Snow depth distribution monitoring method

Abstract

The application relates to a snow depth distribution monitoring method, which utilizes a snow depth distribution map with different resolutions of a sample area and environmental spatial feature data affecting the snow distribution, firstly pretrains a diffusion model by constructing pre-training data, so that the diffusion model learns the correlation between the snow depth parameters of a grid pixel and the environmental spatial features; secondly fine-tunes the pre-trained diffusion model by using constructed fine-tuning training data, so that the model learns the mapping relationship between the snow depth distribution maps with different resolutions; and finally inputs the snow depth distribution map of a region to be monitored into the fine-tuned diffusion model, so that a high-resolution snow depth distribution map is obtained, and high-precision snow depth distribution data is acquired.

Description

Technical Field

[0001] This application relates to the field of snow depth monitoring technology, and in particular to a method for monitoring snow depth distribution. Background Technology

[0002] Snow depth, as a key indicator reflecting the status of snow and ice resources, is of vital importance in many fields such as climate change research, hydrological cycle analysis, and snow and ice economic planning. Accurately understanding changes in snow depth helps to deepen our understanding of the climate system's operating mechanisms, rationally allocate water resources, and promote the sustainable development of the snow and ice industry.

[0003] Currently, snow depth monitoring mainly relies on satellite remote sensing technology, which has the advantages of large-scale and continuous monitoring and plays an important role in snow depth monitoring. However, the data acquired by commonly used snow depth monitoring satellites has significant shortcomings, with low spatial resolution. In relatively small-scale applications such as watersheds and regions, this low-resolution data is insufficient to meet the requirements for high-precision snow depth data.

[0004] To address the aforementioned issues, spatial downscaling techniques are typically used to process remote sensing snow depth data, thereby improving data accuracy and resolution. However, existing spatial downscaling techniques yield snow depth data with limited accuracy, still insufficient to meet the high-precision, high-resolution requirements of practical applications. Therefore, there is an urgent need for a monitoring method that can effectively improve the accuracy of snow depth data to overcome the shortcomings of existing technologies. Summary of the Invention

[0005] Based on this, the purpose of this application is to provide a snow depth distribution monitoring method to solve the problem that existing snow depth monitoring technologies are unable to obtain high-precision, high-resolution snow depth data due to the low spatial resolution of satellite remote sensing data and the limited accuracy of spatial downscaling techniques, thus failing to meet the needs of small-scale applications such as watersheds and regions.

[0006] The snow depth distribution monitoring method of this application includes the following steps:

[0007] A first-resolution snow depth distribution map, environmental spatial feature data affecting snow cover distribution, and a second-resolution snow depth distribution map of the sample area are obtained; wherein, the resolution of the second-resolution snow depth distribution map is greater than the resolution of the first-resolution snow depth map. Based on the first resolution snow depth distribution map and the environmental spatial feature data, pre-training data is constructed; the diffusion model is pre-trained using the pre-training data to obtain a pre-trained diffusion model; the pre-training data is used to enable the diffusion model to learn the correlation between the snow depth parameters of raster pixels and the environmental spatial features. Based on the first resolution snow depth distribution map, the environmental spatial feature data, and the second resolution snow depth distribution map, fine-tuning training data is constructed, and the pre-trained diffusion model is fine-tuned to obtain the fine-tuned diffusion model, so that the diffusion model learns the mapping relationship between the first resolution snow depth distribution map and the second resolution snow depth distribution map. Obtain a first-resolution snow depth distribution map and environmental spatial feature data of the area to be monitored, input the fine-tuned diffusion model, and obtain a second-resolution snow depth distribution map of the area to be monitored; based on the second-resolution snow depth distribution map of the area to be monitored, obtain the snow depth distribution data of the area to be monitored.

[0008] This application's embodiments acquire a first-resolution snow depth distribution map of the sample area, environmental spatial feature data affecting snow cover distribution, and a higher-resolution second-resolution snow depth distribution map, providing a rich and comprehensive information foundation for model training. Through the pre-training phase, the diffusion model deeply learns the intrinsic correlation between the snow depth parameters of raster pixels and environmental spatial features, enabling the model to initially understand the impact of environmental factors on snow depth, laying a theoretical foundation for accurate monitoring. In the fine-tuning training phase, by learning the mapping relationship between the first and second-resolution snow depth distribution maps, the model performance is further optimized, allowing the model to accurately grasp the conversion logic between snow depth data of different resolutions. In practical applications, inputting the first-resolution snow depth distribution map and environmental spatial feature data of the area to be monitored into the fine-tuned diffusion model can quickly generate a high-resolution second-resolution snow depth distribution map, thereby obtaining high-precision snow depth distribution data. This process effectively overcomes the shortcomings of low spatial resolution of satellite remote sensing data and limited accuracy of spatial downscaling techniques in existing technologies. It can meet the urgent needs of small-scale application scenarios such as watersheds and regions for high-precision, high-resolution snow depth data, provide more accurate data support for climate change research, help more rational water resource allocation, and promote the sustainable development of the ice and snow industry based on reliable data. It has demonstrated important application value and positive impact in many related fields.

[0009] To better understand and implement this application, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0010] Figure 1 This is a schematic flowchart of the snow depth distribution monitoring method according to an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. Wherein, when the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements.

[0012] It should be understood that the embodiments described below do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0013] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, in the description of this application, unless otherwise stated, “a plurality” means two or more. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items, for example, A and / or B, which can represent: A alone, A and B together, and B alone; the character “ / ” generally indicates that the preceding and following objects are in an “or” relationship.

[0014] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, this information should not be limited to these terms, and these terms are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. Depending on the context, the word "if" as used in this application can be interpreted as "when," "when," or "in response to determination."

