An adaptive wet troposphere path delay inversion method and system fusing wind speed overlapping partition modeling, multi-scale collaborative sample equalization and model soft fusion

CN122508958APending Publication Date: 2026-08-04NAT SPACE SCI CENT CAS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610437427.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-03
Publication Date
2026-08-04

AI Technical Summary

Technical Problem

[0004]本申请的目的在于,克服现有技术因风速驱动的海表辐射响应非线性变化和样本在风速维度上的非均匀分布,导致难以刻画不同风速条件下的辐射响应特征,在全风速范围内泛化能力受限和存在系统性反演偏差的问题

Benefits of technology

(1)显著降低了高风速条件下的系统性偏差,提升了反演精度。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122508958A_ABST
    Figure CN122508958A_ABST
Patent Text Reader

Abstract

The application relates to the technical field of microwave remote sensing, and proposes an adaptive wet troposphere path delay inversion method and system fusing wind speed overlapping partition modeling, multi-scale collaborative sample balancing and model soft fusion. The method comprises the following steps: acquiring multi-source data and preprocessing the multi-source data to obtain to-be-inverted samples; according to a preset threshold, the full-range wind speed of the to-be-inverted samples is divided into a plurality of wind speed subintervals with overlapping transition zones; the to-be-inverted samples are input into trained inversion submodels, the output results of the inversion submodels are weighted and summed according to final fusion weights, and finally, wet troposphere path delay inversion results are obtained; when abnormal wind speed or invalid submodel prediction output is detected, the output result of a pre-trained global unified model is taken as the final inversion result. While guaranteeing overall inversion precision and stability, the application can significantly suppress systematic deviation under sea conditions of extremely low and extremely high wind speed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of microwave remote sensing technology, and in particular to an adaptive wet tropospheric path delay inversion method and system that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization and model soft fusion. Background Technology

[0002] Wet tropospheric path delay correction (WTC) is a key error source affecting the ranging accuracy of space-based Earth observation systems such as radar altimeters and synthetic aperture radars. Its inversion accuracy directly determines the reliability of sea-level change monitoring and ocean dynamic process analysis. In satellite altimetry missions, using a sub-satellite passive microwave radiometer (MWR) on the same platform for simultaneous observation is currently the most effective WTC technique. Currently, operational WTC inversion methods mainly employ logarithmic empirical formulas and neural networks to establish the mapping relationship between MWR-observed brightness temperature and WTC. However, in practical applications, existing methods generally suffer from decreased inversion accuracy and significant systematic biases under low and high wind speed sea conditions. The reasons for this can be mainly attributed to the following two aspects: First, the nonlinear variation of the sea surface microwave radiation response: the sea surface microwave radiation characteristics are significantly modulated by sea surface wind speed. As wind speed increases, the physical state of the sea surface undergoes an evolution from capillary waves to gravity wave breakup, and then to the formation of white foam, which significantly changes the sea surface emissivity and scattering characteristics, resulting in a significant nonlinear enhancement in the response of observed brightness temperature (WTC) to wind speed. Therefore, the mapping relationship between observed brightness temperature (WTC) and WTC is inherently wind speed-dependent and exhibits regional differences. Although the two existing mainstream inversion methods attempt to reduce inversion bias through empirical wind speed stratification or the introduction of auxiliary parameters, both have certain limitations. On the one hand, although the empirical wind speed stratification method uses a stratification strategy to handle nonlinear problems, its distributed architecture is prone to cascading errors in intermediate variables (such as the initial wind speed) to the final WTC inversion result. In addition, when dealing with discontinuities between adjacent strata, this method only performs simple linear weighting on the output (WTC value), lacking a physical mechanism to support it. On the other hand, neural network methods employ a globally unified regression strategy, attempting to fit observational data across the entire wind speed range with a single model, making it difficult to characterize the local differences and abrupt changes in sea surface radiation response across different wind speed ranges.

