A snow cover monitoring method based on RepFNet network and spatiotemporal filtering

Abstract

The present invention relates to the field of remote sensing image processing and meteorological monitoring technology, and specifically to a snow cover monitoring method based on a RepFNet network. The method comprises the following steps: acquiring FY-4A image data, Landsat satellite image data, and geographic auxiliary data; constructing a data set suitable for deep learning training; generating a high-resolution snow cover image; constructing a RepFNet network model with an encoder-decoder architecture, integrating an improved feature extraction module, a dynamic upsampling module, and an adaptive graph channel attention module; determining an optimal module combination and parameter configuration; training the model using an ADAM optimizer combined with a dynamic learning rate strategy; optimizing model performance by adjusting hyperparameters and conducting comparative experiments; and accurately inverting the snow cover of a study area using the trained RepFNet model, and performing cloud removal processing in combination with a spatiotemporal filtering method to ensure the accuracy and stability of snow monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing and meteorological monitoring, and in particular to a snow cover monitoring method based on a RepFNet network and spatiotemporal filtering. Background Art

[0002] Snow is an important physical element on the Earth's surface, with unique characteristics. It is widely distributed in different climatic zones and has a significant impact on global environmental and climate change research. Fractional snow cover (FSC) is calculated by calculating the percentage of snow within the pixels of a remote sensing imager and provides a more accurate estimate of the snow area than binary snow identification. Traditional snow monitoring methods mainly rely on ground observations and manual field detection, but these methods face many challenges in areas with high altitudes and complex terrain, such as high cost, low efficiency, and difficulty covering large areas. With the development of remote sensing technology, its effective way of perceiving large-scale snow information is also more important for achieving more accurate snow cover.

[0003] In recent years, deep learning technology has made significant progress in image processing and pattern recognition, providing new insights for snow cover inversion. Deep learning models can automatically learn complex features in input data, demonstrating enhanced feature extraction capabilities and robustness. However, existing deep learning-based FSC inversion methods mostly rely on traditional convolutional neural network architectures. These architectures often suffer from information loss due to the increased number of network layers when processing high-resolution remote sensing imagery, compromising inversion accuracy. Furthermore, existing methods fail to fully leverage the strengths of each data source when fusing multi-source data (such as satellite imagery and geographic elevation data), resulting in inaccurate and unreliable inversion results. Therefore, when using remote sensing data for Earth environmental monitoring, deep learning models significantly outperform traditional models, providing relatively accurate snow cover predictions across different regions and seasons.

[0004] Traditional snow cover monitoring methods face two major bottlenecks in complex mountainous terrain. First, the long revisit periods of most satellites make it difficult to capture rapid changes in snow cover (such as snowmelt and snowfall events), limiting dynamic monitoring capabilities. Furthermore, due to the variable climate in mountainous areas, single-phase data are susceptible to cloud contamination, leading to data loss and accumulated inversion errors. Therefore, a snow cover inversion method based on the RepFNet network and spatiotemporal filtering was developed. This method, which efficiently integrates multi-source data (including remote sensing imagery, meteorological data, and terrain data) and removes cloud processing to generate daily snow cover monitoring products, has important practical significance for improving the real-time, accuracy, and robustness of snow cover monitoring, especially in complex mountainous terrain. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of existing technologies in snow cover monitoring and cloud removal in complex mountainous terrain. This method provides a snow cover monitoring method based on the RepFNet network and spatiotemporal filtering. This method uses high-temporal-resolution FY-4A / AGRI meteorological satellite data, combines the efficient feature extraction capabilities of a deep learning model with a spatiotemporal filtering strategy, and improves the accuracy and stability of snow cover inversion.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A snow cover monitoring method based on RepFNet network and spatiotemporal filtering, the method comprising:

[0008] S100: Acquire FY-4A Advanced Geosynchronous Radiation Imager data in good weather conditions, and simultaneously acquire Landsat satellite imagery data with a time difference of no more than 10 minutes to ensure data timeliness. Perform band fusion and resampling operations on the Landsat data to match the spatial resolution of the FY-4A data. Acquire and stitch digital elevation model (DEM) data of the study area, and calculate slope and aspect information based on the DEM as auxiliary terrain factors for snow cover estimation.

