRB, and are spliced with the features extracted by the corresponding feature extraction block FEB to finally achieve cloud mask output or clear sky mask output.
[0065] Exemplarily, the working principle of the STIGNet network model is as follows: the first feature extraction block FEB extracts features from the data set to obtain features , the second feature extraction block FEB Perform feature extraction to obtain features , and so on, the fourth feature extraction block FEB is Perform feature extraction to obtain features ;
[0066] Through four edge learning blocks ELB, the features are , , , Extract edge information to get features , , , , through the feature fusion block FFB, the features , , , The multi-scale edge features are obtained by fusion, and the edge loss function is calculated with the edge truth value for back propagation, which enhances the model's learning of edge features during the training process;
[0067] The first feature recovery block FRB first recovers the feature Deconvolution is performed to obtain features , then the feature With features After concatenation, two rounds of 3 × 3 convolution, batch normalization layer and ReLU activation function are performed to obtain the feature , the second feature recovery block pair Deconvolution is performed to obtain , then and After concatenation, two rounds of 3 × 3 convolution, batch normalization layer and ReLU activation function are performed to obtain the feature , and so on, the fourth feature recovery block is Deconvolution is performed to obtain , then and After concatenation, two rounds of 3 × 3 convolution, batch normalization layer and ReLU activation function are performed to obtain the feature ,feature Finally, the Softmax layer is used to realize the cloud mask output or clear sky mask output.
[0068] Specifically, the cloud detection data set in step S1 specifically includes spectral data, texture data, and geographic data of clouds.
[0069] Specifically, the preprocessing includes converting the radiance of 443nm, 490nm, 565nm, 670nm and 865nm bands in the cloud detection data set into reflectance, and converting the radiance of 11μm and 12μm bands into brightness temperature.
[0070] Specifically, the radiance of the 443nm, 490nm, 565nm, 670nm and 865nm bands in the cloud detection dataset is converted to reflectance by the following formula:
[0071]
[0072] in, is the wavelength, is pi, is the reflectivity of the band, is the radiance of the band, is the cosine function value of the solar zenith angle, is the solar irradiance at the top of the atmosphere.
[0073] Specifically, the brightness temperature is obtained by constructing a lookup table for converting the radiance in the 11 μm and 12 μm bands into brightness temperature.
[0074] Specifically, a lookup table for converting the radiance of the 11μm and 12μm bands to brightness temperature is constructed using the following formula. The specific process is as follows:
[0075] Preset brightness temperature , preset wavelengths within the band range , calculate the wavelength to be The blackbody radiance at :
[0076]
[0077] in, is Planck's constant, is the speed of light, is the Boltzmann constant, is the blackbody radiance, For a wavelength of , brightness temperature is The blackbody radiance at ;
[0078] Then calculate the brightness temperature according to the following formula The radiance corresponding to a wavelength range :
[0079]
[0080] in, is at wavelength The spectral response function of this band is defined above, Indicates wavelength The integral of
[0081] Get the radiance of different bands and brightness temperature The one-to-one correspondence between the radiances of the 11μm and 12μm bands in the cloud detection dataset is found in the lookup table. value.
[0082] Exemplarily, the corresponding relationship between the partial radiance and brightness temperature in the 11 μm band is shown in Table 1.
[0083] Table 1 Correspondence between radiance and brightness temperature in some 11 μm bands
[0084] <![CDATA[Radiance in the 11μm band (mWcm -2 um -1 sr -1 )]]> Brightness temperature (K) 7.3510 282.80676 7.3511 282.80759 7.3512 282.80841 7.3513 282.80920 7.3514 282.81003 7.3515 282.81082 7.3516 282.81165
[0085] For example, , The local standard deviation of the 3×3 area centered on each pixel and the difference between the maximum and minimum values of the brightness temperature in the 11μm band are used as the input of the STIGNet network model to represent cloud detection based on texture features; since the data of the visible light band cannot be obtained at night, the reflectance data of the visible light band is not included in the night input channel. The input features of the cloud detection dataset are shown in Table 2, and the detection effects of cloud detection on the daytime and nighttime datasets are shown in Table 3.
