Accumulated snow coverage estimation method based on residual feature fusion inversion

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a residual feature fusion inversion-based accumulated snow coverage estimation method, which comprises the following steps of: acquiring terrestrial satellite data and performing image data screening according to a screening condition; acquiring meteorological satellite data, preprocessing the meteorological satellite data and splicing the meteorological satellite data with the geographic elevation data; generating a binary accumulated snow map, calculating the accumulated snow coverage according to a definition method and a meteorological satellite resolution, and obtaining an accumulated snow coverage label; cutting the meteorological satellite image into a rectangle according to the accumulated snow coverage label; establishing an accumulated snow coverage estimation model based on deep learning; optimal experiment parameters are found through repeated experiment debugging and comparison; after the prediction is completed, obtaining respective accumulated snow coverage estimation results and splicing the accumulated snow coverage estimation results into a complete accumulated snow coverage estimation result; and the cloud shielding interference is effectively filtered by combining the generated snow coverage estimation result with the satellite data of the front and back time phases, so that the method has higher time resolution and reliability.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a snow cover estimation method based on residual feature fusion inversion. Background Art

[0002] In recent years, with the continuous development of image processing technology, machine learning technology has been successfully applied to remote sensing image detection and has significantly improved the accuracy of meteorological forecasts. Among them, image segmentation technology can divide each pixel in the image into its corresponding category, and regression technology can clarify the specific value of each pixel through fitting. Using these technologies can help improve the efficiency and accuracy of meteorological element detection, effectively alleviate the problem that other snow cover estimation methods are easily affected by complex terrain and underlying surface interference and have low accuracy in the northern plateau area, and improve the temporal resolution of snow cover estimation, thus having broad application prospects and high use value.

[0003] As a traditional method for estimating snow cover, the method based on linear regression and multispectral decomposition can estimate snow cover parameters more accurately in an environment with simple geological conditions. However, the linear regression algorithm has poor applicability, and the detection results for different data and different regions often vary greatly. The multispectral decomposition algorithm assumes that the same components of the pixels have the same spectral characteristics and the spectra are linearly added. Based on this assumption, the spectral decomposition method can effectively estimate the snow cover, which is prone to produce a large number of meaningless negative values. In addition, the above two methods are susceptible to greater interference in an environment with complex terrain, diverse land cover types and scattered snow distribution. There are also large errors in the estimation results during the transition period between snow melting and snow accumulation.

[0004] Artificial neural networks provide a more effective and accurate method for snow cover detection. Artificial neural networks mine the deep features of pixels with different snow cover ranges from the data of each band of satellite multispectral images, and establish a connection with the actual snow cover. In addition, artificial neural networks can easily combine auxiliary factors such as geographic elevation, slope and aspect. Random forest as a regression algorithm can also estimate snow cover very well, and because it adopts the idea of ​​random sampling for a single pixel, it has a strong generalization ability and can achieve good results for snow cover detection. However, the above two algorithms are all for mining the characteristics of different bands of a single pixel, ignoring the association between adjacent pixels. The estimated results not only have no spatial texture features but also have poor robustness.

[0005] The development of deep neural networks has provided an important method for accurate and effective feature extraction. Feature extraction is no longer limited to the features of a single pixel, but the receptive field is expanded to consider the features of adjacent pixels. ResNet_FSC uses the deep residual network ResNet as the backbone network for feature extraction, fully extracts the features of the input image, constructs a deep snow cover estimation method, and achieves results that are superior to artificial neural networks and other methods in a certain area in the north. However, due to the rich high-dimensional abstract features extracted by the deep ResNet, simple splicing and fusion cannot fully utilize the potential correlation and hierarchical information between these features, resulting in the model's performance in local details and global consistency. Therefore, the estimation results for the edge of the snow and sporadic snow still need to be further improved.