[0009] S200: spatially register the pre-processed FY-4A meteorological satellite data, Landsat data, and DEM-derived terrain information, and crop them into 64×64 pixel image blocks to construct a dataset suitable for deep learning training;

[0010] S300 uses the SNOMAP algorithm to identify snow cover in Landsat data using binarization. The snow cover is calculated by combining the Normalized Difference Snow Index (NDSI) with the resolution of meteorological satellites to generate a high-resolution snow cover image, which serves as label data for the deep learning model.

[0011] S400: Using the PyTorch framework, we built and implemented a RepFNet network model with an encoder-decoder architecture. This model integrates an improved feature extraction module, a dynamic upsampling module (DySample), and an adaptive graph channel attention (AGCA) module to enhance the detection capability of snow-covered areas and improve inversion accuracy.

[0012] S500: Deeply optimize and adjust the parameters of the RepFNet network model, adjust key network components such as the feature extraction module and attention mechanism module, and determine the optimal module combination and parameter configuration through extensive experiments to improve the stability and generalization ability of the model;

[0013] S600, designed a new loss function suitable for FSC inversion, and used the ADAM optimizer with a dynamic learning rate strategy for model training. Model performance was optimized through repeated hyperparameter adjustments and comparative experiments. Finally, the trained RepFNet model was used to accurately invert snow cover in the study area. Cloud removal was performed using a spatiotemporal filtering method to ensure accurate and stable snow cover monitoring.

[0014] Preferably, FY-4A / AGRI meteorological satellite data is obtained from the Fengyun Satellite Remote Sensing Data Service Network; land satellite data is obtained from the official website of the United States Geological Survey (USGS); and digital elevation data is obtained from the official website of the Geospatial Data Cloud.

[0015] Preferably, calculating the slope and aspect information based on the digital elevation model in S100 includes:

[0016] Using DEM data, slope and aspect data are calculated using Python scripts to generate terrain auxiliary data layers that are consistent with the FY-4A and Landsat data space, providing terrain feature support for subsequent model training.

[0017] Preferably, S200 includes:

[0018] ENVI 5.3 software was used to define a 64×64 pixel cropping region based on the study area and generate a vector file. FY-4A meteorological satellite data, Landsat data, and geographic auxiliary data (slope and aspect) were loaded separately. Using the Landsat data as the reference data, spatial alignment was performed and each data set was cropped sequentially.

[0019] Preferably, generating a high-resolution snow cover image in S300 includes:

[0020] The normalized difference snow index NDSI is determined according to the formula:

[0021]

[0022] Band 2 is the green light band of Landsat 8 data, and Band 6 is the shortwave infrared band of Landsat 8 data;

[0023] The Landsat data were converted into a binary snow map by selecting the NDSI threshold and then resampled to ensure spatial consistency with the FY-4A data.

[0024] The snow cover within each pixel is calculated based on the number of snow pixels and the total number of pixels to generate a snow cover map.

[0025] Preferably, the RepFNet network model in S400 includes: an improved feature extraction module, a dynamic upsampling module and an adaptive graph channel attention module;

[0026] The network model constructed by the present invention performs feature extraction by improving the RepVGG module, proposes a dynamic residual weight driven by local variance, and realizes the adaptive adjustment of the branch ratio of this module. In order to gradually restore the high-dimensional features output by the feature extraction module to the same spatial resolution as the input data, the present invention also introduces a decoding module on this basis. The decoding module adopts a step-by-step upsampling method, and gradually restores the spatial information of the feature map through the alternating combination of convolutional layers and upsampling layers. The dynamic sampling module DySample is used for upsampling, and the upsampling strategy is adaptively adjusted. In addition, the AGCA module is introduced, which can adaptively adjust the weights of the feature channels and capture the global dependencies between features.

[0027] Preferably, the in-depth optimization and parameter adjustment in S500 include:

[0028] The threshold of network iterations was set to 200 to ensure that the model converged within a limited number of iterations. The initial learning rate was set to 0.001, and the ADAM optimizer was used to update the model parameters.

[0029] After every 10 iterations, the learning rate is dynamically adjusted with a decay coefficient of 0.8 to accelerate model convergence and prevent overfitting.