[0086] Table 2 Input features of cloud detection dataset
[0087]
[0088] Table 3 Detection effects of cloud detection models on daytime and nighttime datasets
[0089]
[0090] Among them, the accuracy rate calculates the "proportion of all correctly predicted pixels in the total pixels", and focuses on the overall prediction accuracy; the recall rate calculates the "proportion of correctly predicted clear sky pixels in the true clear sky pixels", and focuses on how many accurate predictions are in the true clear sky, that is, the situation of cloud over-detection errors; the precision rate calculates the "proportion of correctly predicted clear sky pixels in the predicted clear sky pixels", and focuses on how many accurate predictions are in the predicted clear sky, that is, the situation of cloud omission errors. F1-Score is the harmonic mean of recall rate and precision rate, and is a comprehensive measurement indicator of the two. IoU represents the intersection over Union indicator to evaluate the similarity and difference between sample sets. IoU calculates the ratio of the intersection and union of the predicted clear sky and the clear sky labeled in the data set. The false positive rate refers to the "proportion of the number of cloud omission pixels in the true cloud", which is intended to evaluate how many clouds are mistakenly predicted as clear sky. The cloud over-detection rate refers to the proportion of cloud over-detected pixels in all pixels, and the cloud omission rate refers to the proportion of cloud omission pixels in all pixels.
[0091] Specifically, the preprocessing process further includes constructing classification labels in the cloud detection data set, where the classification labels include cloud and clear sky, wherein the classification label of clear sky is shown in the following formula:
[0092]
[0093] in, is the clear sky condition over the ocean, with the superscript Indicates "observation", superscript represents the “prior background”, represents the state vector of the prior background, represents the observation vector, represents the probability of clear sky over the ocean, represents the probability of clear sky given the prior background state and observation conditions, represents the probability of observations given the prior background state and clear sky conditions, represents the background state probability under clear sky conditions, is the probability of an observation given the prior background state, represents the probability of the prior background state;
[0094] The pixel labels with the calculated Bayesian clear sky probability greater than 0.9 are labeled as clear sky pixels, and this label is added to the cloud detection dataset.
[0095] For example, Bayesian cloud detection performs well in most regions of the world. This dataset manually checks and corrects the misclassification of Bayesian theorem in cloud detection, and adds the correct labels to the cloud detection dataset.
[0096] Specifically, the feature extraction blocks FEB include a 3×3 convolution layer, a batch normalization layer, a rectified linear unit, and a maximum pooling layer;
[0097] Exemplarily, the feature extraction block FEB is used to extract and transform the features of the input data. The 3×3 convolution layer uses a 3×3 convolution kernel to scan the input data and extract local features. The batch normalization layer normalizes each small batch of data to speed up the convergence speed during training, while reducing the sensitivity to initialization parameters, which helps the STIGNet network model learn more effective feature representations. The rectified linear unit is used to introduce nonlinear features into the STIGNet network model, setting the negative part of the input value to 0 and the positive part unchanged, which helps solve the gradient vanishing problem and allows the network to learn more complex features. The spatial dimension of the features is then gradually reduced through the maximum pooling layer, thereby achieving hierarchical abstraction of the features.
[0098] The edge learning blocks ELB each include a 3×3 convolutional layer, a batch normalization layer, a rectified linear unit, and a 1×1 convolutional layer;
[0099] Exemplarily, the edge learning block ELB can extract and enhance edge information, the 3×3 convolution layer is used to capture edge features and is normalized through the batch normalization layer to provide stability and accelerate training; the rectified linear unit is used to introduce nonlinear features, and the 1×1 convolution layer reweights and combines features in the spatial dimension; the feature fusion block FFB is used to fuse the output features from the edge learning block ELB with features from different depths of the encoder; the output features of the feature fusion block FFB will calculate the edge loss function with the edge true value, thereby enhancing the model's learning of edge features during training.
[0100] Since the cloud edge is the main error area, an edge learning module is added to the model to guide the model to learn more edge features. Table 4 shows the comparison before and after the addition. The cloud missed detection rate has been improved after the improvement. The accuracy of all areas has been significantly improved, and the improvement is most obvious in the cloud edge area. At night, the overall accuracy has increased by 0.15%, the accuracy at the cloud edge has increased by 0.26%, and the accuracy at the non-cloud edge has increased by 0.14%. During the day, the overall accuracy has increased by 0.05%, the accuracy at the cloud edge has increased by 0.13%, and the accuracy at the non-cloud edge has increased by 0.03%. This shows that the edge learning module strengthens the model's learning of edge features and has a certain effect on improving the error in the cloud edge area. In addition, this module is only used to guide the learning of the model during the training phase, and does not generate additional overhead in the use of the model, and can still ensure efficiency.
