CN-121808285-B - Snow depth estimation method and system based on hybrid network model

CN121808285BCN 121808285 BCN121808285 BCN 121808285BCN-121808285-B

Abstract

The invention discloses a snow depth estimation method and a system based on a hybrid network model, wherein the method is used for preprocessing and time-space alignment of bright temperature and real snow depth data to construct a standardized data set, the hybrid network model is constructed, an input layer of the hybrid network model adopts convolution partitioning to embed and extract local features, a coding-synergy layer LAYERSCALE mechanism captures global dependence, an output layer integrates deep semantic features and original shallow features through a double-branch residual fusion structure, an average absolute error is used as a loss function, a AdamW optimizer and cosine annealing learning rate scheduling are used for training the model to obtain an optimized snow depth estimation model, and a snow depth estimation value is predicted in real time through the optimized snow depth estimation model. The system has the same technical concept. The invention solves the problems of weak generalization capability, insufficient utilization of local and global characteristics and lack of physical constraint of output of the traditional method, and realizes high-precision and steady inversion of the thickness of the accumulated snow.

Inventors

DONG JIAN
DOU HAOFENG
LIU SHILEI
LI MINGJUN
Xiao Chengwang
LI YINAN
LIU SHUBO

Assignees

中南大学

Dates

Publication Date: 20260508
Application Date: 20260309

Claims (7)