[0006] At present, the temporal resolution of snow cover estimation research has not reached 15 minutes, which is insufficient for monitoring heavy snowfall and snow melt changes that occur in a short period of time. Snow monitoring with high temporal resolution can not only provide a reference for hydrological monitoring, but also has important guiding significance for disaster prediction. Therefore, establishing a snow cover estimation model with high temporal resolution has important scientific and practical significance. Summary of the invention

[0007] The purpose of the present invention is to provide a snow cover estimation method based on residual feature fusion inversion to solve the problems raised in the above background technology.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions: A snow cover estimation method based on residual feature fusion inversion, the method comprising: Step 1: Based on the study area, obtain Landsat data and filter the image data according to the screening criteria; Step 2: Obtain the corresponding meteorological satellite data based on the screened land satellite data, and perform preprocessing and splicing of geographic elevation data; Step 3: Generate a binary snow map based on the filtered Landsat data and the SNOWMAP algorithm, calculate the snow cover according to the definition method and the meteorological satellite resolution, and obtain the snow cover label; The process of calculating snow cover according to the definition method and the meteorological satellite resolution and obtaining the snow cover label is as follows: Since the resolution of Landsat 8 remote sensing images is 30m, and the remote sensing data used in this paper is FY-4A / AGRI 2km resolution, one pixel on the FY-4A / AGRI remote sensing image corresponds to a 67x67 window on the Landsat8 remote sensing image. According to the binary snow map generated by the SNOWMAP algorithm, the snow cover calculation formula for a single pixel is: ; in, represents the number of snow-covered pixels in each window, Represents the total number of pixels contained in each grid; this reciprocating process can obtain the true value label of snow cover with a resolution of 2km; Step 4: Crop the meteorological satellite image into rectangles according to the snow cover label, which are used as the label and input for estimating snow cover; Step 5: Establish a snow cover estimation model based on deep learning; Step 6: Set the model's hyperparameters, total number of training times, learning rate, momentum parameter, and weight decay coefficient. After repeated experimental debugging and comparison, find the best experimental parameters to ensure the best snow cover estimation model. Step 7: Apply the trained optimal snow cover estimation model to test new meteorological satellite data. The new meteorological satellite data needs to be cut into different sizes. After the prediction is completed, the snow cover estimation results of each snow cover are obtained and then spliced ​​into a complete snow cover estimation result. Step 8: Overlay the generated snow cover estimation results onto the file with the shp terrain.

[0009] Preferably, the screening conditions for the land satellite data in step 1 include: 1) Cloud coverage is less than 10%; 2) The error of meteorological satellite shooting time shall not exceed 10 minutes.

[0010] Preferably, obtaining corresponding meteorological satellite data in step 2 includes: Obtain meteorological satellite data from the USGS official website and the Fengyun Satellite Remote Sensing Data Service Network.

[0011] Preferably, step 5 establishes a snow cover estimation model based on deep learning, including: The snow cover estimation model includes a coding layer, a feature fusion layer and a regression output layer; The encoding layer is composed of a Resnet backbone network, which is used to extract the features of the input data and obtain a feature image; The feature fusion layer consists of an upsampling module, an improved coordinated attention mechanism module and a feature refinement module; The regression output layer consists of multi-scale regression modules.

[0012] Preferably, the upsampling module in the feature fusion layer includes: The upsampling module includes 3*3 convolution, PiexlShuffle, 3*3 convolution, batch normalization and Relu activation function; The 3*3 convolution obtains the feature image of the input data. Then for the input feature map , where H is the height, W is the width, and C is the number of channels, then the 3*3 convolution operation in the upsampling module is: ; Among them, W represents the convolution kernel, b represents the bias; The PiexlShuffle, by rearranging the channel data into spatial dimensions, changes the number of feature map channels from Convert to The original number of channels, while upsampling the spatial resolution from H*W to Rh*rW: ; The input feature map , To ensure that the subsequent PiexShuffle is satisfied, r is the upsampling ratio; The batch normalization standardizes the feature distribution to accelerate convergence by performing batch normalization on the convolution results. The normalization parameter is set as and , then we get the formula: ; in , Respectively The mean and standard deviation of the channels, is the two-dimensional coordinate of the feature map pixel; The Relu activation function, the definition of the Relu activation function is according to the formula: ; In summary, the feature map obtained by the upsampling module is : .