[0101] Table 4 Accuracy of cloud edge, non-cloud edge and all areas
[0102]
[0103] The feature recovery blocks FRB include a deconvolution layer, a 3×3 convolution layer, a batch normalization layer and a ReLU activation function.
[0104] Exemplarily, deconvolution is used in the decoder to restore the spatial resolution lost in the encoder. After deconvolution, a 3×3 convolution layer is used to further refine the details of the features. The output of the 3×3 convolution layer is normalized, and the ReLU activation function follows the batch normalization layer to increase the nonlinear expression capability of the network.
[0105] Specifically, in order to guide the encoder to better preserve edge details through back propagation, the edge true value generated from the label is used to calculate the edge loss term, so the loss function is specifically the sum of the focus loss function and the edge loss function;
[0106] The focus loss function is defined as follows:
[0107]
[0108] in, Represents the number of categories, Indicates categories, Representative The true value of each category, represents the predicted value of the network model, represents the weight, Represents the logarithm of the network model prediction value;
[0109] For example, in the training cloud detection dataset, the proportion of clouds is 72%, which means that the proportion of clouds and oceans is seriously unbalanced. The imbalanced data distribution may cause the STIGNet network model to favor classes with a higher proportion of samples during training, resulting in poor cloud detection performance. Therefore, without changing the sample distribution, the focus loss function is used as the loss term to solve the deviation caused by sample imbalance, that is, to minimize the segmentation difference. In addition to minimizing the segmentation difference, the cloud edge constraint is also enforced by minimizing the edge consistency loss. The edge loss function consists of binary cross entropy loss and dice loss.
[0110] The marginal loss function is defined as follows:
[0111]
[0112] in, represents the Laplace smoothing term, =1, represents the marginal truth value, Representative The marginal truth value of each category, Representative The output value of the feature fusion block FFB of each category, Indicates The logarithm of the output value of the feature fusion block FFB of each category, is the binary cross entropy loss, For dice loss.
[0113] For example, as shown in the attached Figure 3Shown are examples of error images for cloud detection, where (a) is the first example of an error image for cloud detection, (b) is the second example of an error image for cloud detection, and (c) is the third example of an error image for cloud detection. In the error images, red represents cloud over-detection by the model, and blue represents cloud omission by the model. The values under the error images represent the accuracy rate (black), cloud omission rate (blue), and cloud over-detection rate (red), respectively.
[0114] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. The protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A deep learning cloud detection method based on HY-1C satellite COCTS data, It is characterized in that The method comprises the following steps: Step S1: construct the HY-1C satellite COCTS global ocean cloud detection dataset; Step S2: preprocessing the cloud detection data set; Step S3: construct and train a STIGNet network model, wherein the STIGNet network model includes an encoder, an edge learning module, and a decoder, and classify the cloud detection data set through the trained STIGNet model to obtain a classification result; The encoder includes four consecutive feature extraction blocks FEB, the edge learning module includes four edge learning blocks ELB and one feature fusion block FFB, and the decoder includes four feature recovery blocks FRB; the feature extraction blocks FEB each include a 3×3 convolution layer, a batch normalization layer, a rectified linear unit, and a maximum pooling layer; the edge learning blocks ELB each include a 3×3 convolution layer, a batch normalization layer, a rectified linear unit, and a 1×1 convolution layer; the feature recovery blocks FRB each include a deconvolution layer, a 3×3 convolution layer, a batch normalization layer, and a ReLU activation function; The preprocessed cloud detection data set is subjected to multi-level feature extraction through four consecutive feature extraction blocks FEB, and the edge information of the features extracted by the feature extraction block FEB is extracted and enhanced through four edge learning blocks ELB to obtain multiple edge features. The multiple edge features obtained by the edge learning block ELB are fused through the feature fusion block FFB to obtain multi-scale edge features. The multi-scale edge features and the edge truth value establish a loss function to enhance the learning of edge features by the feature fusion block FFB. The features extracted by the feature extraction block FEB are restored through four consecutive feature recovery blocks FRB, and are spliced with the features extracted by the corresponding feature extraction block FEB, so as to finally realize cloud mask output or clear sky mask output; The loss function is specifically the sum of a focus loss function and an edge loss function; The focus loss function is defined as follows: Among them, N represents the number of categories, i represents the i-th category, and y i represents the true value of the i-th category, P i represents the predicted value of the network model, γ represents the weight, logP i Represents the logarithm of the network model prediction value; The marginal loss function is defined as follows: Among them, ∈ represents the Laplace smoothing term, y e Represents the marginal truth value, y e i represents the marginal truth value of the i-th category, p e i Represents the output value of the feature fusion block FFB of the i-th category, logp e i Represents the logarithm of the feature fusion block FFB output value of the i-th category.