[0003] Secondly, there is an inherent imbalance in the distribution of observed samples: influenced by the natural distribution patterns of sea states, the observed samples exhibit a typical uneven distribution in the wind speed dimension: samples with medium to low wind speeds dominate, while samples with extremely low and high wind speeds are relatively scarce. For data-driven inversion models, this non-uniform distribution characteristic makes the model prone to being dominated by high-density samples during training, tending to fit samples in the medium to low wind speed range, while failing to adequately learn from the sparse samples in the extremely low and extremely high wind speed ranges, thus reducing the model's generalization ability under extreme wind speed conditions. Summary of the Invention

[0004] The purpose of this application is to overcome the problems of existing technologies, such as difficulty in characterizing the radiation response characteristics under different wind speed conditions, limited generalization ability across the entire wind speed range, and systematic inversion bias, caused by the nonlinear variation of sea surface radiation response driven by wind speed and the non-uniform distribution of samples in the wind speed dimension. This is addressed by providing an adaptive wet tropospheric path delay inversion method and system that integrates wind speed overlapping partition modeling, multi-scale collaborative sample equalization, and model soft fusion.

[0005] To address the aforementioned issues, this invention proposes an adaptive wet tropospheric path delay inversion method that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion, comprising: Acquire multi-source data and preprocess the multi-source data to obtain the sample to be inverted; The full-range wind speed of the sample to be inverted is divided into several wind speed sub-intervals with overlapping transition zones according to a preset threshold. The sample to be inverted is input into the trained inversion sub-model. The output results of each inversion sub-model are weighted and summed according to the final fusion weight of the sample in each wind speed range to obtain the final wet tropospheric path delay inversion result. When abnormal wind speeds are detected or the sub-model prediction output is invalid, the output of the pre-trained global unified model is used as the final wet tropospheric path delay inversion result.

[0006] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the training process of the inversion sub-model is as follows: Acquire multi-source data and preprocess the multi-source data to obtain preprocessed samples; The full-range wind speed of the preprocessed sample is divided into several wind speed sub-intervals with overlapping transition zones according to a preset threshold. A multi-scale collaborative sample equalization strategy is used to divide the wind speed sub-interval into bins, resulting in wind speed bins. The global weights and local rebalancing weights corresponding to the preprocessed samples are calculated and normalized. If the number of preprocessed samples in the wind speed sub-interval meets a preset threshold, the normalized local rebalancing weights are directly used to train the corresponding inversion sub-model. Otherwise, a backoff mechanism is triggered, and the global normalized weights corresponding to the sample are directly used for model training. The samples located in the overlapping transition region of the preprocessed samples also participate in the training of adjacent inversion sub-models; The inversion sub-model is a neural network model, including: an error backpropagation neural network and a recurrent neural network.

[0007] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the acquisition of multi-source data and the preprocessing of the multi-source data include: The multi-source data includes: microwave radiometer observation data and global reanalysis data; The observation data from the microwave radiometer include: observed brightness temperature, satellite latitude, time, rainfall markers, land-sea markers, and sea ice markers; The global reanalysis data includes: sea surface temperature, sea surface wind speed, land-sea markers, sea ice markers, and cloud liquid water content; The acquired observation data and auxiliary parameters are subjected to quality control and validity screening. Input feature vectors are constructed and the wet tropospheric path delay correction value is used as the reference ground value for supervised learning. The annual day accumulation and latitude features are encoded using sine and cosine functions. Finally, the input and output variables are normalized. The wet tropospheric path delay correction value includes: simulated wet path delay and wet tropospheric path delay measurements from other microwave radiometers.

[0008] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the plurality of wind speed sub-intervals with overlapping transition zones include three corresponding inversion sub-models constructed based on the wind speed sub-intervals, including: Medium wind speed baseline model; A low-wind-speed lightweight model, wherein the number of hidden layers and / or the number of hidden layer nodes in the low-wind-speed lightweight model is less than the number of hidden layers or the number of hidden layer nodes in the medium-wind-speed baseline model. A high-wind-speed enhancement model, wherein the number of hidden layers and / or the number of hidden layer nodes in the high-wind-speed enhancement model is greater than the number of hidden layers or the number of hidden layer nodes in the medium-wind-speed baseline model.

[0009] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the multi-scale collaborative sample equalization strategy is as follows: calculate the global weight of the preprocessed sample within the entire wind speed range, and calculate the corresponding local rebalancing weight within each wind speed sub-interval for the allocation of sample weights.

[0010] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the calculation of the global weights involves: performing binning statistics across the entire wind speed range according to a preset physical step size to construct a full wind speed histogram; then calculating the global weights using a power-law smoothing inverse frequency-weighted method with a global smoothing constant and a dynamic hard truncation mechanism; subsequently, the global weights are normalized to a mean of 1. The calculation formula is: ; In the formula, Indicates sample Statistical frequency within the corresponding wind speed compartment; The average density of sample frequencies within the non-air separation chamber across the entire wind speed range; This is the global scaling index. It is the global smoothing constant. This is the global dynamic hard truncation threshold.