[0030] Preferably, in S600, the trained RepFNet model is used to accurately invert the snow cover of the study area, including:

[0031] The model accuracy was verified using the test data set, and the evaluation indicators included the coefficient of determination, root mean square error, correlation coefficient, Kappa coefficient, and explained variance score;

[0032] Among them, the coefficient of determination for:

[0033]

[0034] The root mean square error RMSE is:

[0035]

[0036] Correlation coefficient for:

[0037]

[0038] The Kappa coefficient is:

[0039]

[0040] The explained variance score EVS is:

[0041]

[0042] in, represents the true value, represents the predicted value; represents the mean of the true values, represents the mean of the predicted values, represents the observation accuracy, that is, the amount by which the predicted values match the true values. represents random accuracy, which is the proportion of the model calculated based on the probability of random classification matching the actual category, and n represents the size of the sample;

[0043] Among them, when the coefficient of determination, EVS, correlation coefficient and Kappa value are closer to 1, the model fits the data better; when the RMSE is closer to 0, the model inversion accuracy is higher.

[0044] Preferably, S600 combines a spatiotemporal filtering method to perform cloud removal processing to ensure the accuracy and stability of snow monitoring, including:

[0045] Deploy the trained RepFNet model to the study area data and output the snow cover inversion results;

[0046] The data in the study area are subjected to multi-temporal declouding processing, and the declouding is combined with time series data to monitor the changes in snow cover and ensure the stability and timeliness of the inversion results.

[0047] The computer device described in the present invention includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, the steps of the snow cover inversion method based on the RepFNet network are implemented.

[0048] The computer-readable storage medium of the present invention stores a computer program, and when the computer program is executed by a processor, the steps of the snow cover inversion method based on the RepFNet network are implemented.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The present invention does not require additional hardware equipment. It combines FY-4A / AGRI meteorological satellite data, Landsat land satellite data and terrain auxiliary data (slope, aspect, DEM) to achieve deep fusion of multi-source data and improve the accuracy of snow cover inversion under complex terrain. The improved RepVGG module is used to construct the RepFNet network, which integrates the local variance-driven dynamic residual weight strategy to improve the robustness and adaptability of feature extraction, especially in variable climate conditions and complex backgrounds. The dynamic upsampling module DySample and adaptive global channel attention are introduced to achieve the best performance. The force module AGCA not only enhances the recognition accuracy of snow edge areas and fragmented snow areas, but also adaptively adjusts the upsampling strategy and feature channel weights to capture the spatial details and global dependencies of the snow area, avoiding the problems of over-smoothing and boundary blurring in traditional methods; combining declouding processing with high temporal resolution reconstruction strategy, through multi-temporal declouding technology, it effectively solves the problem of data missing caused by cloud obscuration, ensures the continuity and integrity of the inversion results in space and time, and realizes large-scale, real-time, and high-precision snow cover inversion, providing more reliable data support for snow monitoring, climate research, and disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] 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:

[0052] Figure 1 Schematic diagram of the process of the present invention.

[0053] Figure 2 This is the pre-processed FY-4A / AGRI image data in the embodiment of the present invention.

[0054] Figure 3 This is the landsat8 data in the embodiment of the present invention.

[0055] Figure 4 This is the image data obtained by splicing DEM elevation data in an embodiment of the present invention.

[0056] Figure 5 Schematic diagram of a label in an embodiment of the present invention.

[0057] Figure 6 Schematic diagram of input data in an embodiment of the present invention.

[0058] Figure 7 Schematic diagram of the structure of the snow cover estimation model in an embodiment of the present invention.

[0059] Figure 8 Schematic diagram of a feature extraction network in an embodiment of the present invention.

[0060] Figure 9 Schematic diagram of the structure of the upsampling module in an embodiment of the present invention.

[0061] Figure 10 Schematic diagram of the structure of the attention mechanism module in an embodiment of the present invention.

[0062] Figure 11 Schematic diagram of the high temporal resolution snow cover inversion result and cloud removal result in an embodiment of the present invention.

[0063] Figure 12 This is a schematic diagram of the daily product that is finally successfully constructed in an embodiment of the present invention.

