Machine learning based snow depth detection method, device, medium, and product

Abstract

The application discloses a kind of based on machine learning's snow depth detection method, equipment, medium and product, it is related to machine learning field.The method includes obtaining the multi-source remote sensing data of the region to be measured;Multi-source remote sensing data is spatially registered and resampled to obtain multidimensional data set;For each position point, corresponding observation value and parameter value are extracted from multidimensional data set to construct remote sensing feature vector;Remote sensing feature vector is input into the preset snow depth state classification model, and the snow state of each position point is qualitatively classified to obtain a snow depth category representing that the position point belongs to the preset snow depth interval;According to the snow depth category, from the snow depth estimation regression model library containing a plurality of snow depth estimation regression models corresponding to different snow depth intervals, the matching snow depth estimation regression model is selectively called for the position point;The remote sensing feature vector of position point is input into the snow depth estimation regression model, and the snow depth prediction value of position point is obtained.The application can alleviate the problem that snow depth remote sensing detection precision is not high.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a snow depth detection method, device, medium and product based on machine learning. Background Technology

[0002] Snow depth is a key parameter affecting global climate and water resource cycles. Passive microwave remote sensing has become the mainstream technology for large-scale snow depth monitoring because it can penetrate clouds and dark environments and is sensitive to the internal structure of snow layers.

[0003] Related techniques typically utilize differences in surface brightness temperature observed by satellites across different microwave frequency channels to invert snow depth by establishing empirical or semi-empirical regression models. These methods have achieved, to some extent, macroscopic estimations of snow depth.

[0004] However, the physical relationship between snow depth and microwave brightness temperature is highly nonlinear and is subject to complex interference from various factors such as vegetation, topography, and snow grain size. In particular, the microwave radiation characteristics differ significantly from no snow, thin snow, to deep snow. Related technologies employ a single, universal regression model, attempting to fit all snow cover states with a uniform function. This "one-size-fits-all" approach fails to accurately capture the unique patterns of different snow depth stages, leading to significant estimation errors under complex surface conditions. Summary of the Invention

[0005] To address the aforementioned technical problems and deficiencies, the purpose of this invention is to provide a snow depth detection method, device, medium, and product based on machine learning, which can alleviate the problem of low accuracy in snow depth remote sensing detection.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a snow depth detection method based on machine learning, comprising the following steps: acquiring multi-source remote sensing data of the area to be measured, wherein the multi-source remote sensing data includes satellite passive microwave observation data and auxiliary geographic data, wherein the microwave brightness temperature information is the brightness temperature observation value characterizing the radiation intensity of the land surface in multiple microwave frequency bands, and the auxiliary geographic data is a parameter that can reflect static or slowly changing physical properties of the land surface that affect microwave radiation transmission; performing spatial registration and resampling processing on the multi-source remote sensing data to obtain a multidimensional dataset with spatiotemporal matching on a unified geographic grid; and for each location point on the unified geographic grid, from the multidimensional... The corresponding observation values ​​and parameter values ​​are extracted from the dataset to construct a remote sensing feature vector containing the microwave brightness temperature information and the auxiliary geographic data. The remote sensing feature vector is input into a preset snow depth state classification model to qualitatively classify the snow cover state of each location point, obtaining a snow depth category that represents the location point belonging to a preset snow depth interval. According to the snow depth category, a matching snow depth estimation regression model is selectively called for the location point from a library containing multiple snow depth estimation regression models corresponding to different snow depth intervals. The remote sensing feature vector of the location point is input into the snow depth estimation regression model to obtain the predicted snow depth value of the location point.

[0007] This invention addresses the problem that single regression models in related technologies struggle to accurately fit the complex nonlinear relationship between snow depth and brightness temperature. It proposes an innovative two-tiered "classification-regression" strategy. Instead of attempting a universal model to handle all situations, this invention first utilizes feature vectors integrating multi-source remote sensing information to qualitatively stratify snow cover using a snow depth state classification model, accurately determining its snow depth range. Then, based on this classification result, it intelligently matches and calls the optimal estimation model for the current state from a model library containing multiple specialized regression models. This divide-and-conquer approach decomposes a complex global fitting problem into multiple relatively simple local fitting problems, allowing each model to focus only on the physical laws of a specific snow depth stage, thus more accurately capturing the unique microwave radiation characteristics of each stage. This significantly improves the model's adaptability and estimation accuracy under complex surface conditions such as vegetation and topography, effectively overcoming the technical shortcomings of large estimation errors and achieving more reliable and robust snow depth detection.

[0008] Optionally, in some embodiments, constructing a remote sensing feature vector containing the microwave brightness temperature information and the auxiliary geographic data includes: extracting horizontal polarization brightness temperature values ​​and vertical polarization brightness temperature values ​​from at least two different frequency channels of the satellite passive microwave observation data; extracting topographic elevation values, vegetation index values, and land cover type values ​​from the auxiliary geographic data; and combining the horizontal polarization brightness temperature values, the vertical polarization brightness temperature values, the topographic elevation values, the vegetation index values, and the land cover type values ​​to form the remote sensing feature vector.

[0009] By employing the aforementioned technical solution and clarifying the specific composition of the remote sensing feature vectors, the effectiveness of the method is further enhanced. It stipulates that the feature vectors must include core microwave brightness temperature information (horizontal and vertical polarization brightness temperatures of at least two frequency channels) and key auxiliary geographic data (topographic elevation, vegetation index, and land cover type). This multi-source information fusion construction method enables the feature vectors to more comprehensively and three-dimensionally depict the physical environment of the Earth's surface, providing richer discriminative basis for subsequent machine learning models. It effectively decouples the complex coupling effect between snow depth signals and surface interference factors, thereby directly improving the performance of classification and regression models and forming an important foundation for achieving high-precision detection.

[0010] Optionally, in some embodiments, the training method of the snow depth state classification model includes: obtaining a training sample set containing historical remote sensing feature vectors and their corresponding actual snow depth label values; classifying and labeling each sample in the training sample set according to a preset snow depth threshold to obtain a category label; and using all remote sensing feature vectors of the training sample set and the category label as input to train and generate the snow depth state classification model.

[0011] The above technical solution clarifies the scientific training method for the snow depth classification model, ensuring classification accuracy. By using historical data labeled with actual snow depth and annotating it according to preset thresholds, this method constructs a reliable training supervision signal. This supervised learning process based on a large number of real samples ensures that the trained classification model can learn the intrinsic relationship between remote sensing features and the actual snow depth, rather than blindly fitting. This enables the model to make physically based, high-confidence qualitative judgments about snow cover status during actual detection, providing a solid and reliable guarantee for the correct selection of the regression model, and is a key prerequisite for the success of the entire "classification-regression" strategy.

[0012] Optionally, in some embodiments, the method for obtaining the snow depth estimation regression model library includes: obtaining a training sample set that has been classified and labeled; dividing the training sample set into multiple sample subsets that correspond one-to-one with the snow category according to the category label; for each sample subset, training a regression model separately using the remote sensing feature vector and the actual snow depth label value within the subset; and combining all the trained regression models into the snow depth estimation regression model library.

[0013] The above technical solution elucidates the construction logic of the snow depth estimation regression model library, which is the core step in achieving refined estimation. This method divides training samples according to snow depth categories and trains a dedicated regression model for each subset. This "expert model" strategy ensures that each regression model focuses on fitting unique and relatively simplified physical relationships within a specific snow depth range, avoiding the systematic bias that occurs when a single model fits a wide range of nonlinear curves. It decomposes the complex global modeling task, significantly reducing the learning difficulty of each sub-model, thereby greatly improving the quantitative estimation accuracy at various snow depth stages. This is the key technological innovation of this invention in achieving high-precision results.

[0014] Optionally, in some embodiments, after inputting the remote sensing feature vector of the location point into the snow depth estimation regression model to obtain the predicted snow depth value of the location point, the method further includes: generating single-track snow depth data to characterize the results of a single satellite transit observation based on the predicted snow depth values ​​of all location points in the area to be measured within a single day; projecting the single-track snow depth data onto a unified daily composite spatial grid; calculating the arithmetic mean of all predicted snow depth values ​​falling within the same daily composite spatial grid; and determining the arithmetic mean as the final daily composite snow depth value of the daily composite spatial grid.

[0015] The above-described technical solution provides a post-processing method for generating standardized daily products from instantaneous observations. By spatiotemporally integrating single-track snow depth data and calculating the average value within the daily synthetic spatial grid, this method effectively suppresses random noise and transient anomalies that may exist in a single observation, and integrates observation information from multiple transits within the same day. This not only improves the spatial integrity and continuity of snow depth data but also enhances its temporal representativeness and stability, making it more suitable for applications requiring daily temporal resolution, such as climate and hydrology, and significantly improving the practical value and reliability of snow depth products.

[0016] Optionally, in some embodiments, after generating single-track snow depth data characterizing the results of a single satellite transit observation based on the predicted snow depth values ​​of all locations within the test area within a single day, the method further includes: obtaining the calibrated snow depth values ​​measured by ground stations corresponding to the locations within the same time window; calculating the deviation between the predicted snow depth value and the calibrated snow depth value at the same location; and, based on the deviation value, applying an incremental learning algorithm to fine-tune the model parameters in the snow depth state classification model and / or the snow depth estimation regression model library online to obtain an updated snow depth state classification model and snow depth estimation regression model library.

[0017] By adopting the above technical solution, an online model update mechanism is introduced, endowing the method with dynamic adaptation and self-optimization capabilities. By calculating model bias using real-time ground-calibrated snow depth values ​​and applying an incremental learning algorithm to fine-tune model parameters, this method can continuously learn and adapt to dynamic environmental changes (such as seasonal variations in snow characteristics). This online learning capability avoids the problem of model performance decaying over time, ensuring the long-term stability and high accuracy of the detection system. It transforms the system from a static model application system into a dynamic detection system with intelligent evolution capabilities, greatly enhancing the robustness and lifespan of the method.

