Snow depth estimation method and system based on hybrid network model

Abstract

The application discloses a snow depth estimation method and system based on a hybrid network model, which comprises the following steps: preprocessing and spatio-temporal alignment of brightness temperature and measured snow depth data, and construction of a standardized data set; construction of a hybrid network model, wherein a convolutional block embedding is used in the input layer to extract local features, a coding-collaboration layer LayerScale mechanism is used to capture global dependence, and a double-branch residual fusion structure is used in the output layer to integrate deep semantic features and original shallow features; an average absolute error is used as a loss function, an AdamW optimizer and a cosine annealing learning rate schedule are used to train the model, and an optimized snow depth estimation model is obtained; and real-time prediction of a snow depth estimation value is performed through the optimized snow depth estimation model. The system has the same technical concept. The application overcomes the problems of weak generalization ability, insufficient utilization of local and global features and lack of physical constraints in the output of the traditional method, and realizes high-precision and robust inversion of snow thickness.

Description

Technical Field

[0001] This invention belongs to the field of surface element estimation technology, specifically relating to a snow depth estimation method and system based on a hybrid network model. Background Technology

[0002] Passive microwave brightness temperature (BBT) is a key physical quantity reflecting the dielectric properties and structural information of the land surface and snow cover, and has become an important remote sensing data source for large-scale, long-term snow depth inversion. Traditional mainstream snow depth inversion methods are mainly divided into two categories: radiative transfer models based on physical mechanisms and regression models based on statistical laws. However, physical models typically rely on a series of snow cover parameters (such as density, grain size, and layered structure) that are difficult to obtain precisely, and their generalization ability under different climatic and underlying surface conditions is limited. Statistical models, on the other hand, often fail to fully exploit the complex nonlinear correlations inherent in brightness temperature data over long periods and in multiple frequency channels, leading to limited inversion accuracy, especially under complex terrain and snow cover conditions.

[0003] In recent years, end-to-end deep learning methods have provided a new technical path for snow depth inversion, which can automatically learn high-level features from data and overcome the limitations of traditional methods to some extent. However, existing deep learning-based solutions still have significant shortcomings: if a pure Transformer architecture is used directly, although it can effectively model long-term temporal dependencies, it is easy to ignore the local correlation between brightness and temperature sequences at adjacent time steps or frequency points, resulting in insufficient capture of local physical patterns; at the same time, the training process of deep Transformer networks is unstable and prone to gradient anomalies or convergence difficulties; more importantly, most deep learning models lack physical constraints on the output results, often resulting in situations such as predicting negative snow depths that do not conform to physical reality, seriously affecting the practical usability of the inversion results.

[0004] Therefore, providing a snow depth estimation method and system based on a hybrid network model to solve the above problems is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide a snow depth estimation method and system based on a hybrid network model. This method takes into account both local feature extraction and global dependency modeling, possesses a stable training mechanism, and ensures that the output results strictly conform to physical constraints, thereby achieving high-precision and highly robust estimation of snow depth in polar and complex environments.

[0006] The technical solution provided by this invention is as follows:

[0007] A snow depth estimation method based on a hybrid network model includes the following steps:

[0008] Brightness temperature data and snow depth data are acquired, and the brightness temperature data and snow depth data are processed respectively. A brightness temperature and snow depth dataset is constructed based on the processed brightness temperature data and snow depth data. The brightness temperature and snow depth dataset is then standardized to obtain a brightness temperature and snow depth standard set.

[0009] A convolutional-Transformer hybrid network model is constructed, consisting of an input layer, an encoding-cooperative layer, and an output layer. The input layer is used to extract shallow local features from the brightness temperature and snow depth standard set. The encoding-cooperative layer is used to obtain deep semantic features that characterize global dependencies. The output layer is used to fuse the processed shallow local features and the processed deep semantic features and output the snow depth estimate.

[0010] Based on the brightness temperature and snow depth standard set, the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to train the convolutional-transformer hybrid network model until the loss function converges, thus obtaining the optimized convolutional-transformer hybrid network model.

[0011] The standard set of brightness temperature and snow depth to be estimated is input into the optimized convolutional-transformer hybrid network model to obtain the snow depth estimate corresponding to the standard set of brightness temperature and snow depth to be estimated.

[0012] Preferably, the brightness temperature data is processed by the following steps:

[0013] The brightness temperature data is converted to obtain brightness temperature data in a standard format;

[0014] The brightness temperature data in the standard format is subjected to geometric correction, radiometric correction, error calibration, and cloud filtering operations in sequence to obtain the corrected brightness temperature data.

[0015] Based on the latitude and longitude of the region, the corrected brightness temperature data is selected to obtain the processed brightness temperature data.

[0016] Preferably, the snow depth data is processed, including the following steps:

[0017] Based on the latitude and longitude of the region, the snow depth data is selected by region selection to obtain the snow depth data after region selection;

[0018] The selected area snow depth data is spatially interpolated using bilinear interpolation to obtain the processed snow depth data.

[0019] The spatial interpolation processing is used to align the processed snow depth data with the processed brightness temperature data in spatial dimensions.

[0020] Preferably, the step of constructing a brightness temperature and snow depth dataset based on the processed brightness temperature data and processed snow depth data, and standardizing the brightness temperature and snow depth dataset to obtain a brightness temperature and snow depth standard set, includes the following steps:

[0021] The processed brightness temperature data and the processed snow depth data are time-aligned to obtain spatiotemporally aligned brightness temperature data and snow depth data;

[0022] The spatiotemporally aligned brightness temperature data and snow depth data are matched by latitude and longitude to remove outliers, thus obtaining the brightness temperature and snow depth dataset.

[0023] Normalization is performed on each subset of the brightness temperature and snow depth dataset to obtain the brightness temperature and snow depth standard set containing multiple brightness temperature sequences.

[0024] Preferably, obtaining the deep semantic features used to characterize global dependencies includes the following steps:

[0025] The brightness temperature sequence in the brightness temperature and snow depth standard set is divided into patches by a depthwise separable convolutional module, then mapped to a high-dimensional embedding vector, and a learnable positional code is added to form a token sequence.

[0026] The token sequence is input into a Transformer encoder consisting of several stacked coding blocks; each coding block adopts a Pre-Norm structure and includes: a layer normalization module, a multi-head self-attention module with LayerScale mechanism, and a multi-layer deep residual feedforward network;

[0027] Shallow features are extracted from the Transformer encoder, concatenated with the self-attention weight matrix, and then subjected to layer normalization. Layer-by-layer cumulative residual connection calculation is then performed, and the output of the multi-layer deep residual feedforward network is used as the deep semantic features.

[0028] Preferably, the fusion of the shallow local features and the deep semantic features includes the following steps:

[0029] The shallow local features are abstracted through a first preset multi-layer fully connected network to obtain the processed shallow local features.

[0030] The deep semantic features are flattened and then abstracted through a second preset multi-layer fully connected network to obtain the processed deep semantic features.