[0015] This application provides a snow depth distribution monitoring method to solve the problem that existing snow depth monitoring technologies are unable to obtain high-precision, high-resolution snow depth data due to the low spatial resolution of satellite remote sensing data and the limited accuracy of spatial downscaling techniques, thus failing to meet the needs of small-scale applications such as watersheds and regions.

[0016] Please refer to Figure 1 The snow depth distribution monitoring method of this application includes the following steps: S101: Obtain a first-resolution snow depth distribution map, environmental spatial feature data affecting snow cover distribution, and a second-resolution snow depth distribution map of the sample area; wherein, the resolution of the second-resolution snow depth distribution map is greater than the resolution of the first-resolution snow depth map; S102: Based on the first resolution snow depth distribution map and the environmental spatial feature data, construct pre-training data; pre-train the diffusion model using the pre-training data to obtain a pre-trained diffusion model; the pre-training data is used to enable the diffusion model to learn the correlation between the snow depth parameters of raster pixels and the environmental spatial features. S103: Construct fine-tuning training data based on the first resolution snow depth distribution map, the environmental spatial feature data, and the second resolution snow depth distribution map, and fine-tune the pre-trained diffusion model to obtain the fine-tuned diffusion model, so that the diffusion model learns the mapping relationship between the first resolution snow depth distribution map and the second resolution snow depth distribution map; S104: Obtain the first-resolution snow depth distribution map and environmental spatial feature data of the area to be monitored, input the fine-tuned diffusion model, and obtain the second-resolution snow depth distribution map of the area to be monitored; based on the second-resolution snow depth distribution map of the area to be monitored, obtain the snow depth distribution data of the area to be monitored.

[0017] This application's embodiments acquire a first-resolution snow depth distribution map of the sample area, environmental spatial feature data affecting snow cover distribution, and a higher-resolution second-resolution snow depth distribution map, providing a rich and comprehensive information foundation for model training. Through the pre-training phase, the diffusion model deeply learns the intrinsic correlation between the snow depth parameters of raster pixels and environmental spatial features, enabling the model to initially understand the impact of environmental factors on snow depth, laying a theoretical foundation for accurate monitoring. In the fine-tuning training phase, by learning the mapping relationship between the first and second-resolution snow depth distribution maps, the model performance is further optimized, allowing the model to accurately grasp the conversion logic between snow depth data of different resolutions. In practical applications, inputting the first-resolution snow depth distribution map and environmental spatial feature data of the area to be monitored into the fine-tuned diffusion model can quickly generate a high-resolution second-resolution snow depth distribution map, thereby obtaining high-precision snow depth distribution data. This process effectively overcomes the shortcomings of low spatial resolution of satellite remote sensing data and limited accuracy of spatial downscaling techniques in existing technologies. It can meet the urgent needs of small-scale application scenarios such as watersheds and regions for high-precision, high-resolution snow depth data, provide more accurate data support for climate change research, help more rational water resource allocation, and promote the sustainable development of the ice and snow industry based on reliable data. It has demonstrated important application value and positive impact in many related fields.

[0018] The snow depth distribution monitoring method of this application uses a computer as the execution subject, and the following describes each step in detail.

[0019] For step S101, a first-resolution snow depth distribution map of the sample area, environmental spatial feature data affecting snow distribution, and a second-resolution snow depth distribution map are obtained; wherein, the resolution of the second-resolution snow depth distribution map is greater than the resolution of the first-resolution snow depth distribution map.

[0020] A sample region refers to a specific geographical area used for model training, which contains complete snow depth distribution data and environmental spatial feature data affecting snow cover distribution. In one embodiment, the sample region can be a portion of the area to be monitored, or it can be a region outside the area to be monitored.

[0021] The first-resolution snow depth distribution map represents a snow depth distribution data map with relatively low resolution. It reflects the general distribution of snow depth within the sample area, but the spatial details are not fine enough.

[0022] The environmental spatial feature data encompasses a range of factors influencing snow cover distribution. In one embodiment, environmental spatial features include, but are not limited to, precipitation (the amount of water vapor condensate falling to the ground per unit time), surface temperature (the temperature of the surface at the interface with the atmosphere), snow cover (the degree to which the surface is covered by snow), surface reflectivity (the ability of the surface to reflect solar radiation), vegetation cover (the percentage of the vertically projected area of ​​vegetation on the surface relative to the total area of ​​the statistical area), elevation (the vertical distance of a location on the ground above sea level), latitude (a measure of the north-south distance on the Earth's surface), and land use (the type of land use, such as farmland, forest, and urban land). In one embodiment, the original spatial resolution of the environmental spatial feature data is higher than or equal to the resolution of the second-resolution snow depth distribution map, with some data reaching resolutions of hundreds of meters or even higher. This high-resolution characteristic provides fundamental data support for subsequent models to reconstruct snow depth across scales.

[0023] The second-resolution snow depth distribution map has a higher resolution than the first-resolution snow depth distribution map, and can present the snow depth distribution within the sample area in greater detail. In one embodiment, the second-resolution snow depth distribution map may include several snow depth distribution maps with different resolutions higher than the first resolution, which can present the snow depth distribution within the sample area with different levels of detail, enabling the trained model to have cross-scale generalization ability.

[0024] This step aims to collect the foundational data required for model training. First-resolution snow depth distribution maps of the sample region are obtained, providing a general overview of snow depth in that area. Simultaneously, various environmental spatial feature data influencing snow cover distribution are collected, as these data affect snow formation and distribution from different perspectives. Additionally, a second-resolution snow depth distribution map is acquired, offering higher resolution and providing more detailed snow depth information for the model. By acquiring these three types of data, comprehensive and necessary information is provided for subsequent model training.

[0025] In one embodiment, step S101, obtaining a second-resolution snow depth distribution map of the sample area, includes: Step S1011: Obtain snow distribution probability data for the sample area; wherein the resolution of the snow distribution probability data is consistent with the resolution of the second resolution snow depth distribution map.