2. A deep learning cloud detection method based on HY-1C satellite COCTS data according to claim 1, It is characterized in that The cloud detection data set in step S1 specifically includes spectral data, texture data, and geographic data of clouds.
3. A deep learning cloud detection method based on HY-1C satellite COCTS data according to claim 2, It is characterized in that The preprocessing includes converting the radiance of 443nm, 490nm, 565nm, 670nm and 865nm bands in the cloud detection data set into reflectance, and converting the radiance of 11μm and 12μm bands into brightness temperature.
4. A deep learning cloud detection method based on HY-1C satellite COCTS data according to claim 3, It is characterized in that The radiance of the 443nm, 490nm, 565nm, 670nm and 865nm bands in the cloud detection dataset is converted to reflectance using the following formula: Where λ is the wavelength, π is the circumference, Ref(λ) is the reflectance of the band, Rad(λ) is the radiance of the band, μ 0 is the cosine function value of the solar zenith angle, F 0 (λ) is the solar irradiance at the top of the atmosphere.
5. A deep learning cloud detection method based on HY-1C satellite COCTS data according to claim 4, Features ,The brightness temperature is obtained by constructing a lookup table that converts the radiance of the 11μm and 12μm bands into brightness temperature.
6. A deep learning cloud detection method based on HY-1C satellite COCTS data according to claim 5, It is characterized in that The lookup table for converting the radiance of the 11μm and 12μm bands to brightness temperature is constructed using the following formula. The specific process is as follows: Preset the brightness temperature BT, preset the wavelength λ within the band range, and calculate the blackbody radiance when the wavelength is λ: Where h is Planck's constant, c is the speed of light, and k is B is the Boltzmann constant, L λ is the blackbody radiance, L λ (λ, BT) is the blackbody radiance at wavelength λ and brightness temperature BT; Then calculate the radiance L corresponding to the brightness temperature BT in a wavelength range according to the following formula: Among them, f λ is the spectral response function of the band defined at wavelength λ, and dλ represents the integral over wavelength λ; The one-to-one correspondence between the radiance L and brightness temperature BT in different bands is obtained, and this correspondence is made into a lookup table. The brightness temperature BT value corresponding to the band radiance of the 11μm and 12μm bands in the cloud detection dataset is found according to the lookup table.
7. The deep learning cloud detection method based on HY-1C satellite COCTS data according to claim 1, It is characterized in that The preprocessing process also includes constructing classification labels in the cloud detection data set, wherein the classification labels include cloud and clear sky, wherein the classification label of clear sky is shown in the following formula: Where c is the clear sky state over the ocean, superscript o means "observation", superscript b means "prior background", x b The state vector representing the prior background, y o represents the observation vector, P(c) represents the probability of clear sky over the ocean, P(c|y o ,x b ) represents the probability of clear sky under given prior background state and observation conditions, P(y o |x b ,c) represents the probability of observations under given prior background conditions and clear sky conditions, P(x b |c) represents the background state probability under clear sky conditions, P(y o |x b ) is the probability of observation under a given prior background state, P(x b ) represents the probability of the prior background state; The pixel labels with the calculated Bayesian clear sky probability greater than 0.9 are labeled as clear sky pixels, and this label is added to the cloud detection dataset.