[0031] Based on learnable weight parameters, the processed shallow local features and the processed deep semantic features are weighted and fused using residuals to obtain fused features.

[0032] Preferably, the output snow depth estimate is specifically:

[0033] The fused features are mapped onto the snow depth space using the Softplus activation function of the output layer to output a non-negative snow depth estimate.

[0034] Preferably, the loss function is: mean absolute error;

[0035] The formula for the loss function is:

[0036] ;

[0037] in, The mean absolute error, For the sample size, For the first Measured snow depth values ​​for each sample For the first Snow depth estimates for each sample.

[0038] Preferably, the cosine annealing learning rate scheduling strategy is used to make the learning rate smoothly decrease from the maximum value to the minimum value according to the cosine function curve during iterative training.

[0039] The formula for the cosine annealing learning rate scheduling strategy is as follows:

[0040] ;

[0041] in, The current learning rate, and These represent the initial maximum and final minimum values ​​of the learning rate, respectively. This represents the current iteration number. This represents the total number of iterations.

[0042] A snow depth estimation system based on a hybrid network model includes:

[0043] The standard set acquisition module is used to acquire brightness temperature data and snow depth data, process the brightness temperature data and snow depth data respectively, construct a brightness temperature and snow depth dataset based on the processed brightness temperature data and processed snow depth data, and standardize the brightness temperature and snow depth dataset to obtain a brightness temperature and snow depth standard set.

[0044] The hybrid network model building module is used to construct a convolutional-Transformer hybrid network model consisting of an input layer, an encoding-cooperative layer, and an output layer. The input layer is used to extract shallow local features from the brightness temperature and snow depth standard set, the encoding-cooperative layer is used to obtain deep semantic features to represent global dependencies, and the output layer is used to fuse the processed shallow local features and the processed deep semantic features and output the snow depth estimate.

[0045] The model training module is used to train the convolutional-transformer hybrid network model based on the brightness temperature and snow depth standard set, using the AdamW optimizer and cosine annealing learning rate scheduling strategy, until the loss function converges, and obtain the optimized convolutional-transformer hybrid network model.

[0046] The snow depth estimation module is used to input the standard set of brightness temperature snow depth to be estimated into the optimized convolutional-transformer hybrid network model to obtain the snow depth estimate corresponding to the standard set of brightness temperature snow depth to be estimated.

[0047] The beneficial effects of this invention are as follows: This invention discloses a snow depth estimation method based on a hybrid network model. After acquiring brightness temperature data and snow depth data, the method processes and standardizes them to obtain a brightness temperature and snow depth standard set. A convolutional-Transformer hybrid network model is constructed to extract shallow local features and deep semantic features from the brightness temperature and snow depth standard set. The processed shallow local features and deep semantic features are then fused to output the snow depth estimate. The hybrid network model is optimized according to the brightness temperature and snow depth standard set, the AdamW optimizer, and the cosine annealing learning rate scheduling strategy to obtain an optimized hybrid network model. The brightness temperature and snow depth standard set to be estimated is input into the optimized hybrid network model to obtain the corresponding snow depth estimate, i.e., to predict the snow depth for a certain period in the future.

[0048] The convolutional-Transformer hybrid network model constructed in this invention extracts local feature information from adjacent channels or time steps in the input layer. In the encoding-cooperation layer, it employs a Pre-Norm structure, LayerScale mechanism, multi-head self-attention (MHA), and a seven-layer deep residual network (MLP). In the output layer, it uses a dual-branch residual fusion regression head and Softplus physical constraints, significantly improving the snow depth estimation accuracy. The model is trained using the AdamW optimizer and a cosine annealing learning rate scheduling strategy, improving its generalization and stability. The overall solution provides a technical chain from data acquisition, model construction, model training to practical application, improving operational efficiency.

[0049] In summary, this invention defines an advanced, integrated deep learning framework that systematically addresses the challenges in accuracy, stability, and generalization currently faced in the field of snow depth inversion through a hybrid design in structure and optimization strategies in training.

[0050] This invention also provides a snow depth estimation system based on a hybrid network model. Since it shares the same technical concept and solves the same technical problem as this method, it should have the same beneficial effects, and will not be described in detail here. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a snow depth estimation method based on a hybrid network model provided in an embodiment of the present invention;

[0053] Figure 2 This is a flowchart of the brightness temperature data processing provided in an embodiment of the present invention;

[0054] Figure 3 This is a flowchart of obtaining deep semantic features provided in an embodiment of the present invention;

[0055] Figure 4 This is a flowchart illustrating the fused shallow local features and the fused deep semantic features provided in this embodiment of the invention.

[0056] Figure 5 This is a projection of Arctic sea ice thickness using the standard ViT model provided in this embodiment of the invention;

[0057] Figure 6 This is a projection map of Arctic sea ice thickness based on a hybrid network model provided in this embodiment of the invention;

[0058] Figure 7 This is a projection image of the actual Arctic sea ice thickness provided in an embodiment of the present invention;

[0059] Figure 8 The distribution of prediction errors under different snow depths;

[0060] Figure 9 This is a schematic diagram of the structure of a snow depth estimation system based on a hybrid network model provided in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The embodiments of this invention are written in a progressive manner.

[0063] This invention provides a snow depth estimation method and system based on a hybrid network model. It primarily addresses the technical problems of low prediction accuracy, poor stability, and weak generalization in traditional snow depth inversion methods.

[0064] like Figure 1 As shown, a snow depth estimation method based on a hybrid network model includes the following steps:

[0065] S1. Obtain brightness temperature data and snow depth data, process the brightness temperature data and snow depth data respectively, construct a brightness temperature and snow depth dataset based on the processed brightness temperature data and processed snow depth data, standardize the brightness temperature and snow depth dataset to obtain a brightness temperature and snow depth standard set.

[0066] S2. Construct a convolutional-transformer hybrid network model consisting of an input layer, an encoding-cooperative layer, and an output layer. The input layer is used to extract shallow local features from the brightness, temperature, and snow depth standard set. The encoding-cooperative layer is used to obtain deep semantic features that represent global dependencies. The output layer is used to fuse the processed shallow local features and the processed deep semantic features and output the snow depth estimate.

[0067] S3. Based on the brightness temperature snow depth standard set, the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to train the convolution-transformer hybrid network model until the loss function converges, and the optimized convolution-transformer hybrid network model is obtained.

[0068] S4. Input the standard set of brightness temperature and snow depth to be estimated into the optimized convolutional-transformer hybrid network model to obtain the snow depth estimate corresponding to the standard set of brightness temperature and snow depth to be estimated.