[0013] Preferably, the improved coordinated attention mechanism module in the feature fusion layer includes: The improved coordinated attention mechanism module is composed of a feature matrix that is spliced ​​after maximum pooling and average pooling in the X direction, maximum pooling and average pooling in the Y direction, and then positionally encoded, and finally respectively multiplied by the initial feature matrix after learning the weights in the X direction and the Y direction through 3*1 convolution and 1*3 convolution; The maximum pooling and average pooling in the X direction, for the input feature map , where H is the height, W is the width, and C is the number of channels, then the maximum value pooling in the X direction for: ; Average pooling in the X direction for: ;in, Indicates the channel number; The maximum value pooling in the Y direction , Average Pooling for: ; ; Then the maximum value pooling and average pooling in the X direction, the maximum value pooling and average pooling in the Y direction are spliced ​​to obtain the spliced ​​data ; The position encoding is performed by subjecting the spliced ​​data P to 3*3 convolution, batch normalization, and h-swish activation function, and the formula is obtained: ; in, ; ; in, represents the batch normalization function, Represents a 3*3 convolution function; Then separate it into two features in the X and Y directions: ; ; in, represents the image features in the X direction, Represents the image features in the Y direction; Then will The exchange dimension is , and finally learned through 3*1 convolution and 1*3 convolution respectively In the X direction, The weight in the Y direction is multiplied with the initial feature matrix; Therefore, the feature map calculated by the improved attention mechanism is obtained : ; in, , represents the Sigmoid function.

[0014] Preferably, the feature refinement module in the feature fusion layer includes: The feature refinement module is composed of feature matrices that are respectively subjected to 3*3 convolution, batch normalization, Relu activation function and 1*1 convolution, batch normalization, Relu activation function, and then spliced ​​along the channel direction, and then maximum pooling and average pooling are performed respectively, and then the two groups of pooling results are spliced ​​and 1*1 convolution and Sigmoid activation function are performed to obtain channel weights, and finally the channel weights are multiplied by the input feature matrix to assign high weights to feature maps with an importance higher than a threshold, and then the number of channels is adjusted through 1*1 convolution to discard feature maps with an importance lower than the threshold; For the input feature map , where H is the height, W is the width, and C is the number of channels, then the average pooling in the feature refinement module , Maximum Pooling The calculation formula is: ; ; According to the feature map obtained by the feature refinement module : ; The calculation formula of F is: .

[0015] Preferably, the multi-scale regression module in the regression output layer includes: The multi-scale regression module adjusts the number of channels by 1*1 convolution respectively through the feature maps of three scales, and then restores the resolution through the upsampling module step by step to obtain the multi-scale fusion features. Finally, the multi-scale fusion features are respectively subjected to two 3*3 convolutions, batch normalization, Relu activation function and 1*1 convolution and added to complete the mapping of snow parameters; For the input feature map , , , where H is the height, W is the width, and C is the number of channels, then the output feature map obtained by the multi-scale regression module is for: ; in, .

[0016] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the above-mentioned snow cover estimation method based on residual feature fusion inversion.

[0017] A computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, the steps in the above-mentioned snow cover estimation method based on residual feature fusion inversion are implemented.

[0018] The present invention proposes a snow cover estimation method based on deep learning, which can accurately identify snow-covered areas and estimate snow cover; a snow cover estimation model is established after preprocessing, cutting and sampling, and overlapping sampling of original satellite data, and the optimal model parameters are obtained by continuously training the model, and finally the snow cover can be estimated from satellite images; before training the model, normalizing the preprocessed data is performed, which is more conducive to improving the model convergence speed and can prevent gradient explosion or gradient disappearance caused by inconsistent data scales; the method has a good effect on snow area segmentation and coverage estimation and is simple to use. After the model training is completed, the snow cover estimation time of a single satellite image only takes one to two seconds, and the method has a good application prospect.

[0019] Compared with the prior art, the beneficial effects achieved by the present invention are: The present invention can 1. not require additional hardware equipment; 2. provide almost real-time snow cover estimation results; 3. have higher estimation accuracy than traditional methods; 4. only need to adjust a small number of hyperparameters of the deep learning model during the experimental stage, and provide a snow cover estimation method based on deep learning. According to labels and input data, the deep learning model automatically learns the characteristics of snow cover in different ranges and saves the best model parameters. After the best model is saved, it only needs to call the model to realize snow retrieval and coverage estimation, which is convenient for user operation.