[0018] Optionally, in some embodiments, after constructing the remote sensing feature vector containing the microwave brightness temperature information and the auxiliary geographic data, the method further includes: inputting the remote sensing feature vector into a simplified microwave radiative transfer physical model to calculate a physical baseline snow depth value; inputting the remote sensing feature vector into a pre-trained residual classification model to obtain the residual interval category in which the physical baseline snow depth value is located; selectively calling a matching, pre-trained residual regression model according to the residual interval category; inputting the remote sensing feature vector into the selected residual regression model to calculate a snow depth residual correction value; summing the physical baseline snow depth value and the snow depth residual correction value to obtain the final snow depth prediction value, and using it as the single-track snow depth data.

[0019] The above technical solution provides an alternative approach of "physical model + residual correction," cleverly combining the stability of physical mechanisms with the fitting capabilities of machine learning. This method first uses a simplified physical model to obtain a baseline snow depth with clear physical meaning. Then, a two-level machine learning framework (residual classification and regression) is used to accurately predict and compensate for the systematic biases of the physical model. This approach not only preserves the interpretability and basic accuracy of the physical model but also overcomes the limitations of the physical model under complex conditions through a data-driven approach, achieving refined correction of snow depth predictions. This provides an efficient and robust alternative for high-precision snow depth detection.

[0020] In a second aspect, the present invention provides an electronic device comprising: one or more processors and a memory; the memory being coupled to the one or more processors, the memory being used to store computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the electronic device to perform the methods described in the first aspect and any possible implementation thereof.

[0021] Thirdly, the present invention provides a computer-readable storage medium storing instructions that, when executed on an electronic device, cause the electronic device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, the present invention provides a computer program product comprising instructions that, when the computer program product is run on an electronic device, cause the electronic device to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the electronic device provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided by this invention. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a snow depth detection method based on machine learning according to an embodiment of the present invention.

[0025] Figure 2 This is an example diagram of the spatial distribution of FY3D-MWRI snow depth orbit in an embodiment of the present invention;

[0026] Figure 3 This is an example diagram of another FY3D-MWRI snow depth orbital spatial distribution map in an embodiment of the present invention;

[0027] Figure 4 This is a flowchart illustrating another snow depth detection method based on machine learning according to an embodiment of the present invention.

[0028] Figure 5 This is an example diagram of a spatial distribution map of FY3D-MWRI snow depth diurnal composite in an embodiment of the present invention;

[0029] Figure 6 This is a flowchart of a snow depth orbit algorithm for a snow depth detection method according to an embodiment of the present invention;

[0030] Figure 7This is a scatter plot illustrating the accuracy verification of the predicted and actual snow depth values ​​in this embodiment of the invention.

[0031] Figure 8 This is a schematic diagram of the hardware architecture of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0032] Snow depth detection methods typically rely on multi-frequency channel brightness temperature (Tb) data acquired by spaceborne passive microwave radiometers. The basic principle is based on the volume scattering effect of snow on microwaves. This effect causes the microwave radiation intensity of high-frequency channels to decrease with increasing snow depth, resulting in a correlation between the brightness temperature difference (Tb-difference) between different frequency channels (e.g., 19 GHz and 37 GHz) and snow depth. Based on this, researchers establish empirical or semi-empirical linear / nonlinear regression models to fit the directly observed brightness temperature or brightness temperature difference from satellites with measured ground snow depth data, thereby estimating snow depth in unknown areas. These methods have, to some extent, solved the problem of snow depth estimation on a macroscopic scale and have been applied in some global snow water equivalent (SWE) products.

[0033] However, this type of method has inherent limitations. First, the physical relationship between snow depth and microwave brightness temperature is highly nonlinear and is subject to complex coupling interference from various factors such as surface vegetation cover, topographic complexity, and the microscopic physical structure of snow (e.g., snow grain size, density, and stratification). Second, and more importantly, the microwave radiation mechanism of the surface varies significantly under different snow cover conditions: in the snowless or thin snow (optically thin layer) stage, microwave radiation is mainly controlled by surface emissivity; as snow depth increases, the volume scattering effect within the snow layer gradually becomes dominant; while in the deep snow (optically thick layer) stage, the high-frequency microwave signal tends to saturate, and the sensitivity of brightness temperature difference to snow depth decreases significantly.

[0034] Related technologies generally adopt a single, universal regression model, attempting to construct a unified function to describe the entire process from no snow to deep snow. This "one-size-fits-all" approach ignores the unique physical laws of different snow accumulation stages and cannot accurately capture the nonlinear characteristics of each stage. Therefore, in areas with complex conditions such as vegetation and terrain, its estimation accuracy often fails to meet the needs of practical applications, resulting in large estimation errors.

[0035] Therefore, this invention provides a snow depth detection method based on machine learning, which can solve the problem in related technologies where passive microwave remote sensing is used to detect snow depth, but the single, universal regression model is difficult to adapt to the complex nonlinear relationships under different snow cover conditions, resulting in low estimation accuracy under complex surface conditions.

[0036] The method in this embodiment no longer uses a single model to handle all cases, but introduces a two-level strategy of "classification-regression" to perform refined and differentiated modeling and estimation of different snow cover states.

[0037] Specifically, the first step is to unify and standardize the input data. This involves acquiring multi-source remote sensing data of the area to be measured, including not only core satellite passive microwave brightness temperature observations but also auxiliary geographic data (such as elevation, slope, land cover type, and vegetation index) that reflect static or slowly varying surface physical properties affecting microwave radiation transmission. Through spatial registration and resampling, all data are ensured to be spatiotemporally matched on a unified geographic grid, constructing a comprehensive and dimensionally consistent multidimensional dataset for subsequent analysis.

[0038] Next, for each location point on the grid, all relevant observation and parameter values ​​are extracted from the multidimensional dataset to construct a remote sensing feature vector that can comprehensively characterize the physical environment of that point.

[0039] The innovation of this embodiment lies in that it does not directly use the feature vector for snow depth regression. Instead, it first inputs it into a pre-trained snow depth classification model. This classification model is a qualitative analysis model whose task is to determine the snow cover status at the current location into one of several preset categories, such as "no snow," "light snow," "moderate snow depth," or "deep snow." This classification result, i.e., the snow depth category, provides crucial prior knowledge for subsequent quantitative estimation.

[0040] Subsequently, this embodiment uses this classification result to make a decision. It pre-defines a snow depth estimation regression model library, which contains multiple independent snow depth estimation regression models, each of which is specifically trained and optimized for a particular snow depth category (i.e., snow depth interval). Based on the snow depth category obtained in the previous step, the method selectively calls the regression model that uniquely matches it from the model library.

[0041] Finally, the remote sensing feature vector of this location is used as input and fed into the selected, more specialized snow depth estimation regression model to perform the final quantitative calculation of the snow depth value, thereby obtaining a more accurate snow depth prediction value for this location.

[0042] The following is combined Figure 1 This embodiment describes a snow depth detection method based on machine learning. The execution entity of this snow depth detection method is a snow depth detection system (hereinafter referred to as the system). This snow depth detection system can be deployed on a server, cloud computing platform, or dedicated data processing workstation, and completes the following steps by executing preset programs or instructions:

[0043] Step 101: Obtain multi-source remote sensing data of the area to be measured. The multi-source remote sensing data includes satellite passive microwave observation data and auxiliary geographic data. Microwave brightness temperature information is the brightness temperature observation value that characterizes the radiation intensity of the surface in multiple microwave frequency bands. Auxiliary geographic data is a parameter that can reflect the static or slowly changing physical properties of the surface that affect microwave radiation transmission.

[0044] Specifically, the snow depth detection system first needs to acquire multi-source remote sensing data of the area to be measured within a specific time range from multiple data sources. This data forms the basis for subsequent analysis. The multi-source remote sensing data includes satellite passive microwave observation data and auxiliary geographic data.

[0045] In a preferred embodiment, satellite passive microwave observation data originates from passive microwave radiometers mounted on meteorological or Earth observation satellites, such as AMSR-E / AMSR2 and the Fengyun (FY) series satellite microwave imager (MWRI). The snow depth detection system downloads L1 or L2 level brightness temperature data by accessing official data distribution websites such as satellite meteorological centers. This data includes observations of surface radiation intensity in multiple key microwave frequency bands, specifically including horizontal (H) and vertical (V) polarized brightness temperatures at frequencies such as 10.65 GHz, 18.7 GHz, 23.8 GHz, 36.5 GHz, and 89.0 GHz. These brightness temperature values ​​at different frequencies and polarizations exhibit varying sensitivities to snow physical parameters such as snow depth and snow grain size, serving as the core information source for snow depth retrieval.

[0046] Ancillary geographic data are parameters that reflect static or slowly varying surface physical properties that influence microwave radiation transmission. Specifically, the ancillary geographic data acquired by the snow depth sounding system may include:

[0047] 1. Static geographic data: For example, digital elevation models (DEMs) obtained from products such as the Space Shuttle Radar Topographic Mapping Mission (SRTM) or ASTER GDEM, and topographic factors such as slope and aspect calculated based on the DEM; land cover type information such as forest, grassland, and bare soil obtained from MODIS IGBP (International Geosphere-Biosphere Programme) or other land use / land cover (LULC) products.

[0048] 2. Gradually changing geographic data: For example, Normalized Difference Vegetation Index (NDVI) and Leaf Area Index (LAI) obtained from optical / infrared sensor products such as MODIS or VIIRS can be used to characterize the growth status and cover density of vegetation; at the same time, MODIS snow cover products (such as MOD10A1) can be obtained to obtain the surface snow cover fraction (SCF) as a reference to help determine whether there is snow.

[0049] Step 102: Spatial registration and resampling are performed on the multi-source remote sensing data to obtain a multidimensional dataset with spatiotemporal matching on a unified geographic grid.

[0050] In this embodiment, since the multi-source data acquired above comes from different sensors and has different spatial resolutions, projection methods, and raster systems, it is necessary to perform unification processing. In this step, the snow depth detection system performs spatial registration and resampling on all data to generate a multidimensional dataset that is spatiotemporally and completely matched on a unified geographic grid.

[0051] In practice, the snow depth detection system first determines a target geographic grid. For example, the 25-kilometer resolution EASE-Grid 2.0 (Equal-Area Scalable Earth Grid), widely used in polar research, can be selected as the uniform grid. Then, the system uses tools from Geographic Information Systems (GIS) or remote sensing image processing libraries (such as GDAL) to perform the following operations on all data:

[0052] 1. Projection transformation: Convert the coordinate system of all data to the projection system of the target grid (such as the equal area projection of the Northern or Southern Hemisphere of EASE-Grid 2.0).