[0069] Steps S1 to S4 detail the implementation of the snow depth estimation method based on a hybrid network model. Step S1 involves acquiring daily satellite-observed brightness temperature data and actual measured daily sea ice snow thickness data, processing and constructing a brightness temperature-snow depth dataset corresponding to the marine scene, and standardizing each subset of the complete sea ice brightness temperature-snow depth dataset to obtain a standard brightness temperature-snow depth set. Step S2 involves constructing a snow depth estimation model based on a convolutional-Transformer hybrid network, inputting the brightness temperature sequence into the snow depth estimation model. Local features are extracted through convolutional embedding, global dependencies are captured by the Transformer encoder, and finally, the predicted snow depth data is obtained by fusing high-level semantic representations and original shallow features through a two-branch residual fusion network structure. The convolutional-transformer hybrid network model mainly includes an input layer, an encoding-cooperative layer, and an output layer. Step S3 involves training the convolutional-transformer hybrid network model using the AdamW optimizer and a cosine annealing learning rate scheduling strategy until the loss function (mean absolute error) converges, resulting in an optimized convolutional-transformer hybrid network model. Step S4 involves applying the optimized convolutional-transformer hybrid network model to obtain the corresponding snow depth estimate.

[0070] In one embodiment, brightness temperature data and sea ice / snow cover thickness data from AMSR-E (Advanced Microwave Scanning Radiometer-Earth Observation System) were used as the experimental dataset. The detailed procedure is described below:

[0071] We acquired daily brightness temperature data and daily sea ice and snow thickness data from AMSE-R satellite observations in January of a certain year. We then performed fine preprocessing on the brightness temperature data, including data correction, region selection, and data standardization, to form a continuous brightness temperature dataset with [longitude, latitude, and time] as its structure, ensuring data consistency and usability.

[0072] Brightness temperature data for each subset of the complete sea ice brightness temperature and snow cover thickness dataset were standardized. Specifically, the brightness temperature data was normalized by dividing its original value by a preset physical constant (e.g., 2000.0K) to scale the data to a stable range while preserving its original physical scale characteristics. The standard set is in the shape of [number of grid points, 25], where the first column represents the solar sea ice and snow cover thickness value corresponding to the grid point, and the following 24 columns represent the solar brightness temperature value corresponding to the grid point.

[0073] A convolutional-Transformer hybrid network model is trained using brightness temperature data from a standard set as the input feature set and snow thickness data as the output feature set. A loss function is used between the thickness values ​​output by the network and the thickness values ​​in the target set. The gradient of the loss function with respect to the model parameters is calculated using backpropagation, and the model parameters are updated using an optimization algorithm based on this gradient. Training is stopped when the loss function is minimized, thus determining that the brightness temperature-encoding convolutional co-network model at this point is a suitable ocean snow depth inversion model.

[0074] Using a trained convolutional-transformer hybrid network model, and inputting a brightness temperature data sequence without a trained network structure, we predict the snow depth distribution in February of a certain year and present the predicted snow depth estimation results.

[0075] like Figure 2 As shown, preferably, the brightness temperature data is processed, including the following steps:

[0076] A1. Convert the brightness temperature data to obtain standard format brightness temperature data;

[0077] A2. Perform geometric correction, radiometric correction, error calibration, and cloud filtering operations sequentially on the standard format brightness temperature data to obtain the corrected brightness temperature data;

[0078] A3. Based on the latitude and longitude of the region, select the region from the corrected brightness temperature data to obtain the processed brightness temperature data.

[0079] Steps A1 to A3 detail the specific implementation of brightness temperature data processing. This involves converting the raw brightness temperature data into the NC format (NetCDF), a standard format for scientific data storage and sharing, widely used in meteorology, oceanography, geophysics, and other fields. Subsequently, geometric correction, radiometric calibration, error calibration, and cloud filtering are performed on the converted NC format data to ensure data quality and accuracy. After these operations, the corrected brightness temperature data is obtained. Finally, based on the latitude and longitude of the observed ocean region, the corresponding area is selected from the corrected brightness temperature data to obtain the processed brightness temperature data.

[0080] In one specific embodiment, the present invention uses the Level-3 (L3) daily brightness temperature gridded standard product of the AMSR2 sensor as the raw input. The processing of this data is implemented using the MATLAB R2025b software platform, and the specific process is as follows:

[0081] Data conversion and reading (corresponding to step A1): Since the original downloaded brightness temperature data is in NetCDF (.nc) format, this embodiment converts it using MATLAB's built-in scientific data interface tool.

[0082] Specific steps: Use MATLAB's h5read function to read the brightness temperature dataset from the file. For example, read the 36.5GHz vertical polarization channel data, convert it from the original unsigned integer (uint16) to a double-precision floating-point matrix, and restore the values ​​to the actual physical brightness temperature values ​​(unit: Kelvin K) according to the scale factor (usually 0.1) in the data specification.

[0083] Correction and screening strategy (corresponding to step A2): This embodiment completes the correction step by combining the selection of preferred data sources with quality label filtering.

[0084] Geometric and radiometric correction: This embodiment directly uses the officially pre-processed Level-3 product. Before its release, this level product has undergone geometric correction based on satellite orbit parameters (Geo-reference) and radiometric correction based on blackbody calibration by the data provider (such as the National Snow and Ice Data Center, NSIDC), and the data has been resampled onto the EASE-Grid or Polar Stereographic projection grid to meet the "standard format" requirement of this step.

[0085] Error and Cloud Filtering: The Quality Flag / Mask attached to the data file is used. This Mask matrix is ​​read in MATLAB, and pixels marked as "missing data," "even errors," or "land spillover effects" are removed using logical indexing. For cloud filtering, since passive microwaves have the physical characteristic of penetrating cloud layers (except for extreme precipitation cloud systems), this embodiment primarily removes strong atmospheric interference noise by eliminating brightness temperature anomalies (such as values ​​exceeding the physical range of 50K-320K).

[0086] Region selection and processing (corresponding to step A3): Based on the geographical range of the study area (Arctic Ocean), matrix truncation is performed in MATLAB.

[0087] Specific steps: Obtain the latitude and longitude matrix (Lat / Lon Grid) corresponding to the data. Set the latitude and longitude threshold for the target area (latitude > 60°N). Use MATLAB's logical conditional statements to generate a region mask. Apply this mask to extract the values ​​of the corresponding region from the brightness temperature matrix, remove irrelevant background areas, and complete the region selection.

[0088] Preferably, the snow depth data is processed, including the following steps:

[0089] Based on the latitude and longitude of the region, the snow depth data is selected by region selection to obtain the snow depth data after region selection;

[0090] Bilinear interpolation was used to spatially interpolate the snow depth data after the region selection to obtain the processed snow depth data.

[0091] Spatial interpolation is used to align the processed snow depth data with the processed brightness temperature data in spatial dimensions.

[0092] The above steps detail the specific implementation of snow depth data processing. Based on subsequent alignment considerations, the snow depth data is also region-selected according to the latitude and longitude of the area, choosing snow depth data from regions with the same brightness temperature data to obtain the selected snow depth data. Bilinear interpolation is then used to perform spatial interpolation on this selected snow depth data to obtain the processed snow depth data. Bilinear interpolation is a mathematical method for numerical interpolation on a two-dimensional regular grid. It is used to estimate the approximate value of any target point in the grid based on the values ​​at known discrete points. Its core idea is to perform one-dimensional linear interpolation along two coordinate axes (such as longitude and latitude) within a rectangular region formed by the four nearest neighbor grid points of the target point, and then weighted and synthesized the two interpolation results to obtain the value of the target point. The purpose of this spatial interpolation is to align the processed snow depth data with the processed brightness temperature data in terms of spatial temperature.