[0020] Therefore, the present invention can segment the snow area more accurately and estimate the corresponding snow coverage, thereby increasing the robustness of the snow coverage estimation model, alleviating the poor accuracy of the traditional snow coverage estimation model under complex geological conditions, and effectively increasing the detection rate and accuracy of fragmented snow, thereby obtaining valuable and quantitatively describable snow coverage monitoring information. The result of continuous improvement can be achieved. As time goes by, Fengyun satellite data will become more abundant, so expanding more data can effectively improve the generalization ability and estimation accuracy of the model. The method can quickly obtain the snow coverage estimation result after the real-time Fengyun satellite data is released, and can effectively filter out cloud shielding interference by combining the satellite data of the previous and next phases, and can therefore have higher temporal resolution and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 It is a flow chart of a snow cover estimation method based on residual feature fusion inversion of the present invention; Figure 2is a true color schematic diagram of Landset8 image in this embodiment; Figure 3 This is a schematic diagram of the FY-4A full-staff disk image in this embodiment; Figure 4 This is a schematic diagram of the image after correction and preprocessing of the FY-4A full disk data in the embodiment; Figure 5 This is a schematic diagram of a label in this embodiment; Figure 6 This is a schematic diagram of inputting data in this embodiment; Figure 7 Schematic diagram of the structure of the snow cover estimation model in this embodiment; Figure 8 This is a schematic diagram of the improved coordinated attention mechanism module in this embodiment; Fig. 9 Schematic diagram of the structure of the up-sampling module in this embodiment; Fig.10 Schematic diagram of the structure of the feature refinement module in this embodiment; Fig.11 Schematic diagram of the structure of the regression output module in this embodiment; Fig.12 This is a schematic diagram of the snow cover estimation result in this embodiment; Fig.13 Detailed parameter diagram of each convolutional layer in this embodiment. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] See also Figure 1-Figure 13 , the present invention provides a technical solution:

[0024] Embodiment 1: Step 1: According to the study area, select the Landset8 land satellite images with cloud coverage less than 10% and the error between the time and the shooting time of the meteorological satellite is no more than 10 minutes on the GEE platform to download, and obtain the following Figure 2 30-meter resolution Landsat image shown.

[0025] Step 2: Based on the selected land satellite image data, download the FY4A full-disk meteorological satellite data of the corresponding time range, such as Figure 3Then, the gdal library function gdal.Warp in Python is used to correct and preprocess the full disk data to obtain the meteorological satellite image of the entire study area, as shown in Figure 4 As shown in the figure, the preprocessed FY4A meteorological satellite image is spliced ​​with the geographic elevation data of the study area along the channel direction through the gdal library in python.

[0026] Step 3: The normalized snow cover index (NDSI) is obtained based on the filtered data of the third and sixth bands of the Landsat satellite. Pixels with NDSI greater than or equal to 0.4 are regarded as snow pixels, and pixels with NDSI less than 0.4 are regarded as snow-free pixels. The interference of water bodies on snow recognition is eliminated by judging whether the reflectivity of the fifth band is greater than 0.11, and a binary snow map is obtained. In this embodiment, the FY4A 2KM resolution image is used as the input data. Therefore, the ratio of snow pixels to all pixels in the corresponding grid of the 2KM resolution binary snow map is calculated according to the definition of snow cover, and the snow cover label can be obtained.

[0027] Step 4: Overlay and crop the FY4A image after stitching the geographic elevation data and the snow cover label into rectangles of the same size, which are used as the input and label of the snow cover estimation model respectively. The input data true color image and label are as follows: Figure 4 , Figure 5 As shown; in this embodiment, the value interval of the label image is [0,1], and the size after cropping is 64×64; then execute step 5.