[0011] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the calculation of the local rebalancing weights involves: re-statistically distributing the samples within each wind speed interval, and calculating the weights using a power-law smoothing inverse frequency-weighted method with a local smoothing constant and a dynamic hard truncation mechanism. Subsequently, the local rebalancing weights are normalized. The calculation formula is: ; In the formula, Indicates sample The sample frequency of each box within the corresponding local wind speed zone; The average frequency of effective bins; This is the local scaling index; It is a local smoothing constant; This is the local hard truncation threshold.

[0012] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the sample weight allocation method is as follows: when the number of samples in the wind speed range is greater than or equal to a preset threshold, the local rebalancing weight is used as the sample weight; when the number of samples in the wind speed range is less than the preset threshold, the global weight is used as the sample weight.

[0013] According to the adaptive wet tropospheric path delay inversion method proposed in this invention, the fusion weights of the sample to be inverted in each wind speed range include: The basic weights of each inversion sub-model are calculated using the trapezoidal membership function. The basic weights of each inversion sub-model are normalized to obtain the final fusion weights. These final fusion weights are used to fuse the outputs of each inversion sub-model according to the final fusion weights, generating the final wet tropospheric path delay inversion result. The calculation formula is: ; in, The total number of inversion sub-models fused according to the final fusion weights. For the inversion sample in the th The basic weights within each wind speed sub-interval are calculated using the following formula: ; in, The sea surface wind speed of the sample to be inverted. and Respectively characterize the first The lower and upper limits of the wind speed sub-intervals corresponding to each inversion sub-model. It represents the total bandwidth of the overlapping transition zone between adjacent wind speed sub-intervals.

[0014] This invention also proposes an adaptive wet tropospheric path delay inversion system that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion, for executing any of the methods described above, including: Data acquisition and preprocessing module: used to acquire multi-source data and preprocess the multi-source data to obtain preprocessed samples; Wind speed partitioning module: The preprocessed sample is divided into several wind speed sub-intervals with overlapping transition zones according to preset parameters / thresholds. Multi-scale collaborative sample equalization module: used to divide the wind speed into bins within the wind speed sub-interval, obtain wind speed bins, calculate the global weight of the preprocessed sample in the full wind speed range and the local rebalancing weight in the corresponding wind speed sub-interval, and perform normalization processing to generate the final sample weights for training the inversion sub-model. Inversion sub-model construction and training module: Constructs corresponding inversion sub-models based on the wind speed sub-intervals and trains each sub-model. When the preprocessed sample is within the overlapping transition region, the sample simultaneously participates in the training of adjacent inversion sub-models; and Model output fusion module: Input the sample to be inverted into each trained inversion sub-model, calculate the basic weight of each sub-model corresponding to the sample based on the wind speed information of the sample to be inverted, and dynamically weight and sum the output results of each sub-model based on the final fusion weight obtained after normalization of the basic weight of each sub-model to obtain the final wet tropospheric path delay inversion result; when abnormal wind speed is detected or the sub-model prediction output is invalid, the output result of the pre-trained global unified model is used as the final wet tropospheric path delay inversion result.

[0015] Compared with the prior art, the technical solution proposed in this application has at least the following beneficial effects: (1) It significantly reduced the systematic bias under high wind speed conditions and improved the inversion accuracy.

[0016] By constructing a wind speed zoning modeling mechanism with overlapping transition zones, the problem of insufficient fitting of the nonlinear relationship of radiation response by the global unified model under high wind speed conditions is avoided. This enables the model to characterize the differentiated responses of sea surface emissivity and scattering characteristics in different wind speed ranges, thereby suppressing systematic biases generated under high wind speed conditions and improving inversion accuracy.

[0017] (2) It enhances the generalization ability and robustness of the model under extreme sea conditions (such as extremely low and extremely high wind speeds).

[0018] By introducing a multi-scale collaborative sample equalization mechanism, the weight allocation of sparse samples within each sub-wind speed interval is further strengthened on the basis of global weights. This makes the model training process no longer dominated by high-density wind speed samples, improves the model's feature learning ability for sparse wind speed samples, and thus enhances the overall generalization ability and robustness.