[0093] In one embodiment, the spatial interpolation process is as follows:

[0094] Assume the coordinates of the target grid point to be interpolated are The coordinates of its four nearest neighbor known data points are as follows: bottom left corner bottom right corner Top left corner and the top right corner The corresponding known snow depth values ​​are respectively , , and .

[0095] The formula for calculating the snow depth at a target point using bilinear interpolation is as follows:

[0096] ;

[0097] In the formula, This is the snow depth estimate at the target grid point obtained through bilinear interpolation.

[0098] Preferably, a brightness temperature and snow depth dataset is constructed based on the processed brightness temperature data and processed snow depth data. The brightness temperature and snow depth dataset is then standardized to obtain a standard brightness temperature and snow depth dataset, including the following steps:

[0099] The processed brightness temperature data and processed snow depth data are time-aligned to obtain spatiotemporally aligned brightness temperature data and snow depth data;

[0100] Latitude and longitude matching was performed on the spatiotemporally aligned brightness temperature data and snow depth data to remove outliers and obtain the brightness temperature and snow depth dataset.

[0101] Normalization is performed on each subset of the brightness temperature and snow depth dataset to obtain a standard set of brightness temperature and snow depth containing multiple brightness temperature sequences.

[0102] The above steps detail the specific implementation of constructing the brightness temperature and snow depth standard set. This involves aligning the spatially aligned brightness temperature and snow depth data to the same time base to obtain spatiotemporally aligned brightness temperature and snow depth data. Subsequently, latitude and longitude matching is performed. Based on the matched data, outliers and erroneous data are removed to ensure accuracy and consistency. Each subset of the complete sea ice brightness temperature and snow depth dataset is then standardized. Specifically, the brightness temperature and snow depth data are normalized to the range [0,1] to ensure data consistency and scalability, and to avoid numerical instability during model training. A possible normalization method is the minimax method, with the following formula:

[0103] ;

[0104] in, The original data values, The minimum value of this feature in the dataset. The maximum value of this feature in the dataset. The value is the normalized value, ranging from [0,1].

[0105] like Figure 3 As shown, preferably, obtaining deep semantic features to characterize global dependencies includes the following steps:

[0106] B1. The brightness temperature sequence in the brightness temperature and snow depth standard set is divided into patches by a depthwise separable convolutional module, then mapped to a high-dimensional embedding vector, and a learnable positional code is added to form a token sequence.

[0107] B2. Input the token sequence into a Transformer encoder consisting of several stacked coding blocks; where each coding block adopts a Pre-Norm structure and includes: a layer normalization module, a multi-head self-attention module with LayerScale mechanism, and a multi-layer deep residual feedforward network;

[0108] B3. Extract shallow features from the Transformer encoder, concatenate the shallow features with the self-attention weight matrix, perform layer normalization, and then perform layer-by-layer cumulative residual connection calculation. Use the output of the multi-layer deep residual feedforward network as the deep semantic features.

[0109] Steps B1 to B3 detail the specific implementation of the encoding-cooperative layer for obtaining deep semantic features. This involves inputting a standard set of brightness, temperature, and snow depth data into a convolutional-Transformer hybrid network model. The separable convolutional module in the input layer of the hybrid network model divides the brightness and temperature sequences in the standard set into patch embeddings, linearly maps them to high-dimensional embedding vectors, and adds learnable positional encodings to form token sequences. These token sequences are then input into a Transformer encoder consisting of several stacked encoding blocks. Each encoding block uses a Pre-Norm structure and includes a layer normalization module, a multi-head self-attention module with LayerScale mechanism, and a multi-layer deep residual feedforward network. After processing by the Transformer encoder, shallow features are extracted. These shallow features are concatenated with the self-attention weight matrix and then subjected to layer normalization. Finally, layer-by-layer cumulative residual connection calculations are performed, and the output of the multi-layer deep residual feedforward network is used as the deep semantic features.

[0110] The Pre-Norm structure in step B2 is a layer normalization placement strategy in Transformer neural networks. Its core definition is: before performing computation in each sub-layer (such as a self-attention layer or a feedforward neural network layer), the input data is first normalized, then the core operation of that sub-layer is performed, and finally the result is added to the original input through residual connections. The LayerScale mechanism is an adaptive residual scaling technique used to stabilize the training of deep neural networks (especially Vision Transformer and its variants). Its core idea is: on the residual connection path, a learnable, element-wise (or channel-wise) scaling parameter vector is introduced for each residual branch to dynamically adjust the magnitude of the residual contribution during training, thereby alleviating gradient instability and training divergence problems. The multi-layer deep residual feedforward network is a key sub-module in the Transformer architecture. It is defined as building upon the standard two-layer feedforward network by significantly increasing the number of network layers and introducing a unique layer-by-layer cumulative residual connection mechanism to construct a deep sub-network with extremely strong nonlinear fitting capabilities and gradient fluidity.

[0111] In one embodiment, the implementation process of the above steps is explained in conjunction with the specific structure of the convolutional-Transformer hybrid network model, as follows:

[0112] The input layer in the convolutional-transformer hybrid network model is responsible for processing the 24-dimensional brightness temperature sequence corresponding to H grid points. Convert to a sequence of tokens suitable for Transformer processing ;

[0113] The input layer contains a convolution block embedding module and a learnable positional encoding module;

[0114] This module is used to convert the input brightness temperature sequence into a feature representation suitable for network processing. The module first extracts local features from the brightness temperature sequence, then divides it into fixed-size patches (tokens), and maps these patches to a high-dimensional embedding space D.

[0115] Traditional Transformer models often use simple linear projection for patch embedding when processing sequence data, which easily overlooks the local correlations between adjacent frequency points or time steps. To address this characteristic of brightness temperature sequences, this embodiment employs a convolution patch embedding module, enhanced by introducing a two-layer Stem structure of depthwise separable convolution (Depthwise Conv + Pointwise Conv). This Stem structure acts as a local pattern enhancer before sequence patching, efficiently extracting local feature information from adjacent channels or time steps. This ensures that the tokens sequence processed by the Transformer encoder not only contains global positional information but also richer physically meaningful local contextual information. Finally, the sequence is patched and embedded using a Conv1D projection layer, as follows:

[0116] Assume the sequence dimension of the brightness temperature dataset is W=24. The patch embedding layer divides the sequence into patches of size P. In this invention, the patch size P=4, and the sequence length W=24 is divisible by P. The entire sequence will be divided into N=W / P=6 patches. Each patch, after being flattened, is mapped through a linear projection layer (convolutional layer) to convert it into a fixed embedding dimension D (D=96 in this invention). Formulaically, convolutional block embedding can be expressed as:

[0117] ;

[0118] in, This represents the brightness temperature sequence after extracting local features through the Separate Convolution Stem algorithm. This is the dataset after tile embedding. Finally, all the tile embeddings will form a dataset of shape [H, N, D].