[0064] Figure 13 This is a comprehensive and clear picture of the final daily product in the embodiment of the present invention. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0066] See also Figures 1-13 , the present invention provides a technical solution:

[0067] Example 1:

[0068] like Figure 1 As shown, the present invention designs a snow cover inversion method based on the RepFNet network, which specifically includes the following steps:

[0069] In S100: FY-4A / AGRI image data with good weather conditions is selected from the Fengyun satellite remote sensing data service network. The present invention utilizes bands 1 to 7 of the FY4A / AGRI satellite image, performs radiometric calibration and geometric correction on the FY-4A image data, and obtains the following: Figure 2 Image data of the required area can be obtained by clipping the remote sensing data according to the boundary of the vector data;

[0070] Select land satellite image data that is no older than 10 minutes from FY-4A / AGRI image data through the GEE platform, such as Figure 3As shown, the study area was cropped and band fusion was performed, including bands B3, B4, B5, and B6, to improve the spatial resolution and information content of the image. The fused Landsat image data was resampled to a resolution of 2000m to match the resolution of the FY-4A / AGRI image data.

[0071] In S200: obtain the geographic elevation data of the study area, and splice it to ensure that the data within the study area is complete. The final elevation data is as follows: Figure 4 shown.

[0072] By manually drawing multiple 64×64 pixel rectangular area vector files on the Landsat8 data, we used them as templates for subsequent cropping. We loaded the pre-processed FY-4A / AGRI meteorological satellite data, Landsat8 land satellite data, and geographic auxiliary data, respectively, ensuring that the spatial resolution and projection information of all data were consistent. We used georeferencing to spatially align the data used, and used the cropping template to create a 2000m resolution dataset. We then performed random horizontal and vertical operations on it to expand the dataset. The labels and input data are as follows: Figure 5 、 6 shown.

[0073] In S300, the processed Landsat image data is processed using the SNOMAP algorithm. An NDSI mask is generated based on preset conditions (the NDSI threshold selected in this invention is 0.29). Valid data is screened to create a binary snow map that distinguishes snow-covered and non-snow-covered areas. Because snow and water have similar reflectance characteristics in the short-infrared and visible bands, a factor is introduced to determine whether the reflectance in the fifth band is greater than or equal to 0.11. This eliminates interference from water on snow identification and adjusts the spatial resolution of the binary snow map to be consistent with FY-4A meteorological satellite data. Snow cover is then calculated on the binary snow map using the following formula:

[0074]

[0075] in, Indicates the number of snow pixels in each calculation unit, After calculating the snow cover map, verify that the FSC value range is between 0% and 100% and check for outliers.

[0076] In S400: The Pytorch framework is used to build the RepFNet network model for snow cover inversion tasks. The overall network structure is shown in the figure below: Figure 7The present invention uses the improved RepVGG module to effectively extract the texture and spectral features of the snow-covered area, and uses the residual connection to enhance the training stability and performance of the model, as shown in Figure 8 As shown in the figure, a decoding module is subsequently added based on this structure, and a dynamic upsampling module DySample is introduced for upsampling, and the AGCA module is used to improve the feature expression capability.

[0077] This network is inspired by the residual structure of the ResNet network and is constructed and trained with fixed-ratio identity branches and 1×1 convolution branches. However, the texture complexity of different areas in the snow scene varies significantly, and fixed weights cannot adaptively extract key features. Therefore, this paper introduces dynamic residual weights in the RepVGG module of feature extraction to dynamically adjust branch weights according to the local texture complexity of the input feature map. The formula is as follows:

[0078]

[0079]

[0080] Var(Fin) represents the local lengthening of the feature map Fin of the input data (sliding window 3×3), which represents the complexity of regional texture: when the texture is complex like the edge of the snow, it is a high variance area. Close to 1, dominated by the identity branch, retaining more original feature details; flat areas such as continuous snow layers are low variance areas. Close to 0, dominated by 1×1 convolution, enhancing feature extraction capabilities. The initial value of is set to 0.5 and is automatically optimized after training.

[0081] We also use two-dimensional convolution to extract spatial features from the image. Before the true snow cover label is fed to the encoder, convolution with a stride of 2 is used for downsampling. Subsequent blocks use convolution with a stride of 1 to extract features and compress the feature map into a low-dimensional representation. Before outputting the predicted image, we apply the dynamic upsampling module DySample to the output of the RepVGG decoder. This, combined with the transposed convolutional layer, gradually restores the spatial resolution of the feature map, restoring the encoded feature information to the same spatial resolution as the original image, resulting in high-quality inversion results.