[0019] (3) It ensures the continuity and stability of the inversion results across the entire wind speed range.

[0020] When constructing independent sub-models, samples located within the overlapping transition zone will simultaneously participate in the training and prediction processes of adjacent sub-models. This overlapping and sharing mechanism not only enables each sub-model to accurately capture local nonlinear characteristics within a specific wind speed range, but also encourages adjacent models to fully learn the physical evolution laws of the sea state transition phase, laying the foundation for subsequent smoothing. A soft fusion strategy based on trapezoidal membership functions was adopted to establish a smooth transition mechanism for the output results of adjacent wind speed ranges. This avoids the discontinuity problem in inversion results caused by traditional hard threshold partitioning at wind speed boundaries, ensuring the continuity and stability of inversion results across the entire wind speed range. Attached Figure Description

[0021] Figure 1 The diagram shows the overall framework and implementation process of the adaptive wet tropospheric path delay inversion method described in this invention. Figure 2 The diagram shown is a flowchart of the adaptive wet tropospheric path delay inversion system described in this invention. Figure 3 The diagram shows a trapezoidal membership weighting strategy based on wind speed overlap partitioning according to the present invention. Detailed Implementation

[0022] The technical solutions provided in this application are further illustrated below with reference to the embodiments.

[0023] Example 1 To address the shortcomings of existing wet tropospheric path delay inversion methods in extreme wind speeds and sea states, and their significant system biases, this application proposes an adaptive wet tropospheric path delay inversion method and system that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion. Its technical solution includes at least the following technical features: An adaptive wet tropospheric path delay inversion method integrating wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion includes the following steps: Step 1: Multi-source data acquisition and preprocessing: Step 1-1: Multi-source observation data includes, but is not limited to: brightness temperature observed by the microwave radiometer in multiple channels, sea surface wind speed data, and the corresponding true values ​​of wet path delay. ; Steps 1-2: Perform quality control and preprocessing on the data. Preprocessing includes, but is not limited to: outlier removal and feature normalization. Step 2: Wind speed zoning based on physical thresholds: Step 2-1: Based on the strong phased and highly nonlinear physical characteristics of the sea surface microwave radiation characteristics' response to wind speed, the wind speed samples are divided into zones, at least low-wind-speed, medium-wind-speed, and high-wind-speed zones. The wind speed zone thresholds are determined based on the phased changes in sea surface emissivity, scattering characteristics, and other features with wind speed.

[0024] Step 2-2: Set the overlap zone bandwidth in adjacent wind speed intervals as follows: The overlapping transition zone allows the same wind speed sample located within the transition zone to participate simultaneously in the training and prediction process of adjacent wind speed interval models.

[0025] As an improvement to this embodiment, when dividing wind speed intervals, in addition to dividing based on physical thresholds of wind speed, division can also be based on relevant physical parameters such as sea surface roughness, or data-driven automatic partitioning can be performed using clustering algorithms, and the specific number of partitions can be expanded according to actual needs.

[0026] Step 3: Multi-scale collaborative sample equalization strategy Step 3-1: Global Weight Construction: Within the entire wind speed range, perform binning statistics according to a preset physical step size to construct a global wind speed histogram; then calculate the global weight of the samples. Considering that the ideal reverse frequency can easily introduce anomalous weights and lead to gradient explosion in the extremely sparse region, a global smoothing constant is introduced. and dynamic hard truncation threshold For any observation sample Its global weight The calculation formula is as follows: ; In the formula, Indicates sample Statistical frequency within the corresponding wind speed compartment; This represents the average frequency of samples within the non-air separation chamber across the entire wind speed range. The global scaling exponent is used to avoid the overall scale of the loss function shifting after weighting. Therefore, the global weights of all samples need to be normalized to the mean, so that their expected value is 1.

[0027] Step 3-2, Local Sample Rebalancing within Wind Speed ​​Zones: After completing the wind speed zoning, for the sample subsets within each wind speed zone, a new local wind speed histogram is constructed based on the wind speed distribution within that zone, and the local rebalancing weights of the samples within that zone are calculated. The calculated weights are then locally mean-normalized to a mean of 1. The calculation of local rebalancing weights is consistent with the global weights, but the smoothing parameters, weight upper limit thresholds, and scaling exponents used can be adjusted based on the differences in sample distribution within the interval to avoid excessive amplification of local sample distributions. The calculation formula is: ; in, Indicates sample The frequency of the sub-box within the corresponding local wind speed zone; The average frequency of effective bins; This is the local scaling index; It is a local smoothing constant; This is the local hard truncation threshold.