[0119] By adding positional information to each token using a learnable positional encoding module, the model can better understand the channel / temporal order of elements in the brightness-temperature sequence, thus stabilizing the training process. This embodiment employs learnable positional encoding. Its parameters are continuously updated during training. Compared to traditional positional encoding, learnable positional encoding can better adapt to the spatiotemporal structure of brightness-temperature data. Adding these positional encodings to the feature dataset yields the initial feature set, i.e., the token sequence. The formula is:

[0120] ;

[0121] in, It is a learnable parameter matrix with dimension . .

[0122] The encoding-cooperation layer is used to enhance the Transformer encoder structure;

[0123] The encoding-cooperation layer is the core module for global dependency modeling, consisting of L=4 stacked Transformer blocks. To address the instability issue in deep Transformer training, this invention employs a Pre-Norm structure, LayerScale mechanism, multi-head self-attention (MHA), and a seven-layer deep residual network (MLP) for each block, responsible for extracting complex high-dimensional features from the embedded features.

[0124] To address the instability and convergence difficulties during training of deep Transformer networks, this invention introduces a LayerScale mechanism on the residual connections of the MHA and FFN modules. This mechanism introduces a learnable scaling parameter. The output of the residual branch is adaptively scaled to smoothly connect deep modules.

[0125] The state update calculation for the Lth Transformer block is as follows:

[0126] Multi-head self-attention (MHA) steps:

[0127] ;

[0128] Feedforward Network (FFN) Steps:

[0129] ;

[0130] In the formula, L is the layer index number of the current Transformer coding block;

[0131] The input feature sequence is the Lth Transformer coded block;

[0132] This represents the intermediate feature state after processing by the multi-head self-attention module in the Lth layer;

[0133] This is the output feature sequence of the Lth Transformer coding block (i.e., the input of the L+1th layer).

[0134] This is a multi-head attention module used to capture global dependencies in a sequence;

[0135] This is a feed-forward network module used for non-linear transformation and enhancement of features;

[0136] This is a learnable scaling vector (diagonal matrix) for the LayerScale mechanism in the attention module, used to dynamically adjust the weights of the residual branches;

[0137] This is the learnable scaling vector (diagonal matrix) of the LayerScale mechanism in the feedforward network module.

[0138] This is an element-wise multiplication operation (Hadamard Product).

[0139] This is a stochastic depth regularization operation used to randomly discard parts of the path during training to prevent overfitting.

[0140] The basic operation of multi-head self-attention mechanisms is to map input features X to query (Q), key (K), and value (V) spaces, and then aggregate these values ​​by calculating attention weights to capture the global dependencies of the sequence. The calculation process is as follows:

[0141] Calculate the Q, K, and V space values ​​of the input feature map:

[0142] ;

[0143] ;

[0144] ;

[0145] in, It is the feature set obtained by performing patch embedding and position encoding on the input features (i.e., normalized brightness temperature data), and the token sequence. ), , and It is a trainable weight matrix.

[0146] Calculate attention weights:

[0147] ;

[0148] in, yes , , Dimensions These are the output attention weights of the multi-head self-attention mechanism. Scaled dot product attention is used here to ensure numerical stability when calculating the attention weights;

[0149] Add residual connection and normalization to the output values:

[0150] ;

[0151] in, It is the input feature matrix. It is the output attention weight of the multi-head self-attention mechanism. It is a layer normalization operation. This is the output value.

[0152] In the standard Transformer architecture, the feedforward network (FFN, or multilayer perceptron MLP) typically consists of only two fully connected layers and one activation function (such as GELU or ReLU), which limits its nonlinear expressive power and feature combination depth. To address the extremely high requirements of snow depth inversion tasks for complex nonlinear features, this embodiment employs a seven-layer ReLU-MLP structure in each Transformer block and introduces unique layer-by-layer cumulative residual connections to significantly enhance the network's nonlinear expressive power and multi-scale feature combination capabilities, representing a structurally differentiated improvement.

[0153] Suppose the input to the FFN module is the feature of a layer-normalized LN. Embedding dimension D=96, hidden layer dimension =192.

[0154] The calculation process for the cumulative residual layer by layer is as follows:

[0155] The core of this structure lies in the fact that the computation of each layer (starting from the third layer) accumulates the output of all the previous intermediate layers, thereby preserving multi-scale feature information and effectively alleviating the gradient vanishing problem in deep networks.

[0156] ;

[0157] in, and They are the first Layer weight matrix and bias terms. Final output. via After the transformation, the backbone residual connections will be added via the LayerScale mechanism to complete the process. The calculations are performed to incorporate the higher-order nonlinear features extracted by the MLP into the encoder output. middle.

[0158] like Figure 4 As shown, preferably, fusing the processed shallow local features and the processed deep semantic features includes the following steps:

[0159] C1. The shallow local features are abstracted and processed through a first preset multi-layer fully connected network to obtain the processed shallow local features;

[0160] C2. After flattening the deep semantic features, they are abstracted through a second preset multi-layer fully connected network to obtain the processed deep semantic features;

[0161] C3. Based on learnable weight parameters, the processed shallow local features and processed deep semantic features are weighted and fused to obtain fused features.

[0162] Preferably, the snow depth estimate is output as follows:

[0163] The Softplus activation function of the output layer is used to map the fused features onto the snow depth value space to output a non-negative snow depth estimate.

[0164] Steps C1 to C3 are the specific implementation details of the fusion processing performed at the output layer in the convolutional-transformer hybrid network model. They involve abstracting shallow local features through multiple fully connected networks to obtain processed shallow local features; flattening and abstracting the deep semantic features obtained in the previous steps to obtain processed deep semantic features; and then, using a weighted residual fusion method with preset learnable weight parameters, fusing the processed shallow local features and the processed deep semantic features to obtain the fused features.

[0165] The multi-layer fully connected network abstraction operation in step C3 refers to the process of sequentially passing the input feature vector (or the flattened feature tensor) through multiple fully connected layers (also known as linear layers or dense layers), and applying a non-linear activation function (such as ReLU, GELU, etc.) after each layer to extract, combine, and transform features layer by layer, ultimately obtaining a higher-level and more discriminative abstract representation. Weighted residual fusion is a feature integration strategy. Its core is to combine two feature representations from different sources or at different levels of abstraction by weighting them with learnable weight coefficients, and then use one of the weighted features as a supplement or correction term for the other feature in the form of residual connections, ultimately obtaining a more informative and robust fused feature.

[0166] Specifically, the abstract operation of a multi-layer fully connected network is as follows:

[0167] Given an input feature vector ,one The abstract operations of a fully connected layer network can be formally represented as:

[0168] No. Layer transformation ( ):

[0169] ;

[0170] in, For the first The output of the layer (when) hour, ), This is the weight matrix for this layer. This is the bias vector for this layer. It is a non-linear activation function. For the first Hidden dimensions of layers;

[0171] Output layer transformation :

[0172] If the task is feature abstraction (as described in steps C1 and C2), non-linear activation is usually retained, and the output dimension is... Based on subsequent integration requirements:

[0173] .