[0082] Introduced a dynamic range factor for DySample , to increase the flexibility of the offset. Finally, the output feature map is obtained by bilinear interpolation , where X represents the input and S represents the sampled output graph;

[0083]

[0084] After convolution with the RepVGG block, some fragmented snow areas cannot be accurately captured. By adding an Adaptive Graph Channel Attention (AGCA) module, local and global context information can be aggregated to improve the expressiveness of features.

[0085] like Figure 9 As shown, this module mainly consists of two parts.

[0086] The first part is the feature mapping layer, which converts the feature map dimension to 1×1×C (C represents the number of channels in the image), which also represents the number of feature vertices in the AGCA module.

[0087] Using a linear embedding function As a feature mapping layer, it focuses on the characteristics of important snow areas and fragmented snow, allowing the entire model to more effectively capture key information about snow cover. The module fuses this information to produce improved feature representations. These feature maps can better represent the complex structure and semantic information of the input data. Where W is a weight matrix to be learned through 1×1 convolution.

[0088] The second part dynamically adjusts feature weights through the Adaptive Image Convolution Module (AGCM). This attention mechanism dynamically adjusts feature weights, and finally maps the weights back to the original feature map as channel weights. The AGCA module also utilizes a bottleneck structure, making its feature vertices more sensitive to certain redundancies in snow cover. The introduction of the AGCA module enables the model to more adaptively focus on different areas and features of snow, thereby improving the model's generalization capabilities.

[0089] In S500: To improve the accuracy of the model in processing FSC prediction, the present invention designs a new loss function suitable for this task, which uses a combination of two functions: and ;

[0090] in Huber Loss is a robust loss function. Many regression tasks today use mean squared error (MSE) as a loss function. Compared to MSE, Huber Loss is less susceptible to outliers. Furthermore, compared to mean absolute error (MAE), Huber Loss provides a smoother optimization process, which facilitates convergence of the optimization algorithm. Therefore, Huber Loss balances convergence speed and outlier handling capabilities. The specific formula is as follows:

[0091]

[0092] in, is a hyperparameter (in this paper, =1.3), which is used to control the sensitivity to errors. x represents the true label image of FSC, and y represents the label prediction image of FSC;

[0093] Gradient difference loss (GDL) is a strategy that can sharpen image edges. It directly penalizes the gradient difference of the predicted image in the generated loss function, which can effectively alleviate the phenomenon of image edge blur. It is shown in the following formula:

[0094]

[0095] in, Is an integer. The loss function used in this study is composed of Huber Loss and GDL, as shown in the following formula:

[0096]

[0097] in, and is the weight of the loss function, satisfying .at this time Combining the advantages of both, it is introduced into the RepFNet model, thereby trying to balance different loss characteristics in the inversion of snow cover. In order to achieve the best performance of this loss function in the FSC regression training stage, this study adjusts the loss function in the training stage. and This study selected multiple weight combinations and evaluated the effects of different weights by comparing the performance of the model (as shown in the table below). The final selection results are shown in the table below. This combination is the best combination.

[0098]

[0099] In S600: Adjust the parameters of the RepFNet network model, including learning rate, batch size, number of iterations, number of weight decays, etc.

[0100] The ADAM optimizer was selected for model optimization, and the convergence speed and accuracy of the model were improved by dynamically adjusting the learning rate. During the model training process, by continuously optimizing the parameters, the present invention was able to find the most suitable parameter combination, thereby constructing the optimal snow cover inversion model. Before training the model, to ensure that the data of each channel was at the same order of magnitude and to avoid the problem of gradient explosion during training, the present invention performed normalization preprocessing on the input data. During this process, we used the maximum and minimum normalization method to ensure data stability and model convergence.

[0101] The formula for normalizing the maximum and minimum values is as follows:

[0102]

[0103] Among them, x is the original data, and are the minimum and maximum values in the data channel, respectively. is the normalized data.