[0053] 2. Resampling: Interpolating data at different resolutions to the target grid resolution. For continuous data, such as microwave brightness temperature, DEM, NDVI, etc., bilinear interpolation or cubic convolution interpolation can be used to ensure a smooth transition of values; for categorical data, such as land cover types, the nearest neighbor method is used to maintain the originality of category information.

[0054] After this step, for any grid cell (i.e., location point) in the area to be tested, all data values ​​at the same time point (or within a short time window) can be obtained, forming a structured multidimensional dataset, which lays the foundation for subsequent feature extraction.

[0055] Step 103: For each location point on the unified geographic grid, extract the corresponding observation values ​​and parameter values ​​from the multidimensional dataset, and construct a remote sensing feature vector containing microwave brightness temperature information and auxiliary geographic data.

[0056] In this embodiment, under a unified geographic grid system, the entire area to be measured is divided into grid cells of equal size, with the center of each grid cell being a "location point". After completing the spatial registration and resampling of the multi-source remote sensing data, a spatiotemporally matched multidimensional dataset is obtained. This dataset can be understood as a "data cube", whose spatial dimensions correspond to the rows and columns of the geographic grid, while the third dimension represents different data layers, each layer representing a remote sensing observation or geographic parameter.

[0057] In this step, the process of extracting the corresponding observation values ​​and parameter values ​​from the multidimensional dataset for each location point on the unified geographic grid is implemented as follows:

[0058] The snow depth detection system sequentially traverses every location point (i.e., every grid cell) within the geographic grid. For a given location point, the system precisely locates the corresponding cell within the various data layers of the cube based on its geographic coordinates (e.g., the grid row and column number). The system then reads the numerical values ​​from these cells.

[0059] Specifically, the corresponding observations and parameter values ​​refer to the following two types of information:

[0060] Observed values: These mainly refer to physical quantities that change dynamically over time, measured directly or indirectly by satellite sensors. In this embodiment, the most crucial observed value is microwave brightness temperature information. Specifically, during extraction, the snow depth detection system reads the radiation intensity at a given location in multiple different microwave frequency bands from the microwave brightness temperature information data layer of the multidimensional dataset; these are the brightness temperature observed values. For example, the system extracts the horizontal and vertical polarization brightness temperature values ​​at 18.7 GHz, 36.5 GHz, and potentially other frequencies (such as 10.65 GHz, 23.8 GHz, and 89.0 GHz). Furthermore, vegetation index values ​​(such as NDVI) extracted from the optical remote sensing data layer also belong to dynamically changing observed values.

[0061] Parameter values: These mainly refer to relatively static or slowly changing parameters that reflect the physical properties of the Earth's surface, i.e., auxiliary geographic data. Specifically, during extraction, the snow depth detection system reads the parameter values ​​for that location from the corresponding auxiliary data layer. For example, it reads the topographic elevation value (in meters) for that location from the Digital Elevation Model (DEM) data layer; the slope value (in degrees) from the slope data layer calculated based on the DEM; and the category code representing the land cover type at that location (e.g., an integer, 1 for forest, 2 for grassland, etc.) from the land cover type data layer. These parameters provide necessary geographic background information for understanding microwave signals.

[0062] After extracting all the necessary observation and parameter values ​​for this location, the process of constructing a remote sensing feature vector containing the microwave brightness temperature information and the auxiliary geographic data is as follows:

[0063] The snow depth detection system will arrange all the independent numerical values ​​extracted in the previous step in a predefined, fixed order to form a one-dimensional numerical array. This array is the remote sensing feature vector of that location. This construction process essentially integrates the scattered numerical information describing the various physical characteristics of that location into a structured mathematical object suitable for processing by machine learning models.

[0064] For example, a specific composition of a remote sensing feature vector could be: [18.7GHz horizontal polarization brightness temperature value, 18.7GHz vertical polarization brightness temperature value, 36.5GHz horizontal polarization brightness temperature value, 36.5GHz vertical polarization brightness temperature value, terrain elevation value, slope value, vegetation index value, land cover type code]

[0065] Suppose the extracted values ​​for a certain location are: [255.4K, 260.1K, 240.7K, 248.3K, 1500.5 meters, 12.3 degrees, 0.25, 1]. This array itself constitutes the remote sensing feature vector for that location. In practical applications, to improve model performance, these raw values ​​may be normalized or standardized, and classification features such as land cover type may be one-hot encoded. However, the core principle remains the same: combining this multi-source information into a unique feature vector.

[0066] By repeating the above extraction and construction process for each location point on a unified geographic grid, a corresponding remote sensing feature vector will eventually be generated for all points in the area to be measured, laying the foundation for subsequent classification and regression analysis.

[0067] Step 104: Input the remote sensing feature vector into the preset snow depth state classification model to perform qualitative classification of the snow cover state at each location point, and obtain a snow depth category that represents that the location point belongs to the preset snow depth range.

[0068] In this step, the snow depth detection system inputs the remote sensing feature vector constructed in the previous step into a pre-trained snow depth state classification model to perform qualitative classification of the snow cover state at each location point.

[0069] The snow depth classification model is a multi-class machine learning model, such as a Support Vector Machine (SVM), Random Forest, Gradient Boosting Decision Tree (e.g., XGBoost, LightGBM), or a Multilayer Perceptron (MLP) neural network. This model is trained using a large amount of training data with real-world labels. The features in the training data are remote sensing feature vectors, while the labels are predefined snow depth categories based on measured ground snow depth values.

[0070] For example, the following snow depth ranges can be preset as categories:

[0071] Category 0: No snow (snow depth = 0cm)

[0072] Category 1: Light snow (0cm < snow depth ≤ 5cm)

[0073] Category 2: Moderate snow (5cm < snow depth ≤ 30cm)

[0074] Category 3: Deep snow / Saturated snow (snow depth > 30cm)

[0075] In actual detection, the snow depth detection system inputs the feature vector V(p,t) of the location to be measured into the trained classification model. The model will output a unique category label, such as "Category 2". This label is the qualitative judgment result of the snow cover status of the location at that moment.

[0076] Step 105: Based on the snow depth category, selectively call the matching snow depth estimation regression model for the location point from a library containing multiple snow depth estimation regression models corresponding to different snow depth ranges.

[0077] Based on the snow depth category obtained in the previous step, the snow depth detection system will selectively call a matching snow depth estimation regression model for the location point from a preset snow depth estimation regression model library.

[0078] The snow depth estimation regression model library is a collection of multiple independent regression models. Each regression model in the library is specifically trained for a particular snow depth category. Specifically:

[0079] Model library construction: During the model training phase, all training samples are divided into different subsets based on their actual snow depth labels (corresponding to categories such as no snow, light snow, moderate snow, and deep snow). Then, a separate regression model is trained for each subset. For example, a "moderate snow regression model" is trained using samples from the "moderate snow" subset (actual snow depth between 5-30cm).

[0080] Model types: These regression models can be support vector regression (SVR), random forest regressors, gradient boosting regressors, or neural network regression models, etc.

[0081] Model invocation: During the detection phase, this step is an intuitive conditional judgment and selection process. If the category output in step 104 is "Category 2" (medium snow), the snow depth detection system loads and activates the "medium snow regression model" from the model library; if the category is "Category 1", the "thin snow regression model" is invoked, and so on.

[0082] In this way, the system avoids using a universal model to fit all complex snow depth conditions, and instead prepares an "expert model" for each condition.

[0083] Step 106: Input the remote sensing feature vector of the location point into the snow depth estimation regression model to obtain the predicted snow depth value of the location point.

[0084] Specifically, the snow depth detection system inputs the remote sensing feature vector of the location point into the snow depth estimation regression model selected and invoked in step 105 to calculate the final snow depth prediction value.

[0085] For example, if location point p is classified as "Category 2" and the system invokes the "Medium Snow Regression Model," then the feature vector V(p,t) is input into the model. Since the "Medium Snow Regression Model" only learns the patterns of samples with snow depths in the range of 5-30cm, it can more accurately fit the nonlinear relationship within this interval, thus outputting a more accurate continuous snow depth value, such as 17.8cm. For points classified as "No Snow," the system can directly output a snow depth of 0, or verify this using a dedicated "No Snow Model."

[0086] In summary, the embodiments of this invention employ a two-tiered "classification-regression" strategy. First, multi-source data is used to accurately qualitatively stratify snow cover conditions. Then, optimal specialized models are employed for quantitative estimation of snow cover at different levels. This effectively overcomes the problem of insufficient accuracy of traditional single models under complex surface conditions, achieving high-precision and robust snow depth detection. Those skilled in the art will understand that the above embodiments are illustrative, and various modifications and variations can be made without departing from the spirit and scope of this invention.

[0087] Compared with related technologies, the technical solution of this embodiment has the following significant advantages:

[0088] 1. Significantly improved accuracy and reliability of snow depth detection: This invention employs a two-tiered strategy of "classification first, regression later," decomposing a complex, global nonlinear fitting problem into multiple relatively simple, localized sub-problems. Through a snow depth state classification model, the snow cover state is first accurately qualitatively determined, and then a specially optimized regression model is invoked for different states (e.g., no snow, thin snow, deep snow). Each regression model only needs to focus on fitting the physical relationships within a specific snow depth range, avoiding the problems of partiality and error propagation that occur when a single model fits the entire nonlinear curve. This allows for more accurate capture of the unique microwave radiation characteristics at each stage, especially in the transition and saturation regions where signal sensitivity changes, where the accuracy improvement is particularly significant.

[0089] 2. Enhanced model adaptability and robustness to complex surface conditions: This invention integrates auxiliary geographic data such as vegetation and topography into the feature vector and applies it throughout both the classification and regression stages. The snow depth classification model has comprehensively considered the influence of these interfering factors during qualitative judgment, making the classification results more reliable. Based on this, the regression model for specific snow depth intervals can better decouple the interaction between snow depth signals and other surface factors, effectively suppressing the uncertainty caused by surface heterogeneity, thus making the method more robust in areas where traditional methods struggle, such as forests and mountains.