[0174] In one embodiment, the output layer maps the high-dimensional features output by the encoder to the final predicted snow depth value, employing a dual-branch residual fusion regression head, as detailed below:

[0175] In deep Transformer networks, key linear or weakly nonlinear features in the original input are easily diluted or ignored during the abstraction process of the multi-layer encoder. To address this, this embodiment designs a dual-branch regression head specifically for integrating deep semantic information with the original shallow information to enhance the robustness and prediction accuracy of the model.

[0176] Branch A (High-level semantic features): This branch is responsible for extracting the encoder output. Deep semantic representation in the model. After flattening the encoder output, it is passed through a multi-layer fully connected network ( Further abstraction yields features. .

[0177] Branch B (Original Shallow Features): This branch directly uses the standardized original 24-dimensional brightness temperature input. Processing, through shallow layers Extracting linear or weakly nonlinear representations The purpose is to preserve key information from the original input.

[0178] Subsequently, the features from the two branches are integrated using a weighted residual fusion method to generate the final fused features. The formula is as follows:

[0179] ;

[0180] in, and These are the weights and biases used to project the features of branch B onto the same dimension as branch A and sum the residuals. and These are learnable weighted parameters that allow the model to automatically optimize the ratio of dependence on deep abstract information to shallow raw information, ensuring the optimality of the fused features.

[0181] After obtaining the fused features, the fused features are mapped onto the snow depth space through the Softplus activation function of the output layer to output a non-negative snow depth estimate.

[0182] The Softplus activation function is a smooth, monotonic, and always positive nonlinear activation function. It is often used in the output layer of neural networks to force the output range of the model to be positive, thereby achieving a non-negative physical constraint on the target variable.

[0183] Defined as: ;

[0184] Since the range of this function is always greater than zero, it forces the model's predicted output to be greater than or equal to 0, thus strictly adhering to the physical constraints at the output end, which significantly improves the practicality and reliability of the model inversion results.

[0185] As a physical measurement of surface elements, snow thickness prediction must satisfy a non-negativity constraint (i.e., greater than or equal to 0). Traditional regression models typically use linear outputs, which may result in negative values ​​and lack physical plausibility.

[0186] To address this issue, this embodiment employs the Softplus activation function in the regression head and tail layers to fuse features. Mapping to the final snow depth value space:

[0187] ;

[0188] in, This is an estimate of the snow depth. and These are the weight matrix and bias term of the output layer, respectively.

[0189] Preferably, the loss function is: mean absolute error;

[0190] The formula for the loss function is:

[0191] ;

[0192] in, The mean absolute error, For the sample size, For the first Measured snow depth values ​​for each sample For the first Snow depth estimates for each sample.

[0193] In practical applications, mean absolute error (MAE) is a loss function or evaluation metric used to measure the average absolute deviation between the predicted and actual values ​​of a regression model. It calculates the arithmetic mean of the absolute errors over all samples, is relatively robust to outliers, and has an intuitive physical meaning (a measure of the average prediction error).

[0194] In one embodiment, the loss function is specifically chosen as mean absolute error (MAE). Since the expected model output snow depth value is smaller and smaller than the actual snow depth value, using mean absolute error as the loss function can effectively improve the accuracy of model inversion.

[0195] Preferably, the cosine annealing learning rate scheduling strategy is used to make the learning rate smoothly decrease from the maximum value to the minimum value according to the cosine function curve during iterative training.

[0196] The formula for the cosine annealing learning rate scheduling strategy is:

[0197] ;

[0198] in, The current learning rate, and These represent the initial maximum and final minimum values ​​of the learning rate, respectively. This represents the current iteration number. This represents the total number of iterations.

[0199] In practical application, step S3 uses the AdamW optimizer and cosine annealing learning rate scheduling strategy to train the convolutional-Transformer hybrid network model until the loss function converges. The AdamW optimizer is an improved version of the Adam optimizer. Its core improvement lies in decoupling weight decay (L2 regularization) from the gradient update process, thereby performing regularization more effectively, preventing overfitting, and improving the model's generalization ability.

[0200] The traditional Adam optimizer is used when applied to products with weight decay ( When applying regularized deep learning models, the weight decay effect is often unsatisfactory. This invention employs the AdamW optimizer, which decouples weight decay from gradient updates and applies it independently to model parameters, thereby enabling AdamW to be implemented more accurately and effectively. Regularization. This significantly improves the model's generalization ability, reduces the risk of overfitting, and ensures the efficiency of optimizing complex parameters in the Transformer structure.

[0201] To address the complexity of training deep Transformer structures and the computational resource requirements, this embodiment employs a combination of the following efficient stabilization techniques:

[0202] Cosine Annealing Learning Rate Scheduling Strategy

[0203] In the early stages of model training, a large learning rate is needed to achieve rapid convergence, while a smaller learning rate is required for fine-tuning in the later stages. This embodiment employs a cosine annealing learning rate scheduling strategy, as shown in the formula above, to achieve a learning rate... Smoothly following the cosine function curve from its maximum value Reduce to minimum value This smooth decay method avoids the oscillations caused by a sudden drop in the learning rate, and can better search for the global optimum, ensuring rapid convergence in the early stages of training and fine optimization in the later stages of training.

[0204] Gradient Clipping

[0205] Because the deep Transformer structure (such as LayerScale and seven-layer MLP) used in this embodiment is complex, the risk of gradient explosion is high in the early stages of training or when facing sudden data changes. To ensure the stability and controllability of the training process, this embodiment employs gradient pruning. This technique calculates the L2 norm of the gradient and limits it to a preset maximum threshold max_norm (max_norm = 1.0 in this embodiment). This effectively prevents model crashes or oscillations caused by excessively large parameter updates during training.

[0206] To verify the effectiveness and advancement of the snow depth estimation method based on a hybrid network model proposed in this invention, numerous comparative experiments were conducted using actual observation data from the Arctic region. The experiments used brightness temperature data from the AMSR2 sensor as input and validated sea ice snow cover thickness products as ground truth labels.

[0207] Dataset Construction: Brightness temperature and snow depth data from the Arctic region over six consecutive days were selected, totaling 708,536 sample points. The data was divided into a training set and a test set. The test set contains complete Arctic spatial cover data and is used to evaluate the model's generalization ability.

[0208] Data preprocessing: The Z-Score normalization strategy proposed in this invention is used to process the input brightness temperature data so that its distribution is centered at 0 and has a variance of 1, in order to accelerate model convergence.

[0209] Comparison model settings:

[0210] Comparison scheme (baseline model): Adopts the standard Vision Transformer (ViT) structure. Its characteristic is that it uses convolution with a stride of 4 (Conv1d, stride=4) for patch embedding, dividing the 24 channels into 6 patches, which are then flattened and output through a fully connected layer.