1. The snow depth estimation method based on the hybrid network model is characterized by comprising the following steps of: Acquiring bright temperature data and snow depth data, respectively processing the bright temperature data and the snow depth data, constructing a bright Wen Xueshen data set based on the processed bright temperature data and the processed snow depth data, and carrying out standardized processing on the bright Wen Xueshen data set to obtain a bright Wen Xue deep standard set; Constructing a convolution-transform hybrid network model consisting of an input layer, a coding-cooperative layer and an output layer, wherein the input layer is used for extracting shallow layer local features in the bright Wen Xue deep standard set, the coding-cooperative layer is used for acquiring deep semantic features used for representing global dependency, and the output layer is used for fusing the processed shallow layer local features and the processed deep semantic features and outputting a snow depth estimated value; the method comprises the steps of dividing a bright temperature sequence in a bright Wen Xue deep standard set into blocks by a depth separable convolution module, then mapping the blocks into high-dimensional embedded vectors, adding a learnable position code to form a token sequence, inputting the token sequence into a transform encoder formed by a plurality of stacked encoding blocks, wherein each encoding block adopts a Pre-Norm structure and comprises a layer normalization module, a multi-head self-attention module with a LAYERSCALE mechanism and a multi-layer deep residual feedforward network; The shallow local features after fusion processing and the deep semantic features after processing comprise the following steps of carrying out abstract processing on the shallow local features through a first preset multi-layer full-connection network to obtain processed shallow local features; carrying out abstract processing on the deep semantic features through a second preset multi-layer full-connection network after flattening the deep semantic features to obtain processed deep semantic features; The method comprises the steps of adopting Softplus activation functions of the output layer to map the fusion characteristics in a snow depth value space so as to output the non-negative snow depth estimated value; Training the convolution-transform hybrid network model by adopting a AdamW optimizer and cosine annealing learning rate scheduling strategy based on the bright Wen Xue deep standard set until a loss function converges to obtain an optimized convolution-transform hybrid network model; And inputting the bright Wen Xue deep standard set to be estimated into the optimized convolution-transform mixed network model to obtain a snow depth estimated value corresponding to the bright Wen Xue deep standard set to be estimated.
2. The hybrid network model-based snow depth estimation method of claim 1, wherein the bright temperature data is processed, comprising the steps of: Performing data conversion on the bright temperature data to obtain bright temperature data in a standard format; sequentially performing geometric correction, radiation correction, error correction and cloud screening operation on the standard-format bright temperature data to obtain corrected bright temperature data; and performing region selection on the corrected bright temperature data based on the longitude and latitude of the region to obtain the processed bright temperature data.
3. The method for estimating the depth of snow based on the hybrid network model according to claim 1, wherein the processing of the depth of snow data comprises the steps of: Performing region selection on the snow depth data based on the longitude and latitude of the region to obtain snow depth data after region selection; performing spatial interpolation processing on the snow depth data after the region selection by adopting a bilinear interpolation method to obtain processed snow depth data; the spatial interpolation process is configured to align the processed snow depth data with the processed bright temperature data in a spatial dimension.
4. The hybrid network model-based snow depth estimation method of claim 1, wherein the constructing a bright Wen Xueshen dataset based on the processed bright temperature data and the processed snow depth data, normalizing the bright Wen Xueshen dataset to obtain a bright Wen Xue deep standard set, comprises the steps of: performing time alignment on the processed bright temperature data and the processed snow depth data to obtain bright temperature data and snow depth data after time-space alignment; performing longitude and latitude matching on the light temperature data and the snow depth data after space-time alignment to remove abnormal values in the light temperature data and the snow depth data so as to obtain the light Wen Xueshen data set; And carrying out normalization processing on each subset in the bright Wen Xue deep data set to obtain the bright Wen Xue deep standard set containing a plurality of bright temperature sequences.
5. The method for estimating the depth of snow based on the hybrid network model according to claim 1, wherein the loss function is specifically an average absolute error; The formula of the loss function is: ; Wherein, the As an average of the absolute error of the values, For the number of samples to be taken, Is the first The actual snow depth measurement value of each sample, Is the first Snow depth estimate of individual samples.
6. The hybrid network model-based snow depth estimation method of claim 1, wherein the cosine annealing learning rate scheduling strategy is used for smoothly reducing the learning rate from a maximum value to a minimum value according to a cosine function curve in an iterative training process; The formula of the cosine annealing learning rate scheduling strategy is as follows: ; Wherein, the For the current rate of learning to be the same, And The initial maximum and final minimum of the learning rate respectively, For the current number of iterations, Is the total number of iterations.
7. A hybrid network model-based snow depth estimation system, comprising: the standard set acquisition module is used for acquiring bright temperature data and snow depth data, respectively processing the bright temperature data and the snow depth data, constructing a bright Wen Xueshen data set based on the processed bright temperature data and the processed snow depth data, and carrying out standardized processing on the bright Wen Xueshen data set to obtain a bright Wen Xue deep standard set; The hybrid network model construction module is used for constructing a convolution-transform hybrid network model consisting of an input layer, a coding-cooperative layer and an output layer, wherein the input layer is used for extracting shallow layer local features in the light Wen Xue deep standard set, the coding-cooperative layer is used for acquiring deep semantic features used for representing global dependency, and the output layer is used for fusing the processed shallow layer local features and the processed deep semantic features and outputting a snow depth estimated value; the method comprises the steps of dividing a bright temperature sequence in a bright Wen Xue deep standard set into blocks by a depth separable convolution module, then mapping the blocks into high-dimensional embedded vectors, adding a learnable position code to form a token sequence, inputting the token sequence into a transform encoder formed by a plurality of stacked encoding blocks, wherein each encoding block adopts a Pre-Norm structure and comprises a layer normalization module, a multi-head self-attention module with a LAYERSCALE mechanism and a multi-layer deep residual feedforward network; The shallow local features after fusion processing and the deep semantic features after processing comprise the following steps of carrying out abstract processing on the shallow local features through a first preset multi-layer full-connection network to obtain processed shallow local features; carrying out abstract processing on the deep semantic features through a second preset multi-layer full-connection network after flattening the deep semantic features to obtain processed deep semantic features; The method comprises the steps of adopting Softplus activation functions of the output layer to map the fusion characteristics in a snow depth value space so as to output the non-negative snow depth estimated value; the model training module is used for training the convolution-transform hybrid network model by adopting a AdamW optimizer and cosine annealing learning rate scheduling strategy based on the bright Wen Xue deep standard set until the loss function converges to obtain an optimized convolution-transform hybrid network model; And the snow depth estimation module is used for inputting a bright Wen Xue deep standard set to be estimated into the optimized convolution-transform mixed network model to obtain a snow depth estimated value corresponding to the bright Wen Xue deep standard set to be estimated.