[0028] Step 5: Use the Pytorch deep learning framework and Python language to establish a deep learning-based snow cover estimation model; The snow cover estimation model is constructed as follows: Figure 7 As shown; The model mainly consists of encoding layer, feature fusion layer, and regression output layer; The coding layer consists of 4 residual blocks. The first residual block consists of a downsampled residual module and two basic residual modules. The second residual block consists of a downsampled residual module and three basic residual modules. The third residual block consists of a downsampled residual module and five basic residual modules. The fourth residual block consists of a downsampled residual module and two basic residual modules. The downsampling residual module consists of a main branch and a residual branch. The main branch consists of 3*3 convolution Conv1, batch normalization, Relu activation function, 3*3 convolution Conv2, batch normalization, 1*1 convolution Conv3 and batch normalization. The residual branch does not perform any operation and directly adds the original input of the downsampling module to the result of the main branch. The basic residual module consists of a main branch and a residual branch. The main branch consists of 3*3 convolution Conv4, batch normalization, Relu activation function, 3*3 convolution Conv5, batch normalization. The residual branch does not perform any operation and directly adds the original input of the downsampling module to the result of the main branch. The feature fusion layer consists of an upsampling module, an improved coordinated attention mechanism module, and a feature refinement module; The upsampling module consists of 3*3 convolution Conv6, PiexlShuffle, 3*3 convolution Conv7, batch normalization, and Relu activation function; Improved coordinated attention mechanism: The feature matrix is ​​concatenated after maximum pooling and average pooling in the X direction, maximum pooling and average pooling in the Y direction, and then encoded by 3*3 convolution Conv8, batch normalization and h-swish activation function, and then separated into vectors in the horizontal and vertical directions. Then the vectors in the horizontal and vertical directions are respectively subjected to 1*3 convolution Conv9 and 3*1 convolution Conv10, and then nonlinear mapping is completed through the Sigmoid function, and finally multiplied with the input feature matrix; The feature refinement module first passes the input features through 3*3 convolution Conv11, batch normalization, Relu activation function and 1*1 convolution Conv12, batch normalization, Relu activation function, and then concatenates the two along the channel direction and performs maximum pooling and average pooling respectively. Then, the pooling results are concatenated along the channel direction and passed through 1*1 convolution Conv13, Relu activation function, Conv14, Sigmoid activation function, and finally a 1*1 convolution Conv15 to eliminate redundant channels; The input of the regression output layer is the feature maps of three different dimensions: small scale, medium scale and large scale. First, the number of channels is adjusted by 1*1 convolution Conv16, and then the large-scale feature map is restored to its resolution by the upsampling module, and it is spliced ​​with the medium-scale feature map along the channel direction; then the spliced ​​feature map is restored to its resolution by the upsampling module, and it is spliced ​​with the small-scale feature map along the channel direction; then, the spliced ​​feature map is restored to the final resolution by the upsampling module; finally, the feature map is subjected to two 3*3 convolution Conv17, batch normalization, Relu activation function and 1*1 convolution Conv18, and the two results are added to obtain the final snow cover inversion result; Among them, the detailed parameters of each convolution layer are shown in the attached Fig.13 Detailed parameter diagram of each convolutional layer; Step 6: Set the model's hyperparameters, total number of training times, learning rate, momentum parameter, and weight decay times. After repeated adjustments and comparative experiments, the optimal parameters are obtained to obtain the optimal snow cover estimation model. In this example, the specific model parameter settings are as follows: the total training epoch is 200, the optimizer is Adam, the initial value of the learning rate is set to 0.01, and the learning rate decays to one tenth of the previous value after every 50 epochs, while maintaining the model convergence speed, it will not cause gradient vanishing or gradient explosion. The momentum parameter momentum is set to 0.9 to speed up model convergence, and the weight decay parameter is set to 1e-5 to suppress model overfitting. After experimental comparison, the above parameters are set to the best parameters.