[0028] Step 3-3, Adaptive Weight Selection: When the number of samples within a partition meets the preset threshold, the local rebalancing weights calculated in Step 3-2 are directly used. If the number of samples in a partition is insufficient, the global weight calculated in step 3-1 for that sample is directly inherited. To avoid statistical bias caused by insufficient sample size, the combined weights are used as the final sample weights for training the inversion sub-model within the corresponding wind speed range.

[0029] Step 4: Adaptive Sub-model Construction and Training Step 4-1: Sub-model Construction for Each Wind Speed ​​Zone: Construct corresponding wet path delay inversion sub-models for each wind speed zone. Each sub-model adopts a neural network structure, using the sub-model network structure of the medium wind speed zone as a benchmark. Adjust the network depth or the number of hidden layer nodes according to the sea state complexity within the corresponding wind speed zone (sea state complexity mainly includes changes in sea surface roughness, the proportion of breaking waves, and changes in foam coverage, etc.). Details are as follows: Medium wind speed benchmark model: Construct a medium-sized neural network adapted to normal sea conditions as a benchmark (for example, set the number of hidden layers to M layers and the number of nodes in each layer to N).

[0030] Lightweight model for low wind speed: Since the sea state complexity is low in the low wind speed range (the sea surface is approximately smooth and microwave radiation follows the specular reflection law), a lightweight network structure is adopted, that is, the number of hidden layers is less than that of the medium wind speed benchmark model (such as M-1 layers), or the number of hidden layer nodes is less than that of the benchmark model, in order to suppress the overfitting phenomenon that occurs in low-complexity samples under complex networks.

[0031] High wind speed enhancement model: Due to the high complexity of sea conditions in the high wind speed range (accompanied by drastic changes in sea surface roughness and high white wave coverage, resulting in significant changes in surface emissivity and scattering characteristics), a deeper or wider network structure is adopted, that is, the number of hidden layers is greater than that of the medium wind speed baseline model (such as M+1 or M+2 layers), or the number of hidden layer nodes is greater than that of the baseline model, in order to enhance the model's ability to represent the nonlinear enhancement of sea surface radiation characteristics.

[0032] Step 4-2, Weighted Training Mechanism: During the training of the sub-model, the final sample weights calculated in Step 3-3 are used to construct a weighted loss function, so that sparse samples can make a higher contribution during the parameter update process, thereby enhancing the model's ability to learn sparse sample features and improving the model's generalization ability.

[0033] In other embodiments, the current weighted loss function can be replaced with adaptive weight adjustment, dynamic loss function, data resampling, or other equivalent sample reweighting methods for sample equalization.

[0034] Step 5: Model output fusion based on membership function Step 5-1: Trapezoidal Membership Degree (Basic Weights) Calculation: Based on the sea surface wind speed corresponding to the sample to be inverted, a continuous membership function with overlapping transition bandwidth is introduced at the boundary of each wind speed zone. The trapezoidal membership degree of the sample to each inversion sub-model is calculated, which is the basic weight. The membership function is preferably a trapezoidal or piecewise linear function, taking a value of 1 inside the zone, continuously adjusted with wind speed changes in the transition zone, and taking a value of 0 outside the zone. The trapezoidal function is used as an example below: Assume the sea surface wind speed of the sample to be inverted is... , No. Each inversion sub-model corresponds to a wind speed sub-interval ( The lower and upper limits (corresponding to the sub-models for low, medium, and high wind speeds, respectively) are respectively and The total bandwidth of the overlapping transition zone between adjacent wind speed sub-intervals is set as follows: .

[0035] The sample to be inverted is in the first The basic weights within each wind speed sub-interval are calculated using the following formula: .

[0036] Step 5-2, Weighted Fusion and Backoff Mechanism: The outputs of each wind speed inversion sub-model are weighted and summed according to the final fusion weights to obtain the final wet path delay inversion result. When abnormal wind speeds are detected or the sub-model prediction output is invalid, the output of the pre-trained global unified model is used as the backup inversion result. The final fusion weights are obtained by normalizing the basic weights of each inversion sub-model. The calculation formula is: ; in, This represents the total number of sub-models participating in the fusion. In this embodiment, These correspond to low, medium, and high wind speed sub-models, respectively.