[0104] During the implementation of this invention, through the analysis of a large amount of experimental data and continuous optimization of model training, the model parameter configuration that is most suitable for this example was finally determined. These parameters were adjusted multiple times to ensure that the model achieves optimal performance in the snow cover inversion task. The specific parameter values are as follows: the number of training iterations (epoch) is set to 200 times, the initial learning rate (learning_rate) is 0.001, and the learning rate decay coefficient (gamma) is 0.8. The indicators for model accuracy verification include the determination coefficient ( ), root mean square error (RMSE), correlation coefficient, Kappa coefficient, and explained variance fraction (EVS). The optimal model achieved an accuracy of R² = 0.6763, MSE = 0.0956, Correlation = 0.8539, EVS = 0.6602, and Kappa = 0.4728, demonstrating the effectiveness of this method. Finally, multiple FY-4A / AGRI image data for each day of the desired area were downloaded and inverted using the RepFNet model to generate a snow cover map for the desired area.

[0105] By inverting meteorological image data at multiple hourly time periods within a day, multi-temporal remote sensing images are integrated within the day to achieve cloud removal of the final inversion results. The flow chart is as follows: Figure 11 As shown. The data at the i-th time interval is denoted as Xi(x,y), where (x,y) represents the pixel position in the image. Clouds typically exhibit high reflectivity or brightness, while other ground features (such as snow or land) are relatively stable. The reflectivity of clouds often varies between time intervals. This paper defines a snow accumulation threshold Ts of 0.2. When Xi(x,y) is greater than or equal to 0.2, the pixel is considered to have snow accumulation, denoted by Si(x,y) = 1; otherwise, it is 0.

[0106] For pixels at the same location, the frequency of snow accumulation in a day is calculated as F(x,y):

[0107]

[0108] When F(x,y) is greater than 0.7, the pixel is considered to be a stable snow pixel SN(x,y), otherwise it is marked as a suspicious cloud area.

[0109] At the same time, define the number of pixels within a 3×3 pixel area centered on the pixel (x, y) as N(x, y), and calculate the proportion of snow-covered pixels in this area P(x, y):

[0110]

[0111] If the ratio exceeds the spatial consistency threshold Tp (the threshold is set to 0.6 in the present invention), the central pixel is considered to have snow consistency in space and is determined to be snow.

[0112] After completing the snow accumulation determination, the suspected cloud area is repaired using a multi-time fusion strategy, and valid observations from other times of the same day are selected to fill in the gaps according to the proximity of the observation time to the government, using a weighted average method. Figure 12 As shown in Figure 1, the inversion results of 7 time points in one day were finally de-clouded (as shown in h). Following the above steps, a daily product was successfully constructed, achieving high-precision monitoring of the dynamic changes of daily snow cover, providing strong support for comprehensively and clearly presenting the evolution characteristics of snow cover in time series, as shown in Figure 1. Figure 13 shown.

[0113] Example 2:

[0114] This embodiment provides a snow cover inversion device based on the RepFNet network. The device is deployed on an edge device or in the cloud, and includes:

[0115] The data acquisition module is used to obtain remote sensing image data (such as satellite images, drone images, etc.), pre-process the data, and construct a snow cover inversion dataset;

[0116] A model building module is used to build a snow cover inversion model based on the RepFNet network. The model training process and the process of obtaining the true value of snow cover are the same as those in Example 1.

[0117] The snow cover inversion module is used to obtain real-time remote sensing image data and call the trained RepFNet network model to obtain the current snow cover inversion results.

[0118] Example 3:

[0119] This embodiment provides a non-transitory computer-readable storage medium having computer instructions stored thereon, which enable a computer to execute the snow cover inversion method based on the RepFNet network in Example 1.

[0120] Example 4:

[0121] This embodiment provides an electronic device that may include a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the RepFNet-based snow cover inversion method are implemented. The memory may include internal memory, such as a high-speed random access memory, or may also include non-volatile memory, such as at least one disk drive. The processor, network interface, and memory are interconnected via an internal bus, which may be an industrial standard architecture bus, a peripheral component interconnect standard bus, an extended industrial standard architecture bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. The memory is used to store programs. Specifically, the programs may include program code, and the program code includes computer operating instructions. The memory may include internal memory and non-volatile memory, and provides instructions and data to the processor.