[0029] Before model training, the present invention needs to perform normalization preprocessing data to ensure that the data of each channel is at the same level to avoid gradient explosion. In this embodiment, maximum and minimum normalization is used as the normalization method; Step 7: Apply the trained optimal snow cover estimation model to test new meteorological satellite data. The new meteorological satellite data needs to be superimposed with the geographic elevation data in the channel direction after preprocessing, and then cropped to 64*64 size for prediction. After the prediction is completed, the respective snow cover estimation results are obtained and then spliced ​​into a snow cover estimation result for the complete study area, such as Fig.11 As shown; different colors are set according to different estimated snow coverage ranges, where white is the background and the colors increase from light to dark snow coverage; The processing flow of the optimal snow cover estimation model is as follows: input a 64*64*8 matrix, output a 64*32*32 feature map after feature extraction by the first residual block, output a 128*16*16 feature map after feature extraction by the second residual block, output a 256*8*8 feature map after feature extraction by the third residual block, and output a 512*4*4 feature map after feature extraction by the fourth residual block; Then, the feature map output by the second residual block is passed through the upsampling module to output a 128*32*32 feature map and is spliced ​​with the feature map output by the first residual block along the channel direction, and the first-level small-scale fusion feature of 96*32*32 is obtained through the improved coordinated attention mechanism and feature refinement module in turn; the feature map output by the third residual block is passed through the upsampling module to output a 256*16*16 feature map and is spliced ​​with the feature map output by the second residual block along the channel direction, and the first-level medium-scale fusion feature of 192*16*16 is obtained through the improved coordinated attention mechanism and feature refinement module in turn; the feature map output by the fourth residual block is passed through the upsampling module to output a 512*8*8 feature map and is spliced ​​with the feature map output by the third residual block along the channel direction, and the first-level large-scale fusion feature of 384*8*8 is obtained through the improved coordinated attention mechanism and feature refinement module in turn; the first-level medium-scale fusion feature is passed through the upsampling module to output a 192*32*32 feature map and is spliced ​​along the channel direction. The first-level large-scale fusion feature is spliced ​​with the first-level small-scale fusion feature, and the second-level small-scale fusion feature of 144*32*32 is obtained in turn through the improved coordinated attention mechanism and the feature refinement module; the first-level large-scale fusion feature is subjected to the upsampling module to output a 384*16*16 feature map and is spliced ​​with the first-level medium-scale fusion feature along the channel direction, and the second-level medium-scale fusion feature of 288*16*16 is obtained in turn through the improved coordinated attention mechanism and the feature refinement module; the second-level medium-scale fusion feature is subjected to the upsampling module to output a 288*32*32 feature map and is spliced ​​with the second-level small-scale fusion feature along the channel direction, and the third-level small-scale fusion feature of 216*32*32 is obtained in turn through the improved coordinated attention mechanism and the feature refinement module; the second-level medium-scale fusion feature and the first-level large-scale fusion feature are respectively subjected to the improved coordinated attention mechanism and the feature refinement module to obtain the third-level medium-scale fusion feature of 144*16*16 and the second-level large-scale fusion feature of 192*8*8; The third-level small-scale fusion features, the third-level medium-scale fusion features, and the second-level large-scale fusion features are used as the input features of the regression output layer, and the number of feature channels is adjusted to 144*32*32, 144*16*16, and 144*8*8 respectively through 1*1 convolution; the 144*8*8 large-scale features are output as 72*16*16 features through the upsampling module and are concatenated with the 144*16*16 medium-scale features along the channel direction to obtain a 216*16*16 feature map, which is then passed through The upsampling module outputs a small-scale feature map of 108*32*32 and concatenates it with the small-scale fusion feature of 144*32*32 to obtain a feature map of 252*32*32; then the upsampling module obtains a final small-scale feature map of 126*64*64, and then it undergoes 1*1 convolution, batch normalization, Sigmoid activation function and two 3*3 convolutions, batch normalization, and Sigmoid activation function to complete nonlinear mapping, and then the two results are added together to obtain the final snow cover inversion result.

[0030] The improved attention mechanism in the snow cover estimation method proposed in the present invention strengthens the feature differences between snow-covered and snow-free pixels and pixels with different snow coverage, and provides a more prominent feature map for the identification of snow pixels and the estimation of snow coverage; the feature refinement module can further extract and optimize the spatial and spectral characteristics of the snow-covered area, effectively eliminate the interference of noise and non-snow pixels, remove redundant information, and at the same time improve the ability to distinguish different snow coverage degrees, thereby providing a guarantee for the robustness of the model in complex scenarios; the feature fusion layer corresponds to the encoding layer, and the downsampled multi-scale feature maps are fused and gradually restored to the target resolution; the regression output layer finally fully aggregates the fused features of each scale for comprehensive regression prediction, and finally generates a high-precision snow cover estimation result, with the R2 score and root mean square error RMSE on the test set reaching 0.6744 and 0.0663, respectively.

[0031] Step 8: Stitch the cropped snow cover estimation results back to their original size and overlay them on the file with the shp terrain, as shown in Fig.12 shown.

[0032] Therefore, the present invention can retrieve snow pixels more accurately and estimate the corresponding snow coverage. The receptive field is expanded by the convolution-based feature extraction method, the accuracy and stability of snow coverage estimation are improved, the precision interference caused by complex geological conditions is alleviated, and the sensitivity to sporadic snow is increased. By using FY4A meteorological satellite image data, the temporal resolution of snow coverage estimation is improved, and the interference caused by cloud cover is effectively alleviated. Finally, the snow pixels are successfully retrieved and the snow coverage is estimated, and valuable and quantitatively describable snow parameters are obtained.