[0026] Snow cover distribution probability data reflects the probability of differences in snow cover distribution on the ground surface. Its resolution is consistent with that of the second-resolution snow depth distribution map, providing a crucial basis for subsequent sub-pixel decomposition.

[0027] Snow cover distribution probability data can reflect the probability of snow cover appearing in each raster cell within the sample area. Its resolution is the same as that of the target second-resolution snow depth distribution map, ensuring consistency and matching in spatial scale when subsequent operations are performed based on this data, thus laying the foundation for accurately generating the second-resolution snow depth distribution map.

[0028] Step S1012: Based on the snow distribution probability data, perform sub-pixel decomposition on the first resolution snow depth distribution map to obtain the second resolution snow depth distribution map.

[0029] Sub-pixel decomposition is a technique that decomposes pixel information in a low-resolution image onto a higher-resolution raster. By combining other relevant data, it refines the regional information represented by low-resolution pixels into high-resolution pixels to generate a high-resolution image.

[0030] Because the first-resolution snow depth distribution map has a low resolution, it is difficult to meet the needs of practical applications for high-precision snow depth data. Snow cover distribution probability data, however, provides more detailed information about the probability of snow cover distribution within the sample area. Through sub-pixel decomposition technology, each low-resolution pixel in the first-resolution snow depth distribution map is decomposed into a higher-resolution raster based on the probability of snow cover reflected in the snow cover distribution probability data. Specifically, based on the probability of snow cover appearing in each high-resolution raster pixel in the snow cover distribution probability data, the snow depth information represented by the first-resolution pixel is redistributed and refined, thereby generating a second-resolution snow depth distribution map with a higher resolution, making the snow depth information more refined and accurate spatially.

[0031] In this embodiment, during the acquisition of the second-resolution snow depth distribution map of the sample area, snow cover distribution probability data with the same resolution as the second-resolution snow depth distribution map is first acquired. Then, sub-pixel decomposition is performed on the first-resolution snow depth distribution map based on this data. This approach fully utilizes the probability information of fine snow distribution within the sample area contained in the snow cover distribution probability data, overcoming the shortcomings of directly using the low resolution of the first-resolution snow depth distribution map. Through sub-pixel decomposition technology, the low-resolution snow depth information is refined onto a high-resolution raster, generating a more accurate second-resolution snow depth distribution map. This high-resolution snow depth distribution map provides more accurate target data for the subsequent training of the diffusion model, enabling the model to learn more refined snow depth distribution patterns and the mapping relationship between snow depth distribution maps of different resolutions.

[0032] In one embodiment, step S1011, obtaining the snow cover distribution probability data of the sample area, includes: Step S10111: Obtain snow cover score data and clear sky index data for the sample area.

[0033] Snow cover score data is a type of data used to describe the snow cover situation in a sample area. It is usually presented in raster form, where the value of each raster cell represents the degree of snow cover in that area. For example, it can be represented by a value between 0 and 1, where 0 means no snow cover and 1 means completely covered by snow.

[0034] Sunnyness index data is an indicator that reflects the degree of sunshine in a sample area. It is also in raster form, and the value of each raster cell represents the degree of sunshine in that area at a specific time. It can be used to measure the impact of factors such as light conditions on snow cover distribution.

[0035] In this step, the snow cover score data directly reflects the snow cover situation within the sample area and is fundamental information for understanding snow distribution. The sunshine index data reflects weather conditions such as sunlight in the sample area, which affect the melting and accumulation of snow, thus influencing snow distribution. Obtaining these two types of data provides necessary information for subsequent calculations of snow distribution probability.

[0036] Step S10112: Calculate the ratio between the snow cover score data and the corresponding clear skies index data to obtain the snow distribution probability data.

[0037] The snow cover score reflects the actual snow cover, while the sunshine index reflects environmental factors influencing snow distribution. Calculating the ratio between the two allows for a comprehensive consideration of the impact of both actual snow cover and weather conditions on its distribution. For example, in areas with a high snow cover score but also a high sunshine index (i.e., ample sunlight that could lead to snow melting), the ratio calculation yields a relatively reasonable probability value for snow distribution, which more accurately reflects the likelihood of snow accumulation in that area under current conditions.

[0038] In this embodiment, when acquiring snow cover distribution probability data for the sample area, snow cover fraction data and clear weather index data are first obtained. These two types of data reflect relevant information about snow cover distribution from two aspects: actual snow cover and the influence of weather conditions, respectively. By calculating the ratio between the snow cover fraction data and the corresponding clear weather index data, the actual presence of snow and the influence of environmental factors on its distribution are comprehensively considered, resulting in more accurate and comprehensive snow cover distribution probability data. This data not only provides a reliable basis for subsequent sub-pixel decomposition based on this data, enabling the generated second-resolution snow depth distribution map to more accurately reflect the snow depth distribution within the sample area, but also lays a solid foundation for subsequent model training in the entire snow depth distribution monitoring method.

[0039] In one embodiment, step S1012, which involves performing sub-pixel decomposition on the first resolution snow depth distribution map based on the snow distribution probability data to obtain a second resolution snow depth distribution map, includes: Step S10121: For each grid cell of the first resolution snow depth distribution map, determine a number of second resolution sub-pixels in the snow distribution probability data.

[0040] The second-resolution sub-pixel is a higher-resolution raster unit that further subdivides the pixels of the first-resolution snow depth distribution map during the sub-pixel decomposition process. Its resolution is the same as that of the snow cover distribution probability data.

[0041] For each pixel in the first-resolution snow depth distribution map, several corresponding second-resolution sub-pixels in the snow cover distribution probability data are identified. Since the resolution of the snow cover distribution probability data is the same as that of the second-resolution snow depth distribution map, while the resolution of the first-resolution snow depth distribution map is relatively low, each first-resolution pixel spatially corresponds to multiple second-resolution sub-pixels. This step clarifies the object and scope of the decomposition, laying the foundation for subsequent calculations and allocation.