[0211] The proposed solution (optimized hybrid model) uses the convolutional-Transformer hybrid network constructed in this invention, introduces the ConvStem convolutional feature extractor (two convolutional layers extract local gradients and deep features), sets the Patch Size to 1 to preserve the fine-grained relationships of all 24 channels, uses global average pooling (GAP) to replace the flattening operation, and uses SmoothL1Loss as the loss function.

[0212] Quantitative results analysis:

[0213] The prediction results based on the test set data were evaluated using metrics such as root mean square error (RMSE), coefficient of determination (R²), and average bias. The experimental results are shown in Table 1.

[0214] Table 1 Comparative Analysis of Inversion Methods

[0215]

[0216] Significantly improved accuracy (RMSE reduced by 11.1%): Existing technologies directly segment brightness temperature data, easily losing local correlations between adjacent frequency points. This invention introduces a convolution block embedding module (ConvStem) to effectively extract shallow local features of the brightness temperature sequence. Combined with the global modeling capabilities of Transformer, the root mean square error (RMSE) is reduced from 0.0836m to 0.0743m. This means that in practical applications, the snow depth estimate of this invention is closer to the true physical value.

[0217] Significantly Enhanced Fit (R² Improvement of 25.0%): The coefficient of determination (R²) reflects the model's ability to explain data variations. This invention significantly improves R² from 0.4541 to 0.5676, indicating that the model can not only capture the average trend of snow depth but also more accurately characterize the spatiotemporal variations of snow depth. This is attributed to the LayerScale mechanism and deep residual feedforward network introduced in the encoding-cooperative layer of this invention, which significantly enhances the nonlinear expressive power of deep networks and solves the problem that traditional Transformers are difficult to train deep networks.

[0218] Visual verification

[0219] To visually demonstrate the predictive effect of this invention, a comparison between the predicted snow depth distribution of this invention model over multiple consecutive days in the Arctic region and the actual values ​​is shown using a specific day as an example.

[0220] Spatial distribution consistency: such as Figures 5 to 7 As shown, the snow depth distribution map generated by the model (Prediction) of this invention is highly consistent with the actual observation map (Actual) in terms of spatial texture.

[0221] Detail capture capability: In areas of heavy snow in northern Greenland and near the Canadian archipelago (yellow and red areas in the figure, snow depth > 0.4m), the model can accurately predict the high value center; in areas of light snow in the marginal sea (blue area), the model also shows good transition characteristics.

[0222] Temporal stability: Continuous prediction results from Day 1 to Day 6 show that the model output changes smoothly over time without any abrupt noise spots, demonstrating the model's robustness in the time dimension.

[0223] To further investigate the model's performance under different snow depths, the test data was divided into four intervals (0-0.1m, 0.1-0.2m, 0.2-0.3m, >0.3m), and the distribution of the predicted-observed error was calculated. Figure 8 As shown.

[0224] Depend on Figure 8 As can be seen, the horizontal axis represents different snow depth classification intervals (from thin snow to thick snow); the vertical axis represents the prediction error (predicted value - measured value), and the dashed line (y=0) represents the ideal state of zero error. Within all snow depth intervals, the box height (i.e., interquartile range, IQR) of the model of this invention (red) is significantly smaller than that of the prior art model (blue). This indicates that the prediction results of the model of this invention are more concentrated, with less dispersion, and have higher stability. Especially in the thick snow interval (greater than 0.3m), the error distribution range of the prior art (blue) is extremely wide (the box is elongated and has long upper and lower dimensions), exhibiting significant uncertainty. In contrast, the model of this invention (red) maintains a narrower error distribution in this interval, significantly suppressing the divergence phenomenon in thick snow inversion. Observing the median line (the horizontal line in the middle of the box) within the box, the median line of the model of this invention is closer to the 0 mark, indicating that this invention effectively corrects the systematic biases present in the prior art (such as the tendency to underestimate in shallow snow areas and overestimate in thick snow areas).

[0225] like Figure 9 As shown, a snow depth estimation system based on a hybrid network model includes:

[0226] The standard set acquisition module is used to acquire brightness temperature data and snow depth data, process the brightness temperature data and snow depth data respectively, construct a brightness temperature and snow depth dataset based on the processed brightness temperature data and processed snow depth data, and standardize the brightness temperature and snow depth dataset to obtain the brightness temperature and snow depth standard set.

[0227] The hybrid network model building module is used to construct a convolutional-Transformer hybrid network model consisting of an input layer, an encoding-cooperative layer, and an output layer. The input layer is used to extract shallow local features from the brightness, temperature, and snow depth standard set, the encoding-cooperative layer is used to obtain deep semantic features to represent global dependencies, and the output layer is used to fuse the processed shallow local features and the processed deep semantic features and output the snow depth estimate.

[0228] The model training module is used to train the convolutional-transformer hybrid network model based on the brightness, temperature and snow depth standard set, using the AdamW optimizer and cosine annealing learning rate scheduling strategy, until the loss function converges, and obtain the optimized convolutional-transformer hybrid network model.

[0229] The snow depth estimation module is used to input the standard set of brightness temperature snow depth to be estimated into the optimized convolutional-transformer hybrid network model to obtain the snow depth estimate corresponding to the standard set of brightness temperature snow depth to be estimated.

[0230] This invention also discloses a snow depth estimation system based on a hybrid network model. The system aims to achieve snow depth estimation using a modular approach. A standard set acquisition module acquires brightness temperature and snow depth data, processes and standardizes them respectively, generating a brightness temperature and snow depth standard set. A hybrid network model construction module constructs a convolutional-transformer hybrid network model consisting of an input layer, an encoding-cooperation layer, and an output layer. This model extracts shallow local features from the brightness temperature and snow depth standard set, obtains deep semantic features representing global dependencies, fuses the processed shallow local features and processed deep semantic features, and outputs the snow depth estimate. A model training module trains and optimizes the convolutional-transformer hybrid network model based on the brightness temperature and snow depth standard set using the AdamW optimizer and a cosine annealing learning rate scheduling strategy, resulting in an optimized convolutional-transformer hybrid network model. Finally, a snow depth estimation module obtains the snow depth estimate corresponding to the brightness temperature and snow depth standard set to be estimated based on the optimized convolutional-transformer hybrid network model.