[0090] 3. Improved interpretability and physical significance: Compared to a single "black box" regression model, the two-stage structure of this invention provides a clear intermediate product—snow depth category. This classification result itself has clear physical significance and application value, enabling technicians to intuitively understand the spatial distribution pattern of surface snow cover. This divide-and-conquer modeling approach is closer to the physical process of snow microwave radiation transmission, making the entire detection process more transparent and easier to understand, facilitating further mechanism analysis and model improvement.

[0091] The predicted snow depth obtained through the method in this embodiment is ultimately used to form an orbital snow depth product, such as... Figure 2 and Figure 3 As shown.

[0092] Specifically, Figure 2 An example of the FY3D-MWRI snow depth orbital spatial distribution map generated in this embodiment shows the results of a satellite overpass observation at 14:11 Beijing time on February 1, 2025. This map visually presents the snow depth distribution in northern and northeastern China covered by a single satellite orbital scan. Different colors in the map represent different snow depth values ​​(unit: millimeters), with colors grading from dark blue and green to yellow and red, corresponding to the increasing snow depth from light to dark. Specific values ​​can be seen on the color scale in the lower left corner of the reference map. Figure 2 As shown, most areas of eastern Inner Mongolia, Heilongjiang, and Jilin provinces are covered by snow, with the snow depth in the Northeast region generally being thicker, appearing as large areas of yellow and red. This image has a spatial resolution of 10 kilometers and is an unsynthesized, raw orbital product; its banded coverage clearly reflects the scanning trajectory of the satellite during its transit.

[0093] Figure 3 Another example of the FY3D-MWRI snow depth orbital spatial distribution map generated in this embodiment shows the results of another satellite transit observation at 04:33 Beijing time on February 1, 2025. This map mainly covers western and southwestern China, including Qinghai, Gansu, Tibet, and Sichuan. (See attached image.) Figure 1Similarly, this map also uses color rendering to show the continuous spatial distribution of snow depth. It can be seen from the map that there is a large area of ​​deep snow cover in southeastern Tibet and northwestern Sichuan, represented by concentrated orange and red areas, while the surrounding areas have shallower snow cover or no snow at all. This map is also an orbital product with a spatial resolution of 10 kilometers, and its observation strips and attached... Figure 1 The bands differ in both geographical location and time, and together they constitute part of the multi-orbit observations on that day.

[0094] The following is combined Figure 4 To further illustrate the snow depth detection method provided in this embodiment, the specific steps are as follows: Step 201, obtain a training sample set containing historical remote sensing feature vectors and their corresponding actual snow depth label values.

[0095] The purpose of this step is to prepare the necessary data foundation for subsequent machine learning model training. The snow depth detection system needs to collect and organize a large number of training samples from historical data archives. Each training sample represents a specific geographical location and time point, and contains two core pieces of information: a historical remote sensing feature vector and an actual snow depth label value that precisely matches it.

[0096] The method for constructing historical remote sensing feature vectors is completely consistent with the method for constructing remote sensing feature vectors during subsequent actual explorations, ensuring the consistency of data features between model training and application. The actual snow depth label value is the learning target of the model, and its source is usually high-precision ground observation data, such as snow depth reports from weather stations of the Global Weather Exchange System (GTS), snow pillow (a device that can automatically and continuously measure the weight of snow accumulation and convert it into snow depth or snow water equivalent) data deployed in a specific study area, or snow depth profile data obtained through manual field surveys.

[0097] The snow depth detection system uses a spatiotemporal matching algorithm to precisely pair historical remote sensing observation data with these measured snow depth data on the ground, forming a large training sample set.

[0098] Step 202: Classify and label each sample in the training sample set according to the preset snow depth threshold to obtain the category label.

[0099] This step aims to assign a qualitative snow depth category to each training sample in order to train a snow depth classification model. The snow depth detection system first defines a set of snow depth thresholds to divide different snow depth intervals. For example, thresholds of 0 cm, 5 cm, and 30 cm can be set to define four snow depth intervals, each corresponding to a different snow depth category: Category 0 (no snow: snow depth equals 0 cm), Category 1 (light snow: 0 cm < snow depth ≤ 5 cm), Category 2 (moderate snow: 5 cm < snow depth ≤ 30 cm), and Category 3 (deep snow / saturated snow: snow depth > 30 cm).

[0100] Subsequently, the snow depth detection system traverses each sample in the training sample set, reads its corresponding actual snow depth label value, and assigns a unique category label to the sample according to the snow depth threshold defined above. For example, a sample with an actual snow depth label value of 18 cm will be assigned the category label "Category 2".

[0101] After this step, each sample in the training sample set has both a continuous actual snow depth label value and a discrete class label.

[0102] Step 203: Use all remote sensing feature vectors of the training sample set and the category labels as input to train and generate the snow depth state classification model.

[0103] The goal of this step is to train a classification model that can automatically determine snow cover status based on remote sensing features. The snow depth detection system uses a machine learning classification algorithm suitable for handling multidimensional data and nonlinear relationships, such as a random forest classifier, support vector machine (SVM), or gradient boosting decision tree (e.g., XGBoost).

[0104] Then, the training sample set processed in step 202 is used as input, where the historical remote sensing feature vector of each sample serves as the model's input feature (X), and the corresponding class label serves as the model's learning target (Y). By performing the algorithm's training process on the entire training sample set, the model learns a complex mapping function from remote sensing feature vectors to snow depth class labels. The training process typically includes cross-validation and hyperparameter tuning to ensure the model's generalization ability.

[0105] After training, the final model obtained is the preset snow depth state classification model, which is saved for use in subsequent actual detection.

[0106] Step 204: Obtain the training sample set that has been classified and labeled.

[0107] This step is the preparatory work for training the snow depth estimation regression model. The snow depth detection system directly calls the training sample set that has been processed in step 202. At this point, each sample in the training sample set has three pieces of information: historical remote sensing feature vector, actual snow depth label value, and category label, laying the foundation for subsequent data segmentation by category.

[0108] Step 205: Divide the training sample set into multiple sample subsets that correspond one-to-one with the snow category according to the category label.

[0109] This step aims to divide the highly heterogeneous full training data into multiple subsets of samples with more consistent internal characteristics. The snow depth detection system groups the training sample set according to category labels. For example, all samples labeled "Category 1" (thin snow) are selected to form the "thin snow sample subset"; all samples labeled "Category 2" (moderate snow) form the "moderate snow sample subset," and so on. Typically, samples in the no-snow category (Category 0) are not needed for training the regression model because their snow depth value is a fixed 0. Therefore, this step ultimately generates multiple subsets of samples corresponding to different snow depth ranges.

[0110] Step 206: For each sample subset, use the remote sensing feature vector and actual snow depth label value within that subset to train a separate regression model.

[0111] This step is the core of the divide-and-conquer strategy of this invention, namely, customizing a specific regression model for each snow depth interval. The snow depth detection system independently performs regression model training once for each sample subset obtained in step 205.

[0112] Taking the “medium snow sample subset” as an example, the system selects a regression algorithm (such as support vector regression SVR, random forest regressor or neural network), uses all remote sensing feature vectors in the subset as input features (X), and uses the corresponding actual snow depth label value (which is a continuous value, such as 18 cm) as the learning target (Y).

[0113] Through training, the model will specifically learn and fit a specific nonlinear relationship between remotely sensed features and snow depth within the range of 5 cm to 30 cm snow depth. The same operation is performed on other sample subsets (such as the "thin snow sample subset" and the "deep snow sample subset"), thereby training a dedicated, high-performance regression model for each snowy category.

[0114] Step 207: Combine all the trained regression models into the snow depth estimation regression model library.

[0115] This step aims to systematically manage the multiple trained specialized regression models, forming a model library that can be easily accessed. The snow depth detection system will serialize and save each regression model trained in step 206 (e.g., "thin snow regression model," "medium snow regression model," etc.), that is, convert the model object into a file format. Simultaneously, an index or mapping table is created to record the association between each category label and its corresponding regression model file. This collection, containing multiple model files and an index table, constitutes the snow depth estimation regression model library.

[0116] After the model training and preparation are completed, the snow depth detection system begins to execute the actual snow depth detection process, proceeding to step 208.

[0117] Step 208: Obtain multi-source remote sensing data for the area to be measured. This step can be referred to in the previous embodiment and will not be repeated here.

[0118] Step 209 involves spatial registration and resampling of the multi-source remote sensing data to obtain a spatiotemporally matched multidimensional dataset on a unified geographic grid. This step can be referenced from the previous embodiments and will not be repeated here.

[0119] Step 210: For each location point on the unified geographic grid, extract the corresponding observation and parameter values ​​from the multidimensional dataset. This step can be referred to in the previous embodiment and will not be repeated here.

[0120] Step 211: Extract the horizontal polarization brightness temperature and vertical polarization brightness temperature values ​​of at least two different frequency channels from the satellite passive microwave observation data.

[0121] The horizontal polarization brightness temperature refers to the equivalent temperature value of microwave radiation intensity received by a passive microwave remote sensor when the electric field vector vibration direction is parallel to the incident surface of the Earth's surface. The vertical polarization brightness temperature refers to the equivalent temperature value of microwave radiation intensity received by a passive microwave remote sensor when the electric field vector vibration direction is perpendicular to the incident surface of the Earth's surface.

[0122] Specifically, when processing a location point on a unified geographic grid, the snow depth detection system first locates multiple data layers corresponding to satellite passive microwave observation data within the spatiotemporally matched multidimensional dataset. These data layers are clearly divided according to frequency and polarization, such as the "18.7 GHz horizontal polarization brightness temperature layer," "18.7 GHz vertical polarization brightness temperature layer," "36.5 GHz horizontal polarization brightness temperature layer," and "36.5 GHz vertical polarization brightness temperature layer."

[0123] Subsequently, based on the geographic grid coordinates of the current location, the system locates the corresponding pixels in these specific data layers and directly reads the values ​​stored in these pixels. This value is the brightness temperature observation value for that location at that frequency and polarization mode.

[0124] By repeating this read operation on at least two different frequency (e.g., 18.7 GHz and 36.5 GHz) horizontal and vertical polarization data layers, the system can successfully extract a complete set of core microwave brightness temperature information for that location point for subsequent analysis.