[0042] Step S10122: Calculate the sum of probabilities of the several second-resolution sub-pixels based on the snow distribution probability parameters of the several second-resolution sub-pixels.

[0043] The snow cover distribution probability parameter is the value corresponding to each second-resolution sub-pixel in the snow cover distribution probability data, representing the probability of snow cover appearing in that sub-pixel.

[0044] Step S10123: Divide the snow distribution probability parameter of each of the plurality of second resolution sub-pixels by the sum of probabilities to obtain the weight value of each of the plurality of second resolution sub-pixels.

[0045] The weight value is obtained by dividing the snow distribution probability parameter of each second-resolution sub-pixel by the corresponding sum of probabilities, and is used to measure the importance of the sub-pixel in the snow depth parameter allocation process of the first-resolution pixel.

[0046] In this step, the weight values ​​reflect the importance of each second-resolution sub-pixel to the snow cover distribution within the region represented by the first-resolution pixel. Sub-pixels with larger snow cover distribution probability parameters have larger weight values, indicating that these sub-pixels should play a more important role in the snow depth parameter allocation process.

[0047] Step S10124: Based on the weight value of each second resolution sub-pixel in the plurality of second resolution sub-pixels, the snow depth parameters of the corresponding raster pixels in the first resolution snow depth distribution map are weighted and allocated to obtain the second resolution snow depth distribution map.

[0048] The snow depth parameters of the first-resolution pixels are allocated according to the weight values ​​of each second-resolution sub-pixel, allowing the snow depth information to be more reasonably refined to the high-resolution sub-pixels. In this way, the low-resolution snow depth distribution map is transformed into a high-resolution snow depth distribution map, making the snow depth information more refined and accurate in space.

[0049] In this embodiment, during the sub-pixel decomposition of the first-resolution snow depth distribution map based on snow cover distribution probability data, a series of precise calculation and allocation steps effectively transform low-resolution snow depth information into high-resolution snow depth information. First, the second-resolution sub-pixel corresponding to each first-resolution pixel is clearly defined to ensure the accuracy and relevance of the decomposition. Then, the probability sum and weight values ​​of the snow cover distribution probability parameters are calculated, fully considering the influence of snow cover distribution probability on the allocation of snow depth parameters, making the allocation process more reasonable. Finally, a weighted allocation is performed based on the weight values, resulting in a second-resolution snow depth distribution map that more accurately reflects the snow depth distribution within the sample area. This high-resolution snow depth distribution map provides more accurate target data for the subsequent training of the diffusion model, enabling the model to learn more refined snow depth distribution patterns and the mapping relationships between snow depth distribution maps of different resolutions.

[0050] In one embodiment, the environmental spatial characteristic data includes at least one of precipitation data, surface temperature data, snow cover data, surface reflectance data, vegetation cover data, elevation data, latitude data, and land use data.

[0051] Precipitation data is the primary source of snow cover; its amount and intensity affect snow thickness. Precipitation patterns differ across regions, and incorporating it into the model helps understand initial snow accumulation and predict snow depth. Surface temperature data influences snow retention and melting, and is affected by atmospheric temperature and surface cover, reflecting actual surface temperature and helping the model predict dynamic changes in snow cover. Snow cover data directly reflects the extent and degree of surface snow cover, acquired through remote sensing, providing the model with initial snow distribution information and predicting its changes and snow depth evolution. Surface reflectivity data refers to the surface's ability to reflect solar radiation; high reflectivity in snow and ice slows melting, and incorporating it into the model allows for accurate simulation of snow energy balance, improving prediction accuracy. Vegetation cover data reflects the degree of surface vegetation cover; vegetation intercepts snowfall and provides shade, influencing snow distribution and melting, and helping the model consider the impact of vegetation on snow depth. Elevation data provides topographic information for the model, predicting differences in snow depth at different altitudes. Because altitude affects temperature and precipitation patterns, higher altitudes typically experience snow accumulation and longer retention. Latitude data determines a region's geographical location and climate zone, influencing temperature and precipitation distribution, providing a macro-geographical reference for the model, and considering the impact of climate differences on snow depth. Land use data describes human land use patterns; different types have different impacts on snow cover, helping the model account for the influence of human activities on snow depth.

[0052] This embodiment incorporates various environmental spatial characteristic data, reflecting the factors influencing snow cover distribution and snow depth changes from multiple dimensions. The model thus more comprehensively and accurately captures the complex processes of snow formation and environmental impact mechanisms, significantly improving the accuracy and reliability of snow depth prediction, enhancing its adaptability to the snow depth distribution characteristics of different regions, and providing strong data support for applications in multiple fields such as climate change research, hydrological cycle analysis, and ice and snow economic planning.

[0053] For step S102, pre-training data is constructed based on the first resolution snow depth distribution map and the environmental spatial feature data; the diffusion model is pre-trained using the pre-training data to obtain a pre-trained diffusion model; the pre-training data is used to enable the diffusion model to learn the correlation between the snow depth parameters of raster pixels and the environmental spatial features.

[0054] The diffusion model is a probability-based generative model that gradually adds noise to the data through a forward diffusion process, transforming it into a pure noise distribution. Then, a reverse denoising process is used to gradually remove the noise using a deep neural network, restoring the spatial structure of the original data and thus learning the true distribution characteristics of the data.

[0055] A raster cell refers to a regular grid unit that divides geographic space; each unit is a raster cell. Raster cells are the basic unit for storing and analyzing remote sensing data; data such as snow depth parameters are stored in corresponding raster cells.