[0122] In the computer device, when the processor executes the computer program, the following steps are specifically implemented:

[0123] Acquire remote sensing image data through the data acquisition module, preprocess the data, and construct a snow cover inversion dataset;

[0124] The AGCA module and Dysample module are used to enhance data features and dynamically sample data to optimize data distribution.

[0125] Build and train a snow cover inversion model based on the RepFNet network;

[0126] Call the trained RepFNet network model to perform snow cover inversion on real-time remote sensing image data and output the inversion results.

[0127] Example 5:

[0128] This embodiment provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the snow cover inversion method based on the RepFNet network in Example 1.

[0129] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0130] Based on this understanding, the above technical solution can essentially or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0131] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are 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 will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A snow cover monitoring method based on RepFNet network and spatiotemporal filtering, characterized by: The method comprises: S100, obtaining FY-4A image data and Landsat satellite image data that is no older than 10 minutes from the image data, and performing band fusion and resampling operations on the Landsat satellite image data; obtaining and splicing digital elevation model data of the study area, and calculating slope and aspect information based on the digital elevation model as geographic auxiliary data for snow cover estimation; S200, spatially registering the pre-processed FY-4A image data, Landsat data, and geographic auxiliary data using Landsat data as the reference data, and cropping them into 64×64 pixel image blocks to construct a dataset suitable for deep learning training; S300 uses the SNOMAP algorithm to identify snow cover in Landsat data using binarization. The snow cover is calculated by combining the Normalized Difference Snow Index with the resolution of meteorological satellites, generating a high-resolution snow cover image as label data for the deep learning model. S400. Using the PyTorch framework, build and implement the RepFNet network model with an encoder-decoder architecture; this model integrates an improved feature extraction module, a dynamic upsampling module, and an adaptive graph channel attention module; S500, determining the optimal module combination and parameter configuration based on deep optimization and parameter adjustment of the RepFNet network model; S600, designed a new loss function suitable for FSC inversion and used the ADAM optimizer combined with a dynamic learning rate strategy for model training; optimized model performance by adjusting hyperparameters and conducting comparative experiments; used the trained RepFNet model to accurately invert snow cover in the study area, and combined it with spatiotemporal filtering methods for cloud removal to ensure the accuracy and stability of snow monitoring; The RepFNet network model in S400 includes: an improved feature extraction module, a dynamic upsampling module, and an adaptive graph channel attention module; The improved feature extraction module is used to propose a dynamic residual weight driven by local variance to achieve adaptive adjustment of module branch ratio; The dynamic upsampling module is used to gradually restore the high-dimensional features output by the feature extraction module to the same spatial resolution as the input data; it uses a step-by-step upsampling method to gradually restore the spatial information of the feature map through the alternating combination of convolutional layers and upsampling layers; The adaptive graph channel attention module is used to adaptively adjust the weights of feature channels while capturing the global dependencies between features; Among them, in S400: using the Pytorch framework to build a RepFNet network model for the snow cover inversion task, including: using the improved feature extraction module RepVGG to extract the texture and spectral features of the snow-covered area, and using residual connections to enhance the training stability and performance of the model; subsequently adding a decoding module based on this structure, introducing a dynamic upsampling module DySample for upsampling, and improving the feature expression capability through the adaptive graph channel attention module AGCA; Among them, dynamic residual weights are introduced into the improved feature extraction module RepVGG, and the branch weights are dynamically adjusted according to the local texture complexity of the input feature map. The formula is as follows: ； ； in, The initial value of is set to 0.5 and automatically optimized after training; Represents the feature map of the calculated input data The local length of the sliding window is 3×3, which is used to characterize the regional texture complexity: when it belongs to a high variance area with complex texture, Close to 1, dominated by the identity branch, retaining more original feature details; when it belongs to the low variance area of the flat area, Close to 0, dominated by 1×1 convolution, enhancing feature extraction capabilities; At the same time, two-dimensional convolution is selected to extract spatial features from the image. Before the true value of the snow cover label is sent to the encoder, a convolution with a stride of 2 is used for downsampling. The following block uses a convolution with a stride of 1 to extract features and compress the feature map into a low-dimensional representation. Before outputting the prediction map, the output of the RepVGG decoder is upsampled using the dynamic upsampling module DySample. This and the transposed convolution layer gradually restore the spatial resolution of the feature map, restoring the encoded feature information to the spatial resolution of the original image to produce a high-quality inversion result. Introducing a dynamic range factor in the dynamic upsampling module DySample , to increase the flexibility of the offset; and obtain the output feature map through bilinear interpolation , where X represents the input and S represents the sampled output graph; ； Then, the adaptive graph channel attention module AGCA is added to aggregate local and global context information to improve the expressiveness of features; Furthermore, the RepFNet network module mainly consists of two parts; The first part is the feature mapping layer, where the feature map dimension is converted to 1×1×C, where C represents the number of channels in the image and also the number of feature vertices in the adaptive graph channel attention module AGCA; Using linear embedding function As a feature mapping layer, it captures key features including important snow areas and fragmented snow, and fuses the key feature information to produce an improved feature representation; wherein, Represents the weight matrix obtained by 1×1 convolution; The second part is to dynamically adjust the feature weights through a designed attention mechanism, and finally map the weights back to the original feature map as channel weights. At the same time, the adaptive graph channel attention module AGCA also uses a bottleneck structure to make its feature vertices more sensitive to the feature redundancy of snow accumulation. The in-depth optimization and parameter adjustment in S500 include: The threshold of network iterations was set to 200 to ensure that the model converged within a limited number of iterations. The initial learning rate was set to 0.001, and the ADAM optimizer was used to update the model parameters. After every 10 iterations, the learning rate is dynamically adjusted with a decay coefficient of 0.8 to accelerate model convergence and prevent overfitting.