[0125] Step 212: Extract topographic elevation values, vegetation index values, and land cover type values ​​from the auxiliary geographic data.

[0126] This step aims to extract auxiliary information reflecting the surface environment to help the model distinguish microwave signals under different surface conditions. The snow depth detection system continues to extract auxiliary geographic data from the multidimensional dataset at this location, which may include: topographic elevation values, derived from a digital elevation model (DEM), used to characterize the impact of altitude on temperature and snowfall; vegetation index values, such as the Normalized Difference Vegetation Index (NDVI), used to quantify the density and growth status of vegetation cover, as vegetation is an important factor affecting microwave radiation; and land cover type values, derived from land cover products, representing the basic type of the land surface in the form of category codes (such as forest, grassland, water body, etc.).

[0127] This step follows the same logic as searching for data at a specific location point in a cube. When the snow depth detection system processes this location point, it accesses other data layers corresponding to the auxiliary geographic data in the cube in parallel. The system first locates the "topographic elevation data layer" (usually derived from a Digital Elevation Model, DEM) and reads the corresponding pixel value based on the grid coordinates of the location point; this value is the topographic elevation value. In raster data (such as remote sensing imagery), a pixel is the smallest indivisible basic unit that constitutes an image, typically represented as a square or rectangular area. The pixel value is the numerical value stored in a single pixel, representing a quantitative measurement of a certain physical attribute (such as brightness temperature, elevation, or vegetation index) within the geographic area covered by that pixel.

[0128] Next, the system will switch to the "vegetation index data layer" (usually derived from optical remote sensing products such as MODIS NDVI), read the cell values ​​at the same grid coordinates, and obtain the vegetation index value at that location.

[0129] Finally, the system accesses the "Land Cover Type Data Layer" and reads the cell value at the same coordinate location. This value is usually an integer code representing a specific land surface type (e.g., 1 represents forest, 2 represents grassland), thus obtaining the land cover type value for that location.

[0130] Through this series of pixel-level search and read operations, the system can efficiently extract all the auxiliary geographic data needed to construct the remote sensing feature vector for that location.

[0131] Step 213: Combine the horizontal polarization brightness temperature value, vertical polarization brightness temperature value, topographic elevation value, vegetation index value, and land cover type value to construct a remote sensing feature vector containing microwave brightness temperature information and auxiliary geographic data.

[0132] This step integrates all extracted features into a unified vector, which serves as input to the machine learning model. The snow depth detection system arranges all the values ​​extracted in steps 211 and 212 (e.g., Tb_18.7H, Tb_18.7V, Tb_36.5H, Tb_36.5V, DEM value, NDVI value, land cover type code) in a pre-defined order, forming a one-dimensional array or vector. This vector is the remote sensing feature vector for that location, comprehensively describing the microwave radiation characteristics and surface physical environment of that point.

[0133] Step 214: Input the remote sensing feature vector into the preset snow depth classification model to qualitatively classify the snow cover state at each location point, obtaining a snow depth category that represents whether the location point belongs to a preset snow depth range. This step can be referred to in the previous embodiment and will not be repeated here.

[0134] Step 215: Based on the snow depth category, selectively call the matching snow depth estimation regression model for the location point from a library containing multiple snow depth estimation regression models corresponding to different snow depth intervals. This step can be referred to in the previous embodiment and will not be repeated here.

[0135] Step 216: Input the remote sensing feature vector of the location point into the corresponding snow depth estimation regression model to obtain the predicted snow depth value of the location point. This step can be referred to in the previous embodiment and will not be repeated here.

[0136] Step 217: Based on the predicted snow depth values ​​of all locations within the area to be measured within a single day, generate single-track snow depth data to characterize the results of a single satellite transit observation.

[0137] This step aims to organize discrete prediction results into a spatially structured data product. Single-track snow depth data refers to a two-dimensional raster image composed of predicted snow depth values ​​for all locations (pixels) within a strip area covered by a single satellite pass. The snow depth detection system organizes the predicted snow depth values ​​for all locations detected during each satellite pass (typically lasting several minutes to tens of minutes) within a single day according to their geographical location on the satellite's scan trajectory, generating a single-track snow depth image. Within a day, a polar-orbiting satellite may conduct multiple observations of the same area (ascending and descending), thus generating multiple sets of single-track snow depth data.

[0138] Then proceed to step 218 or step 221.

[0139] Step 218: Project the single-track snow depth data onto a unified daily synthetic spatial grid.

[0140] This step aims to integrate observations from different times and orbits onto a standardized daily product grid. The snow depth sounding system establishes a standard daily composite spatial grid (e.g., a fixed-resolution geographic grid covering the entire area to be measured). This daily composite spatial grid is a standardized geographic reference grid covering a specific geographic area, used to integrate observation data from different satellite orbits and times throughout the day to generate a seamless daily data product.

[0141] Then, each single-track snow depth data generated in step 217 is projected onto this unified daily synthetic spatial grid through geographic coordinate transformation.

[0142] Step 219: Calculate the arithmetic mean of all predicted snow depth values ​​falling within the same daily composite spatial grid.

[0143] Since satellites may conduct multiple observations of the same region within a single day (both ascending and descending), a single daily composite spatial grid cell may correspond to multiple snow depth predictions from different single-track snow depth data. This step aims to fuse these multiple observation results. The snow depth sounding system traverses every grid cell on the daily composite spatial grid, identifies all snow depth predictions whose projections fall within that cell, and then calculates the arithmetic mean of these values.

[0144] Step 220: Determine the arithmetic mean as the final daily synthetic snow depth value for the daily synthetic spatial grid.

[0145] This step produces the final daily snow depth product. The snow depth detection system assigns the arithmetic mean of the snow depth predictions calculated above to the corresponding daily synthetic spatial grid cell, which serves as the final snow depth estimate for that cell on that day.

[0146] By performing this operation on all grid cells, a seamless daily synthetic snow depth distribution map covering the entire area under test is generated. This product integrates all valid observation information for the day, resulting in higher data integrity and stability.

[0147] Step 221: Obtain the calibrated snow depth value measured by the ground station corresponding to the location point within the same time window.

[0148] The purpose of this step is to introduce new, real-time ground truth data for online model updates. The snow depth sounding system connects to the ground observation network (such as a weather station network) in real-time or near real-time to acquire ground station-measured snow depth data that matches the satellite transit time window (e.g., within one hour before or after transit). This data is considered as high-precision calibration snow depth values, used to evaluate and correct model performance.

[0149] Step 222: Calculate the deviation between the predicted snow depth and the calibrated snow depth at the same location.

[0150] This step quantifies the model's prediction error at the current moment. For each location covered by a ground station, the snow depth detection system compares the predicted snow depth at that location with the calibrated snow depth value obtained in step 221, and calculates the difference between the two. This difference is the bias value. This bias value directly reflects the model's prediction performance under the current conditions.

[0151] Step 223: Based on the deviation value, apply the incremental learning algorithm to fine-tune the model parameters in the snow depth state classification model and / or the snow depth estimation regression model library online, so as to obtain the updated snow depth state classification model and snow depth estimation regression model library.

[0152] This step enables the model to self-evolve and continuously optimize. Incremental learning is a technique that allows machine learning models to continuously learn and update using new data without reusing all historical data. The snow depth detection system triggers the incremental learning process based on the bias value calculated in step 222. For example, if it finds that the predicted values ​​for a certain area are generally too high, the incremental learning algorithm will use this new bias information to make minor adjustments to the weights and parameters of the relevant regression models in the snow depth state classification model or snow depth estimation regression model library to reduce future prediction bias.

[0153] Through this online fine-tuning mechanism, the model can adapt to dynamic factors such as seasonal changes and variations in snow characteristics, maintaining high accuracy performance over the long term.

[0154] In some implementations, the daily synthesis algorithm of the snow depth detection method in this embodiment can synthesize microwave snow depth observations from multiple orbits on a unified grid to generate a daily-scale snow depth distribution product, reducing gaps and improving spatial continuity. The specific algorithm flow is as follows:

[0155] (1) Spatial matching: Map the orbital snow depth to a target spatial grid with a resolution of 15km;

[0156] (2) Multiple observations in the same grid: If multiple orbital observations fall into the same grid on the same day, the arithmetic mean of them is taken as the daily composite snow depth of that grid point;

[0157] (3) Single observation case: If only one orbital is covered, then this value is used directly;

[0158] (4) No observation: If no orbit falls into the grid, it is marked as -9999 null value.

[0159] Among them, grid and resolution selection:

[0160] The original orbital data had a spatial resolution of 10km. A resolution of 15km was chosen for the composite data because sparse orbital coverage and spatial projection matching issues resulted in a large number of grid points lacking effective observations. To improve the spatial coverage of the composite data, this embodiment uses a 15km*15km latitude and longitude grid as the output resolution. The resolution is achieved by aggregating adjacent 10km grid points, which can smooth out observation noise to some extent and significantly reduce blank grid points.

[0161] See Figure 5 This example shows the FY3D-MWRI snow depth diurnal composite spatial distribution map generated in this embodiment, demonstrating the final diurnal product after synthesizing all satellite transit orbit data for the entire day of February 1, 2025. The spatial resolution of this map is 15 kilometers. Through diurnal composite processing, the... Figure 2 and Figure 3 The two orbital observations shown, along with other unshown data from the same day, were integrated onto a unified geographic grid. From Figure 5 As can be clearly seen, the originally separate, strip-shaped observation areas have been stitched together into a more spatially continuous and complete snow depth cover map. For example, the snow cover areas of Northeast and West China are fully presented on the same map, effectively reducing data gaps caused by incomplete coverage in a single orbital observation. For orbital overlap areas, the snow depth values ​​are the average results of multiple observations, which makes the spatial continuity of the daily composite product better, and the data more stable and reliable, comprehensively reflecting the macroscopic distribution of major snow cover areas in China on that day.

[0162] In some implementations, after the remote sensing feature vector is constructed (i.e., after step 213), the snow depth detection method of this embodiment can employ a strategy that integrates the physical model and machine learning residual correction to calculate the predicted snow depth. This strategy no longer uses the "classification-regression" two-level model described in steps 214 to 216, but instead replaces it with the following series of steps, specifically including:

[0163] Step 301: Input the remote sensing feature vector into a simplified microwave radiative transfer physical model to calculate a physical baseline snow depth value.