[0056] The pre-training data is used to initially train the diffusion model. It is constructed from the first-resolution snow depth distribution map and environmental spatial feature data that affect snow cover distribution. The purpose is to allow the model to learn the correlation between the snow depth parameters of the raster pixels and the environmental spatial features.

[0057] Pre-training data was constructed based on the acquired first-resolution snow depth distribution map and environmental spatial feature data. This data was then input into a diffusion model for pre-training. During training, the diffusion model gradually learned the intrinsic correlation between snow depth parameters in raster pixels and environmental spatial features through forward diffusion and backward denoising. For example, the model can understand how snow depth might change under specific environmental conditions such as precipitation and surface temperature, laying the foundation for more accurate snow depth predictions in the future.

[0058] In one embodiment, the training samples of the pre-training data are the snow depth parameters of the raster pixels in the third-resolution snow depth distribution map and the corresponding environmental spatial features; the training labels of the pre-training data are the snow depth parameters of the corresponding raster pixels in the first-resolution snow depth distribution map; wherein, the third-resolution snow depth distribution map is a snow depth distribution map of the same resolution obtained by downsampling and then interpolating the first-resolution snow depth distribution map.

[0059] In other words, the third-resolution snow depth distribution map is obtained by degrading the accuracy of the first-resolution snow depth distribution map. Accuracy degradation is achieved by downsampling the first-resolution snow depth distribution map and then interpolating to restore it to its original size. Downsampling can be achieved using the neighborhood averaging method, and interpolation can be performed using bilinear interpolation to restore the downsampled snow depth data to its original raster size. The final result is a snow depth distribution map with the same size and number of raster cells as the first-resolution map, but with blurred snow depth texture information and missing details, used to simulate low-quality remote sensing snow depth data. In other words, the third-resolution snow depth distribution map maintains the resolution of the first-resolution snow depth distribution map, but the snow depth parameters of each pixel are blurred, making them less precise than those of the first-resolution snow depth distribution map. This blurred snow depth parameters are then used as one of the input features for training samples. The snow depth parameters of the original first-resolution snow depth distribution map pixels serve as training labels.

[0060] Specifically, training samples are the input units in the pre-training data, containing snow depth parameters of raster pixels after accuracy degradation and spatial environmental features affecting snow cover (such as spatialized data of topography, temperature, and precipitation). Training labels are the target values ​​in the pre-training data, i.e., the snow depth parameters of the corresponding raster pixels in the first-resolution snow depth distribution map, used to measure the difference between the model's prediction results and the true values.

[0061] Step S102, which involves pre-training the diffusion model using the pre-training data to obtain a pre-trained diffusion model, includes: Step S1021: Input the pre-trained data into the diffusion model to obtain the predicted snow depth parameters corresponding to each training sample; calculate the loss value based on the predicted snow depth parameters and corresponding training labels of each training sample; when the loss value is greater than the preset loss threshold, adjust the training parameters of the diffusion model and retrain until the loss value is less than or equal to the preset loss threshold, and determine that the pre-training of the diffusion model is complete.

[0062] The loss value is an error metric calculated by comparing the predicted snow depth parameters output by the model with the training labels, reflecting the degree of deviation between the model's current prediction and the true value.

[0063] The preset loss threshold is a pre-defined convergence condition for the model. When the loss value drops below this threshold, pre-training is considered complete, and the model has fully learned the correlation between environmental features and snow depth parameters. The preset loss threshold can be adaptively adjusted according to the snow depth data characteristics of the sample area, with a typical range of 0.01-0.1. The initial learning rate of the model training parameters ranges from 1e-4 to 1e-3, and gradually decreases with the number of training iterations.

[0064] This step first inputs pre-trained data into the diffusion model, which then generates predicted snow depth parameters for each training sample based on the current parameters. Next, it calculates the loss value between the predicted value and the training label. If this value exceeds a preset loss threshold, the model parameters (such as weights and biases) are adjusted, and the pre-trained data is re-inputted for training. This process is repeated until the loss value falls below the threshold, at which point the diffusion model has completed pre-training. This step, through supervised learning, enables the model to gradually learn the mapping relationship between "environmental spatial features → snow depth parameters," laying the foundation for learning the mapping relationship between snow depth distribution maps at different resolutions in the subsequent fine-tuning stage.

[0065] This embodiment, through the design of pre-training data and the iterative training mechanism in step S1021, enables the diffusion model to efficiently capture the intrinsic correlation between environmental spatial features and snow depth parameters during the pre-training stage. Specifically, the joint input of environmental spatial features and snow depth parameters in the training samples helps the model understand how environmental factors such as terrain and climate affect snow depth; while the loss calculation of training labels and predicted values ​​and parameter adjustment ensure that the model's prediction results gradually approach the true snow depth values, forming a reliable basic mapping capability. This pre-training mechanism not only improves the model's representation accuracy of the environment-snow depth relationship, but also provides more accurate initial parameters for learning the conversion relationship from low-resolution to high-resolution snow depth distribution maps in the subsequent fine-tuning stage.

[0066] In one embodiment, the diffusion model includes a forward diffusion module, a spatiotemporal feature encoding module, and a denoising and reconstruction module; The forward diffusion module is used to gradually add Gaussian noise to the first resolution snow depth distribution map in time steps according to a preset noise scaling factor, so that it gradually forms a pure noise distribution and obtains the intermediate snow depth data at each time step. The spatiotemporal feature encoding module is used to extract and enhance the spatiotemporal features of the snow depth parameters and environmental spatial features of the first resolution snow depth distribution map, and obtain a spatiotemporal enhanced feature tensor that integrates snow depth and environmental information. The denoising and reconstruction module uses U-Net as the backbone network to receive intermediate snow depth data and the spatiotemporal enhancement feature tensor at each time step, gradually remove noise from the intermediate snow depth data, and reconstruct the intermediate snow depth data into a second-resolution snow depth distribution map based on the spatiotemporal enhancement feature tensor.