Description

Snow depth estimation method and system based on hybrid network model Technical Field The invention belongs to the technical field of surface element estimation, and particularly relates to a snow depth estimation method and system based on a hybrid network model. Background The passive microwave bright temperature (Brightness Temperature) is a key physical quantity reflecting the dielectric characteristics and structural information of the ground surface and snow, and has become an important remote sensing data source for large-range and long-time sequence snow depth inversion. The traditional mainstream snow depth inversion method is mainly divided into two types, namely a radiation transmission model based on a physical mechanism and a regression model based on a statistical rule. However, the physical model usually depends on a series of snow cover parameters (such as density, particle size, lamellar structure and the like) which are difficult to accurately obtain, and the generalization capability under different climates and under-pad conditions is limited, and the statistical model often has difficulty in fully mining complex nonlinear correlations of the bright temperature data in a long time sequence and a multi-frequency channel, so that inversion accuracy is limited, and the inversion is unstable especially under complex topography and snow cover conditions. In recent years, the end-to-end deep learning method provides a new technical path for snow depth inversion, can automatically learn high-level features from data, and breaks through the limitation of the traditional method to a certain extent. However, the existing deep learning-based scheme still has obvious defects that if a pure transducer architecture is directly adopted, long time sequence dependence can be effectively modeled, local correlation of bright temperature sequences between adjacent time steps or frequency points is easily ignored, so that local physical mode capture is insufficient, meanwhile, the deep transducer network training process is unstable, gradient abnormality or convergence difficulty is easily caused, more importantly, most of deep learning models lack of physical constraint on output results, the situation that predicted snow depth is negative and the like does not accord with physical reality often occurs, and the actual usability of inversion results is seriously affected. Therefore, it is a technical problem to be solved by those skilled in the art to provide a hybrid network model-based snow depth estimation method and system for solving the above problems. Disclosure of Invention In order to solve the technical problems, the invention aims to provide a snow depth estimation method and a system based on a hybrid network model, wherein the method combines local feature extraction and global dependence modeling, has a stable training mechanism, and the output result strictly accords with physical constraint so as to realize high-precision and high-robustness estimation of the snow depth in polar and complex environments. The technical scheme provided by the invention is as follows: A snow depth estimation method based on a hybrid network model comprises the following steps: Acquiring bright temperature data and snow depth data, respectively processing the bright temperature data and the snow depth data, constructing a bright Wen Xueshen data set based on the processed bright temperature data and the processed snow depth data, and carrying out standardized processing on the bright Wen Xueshen data set to obtain a bright Wen Xue deep standard set; Constructing a convolution-transform hybrid network model consisting of an input layer, a coding-cooperative layer and an output layer, wherein the input layer is used for extracting shallow layer local features in the bright Wen Xue deep standard set, the coding-cooperative layer is used for acquiring deep semantic features used for representing global dependency, and the output layer is used for fusing the processed shallow layer local features and the processed deep semantic features and outputting a snow depth estimated value; Training the convolution-transform hybrid network model by adopting a AdamW optimizer and cosine annealing learning rate scheduling strategy based on the bright Wen Xue deep standard set until a loss function converges to obtain an optimized convolution-transform hybrid network model; And inputting the bright Wen Xue deep standard set to be estimated into the optimized convolution-transform mixed network model to obtain a snow depth estimated value corresponding to the bright Wen Xue deep standard set to be estimated. Preferably, the bright temperature data is processed, which comprises the following steps: Performing data conversion on the bright temperature data to obtain bright temperature data in a standard format; sequentially performing geometric correction, radiation correction, error correction and cloud screening operation on the st