[0164] The goal of this step is to quickly obtain a preliminary snow depth estimate with clear physical meaning, although it may contain systematic biases, using well-established physical mechanisms. The snow depth detection system first invokes a simplified microwave radiative transfer physical model. Here, "simplified microwave radiative transfer physical model" typically refers to an analytical or semi-empirical formula derived from radiative transfer theory that can directly invert snow depth from brightness temperature, such as the classic Chang algorithm or its improved forms. These models are usually based on specific assumptions (such as uniform snow layer and fixed snow grain size) to establish a direct functional relationship between the observed brightness temperature differences at different frequencies (e.g., the brightness temperature difference between 18.7 GHz and 36.5 GHz) and snow depth.

[0165] In practice, the snow depth detection system extracts specific inputs required by the model from remote sensing feature vectors, primarily microwave brightness temperature values ​​at different frequencies. These brightness temperature values ​​are then substituted into the formulas of the physical model for calculation. For example, a simplified model might be: Snow Depth = C * (Tb_18.7V - Tb_36.5V), where C is an empirical coefficient. The calculated result is the "physical baseline snow depth value," which can be considered a "first-principles" approximation of the snow depth.

[0166] Step 302: Input the remote sensing feature vector into a pre-trained residual classification model to obtain the residual interval category in which the physical baseline snow depth value is located.

[0167] This step aims to qualitatively classify the prediction errors (i.e., residuals) of the physical model. Since the simplifying assumptions of the physical model can introduce systematic, nonlinear errors under complex surface conditions (such as the presence of vegetation cover, topographic relief, or variations in snow grain size), this invention proposes to learn and compensate for these errors through machine learning.

[0168] First, a "residual classification model" needs to be pre-trained. The training process is as follows: Using the historical training sample set, a physical baseline snow depth value is calculated for each sample using the physical model from step one. Then, this physical baseline snow depth value is subtracted from the sample's actual snow depth label value to obtain the "true residual value" (true residual value = actual snow depth - physical baseline snow depth). Next, based on preset residual thresholds (e.g., -10 cm, 0 cm, 10 cm), these true residual values ​​are divided into different "residual interval categories," such as "large negative deviation," "close to accurate," and "large positive deviation." Finally, using remote sensing feature vectors as input and the corresponding residual interval categories as labels, a multi-class machine learning model (such as random forest) is trained, resulting in the residual classification model.

[0169] During actual detection, the snow depth detection system inputs the remote sensing feature vector of the current location point into this trained residual classification model. The model outputs a category that predicts the possible error range and direction of the physical baseline snow depth value calculated by the physical model.

[0170] Step 303: Selectively invoke a pre-trained residual regression model that matches the residual interval category.

[0171] This step is similar to the model selection logic in the previous embodiments, but it targets the residuals rather than the snow depth itself. The snow depth detection system maintains a residual regression model library, which stores multiple dedicated regression models for different residual intervals.

[0172] The residual regression model library is constructed as follows: During the training phase, historical training samples are divided into multiple sample subsets based on the residual interval category to which their "true residual values" belong. Then, a separate regression model is trained for each subset, and the learning objective of this model is to accurately predict the "true residual values" of the samples within that subset. For example, using samples from the "negative large deviation" subset, a "negative large deviation residual regression model" is trained specifically to predict the specific values ​​of negative large deviations.

[0173] During the detection process, the snow depth detection system accurately finds and loads a matching residual regression model from the residual regression model library based on the residual interval category output in step two.

[0174] Step 304: Input the remote sensing feature vector into the selected residual regression model to calculate a snow depth residual correction value.

[0175] The goal of this step is to quantitatively calculate the specific amount of correction needed to the physical baseline snow depth value. The snow depth detection system takes the remote sensing feature vector of the current location point (the same vector input into the residual classification model) as input and feeds it into the dedicated residual regression model selected in step three. Because this residual regression model specifically learns the error patterns that the physical model may produce in specific intervals under the current surface conditions (represented by the remote sensing feature vector), it can output a more accurate error prediction value, which is the snow depth residual correction value.

[0176] Step 305: Sum the physical baseline snow depth value and the snow depth residual correction value to obtain the final snow depth prediction value, and use it as the single-track snow depth data.

[0177] This step is the final fusion and correction process. The snow depth detection system performs a simple algebraic summation operation on the physical baseline snow depth value calculated in step one and the snow depth residual correction value calculated in step four (final snow depth = physical baseline snow depth + snow depth residual correction value). Since the snow depth residual correction value accurately compensates for the systematic bias of the physical model, the sum of the two is the more accurate final snow depth prediction value after fine correction by machine learning. This final snow depth prediction value is then used to generate single-track snow depth data, thereby completing the snow depth detection at this location.

[0178] By combining "physical driving + data driving" approaches, this embodiment fully leverages the stability of the physical model and the fitting capability of the machine learning model to achieve high-precision inversion of snow depth.

[0179] In some embodiments, the “classification-regression” strategy, while superior to a single model, is based on a “hard classification” process. This involves forcibly categorizing a continuously changing snow accumulation process (e.g., snow depth smoothly increasing from 4.9 cm to 5.1 cm) into discrete, clearly defined categories (“light snow” vs. “moderate snow”). This can raise two problems at the category boundaries:

[0180] 1) Prediction discontinuity: Two points that are geographically or temporally close and have almost identical physical conditions may be assigned to two completely different regression models because they cross the classification threshold, resulting in unreasonable "jumps" in the final snow depth prediction value and destroying the physical reality of the spatial distribution of snow depth.

[0181] 2) Disastrous consequences of misjudgment: Once the classification model makes a misjudgment (for example, misclassifying deep snow as thin snow), the system will call an incorrect "expert model", whose output will differ significantly from the true value, resulting in outlier error.

[0182] Therefore, the method in this embodiment also provides an adaptive model fusion framework for "soft decision-making". The specific implementation is as follows:

[0183] 1. Constructing a probabilistic state membership classification model: First, modify the existing snow depth state classification model. Instead of outputting a unique, definite category label, train a model that outputs a probability distribution. For example, use a neural network classifier with a softmax activation function. For any input remote sensing feature vector, the model will output a probability vector, such as P = [P(no snow), P(light snow), P(moderate snow), P(deep snow)], where P(category A) + P(category B) + ... = 1. This probability vector (e.g., [0.0, 0.3, 0.7, 0.0]) represents the confidence or probability that the current location point simultaneously "belongs" to each snow depth state; we call this the state membership.

[0184] 2. Parallel computation and weighted fusion of the entire model: After obtaining the state membership probability vector, the snow depth detection system no longer selects a single model, but instead simultaneously inputs the remote sensing feature vector of the location point into all regression models in the snow depth estimation regression model library (e.g., "thin snow regression model", "medium snow regression model", etc.) to obtain a set of parallel preliminary snow depth prediction values ​​under different state assumptions (SD_thin snow, SD_medium snow, SD_deep snow...).

[0185] 3. Dynamic weighted synthesis of final snow depth value: Finally, the state membership probability vector obtained in the previous step is used as weights to perform a weighted average of all the preliminary snow depth predictions calculated in parallel, resulting in the final snow depth prediction value. The calculation formula is: Final snow depth = P(thin snow) * SD_thin snow + P(medium snow) * SD_medium snow + P(deep snow) * SD_deep snow + ...

[0186] When encountering boundary problems, those skilled in the art might conventionally approach the issue through post-processing smoothing or introducing fuzzy logic. However, this embodiment fundamentally solves this problem at the model architecture level. It transforms the classifier's role from a "decision-maker" to a "weight allocator," and the regression model library from a group of "mutually exclusive experts" to a "cooperative committee." This dynamic model weighted fusion based on probabilistic membership allows for a completely smooth and continuous transition between different states in snow depth predictions, completely eliminating boundary jumps. Furthermore, even with minor deviations in classification probabilities, the final result is a smooth fusion of predictions from multiple models, rather than a catastrophic switch to a single erroneous model, greatly enhancing the system's stability and robustness. This is a novel, more elegant, and physically complete snow depth inversion paradigm.

[0187] The snow depth detection methods described above predict snow depth by learning correlations from historical data, but the models themselves are "completely unaware" of the physical processes of snow microwave radiation. In some cases, this can lead to two problems:

[0188] 1) Lack of physical consistency: The model's predictions may be mathematically optimal, but not necessarily physically reasonable. For example, under certain extreme combinations of features, the model may output a result that contradicts the laws of microwave physics.

[0189] 2) Vulnerable extrapolation ability: When encountering novel surface conditions or extreme weather events that have never appeared in the training data (e.g., abnormal snow crystal morphology, sudden snow melting and refreezing), the pure data-driven model is likely to make completely wrong extrapolation predictions due to the lack of guidance from physical mechanisms, leading to model failure.

[0190] Therefore, embodiments of the present invention provide a neural inverse modeling scheme based on a differentiable physics engine.

[0191] This approach overturns the traditional inversion method of "observation -> snow depth" and proposes a neural inverse modeling framework based on "forward simulation + reverse optimization".

[0192] The specific implementation steps are as follows:

[0193] 1. Constructing a Differentiable Microwave Radiative Transfer (RTM) Physics Engine: First, instead of using a physical model as a simple baseline, a differentiable RTM is constructed. This can be achieved in two ways: a) rewriting the existing RTM computation process using a deep learning framework that supports automatic differentiation (such as PyTorch, TensorFlow), making all computational steps differentiable with respect to input parameters (snow depth, snow grain size, density, etc.); b) training a deep neural network (Physical Information Neural Network PINN) to accurately fit the input-output relationship of the complex RTM; this neural network itself is naturally differentiable. This "differentiable physics engine" can simulate: given a set of snow layer physical parameters, what microwave brightness temperature observations will be generated on the Earth's surface.

[0194] 2. Constructing an implicit snow parameter state encoder: Train a deep autoencoder network that can compress the input remote sensing feature vector (containing microwave brightness temperature and surface auxiliary information) into a low-dimensional implicit state vector. This implicit state vector aims to capture a comprehensive representation of all key snow physical properties (snow depth, grain size, density, temperature, humidity, etc.) that affect microwave radiation.