[0067] The diffusion model in this embodiment achieves efficient conversion from low-resolution snow depth data to high-resolution distribution maps through a collaborative mechanism of forward diffusion, spatiotemporal feature encoding, and denoising reconstruction. While preserving the correlation between the environment and snow depth, the model significantly improves the spatial detail restoration and resolution of snow depth data through spatiotemporal feature enhancement and the fine denoising capabilities of U-Net. This provides high-precision, high-resolution technical support for multi-scale snow depth monitoring, effectively meeting the demand for fine snow depth data in fields such as climate change research, hydrological analysis, and snow and ice economic planning.

[0068] In one embodiment, the spatiotemporal feature encoding module includes a spatial feature encoding unit, a temporal feature encoding unit, and a convolutional block attention module; The spatial feature encoding unit is used to extract and analyze the spatial dimension features of the snow depth parameters and environmental spatial features of the first resolution snow depth distribution map to obtain the snow depth-environment spatial coupling features. The temporal feature encoding unit is used to perform temporal feature mapping on the snow depth-environment spatial coupling features, adapt to the time step noise change law of the diffusion model, and obtain the preliminary spatiotemporal fusion features of snow depth-environment. The convolutional block attention module is used to assign attention weights to the initial spatiotemporal fusion features of snow depth and environment, strengthen key features related to snow depth, suppress redundant interference features, and obtain a spatiotemporal enhanced feature tensor that fuses snow depth and environment information.

[0069] The spatiotemporal feature encoding module in this embodiment deeply integrates the spatiotemporal correlation information of snow depth and environmental features through a three-level collaborative approach of space, time, and attention. Spatial encoding captures static spatial patterns, temporal encoding adapts to dynamic noise processes, and the attention mechanism strengthens the expression of core features. Together, these three mechanisms enhance the model's ability to represent complex spatiotemporal relationships, enabling the reconstructed high-resolution snow depth distribution map to more accurately reflect the spatial details of snow depth changes driven by the environment, and enhancing the model's applicability and reliability in multi-scale, multi-temporal snow depth monitoring.

[0070] For step S103, fine-tuning training data is constructed based on the first resolution snow depth distribution map, the environmental spatial feature data, and the second resolution snow depth distribution map. The pre-trained diffusion model is then fine-tuned to obtain the fine-tuned diffusion model, so that the diffusion model learns the mapping relationship between the first resolution snow depth distribution map and the second resolution snow depth distribution map.

[0071] The fine-tuning training data is used to further optimize the pre-trained diffusion model. It is constructed from the first-resolution snow depth distribution map, environmental spatial feature data, and the second-resolution snow depth distribution map, enabling the model to learn the mapping relationship between snow depth distribution maps of different resolutions.

[0072] Fine-tuning training data is constructed using a first-resolution snow depth distribution map, environmental spatial feature data, and a second-resolution snow depth distribution map. This data is then input into a pre-trained diffusion model for fine-tuning, enabling the model to further learn the mapping relationship between the first-resolution and second-resolution snow depth distribution maps. Through this step, the model learns how to convert low-resolution snow depth distribution maps into high-resolution ones, improving its ability to process data at different resolutions and its prediction accuracy.

[0073] Since the second spatial resolution snow depth data is generated from the probability of snow cover distribution, it contains a certain amount of error and is essentially a noisy training label. Therefore, during fine-tuning training, the confidence weight of the second spatial resolution snow depth data is introduced to reduce the corresponding error.

[0074] In one embodiment, Normalized Differential Snow Index (NDSI) data with a resolution higher than the second resolution is acquired within the sample area. Based on the NDSI data, multi-threshold progressive testing is performed on each raster pixel of the second-resolution snow depth distribution map to obtain the confidence weight of each raster pixel. The confidence weight and fine-tuning training data are input into the pre-trained diffusion model, and the loss value is calculated by combining the confidence weight, predicted snow depth parameters, and training labels to complete the fine-tuning training.

[0075] The confidence weight is calculated using the following formula:

[0076] In the formula, Represents a cell index. This is the NDSI value of the pixel. For the first A threshold, For second spatial resolution snow depth data in pixels The false label snow depth value at the location; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The threshold is the total number.

[0077] This embodiment uses NDSI data to construct confidence weights to achieve weakly supervised training, accurately quantifies the error level of the second-resolution snow depth pseudo-label, and allows high-confidence pixels to have higher weights in training, effectively reducing the error interference of noisy labels, improving the model's learning accuracy of the real snow distribution, and making the snow depth prediction of the fine-tuned model more consistent with the actual spatial characteristics.

[0078] In one embodiment, the NDSI data value ranges from 0 to 100, the incremental NDSI threshold value ranges from 20 to 80, and the total number of thresholds K ranges from 3 to 5; the larger the confidence weight value, the higher the weight ratio of the corresponding raster pixel in the loss value calculation.

[0079] This embodiment clarifies the range of NDSI data and threshold values, providing a specific practical standard for the calculation of confidence weights, avoiding arbitrariness in threshold selection, ensuring the scientific nature and consistency of weight evaluation, further improving the stability of weakly supervised training, and making the weight construction results of different sample regions comparable and adaptable.

[0080] In one embodiment, the training samples of the fine-tuning training data are the snow depth parameters of a single raster cell in the first-resolution snow depth distribution map and their corresponding environmental spatial features; the training labels of the fine-tuning training data are the snow depth parameters of several raster cells in the second-resolution snow depth distribution map corresponding to the geographical range of a single raster cell in the first-resolution snow depth distribution map.