[0037] As an improvement to this embodiment, in terms of model fusion, the membership function can be replaced with other continuous weight functions or smooth transition strategies.

[0038] Example 2 To address the shortcomings of existing wet tropospheric path delay inversion methods in extreme wind speed and sea state conditions (extremely low or extremely high wind speeds) and their significant system biases, this application proposes an adaptive wet tropospheric path delay inversion system that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion. Its technical solution includes at least the following technical features: A wet tropospheric path delay inversion system includes: Data acquisition and preprocessing module: This module is used to acquire satellite observation data and corresponding auxiliary environmental data, and to preprocess the data. It is used to execute step 1 in Example 1.

[0039] Wind speed zoning module: This module is used to divide the sample into at least two overlapping wind speed sub-intervals according to the acquired sea surface wind speed parameters and preset wind speed zoning rules. This module is used to execute step 2 in Example 1.

[0040] The multi-scale collaborative sample equalization module is used to perform step 3 in Example 1.

[0041] The multi-scale collaborative sample equalization module includes: Global weight calculation unit: used to calculate the global sample weight of samples across the entire wind speed range; Local rebalancing weight calculation unit: Calculates the local sample weights of the samples within each wind speed sub-interval; Weight combination unit: Based on preset conditions, it combines global weights and local rebalancing weights to generate the final weights for model training.

[0042] The adaptive sub-model construction and training module is used to perform step 4 in Example 1.

[0043] The adaptive sub-model construction and training module includes: Sub-model building unit: used to build corresponding wet path delay inversion sub-models for different wind speed sub-intervals; Model training unit: The sub-model is trained using training samples and sample weights within the corresponding wind speed range.

[0044] The model output fusion module is used to perform step 5 in Example 1. Specifically, it includes: Membership calculation unit: used to calculate the model membership degree of the sample to be inverted in each wind speed sub-interval based on the wind speed information; Result fusion unit: The output results of multiple inversion sub-models are weighted and fused according to the membership weights to obtain the final wet path delay inversion result; Anomaly detection and fallback unit: When anomaly wind speed is detected or the sub-model prediction output is invalid, the output of the pre-trained global unified model is used as a backup inversion result.

[0045] Example 3 The invention will now be described below using a three-band spaceborne microwave radiometer (MWR) as an example, with reference to the accompanying drawings, from both the methodological and system perspectives. It is assumed that the MWR operates in an orbit at an altitude of 970 km and an inclination of 66°, with observation frequencies of 18.7 GHz, 23.8 GHz, and 37 GHz. The MWR data products are released on a 10-day cycle.

[0046] This embodiment provides an adaptive wet tropospheric path delay inversion method that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion. It is applicable to atmospheric wet path delay correction based on spaceborne microwave radiation. Figure 1 As shown, the main steps of the method include: multi-source data acquisition and preprocessing, wind speed zoning based on physical thresholds, multi-scale collaborative sample equalization, construction and training of adaptive sub-models, and model output fusion based on membership functions.

[0047] The specific implementation steps are as follows: Step 1: Multi-source data acquisition and preprocessing, including: (1) Quality control: Quality control and validity screening are carried out on the acquired MWR observation data and auxiliary parameters. Specifically, this includes: removing missing values, non-numerical (NaN / Inf) data, and data records that do not meet the requirements for product quality representation.

[0048] (2) Input and output feature construction: To characterize the variation characteristics of WTC under different air-sea conditions, the model constructs a multi-source joint input feature vector, including multi-channel observed brightness temperature, sea surface temperature, sea surface wind speed, satellite latitude, and day of year (DOY). The model output uses the WTC simulated by the numerical model as the ground truth value for supervised learning.

[0049] (3) Spatiotemporal feature periodic encoding: To avoid numerical jumps in the numerical representation of temporal and spatial features, this example uses sine and cosine functions to encode the DOY and latitude features, and their expressions are as follows: ; ; ; .

[0050] (4) Data normalization: Considering the differences in the units and numerical ranges of different input and output variables, the maximum-minimum normalization method is uniformly adopted for all input and output variables to map them to the [0,1] interval, so as to improve the numerical stability and convergence of the model training process.