[0195] 3. Real-time inverse optimization based on gradient descent: This is the core of this scheme. For any point to be tested:

[0196] a) Input the remote sensing feature vector of the location point into the implicit snow parameter state encoder to obtain an initial implicit state vector.

[0197] b) This implicit state vector is used as input and fed into a differentiable physics engine to positively simulate the microwave brightness temperature that the satellite should observe in this state.

[0198] c) Compare the simulated brightness temperature with the actual brightness temperature observed by the satellite, and calculate the difference between the two (loss function).

[0199] d) Key steps: Since the entire computation chain (encoder -> physics engine) is end-to-end differentiable, we can use the backpropagation algorithm to calculate the gradient of the loss function with respect to the implicit state vector.

[0200] e) Based on this gradient, use an optimizer (such as Adam) to iteratively update the implicit state vector at this location, with the goal of minimizing the difference between the simulated brightness temperature and the actual brightness temperature.

[0201] f) After the iteration converges, one or more dimensions in the optimal implicit state vector correspond to the final snow depth prediction value.

[0202] Those skilled in the art typically address the physical consistency problem by adding physical constraint terms to the loss function of machine learning. This embodiment, however, takes a completely different approach. Instead of directly retrieving snow depth, it reconstructs the snow depth inversion problem into a real-time optimization problem based on a physical model. It dynamically searches for an intrinsic physical state that best "interprets" the satellite observation data for each pixel. This "neural inverse modeling" paradigm represents a deep integration of physics and cutting-edge artificial intelligence.

[0203] The innovation of this embodiment lies in: 1) It ensures that every prediction result strictly adheres to the physical laws of microwave radiation transmission, solving the problem of physical consistency. 2) When encountering novel conditions, it does not blindly extrapolate, but rather searches and optimizes within a reasonable physical parameter space under the guidance of a physics engine, possessing generalization and extrapolation capabilities far exceeding traditional methods. This represents a fundamental leap from "pattern recognition" to "physical reasoning."

[0204] In some implementations, the snow depth orbit algorithm flow of the snow depth detection method of this embodiment can be referred to Figure 6 As shown, it can specifically include the following:

[0205] (1) Data fusion:

[0206] Utilizing H / V polarization brightness temperature information from four key channels (10.65 GHz, 18.7 GHz, 36.5 GHz, and 89 GHz) of the FY3D-MWRI satellite microwave radiometer, and combining it with topographic elevation data (USGSDEM), land cover data (CLCD LAND), and ERA5 reanalysis snow depth data, high-precision spatial matching of multi-source data was achieved through fusion methods such as spatial registration, resampling, and gridding, providing consistent input for subsequent feature extraction and model training.

[0207] (2) Automatic feature variable selection:

[0208] During model training, an embedded feature selection method is introduced. This method utilizes L1 regularization and gradient boosting tree-based feature importance evaluation to automatically select the input variables most discriminative for snow depth inversion. This approach not only reduces the interference of redundant features on the model but also improves training efficiency and generalization ability.

[0209] (3) Hierarchical classification modeling:

[0210] The samples were stratified based on snow depth thresholds: no snow (SD < 5 mm), low snow depth (5–500 mm), medium snow depth (500–2000 mm), and high snow depth (SD > 2000 mm). The LightGBM classifier (LgbmClassifier) ​​was used in the classification modeling stage, leveraging its efficient gradient boosting framework to handle large-scale training samples, thereby improving classification accuracy and computational efficiency.

[0211] (4) Partition regression modeling:

[0212] For samples in snowy areas (SD≥5mm), a partitioned regression strategy was further adopted. A suitable regression model was selected based on the snow depth range, and the optimal model structure was chosen through cross-validation to improve the accuracy and stability of snow depth estimation.

[0213] (5) Model prediction and inversion:

[0214] The optimized classification-regression joint model is fixed as a static parameter set and then used as input for real-time snow depth inversion with new FY3D_MWRI observation data. The final output is a grid data of snow depth distribution covering the area to be measured.

[0215] To implement the snow depth detection method in this embodiment, the snow depth detection system is configured with clear data input and output interfaces, ensuring the automation of the entire process and the standardization of the product.

[0216] 1. Input Data Interface: The normal operation of the system relies on the stable input of the following multi-source data: FY3D_MWRI satellite L1-level orbit data: This is the core data source of the method in this embodiment, providing horizontal (H) and vertical (V) polarization brightness temperature observations for four key frequency channels: 10.65GHz, 18.7GHz, 36.5GHz, and 89GHz. Simultaneously, the data also includes necessary geographical and observational geometric information such as the latitude and longitude of the corresponding pixels, land and sea masking codes, satellite zenith angle, and solar zenith angle. In the implementation of this invention, long-term series (e.g., 2020 to 2024) ascending orbit (daytime transit) and descending orbit (nighttime transit) data are typically collected to construct a sufficiently representative training sample set and perform real-time inversion.

[0217] The ERA5 global atmospheric reanalysis dataset provides snow depth data, which is a key source for constructing the "actual snow depth label values" in the training sample set. When acquiring the data, a daily average (day_mean) product is typically used. This is advantageous because it smooths out short-term fluctuations and local noise within a single day, providing a more robust ground truth reference for the stable training of machine learning models.

[0218] MODIS Vegetation Index Static Data: To quantify the impact of vegetation on microwave signals, the system introduces the Normalized Difference Vegetation Index (NDVI). Specifically, a 16-day composite product (MOD13Q1) from the MODIS / Terra satellite is selected. This product passes over the area in the morning, ensuring stable lighting conditions and high data quality. Monthly NDVI data from several consecutive years (e.g., 2022 to 2024) are averaged to generate a 12-month static NDVI dataset covering the entire year. This dataset undergoes efficient preprocessing using a geographic information platform (such as Google Earth Engine), including orbit stitching, projection unification, and cloud masking, to ensure data quality.

[0219] DEM elevation data: used to characterize topographic effects. Data can be obtained from public platforms (such as USGS EarthExplorer) and processed using geographic information system software (such as ArcGIS), resampled to a spatial resolution matching microwave data (e.g., 10 km), and cropped to the target area.

[0220] LandCover surface type data: used to distinguish different underlying surface types. This invention can use high-precision regional land cover products (e.g., the China Regional Land Cover Dataset developed by the Chinese Academy of Sciences), which are also resampled and projected to ensure spatial consistency with other input data.

[0221] 2. Output Data Interface: The method in this embodiment ultimately produces a series of standardized snow depth products, specifically including: FY3D-MWRI Snow Depth Orbit Products: Snow depth data generated in real-time, maintaining the original satellite observation path and resolution (e.g., 10 km). Each orbit product contains an HDF format data file, internally storing scientific datasets such as snow depth, longitude, and latitude, and includes a PNG format thematic preview image and a set of GIS texture files supporting multiple standard projections (such as iso-latitude, Mercator, Lambert, etc.), facilitating rapid visualization and analysis on different platforms.

[0222] FY3D-MWRI Snow Depth Daily Composite Product: This daily product is generated by spatiotemporally composited all orbital products within a single day, resulting in higher spatial coverage. Its spatial resolution can be set according to requirements (e.g., 15 km). The product format is similar to that of the orbital product, also including HDF data files, PNG thematic maps, and multi-projection GIS textures.

[0223] In some implementations of this embodiment, the specific snow depth track product inputs and outputs can be referred to Table 1:

[0224] Table 1 Input / Output Description for Snow Depth Track Products

[0225]

[0226] To verify the effectiveness and accuracy of the method in this embodiment, a series of model performance evaluation and result verification experiments were also conducted.

[0227] 1. Model performance evaluation: In some embodiments, the constructed machine learning model exhibits excellent performance.

[0228] (1) Accuracy of the snow depth classification model: When distinguishing between four snow depth states—no snow, low snow depth, medium snow depth, and high snow depth—the model achieved an overall classification accuracy of 87% and a weighted F1 score of 0.88. Specifically, the model demonstrated high accuracy and recall for the "no snow," "low snow depth," and "medium snow depth" categories, with F1 scores all above 0.83. For the "high snow depth" category, the model exhibited extremely high recall (97%), indicating that it could effectively identify the vast majority of deep snow pixels. However, the accuracy (71%) was relatively low, suggesting that the model might misclassify non-high snow depth pixels as high snow depth pixels in certain situations. This is a common technical challenge when making distinctions in areas with saturated deep snow signals.

[0229] The accuracy evaluation results of the snow depth state classification model can be found in Table 2.

[0230] Table 2 Accuracy Evaluation Table of Snow Depth Classification Model

[0231] Classification accuracy Recall F1_score Data volume No snow 0.95 0.88 0.91 90,0000 Low snow depth 0.82 0.84 0.83 60,0000 Medium snow depth 0.84 0.88 0.86 30,0000 Gao Xueshen 0.71 0.97 0.82 9,5372 Overall accuracy 0.87 189,5372 macro average 0.83 0.89 0.86 189,5372 weighted average 0.88 0.87 0.88 189,5372

[0232] The accuracy evaluation results of the snow depth estimation regression model are shown in Table 3:

[0233] Model MAE(mm) RMSE (mm) R2 Low snow depth LightGBM 27.52 3125.73 0.746 Medium snow depth LightGBM 112.28 29749.47 0.823 Gao Xueshen Catboost 1901.70 7693407.30mm 0.881

[0234] (2) Accuracy of Snow Depth Estimation Regression Model: The performance of a dedicated regression model trained for different snow depth ranges is measured by mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 The model's R-value was evaluated. In the "low snow depth" range, the model's R-value was... 2 Reaching 0.746; in the "medium snow depth" range, R 2 Increased to 0.823; in the "high snow depth" range, R 2 The R² value can reach 0.881. The results show that the partitioned regression strategy used in this invention can effectively capture the patterns at different snow depth stages, with R² reaching 0.881. 2 The value increases with increasing snow depth, demonstrating good fitting ability.