[0081] The training samples are the input units for the fine-tuning stage, containing snow depth values ​​of a single grid at the first resolution and associated environmental spatial features (such as spatial indicators like terrain and precipitation). The training labels are the target values ​​for the fine-tuning stage, which are the set of all high-precision snow depth values ​​at the second resolution within the region corresponding to the single grid at the first resolution. These labels are used to measure the degree of matching between the high-resolution snow depth distribution generated by the model and the true values.

[0082] Step S103, which involves fine-tuning the pre-trained diffusion model using the fine-tuning training data to obtain the fine-tuned diffusion model, includes: Step S1031: Input the fine-tuning training data into the pre-trained diffusion model to obtain the predicted snow depth parameters of all second-resolution raster pixels corresponding to each training sample; calculate the loss value based on the predicted snow depth parameters of all second-resolution raster pixels corresponding to each training sample and the corresponding training label; when the loss value is greater than the preset fine-tuning loss threshold, adjust the training parameters of the diffusion model, re-input the fine-tuning training data into the diffusion model, until the loss value is less than or equal to the preset fine-tuning loss threshold, determine that the fine-tuning training is complete, and obtain the fine-tuned diffusion model.

[0083] The loss value is an error metric calculated by comparing the high-resolution snow depth parameter set output by the model with the training labels, reflecting the deviation between the model's current high-resolution generation capability and the real situation.

[0084] The preset fine-tuning loss threshold is a pre-defined convergence standard for fine-tuning. When the loss value drops below this threshold, fine-tuning is considered complete, and the model has mastered the accurate conversion rules from low resolution to high resolution. In one embodiment, the preset fine-tuning loss threshold ranges from 0.005 to 0.05, which is less than the loss threshold in the pre-training stage, to improve the prediction accuracy of the model after fine-tuning. The initial learning rate for fine-tuning training is set to 1e-5-1e-4, which is lower than the learning rate in the pre-training stage, to avoid overfitting in training with noisy labels.

[0085] This step first inputs the fine-tuning training data into the pre-trained model. Based on the learned environment-snow depth relationship, the model generates predicted snow depth values ​​for all second-resolution pixels within the corresponding region of a first-resolution single-grid. Then, it calculates the loss value between the predicted value set and the training labels (the true high-resolution snow depth set). If this value exceeds a preset fine-tuning loss threshold, the model parameters are adjusted (e.g., optimizing feature fusion weights, enhancing detail generation capabilities), and the fine-tuning training data is re-inputted. This process is repeated until the loss value falls below the threshold, at which point the diffusion model has completed fine-tuning. This step, through high-resolution target constraints, enables the model to accurately grasp the transformation rule of "low-resolution snow depth + environmental features → high-resolution snow depth set," strengthening the model's spatial detail restoration capabilities.

[0086] This embodiment employs a fine-tuning training data design and precise optimization mechanism to enable the diffusion model to deeply learn the conversion rules between low-resolution snow depth parameters and environmental spatial features into high-resolution snow depth distribution maps during the fine-tuning stage. Specifically, the joint input of low-resolution single-grid snow depth and environmental features in the training samples, combined with the constraints of high-resolution snow depth sets in the training labels, prompts the model to focus on improving its spatial detail generation capabilities while maintaining the basic environment-snow depth correlation. Meanwhile, the precise calculation of the loss value and parameter adjustment ensure that the high-resolution snow depth sets generated by the model gradually approach the true values, forming a reliable high-resolution conversion capability. This fine-tuning mechanism synergizes with the pre-training stage: pre-training establishes the basic environment-snow depth mapping, while fine-tuning enhances the resolution conversion details. Together, they enable the model to more accurately fuse low-resolution snow depth data and environmental features of the monitored area, generating high-resolution snow depth distribution maps with high spatial detail.

[0087] For step S104, the first resolution snow depth distribution map and environmental spatial feature data of the area to be monitored are obtained, and the fine-tuned diffusion model is input to obtain the second resolution snow depth distribution map of the area to be monitored; the snow depth distribution data of the area to be monitored is obtained based on the second resolution snow depth distribution map of the area to be monitored.

[0088] The area to be monitored refers to a specific geographical area where high-precision snow depth distribution data needs to be obtained. It may be different from the sample area, or it may be the main area where the sample area is located (i.e., the sample area is a part of the area selected from the area to be monitored).

[0089] When snow depth distribution data for a monitored area is required, a first-resolution snow depth distribution map and environmental spatial feature data for that area are first acquired. These data are then input into a finely tuned and trained diffusion model. Based on previously learned correlation patterns and mapping relationships, the model outputs a second-resolution snow depth distribution map for the monitored area. Finally, based on this high-resolution snow depth distribution map, the snow depth distribution data for the monitored area can be obtained, meeting the demand for high-precision snow depth data in practical applications.

[0090] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for monitoring snow depth distribution, characterized in that, Includes the following steps: A first-resolution snow depth distribution map, environmental spatial feature data affecting snow cover distribution, and a second-resolution snow depth distribution map of the sample area are obtained; wherein, the resolution of the second-resolution snow depth distribution map is greater than the resolution of the first-resolution snow depth map. Based on the first resolution snow depth distribution map and the environmental spatial feature data, pre-training data is constructed; The diffusion model is pre-trained using the pre-training data to obtain a pre-trained diffusion model; the pre-training data is used to enable the diffusion model to learn the correlation between the snow depth parameters of raster pixels and environmental spatial features. Based on the first resolution snow depth distribution map, the environmental spatial feature data, and the second resolution snow depth distribution map, fine-tuning training data is constructed, and the pre-trained diffusion model is fine-tuned to obtain the fine-tuned diffusion model, so that the diffusion model learns the mapping relationship between the first resolution snow depth distribution map and the second resolution snow depth distribution map. Obtain a first-resolution snow depth distribution map and environmental spatial feature data of the area to be monitored, input the fine-tuned diffusion model, and obtain a second-resolution snow depth distribution map of the area to be monitored; based on the second-resolution snow depth distribution map of the area to be monitored, obtain the snow depth distribution data of the area to be monitored.