[0033] Embodiment 2: The computer-readable storage medium of this embodiment stores a computer program thereon, and when the program is executed by a processor, the steps in a snow cover estimation method based on residual feature fusion inversion in embodiment 1 are implemented.

[0034] The computer-readable storage medium of this embodiment may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal; the computer-readable storage medium of this embodiment may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash memory card, etc. equipped on the terminal; further, the computer-readable storage medium may also include both an internal storage unit of the terminal and an external storage device.

[0035] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or is to be output.

[0036] Embodiment 3: The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of a snow cover estimation method based on residual feature fusion inversion in Example 1 are implemented.

[0037] In this embodiment, the processor may be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, readily available programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0038] Those skilled in the art will appreciate that the contents disclosed in the embodiments may be provided as methods, systems, or computer program products. Therefore, the present solution may take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware. Moreover, the present solution may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program codes.

[0039] The present solution is described with reference to the method according to the embodiment of the present solution and the flowchart and / or block diagram of the computer program product. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions; these computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 one or more processes and / or methods Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0040] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 one or more processes and / or methods Figure 1 A function specified in one or more boxes.

[0041] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 one or more processes and / or methods Figure 1 The steps for the functions specified in one or more boxes.

[0042] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0043] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A snow cover estimation method based on residual feature fusion inversion, characterized by: The method comprises: Step 1: Based on the study area, obtain Landsat data and filter the image data according to the screening criteria; Step 2: Obtain the corresponding meteorological satellite data based on the screened land satellite data, and perform preprocessing and splicing of geographic elevation data; Step 3: Generate a binary snow map based on the filtered Landsat data and the SNOWMAP algorithm, calculate the snow cover according to the definition method and the meteorological satellite resolution, and obtain the snow cover label; The process of calculating snow cover according to the definition method and the meteorological satellite resolution and obtaining the snow cover label is as follows: Since the resolution of Landsat 8 remote sensing images is 30m, and the remote sensing data used in this paper is FY-4A / AGRI 2km resolution, one pixel on the FY-4A / AGRI remote sensing image corresponds to a 67x67 window on the Landsat8 remote sensing image. According to the binary snow map generated by the SNOWMAP algorithm, the snow cover calculation formula for a single pixel is: ； in, represents the number of snow-covered pixels in each window, Represents the total number of pixels contained in each grid; this reciprocating process can obtain the true value label of snow cover with a resolution of 2km; Step 4: Crop the meteorological satellite image into rectangles according to the snow cover label, which are used as the label and input for estimating snow cover; Step 5: Establish a snow cover estimation model based on deep learning; Step 6: Set the model's hyperparameters, total number of training times, learning rate, momentum parameter, and weight decay coefficient. After repeated experimental debugging and comparison, find the best experimental parameters to ensure the best snow cover estimation model. Step 7: Apply the trained optimal snow cover estimation model to test new meteorological satellite data. The new meteorological satellite data needs to be cut into different sizes. After the prediction is completed, the snow cover estimation results of each snow cover are obtained and then spliced ​​into a complete snow cover estimation result. Step 8: Overlay the generated snow cover estimation results onto the file with the shp terrain.

2. The snow cover estimation method based on residual feature fusion inversion according to claim 1 is characterized in that: The screening conditions of the land satellite data in step 1 include: 1) Cloud coverage is less than 10%; 2) The error of meteorological satellite shooting time shall not exceed 10 minutes.

3. The snow cover estimation method based on residual feature fusion inversion according to claim 1 is characterized in that: The step 5 establishes a snow cover estimation model based on deep learning, including: The snow cover estimation model includes a coding layer, a feature fusion layer and a regression output layer; The encoding layer is composed of a Resnet backbone network, which is used to extract the features of the input data and obtain a feature image; The feature fusion layer consists of an upsampling module, an improved coordinated attention mechanism module and a feature refinement module; The regression output layer consists of multi-scale regression modules.