2. The snow cover monitoring method based on RepFNet network and spatiotemporal filtering according to claim 1, characterized in that: The calculation of slope and aspect information based on the digital elevation model in S100 includes: Using digital elevation model data, slope and aspect data are calculated through Python scripts to generate terrain auxiliary data layers that are consistent with the FY-4A and Landsat data space, providing terrain feature support for subsequent model training.

3. The snow cover monitoring method based on RepFNet network and spatiotemporal filtering according to claim 1, characterized in that: The step S300 generates a high-resolution snow cover image, including: The normalized difference snow index NDSI is determined according to the formula: ； Band 2 is the green light band of Landsat 8 data, and Band 6 is the shortwave infrared band of Landsat 8 data; The Landsat data were converted into a binary snow map by selecting the NDSI threshold and then resampled to ensure spatial consistency with the FY-4A data. The snow cover within each pixel is calculated based on the number of snow pixels and the total number of pixels to generate a snow cover map.

4. The snow cover monitoring method based on RepFNet network and spatiotemporal filtering according to claim 1, characterized in that: In S600, the trained RepFNet model is used to accurately invert the snow cover of the study area, including: The model accuracy was verified using the test data set, and the evaluation indicators included the coefficient of determination, root mean square error, correlation coefficient, Kappa coefficient, and explained variance score; Among them, the coefficient of determination for: ； The root mean square error RMSE is: ； Correlation coefficient for: ； The Kappa coefficient is: ； The explained variance score EVS is: ； in, represents the true value, represents the predicted value; represents the mean of the true values, represents the mean of the predicted values, represents the observation accuracy, that is, the amount by which the predicted values match the true values. represents random accuracy, which is the proportion of the model calculated based on the probability of random classification matching the actual category, and n represents the size of the sample; Among them, when the coefficient of determination, EVS, correlation coefficient and Kappa value are closer to 1, the model fits the data better; when the RMSE is closer to 0, the model inversion accuracy is higher.

5. The snow cover monitoring method based on RepFNet network and spatiotemporal filtering according to claim 1, characterized in that: The S600 combines spatiotemporal filtering methods to perform cloud removal to ensure the accuracy and stability of snow monitoring, including: Deploy the trained RepFNet model to the study area data and output the snow cover inversion results; The data in the study area are subjected to multi-temporal declouding processing, and the declouding is combined with time series data to monitor the changes in snow cover and ensure the stability and timeliness of the inversion results.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the snow cover monitoring method based on the RepFNet network and spatiotemporal filtering are implemented as described in any one of claims 1 to 5.

7. 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 of the snow cover monitoring method based on RepFNet network and spatiotemporal filtering are implemented as described in any one of claims 1-5.

Images