[0235] 2. Result Validation: To evaluate the actual accuracy of the final snow depth product, snow depth prediction results from multiple historical time periods (e.g., January 1-5, 2019; May 25-29, 2024) were selected and compared with ERA5 reanalysis snow depth data at the pixel level for validation. The results show that in over 90% of the validation pixels, the prediction results of the method of this invention exhibit a high degree of consistency with the reference data. The overall root mean square error (RMSE) stabilized at around 5 cm (50 mm), and the correlation coefficient (R0) was [not specified]. 2 It can reach around 0.80, and in some cases, R... 2 It even exceeds 0.90.

[0236] Furthermore, from the perspective of spatial distribution, the spatial distribution pattern of snow depth inverted in this embodiment is highly consistent with the spatial distribution characteristics of snow depth in ERA5 reanalysis, and can accurately reproduce the spatial morphology of snow cover range and high value area, proving that the present invention has high reliability and practical value on a macro scale.

[0237] For specific results verification, please refer to Table 4 and... Figure 7 As shown:

[0238] Table 4 Summary of Results Verification

[0239] Figure 7 Part (a) shows a scatter plot of the verification results of the snow depth prediction accuracy of the embodiment of the present invention on January 1, 2025. The horizontal axis of the plot is the reference true snow depth, and the vertical axis is the predicted snow depth by the method of the present invention, both in millimeters (mm). Each blue scatter point in the plot represents an independent pair of verification pixels. The red dashed line is a 1:1 reference line; the closer the scatter points are to this reference line, the closer the predicted value is to the true value. The text box in the upper left corner shows the statistical indicators of this verification: the total number of verification sample points (Count) is 55,950; the mean absolute error (MAE) is 31.61 mm; the root mean square error (RMSE) is 55.81 mm; and the coefficient of determination (R²) is 100%. 2 The mean squared error (R²) was 0.89; the mean deviation (Bias) was 2.37 mm. This figure shows that, on the verification date, the predicted results of the method of this invention have a very high correlation (R²) with the actual values. 2 =0.89), and the error is small.

[0240] Figure 7Part (b) is a scatter plot of the verification results of snow depth prediction accuracy of the present invention on January 2, 2025. According to the statistical information shown in the figure, the total number of sample points in this verification was 60,336. Evaluation indicators show that the mean absolute error (MAE) was 26.07 mm, the root mean square error (RMSE) was 50.59 mm, and the coefficient of determination (R²) was... 2 The mean squared error (R²) was 0.82, and the average bias (Bias) was -1.75 mm. The results further confirm that the method of this invention maintains high prediction accuracy and stability across different dates, and the correlation between the predicted and actual values ​​is good (R²). 2 =0.82), and the average deviation is very small.

[0241] Figure 7 Part (c) is a scatter plot of the verification results of snow depth prediction accuracy of the embodiment of the present invention on January 3, 2025. As shown in the figure, the total number of sample points verified on that day was 64,439. The statistical results show that the mean absolute error (MAE) was 23.45 mm, the root mean square error (RMSE) was 47.24 mm, and the coefficient of determination (R²) was... 2 The mean squared error (MSE) was 0.79, and the average bias was 1.35 mm. The scatter plot distribution shows that most data points closely clustered around the 1:1 reference line, especially in the light snow area, demonstrating a strong consistency between the predicted results and the actual values, thus proving the reliability of the method.

[0242] Figure 7 The scatter plot of the verification results of snow depth prediction accuracy of the present invention embodiment on January 4, 2025, is shown in section (d). This verification included a total of 59,410 sample points. The accuracy indicators are as follows: mean absolute error (MAE) of 24.80 mm, root mean square error (RMSE) of 47.89 mm, and coefficient of determination (R²) of 100 mm. 2 The value was 0.80, and the average deviation was 8.86 mm.

[0243] Figure 7 Part (d) together with the first three parts constitutes a multi-day accuracy verification sequence, comprehensively demonstrating that the method of the present invention can continuously and stably provide high-precision snow depth prediction products in practical applications, and its prediction results always maintain a high correlation and a low error level with the reference benchmark data.

[0244] The snow depth detection method proposed in this invention not only performs excellently in point-to-point numerical accuracy, but also accurately and reliably reproduces the true spatial distribution characteristics of large-scale snow cover.

[0245] The method provided in the above embodiments can be executed by a snow depth detection system, which specifically includes electronic equipment. The electronic equipment in this embodiment of the invention is described below from a hardware processing perspective; please refer to [link to relevant documentation]. Figure 8 This is a schematic diagram of the physical device structure of an electronic device in an embodiment of the present invention.

[0246] It should be noted that, Figure 8 The structure of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0247] like Figure 8 As shown, the electronic device includes a Central Processing Unit (CPU) 401, which can perform various appropriate actions and processes based on a program stored in Read-Only Memory (ROM) 402 or a program loaded from storage portion 408 into Random Access Memory (RAM) 403, such as performing the methods described in the above embodiments. The RAM 403 also stores various programs and data required for system operation. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An Input / Output (I / O) interface 405 is also connected to the bus 404.

[0248] The following components are connected to I / O interface 405: input section 406 including audio input devices, push-button switches, etc.; output section 407 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 408 including a hard disk, etc.; and communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0249] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit (CPU) 401, it performs the various functions defined in the present invention.

[0250] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0251] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0252] Specifically, the electronic device of this embodiment includes a processor and a memory. The memory is coupled to one or more processors and is used to store computer program code. The computer program code includes computer instructions. One or more processors call the computer instructions to cause the electronic device to perform the method provided in the above embodiment.

[0253] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The storage medium carries one or more computer programs that, when executed by a processor of the electronic device, cause the electronic device to implement the methods provided in the above embodiments.

[0254] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

[0255] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A snow depth detection method based on machine learning, characterized in that, Includes the following steps: Acquire multi-source remote sensing data of the area to be measured. The multi-source remote sensing data includes satellite passive microwave observation data and auxiliary geographic data. The microwave brightness temperature information is the brightness temperature observation value that characterizes the radiation intensity of the surface in multiple microwave frequency bands. The auxiliary geographic data is a parameter that can reflect the static or slowly changing physical properties of the surface that affect microwave radiation transmission. Spatial registration and resampling are performed on the multi-source remote sensing data to obtain a spatiotemporally matched multidimensional dataset on a unified geographic grid. For each location point on the unified geographic grid, the corresponding observation values ​​and parameter values ​​are extracted from the multidimensional dataset to construct a remote sensing feature vector containing the microwave brightness temperature information and the auxiliary geographic data. The remote sensing feature vector is input into a preset snow depth state classification model to qualitatively classify the snow cover state of each location point, thereby obtaining a snow depth category that represents the location point as belonging to a preset snow depth range. Based on the snow depth category, a matching snow depth estimation regression model is selectively invoked for the location point from a library containing multiple snow depth estimation regression models corresponding to different snow depth ranges. The remote sensing feature vector of the location point is input into the snow depth estimation regression model to obtain the predicted snow depth value of the location point.

2. The method according to claim 1, characterized in that, The construction of the remote sensing feature vector containing the microwave brightness temperature information and the auxiliary geographic data includes: Extract at least two different frequency channels' horizontal polarization brightness temperature values ​​and vertical polarization brightness temperature values ​​from the satellite's passive microwave observation data; Extract topographic elevation values, vegetation index values, and land cover type values ​​from the auxiliary geographic data; The horizontal polarization brightness temperature value, the vertical polarization brightness temperature value, the terrain elevation value, the vegetation index value, and the land cover type value are combined to form the remote sensing feature vector.

3. The method according to claim 1, characterized in that, The training methods for the snow depth classification model include: Obtain a training sample set containing historical remote sensing feature vectors and their corresponding actual snow depth label values; Each sample in the training sample set is classified and labeled according to a preset snow depth threshold to obtain a category label; The snow depth classification model is trained by taking all remote sensing feature vectors of the training sample set and the category labels as input.

4. The method according to claim 3, characterized in that, The methods for obtaining the snow depth estimation regression model library include: Obtain the training sample set that has been classified and labeled; The training sample set is divided into multiple sample subsets corresponding one-to-one with the snow category according to the category label; For each of the sample subsets, a regression model is trained separately using the remote sensing feature vectors and actual snow depth label values ​​within the sample subset; All the trained regression models are combined into the snow depth estimation regression model library.

5. The method according to any one of claims 1-4, characterized in that, After inputting the remote sensing feature vector of the location point into the snow depth estimation regression model to obtain the predicted snow depth value of the location point, the method further includes: Based on the predicted snow depth values ​​of all locations within the test area within a single day, single-track snow depth data is generated to characterize the results of a single satellite transit observation. The single-track snow depth data is projected onto a unified daily synthetic spatial grid; Calculate the arithmetic mean of all predicted snow depth values ​​falling within the same daily composite spatial grid; The arithmetic mean is determined as the final daily synthetic snow depth value of the daily synthetic spatial grid.

6. The method according to claim 5, characterized in that, After generating single-track snow depth data to characterize the results of a single satellite transit observation based on the predicted snow depth values ​​of all locations within the area to be measured within a single day, the method further includes: Obtain the actual measured calibration snow depth values ​​of the ground stations corresponding to the location points within the same time window; Calculate the deviation between the predicted snow depth value and the calibrated snow depth value at the same location point; Based on the deviation value, the incremental learning algorithm is applied to fine-tune the model parameters in the snow depth state classification model and / or the snow depth estimation regression model library online, so as to obtain the updated snow depth state classification model and snow depth estimation regression model library.

7. The method according to claim 1, characterized in that, After constructing the remote sensing feature vector containing the microwave brightness temperature information and the auxiliary geographic data, the method further includes: The remote sensing feature vector is input into a simplified microwave radiative transfer physical model to calculate a physical baseline snow depth value. The remote sensing feature vector is input into a pre-trained residual classification model to obtain the residual interval category in which the physical baseline snow depth value is located; Based on the residual interval category, a pre-trained residual regression model that matches it is selectively invoked; The remote sensing feature vector is input into the selected residual regression model to calculate a snow depth residual correction value; The physical baseline snow depth value is summed with the snow depth residual correction value to obtain the final snow depth prediction value, which is then used as the single-track snow depth data.

8. An electronic device, characterized in that, Includes one or more processors and memory; The memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed on the electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on an electronic device, it causes the electronic device to perform the method as described in any one of claims 1-7.

Images