4. The snow cover estimation method based on residual feature fusion inversion according to claim 3 is characterized in that: The upsampling module in the feature fusion layer includes: The upsampling module includes 3*3 convolution, PiexlShuffle, 3*3 convolution, batch normalization and Relu activation function; The 3*3 convolution obtains the feature image of the input data. Then for the input feature map , where H is the height, W is the width, and C is the number of channels, then the 3*3 convolution operation in the upsampling module is: ； Among them, W represents the convolution kernel, b represents the bias; The PiexlShuffle, by rearranging the channel data into spatial dimensions, changes the number of feature map channels from Convert to The original number of channels, while upsampling the spatial resolution from H*W to Rh*rW: ； The input feature map , To ensure that the subsequent PiexShuffle is satisfied, r is the upsampling ratio; The batch normalization standardizes the feature distribution to accelerate convergence by performing batch normalization on the convolution results. The normalization parameter is set as and , then we get the formula: ； in , Respectively The mean and standard deviation of the channels, is the two-dimensional coordinate of the feature map pixel; The Relu activation function, the definition of the Relu activation function is according to the formula: ； In summary, the feature map obtained by the upsampling module is : 。 5. The snow cover estimation method based on residual feature fusion inversion according to claim 3 is characterized in that: The improved coordinated attention mechanism module in the feature fusion layer includes: The improved coordinated attention mechanism module is composed of a feature matrix that is spliced ​​after maximum pooling and average pooling in the X direction, maximum pooling and average pooling in the Y direction, and then positionally encoded, and finally respectively multiplied by the initial feature matrix after learning the weights in the X direction and the Y direction through 3*1 convolution and 1*3 convolution; The maximum pooling and average pooling in the X direction, for the input feature map , where H is the height, W is the width, and C is the number of channels, then the maximum value pooling in the X direction for: ； Average pooling in the X direction for: ;in, Indicates the channel number; The maximum value pooling in the Y direction , Average Pooling for: ； ； Then the maximum value pooling and average pooling in the X direction, the maximum value pooling and average pooling in the Y direction are spliced ​​to obtain the spliced ​​data ; The position encoding is performed by subjecting the spliced ​​data P to 3*3 convolution, batch normalization, and h-swish activation function, and the formula is obtained: ； in, ; ; in, represents the batch normalization function, Represents a 3*3 convolution function; Then separate it into two features in the X and Y directions: ； ； in, represents the image features in the X direction, Represents the image features in the Y direction; Then will The exchange dimension is , and finally learned through 3*1 convolution and 1*3 convolution respectively In the X direction, The weight in the Y direction is multiplied with the initial feature matrix; Therefore, the feature map calculated by the improved attention mechanism is obtained : ； in, , represents the Sigmoid function.

6. The snow cover estimation method based on residual feature fusion inversion according to claim 3 is characterized in that: The feature refinement module in the feature fusion layer includes: The feature refinement module is composed of feature matrices that are respectively subjected to 3*3 convolution, batch normalization, Relu activation function and 1*1 convolution, batch normalization, Relu activation function, and then spliced ​​along the channel direction, and then maximum pooling and average pooling are performed respectively, and then the two groups of pooling results are spliced ​​and 1*1 convolution and Sigmoid activation function are performed to obtain channel weights, and finally the channel weights are multiplied by the input feature matrix to assign high weights to feature maps with an importance higher than a threshold, and then the number of channels is adjusted through 1*1 convolution to discard feature maps with an importance lower than the threshold; For the input feature map , where H is the height, W is the width, and C is the number of channels, then the average pooling in the feature refinement module , Maximum pooling The calculation formula is: ； ； According to the feature map obtained by the feature refinement module : ； The calculation formula of F is: .

7. The snow cover estimation method based on residual feature fusion inversion according to claim 3 is characterized in that: The multi-scale regression module in the regression output layer includes: The multi-scale regression module adjusts the number of channels by 1*1 convolution respectively through the feature maps of three scales, and then restores the resolution through the upsampling module step by step to obtain the multi-scale fusion features. Finally, the multi-scale fusion features are respectively subjected to two 3*3 convolutions, batch normalization, Relu activation function and 1*1 convolution and added to complete the mapping of snow parameters; For the input feature map , , , where H is the height, W is the width, and C is the number of channels, then the output feature map obtained by the multi-scale regression module is for: ； in, .

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a snow cover estimation method based on residual feature fusion inversion as described in any one of claims 1 to 7 are implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the steps in the snow cover estimation method based on residual feature fusion inversion as described in any one of claims 1-7 are implemented.

Images