2. The snow depth distribution monitoring method according to claim 1, characterized in that, The training samples of the pre-training data are the snow depth parameters of the raster pixels in the third-resolution snow depth distribution map and the corresponding environmental spatial features; the training labels of the pre-training data are the snow depth parameters of the corresponding raster pixels in the first-resolution snow depth distribution map; wherein, the third-resolution snow depth distribution map is a snow depth distribution map of the same resolution obtained by downsampling and then interpolating the first-resolution snow depth distribution map. The step of pre-training the diffusion model using the pre-training data to obtain a pre-trained diffusion model includes: The pre-trained data is input into the diffusion model to obtain the predicted snow depth parameters corresponding to each training sample. The loss value is calculated based on the predicted snow depth parameters and corresponding training labels of each training sample. When the loss value is greater than a preset loss threshold, the training parameters of the diffusion model are adjusted and retrained until the loss value is less than or equal to the preset loss threshold, and the pre-training of the diffusion model is determined to be complete.

3. The snow depth distribution monitoring method according to claim 1, characterized in that, The training samples of the fine-tuning training data are the snow depth parameters of a single raster cell in the first resolution snow depth distribution map and its corresponding environmental spatial features; the training labels of the fine-tuning training data are the snow depth parameters of several raster cells in the second resolution snow depth distribution map within the area corresponding to the single raster cell. The step of fine-tuning the pre-trained diffusion model using the fine-tuning training data to obtain the fine-tuned diffusion model includes: The fine-tuned training data is input into the pre-trained diffusion model to obtain the predicted snow depth parameters of all second-resolution raster pixels corresponding to each training sample. The loss value is calculated based on the predicted snow depth parameters of all second-resolution raster pixels corresponding to each training sample and the corresponding training label. When the loss value is greater than the preset fine-tuning loss threshold, the training parameters of the diffusion model are adjusted, and the fine-tuned training data is re-inputted into the diffusion model until the loss value is less than or equal to the preset fine-tuning loss threshold. The fine-tuning training is then completed, and the fine-tuned diffusion model is obtained.

4. The snow depth distribution monitoring method according to claim 1, characterized in that, The steps for obtaining a second-resolution snow depth distribution map of the sample area include: Obtain snow cover distribution probability data for the sample area; wherein the resolution of the snow cover distribution probability data is consistent with the resolution of the second resolution snow depth distribution map; Based on the snow cover distribution probability data, the first resolution snow depth distribution map is decomposed into sub-pixel values ​​to obtain the second resolution snow depth distribution map.

5. The snow depth distribution monitoring method according to claim 4, characterized in that, The steps to obtain snow cover distribution probability data for a sample area include: Obtain snow cover score data and clear sky index data for the sample area; The snow cover percentage data is compared with the corresponding clear skies index data to obtain the snow distribution probability data.

6. The snow depth distribution monitoring method according to claim 4, characterized in that, The step of performing sub-pixel decomposition on the first resolution snow depth distribution map based on the snow distribution probability data to obtain a second resolution snow depth distribution map includes: For each grid cell in the first resolution snow depth distribution map, determine a number of corresponding second resolution sub-cells in the snow cover distribution probability data; The sum of probabilities is calculated based on the snow distribution probability parameters of the aforementioned sub-pixels of second resolution. Divide the snow distribution probability parameter of each of the plurality of second resolution sub-pixels by the sum of the probabilities to obtain the corresponding weight value; Based on the weight value of each second resolution sub-pixel in the plurality of second resolution sub-pixels, the snow depth parameters of the corresponding raster pixels in the first resolution snow depth distribution map are weighted and allocated to obtain the second resolution snow depth distribution map.

7. The snow depth distribution monitoring method according to claim 1, characterized in that, The diffusion model includes a forward diffusion module, a spatiotemporal feature encoding module, and a denoising and reconstruction module; The forward diffusion module is used to gradually add Gaussian noise to the first resolution snow depth distribution map in time steps according to a preset noise scaling factor, so that it gradually forms a pure noise distribution and obtains the intermediate snow depth data at each time step. The spatiotemporal feature encoding module is used to extract and enhance the spatiotemporal features of the snow depth parameters and environmental spatial features of the first resolution snow depth distribution map, and obtain a spatiotemporal enhanced feature tensor that integrates snow depth and environmental information. The denoising and reconstruction module uses U-Net as the backbone network to receive intermediate snow depth data and the spatiotemporal enhancement feature tensor at each time step, gradually remove noise from the intermediate snow depth data, and reconstruct the intermediate snow depth data into a second-resolution snow depth distribution map based on the spatiotemporal enhancement feature tensor.

8. The snow depth distribution monitoring method according to claim 7, characterized in that, The spatiotemporal feature coding module includes a spatial feature coding unit, a temporal feature coding unit, and a convolutional block attention module; The spatial feature encoding unit is used to extract and analyze the spatial dimension features of the snow depth parameters and environmental spatial features of the first resolution snow depth distribution map to obtain the snow depth-environment spatial coupling features. The temporal feature encoding unit is used to perform temporal feature mapping on the snow depth-environment spatial coupling features, adapt to the time step noise change law of the diffusion model, and obtain the preliminary spatiotemporal fusion features of snow depth-environment. The convolutional block attention module is used to assign attention weights to the initial spatiotemporal fusion features of snow depth and environment, strengthen key features related to snow depth, suppress redundant interference features, and obtain a spatiotemporal enhanced feature tensor that fuses snow depth and environment information.

9. The snow depth distribution monitoring method according to claim 1, characterized in that, The environmental spatial characteristic data includes at least one of precipitation data, surface temperature data, snow cover data, surface reflectance data, vegetation cover data, elevation data, latitude data, and land use data.

Images