[0231] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

[0232] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0233] If a flowchart is used in this application, it is used to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0234] The foregoing has provided a detailed description of a method and system for snow depth estimation based on a hybrid network model, as provided in this application. The above description of the disclosed embodiments enables those skilled in the art to implement or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A snow depth estimation method based on a hybrid network model, characterized in that, Includes the following steps: Brightness temperature data and snow depth data are acquired, and the brightness temperature data and snow depth data are processed respectively. A brightness temperature and snow depth dataset is constructed based on the processed brightness temperature data and snow depth data. The brightness temperature and snow depth dataset is then standardized to obtain a brightness temperature and snow depth standard set. A convolutional-Transformer hybrid network model is constructed, consisting of an input layer, an encoding-cooperative layer, and an output layer. The input layer is used to extract shallow local features from the brightness temperature and snow depth standard set. The encoding-cooperative layer is used to obtain deep semantic features that characterize global dependencies. The output layer is used to fuse the processed shallow local features and the processed deep semantic features and output the snow depth estimate. The acquisition of deep semantic features for characterizing global dependencies includes the following steps: dividing the brightness temperature sequence in the brightness temperature-snow depth standard set into patch embeddings using a depthwise separable convolutional module, mapping them to high-dimensional embedding vectors, and adding learnable positional encodings to form token sequences; inputting the token sequences into a Transformer encoder consisting of several stacked encoding blocks; wherein each encoding block adopts a Pre-Norm structure and includes: a layer normalization module, a multi-head self-attention module with LayerScale mechanism, and a multi-layer deep residual feedforward network; extracting shallow features from the Transformer encoder, concatenating the shallow features with the self-attention weight matrix, performing layer normalization processing, and then performing layer-by-layer cumulative residual connection calculation; and using the output of the multi-layer deep residual feedforward network as the deep semantic features. The fusion of the shallow local features and the deep semantic features includes the following steps: abstracting the shallow local features through a first preset multi-layer fully connected network to obtain processed shallow local features; flattening the deep semantic features and abstracting them through a second preset multi-layer fully connected network to obtain processed deep semantic features; and performing weighted residual fusion of the processed shallow local features and the processed deep semantic features based on learnable weight parameters to obtain fused features. The output snow depth estimate is specifically achieved by using the Softplus activation function of the output layer to map the fused features onto the snow depth value space, so as to output a non-negative snow depth estimate. Based on the brightness temperature and snow depth standard set, the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to train the convolutional-transformer hybrid network model until the loss function converges, thus obtaining the optimized convolutional-transformer hybrid network model. The standard set of brightness temperature and snow depth to be estimated is input into the optimized convolutional-transformer hybrid network model to obtain the snow depth estimate corresponding to the standard set of brightness temperature and snow depth to be estimated.

2. The snow depth estimation method based on a hybrid network model as described in claim 1, characterized in that, The brightness temperature data is processed, including the following steps: The brightness temperature data is converted to obtain brightness temperature data in a standard format; The brightness temperature data in the standard format is subjected to geometric correction, radiometric correction, error calibration, and cloud filtering operations in sequence to obtain the corrected brightness temperature data. Based on the latitude and longitude of the region, the corrected brightness temperature data is selected to obtain the processed brightness temperature data.

3. The snow depth estimation method based on a hybrid network model as described in claim 1, characterized in that, Processing snow depth data includes the following steps: Based on the latitude and longitude of the region, the snow depth data is selected by region selection to obtain the snow depth data after region selection; The selected area snow depth data is spatially interpolated using bilinear interpolation to obtain the processed snow depth data. The spatial interpolation process is used to align the processed snow depth data with the processed brightness temperature data in spatial dimensions.

4. The snow depth estimation method based on a hybrid network model as described in claim 1, characterized in that, The process of constructing a brightness temperature and snow depth dataset based on processed brightness temperature and snow depth data, and then standardizing the brightness temperature and snow depth dataset to obtain a standard brightness temperature and snow depth dataset includes the following steps: The processed brightness temperature data and the processed snow depth data are time-aligned to obtain spatiotemporally aligned brightness temperature data and snow depth data; The spatiotemporally aligned brightness temperature data and snow depth data are matched by latitude and longitude to remove outliers, thus obtaining the brightness temperature and snow depth dataset. Normalization is performed on each subset of the brightness temperature and snow depth dataset to obtain the brightness temperature and snow depth standard set containing multiple brightness temperature sequences.

5. The snow depth estimation method based on a hybrid network model as described in claim 1, characterized in that, The loss function is specifically: mean absolute error; The formula for the loss function is: ； in, The mean absolute error, For the sample size, For the first Measured snow depth values ​​for each sample. For the first Snow depth estimates for each sample.

6. The snow depth estimation method based on a hybrid network model as described in claim 1, characterized in that, The cosine annealing learning rate scheduling strategy is used to make the learning rate smoothly decrease from the maximum value to the minimum value according to the cosine function curve during iterative training. The formula for the cosine annealing learning rate scheduling strategy is as follows: ； in, The current learning rate, and These represent the initial maximum and final minimum values ​​of the learning rate, respectively. This represents the current iteration number. This represents the total number of iterations.

7. A snow depth estimation system based on a hybrid network model, characterized in that, include: The standard set acquisition module is used to acquire brightness temperature data and snow depth data, process the brightness temperature data and snow depth data respectively, construct a brightness temperature and snow depth dataset based on the processed brightness temperature data and processed snow depth data, and standardize the brightness temperature and snow depth dataset to obtain a brightness temperature and snow depth standard set. The hybrid network model building module is used to construct a convolutional-Transformer hybrid network model consisting of an input layer, an encoding-cooperative layer, and an output layer. The input layer is used to extract shallow local features from the brightness temperature and snow depth standard set, the encoding-cooperative layer is used to obtain deep semantic features to represent global dependencies, and the output layer is used to fuse the processed shallow local features and the processed deep semantic features and output the snow depth estimate. The acquisition of deep semantic features for characterizing global dependencies includes the following steps: dividing the brightness temperature sequence in the brightness temperature-snow depth standard set into patch embeddings using a depthwise separable convolutional module, mapping them to high-dimensional embedding vectors, and adding learnable positional encodings to form token sequences; inputting the token sequences into a Transformer encoder consisting of several stacked encoding blocks; wherein each encoding block adopts a Pre-Norm structure and includes: a layer normalization module, a multi-head self-attention module with LayerScale mechanism, and a multi-layer deep residual feedforward network; extracting shallow features from the Transformer encoder, concatenating the shallow features with the self-attention weight matrix, performing layer normalization processing, and then performing layer-by-layer cumulative residual connection calculation; and using the output of the multi-layer deep residual feedforward network as the deep semantic features. The fusion of the shallow local features and the deep semantic features includes the following steps: abstracting the shallow local features through a first preset multi-layer fully connected network to obtain processed shallow local features; flattening the deep semantic features and abstracting them through a second preset multi-layer fully connected network to obtain processed deep semantic features; and performing weighted residual fusion of the processed shallow local features and the processed deep semantic features based on learnable weight parameters to obtain fused features. The output snow depth estimate is specifically achieved by using the Softplus activation function of the output layer to map the fused features onto the snow depth value space, so as to output a non-negative snow depth estimate. The model training module is used to train the convolutional-transformer hybrid network model based on the brightness temperature and snow depth standard set, using the AdamW optimizer and cosine annealing learning rate scheduling strategy, until the loss function converges, and obtain the optimized convolutional-transformer hybrid network model. The snow depth estimation module is used to input the standard set of brightness temperature snow depth to be estimated into the optimized convolutional-transformer hybrid network model to obtain the snow depth estimate corresponding to the standard set of brightness temperature snow depth to be estimated.

Images