[0051] Step 2: Wind speed zoning based on physical thresholds: The sea surface microwave radiation response exhibits significant stage-dependent and nonlinear characteristics in response to wind speed changes: under low wind speed conditions, the sea surface radiation response is predominantly characterized by near-specular reflection; as wind speed increases, the generation of tension waves, breaking waves, and foam causes significant changes in sea surface emissivity and scattering characteristics, resulting in a significant nonlinear enhancement in the response of observed brightness temperature (WTC) to wind speed. Therefore, the mapping relationship between observed brightness temperature (WTC) and WTC is inherently wind-speed dependent and varies significantly across different wind speed ranges.

[0052] Based on the above physical mechanism, this embodiment uses sea surface wind speed as the zoning basis for inversion modeling. At the two key wind speed thresholds of 6 m / s and 12 m / s, such as... Figure 3 As shown, an overlap buffer with a width of δ=2m / s is set to divide the samples into low-wind-speed, medium-wind-speed, and high-wind-speed intervals. The wind speed partitioning unit marks and classifies the samples according to the wind speed interval, which is used to construct the wet path delay inversion sub-models for the corresponding wind speed intervals and to provide a basis for model output fusion.

[0053] As an improvement to the above method, when dividing wind speed ranges, in addition to dividing based on wind speed thresholds, division can also be based on relevant physical parameters such as sea surface roughness, or data-driven automatic partitioning can be performed using clustering algorithms, and the specific number of partitions can be expanded according to actual needs.

[0054] Step 3: Multi-scale collaborative sample equalization: To mitigate the impact of uneven sample distribution on the inversion results, this embodiment employs a multi-scale collaborative sample balancing strategy consisting of global weights and local rebalancing weights.

[0055] It should be noted that the following sample weights only apply to the loss function weighting during the model training phase, and their physical meaning differs from the model membership weights used in the model output fusion phase.

[0056] (1) Global Weights: Wind speed samples are binned and statistically analyzed over the entire wind speed range with a physical step size of 0.5 m / s. A power-law smoothing inverse frequency method with a truncation mechanism is used to calculate the global weight of each sample. The global smoothing constant is... K global Set to 10, global scaling index Set to 0.5, and the global dynamic hard truncation threshold. Limit global Set it to 50. Then, normalize the weights of all global samples so that their mean is 1.

[0057] (2) Local rebalancing weights: The sample distribution is recalculated and the local sample weights are calculated within each wind speed sub-interval. The local smoothing constant is used as an example. K localSet to 5, local scaling index α local Set to 0.5, and the local dynamic hard truncation threshold. Limit local The value is set to 20. Subsequently, all local rebalancing weights are normalized so that their mean is kept at 1.

[0058] (3) Weight combination: When the number of samples in a partition is sufficient (≥1000), the local rebalancing weight is used as the weight of that sample; if the number of samples in a partition is insufficient (<1000), the global sample weight is used as the weight of that sample.

[0059] Step 4: Construction and training of adaptive sub-models A Back Propagation Neural Network (BPNN) is used as the basic framework for the inversion model, and the sub-model structure is adaptively configured according to the physical complexity of different wind speed ranges. Among them, the low wind speed sub-model adopts a lightweight structure (hidden layer nodes [16,8]) to prevent overfitting; the medium wind speed sub-model adopts a medium-sized structure ([24,12]); and the high wind speed sub-model adopts a deep structure ([32,16]) to enhance the characterization of complex nonlinearities.

[0060] During training, the weighted mean squared error (WMSE) is used as the loss function, which is defined as:

[0061] in, The number of samples; and The first The predicted value and the true value of each sample; These are the training weights corresponding to this sample. Then, the Levenberg-Marquardt algorithm is used to minimize the WMSE function to obtain the trained sub-model parameters.

[0062] Step 5: Model output fusion based on membership function In the prediction phase, based on the wind speed information of the sample to be inverted, the model weights in each wind speed sub-interval are calculated using the trapezoidal membership function, and the output results of each sub-model are weighted and fused to obtain continuous and smooth wet tropospheric path delay inversion results.

[0063] The core idea of ​​this strategy is as follows: within the non-overlapping core wind speed range corresponding to each sub-model, the model is strictly assigned a unit weight to maintain its dominant advantage; when the wind speed enters the overlapping transition zone, the linear change characteristics of the trapezoidal membership function are used to continuously and dynamically weight the outputs of adjacent sub-models, thereby achieving a smooth transition of the inversion results with wind speed changes. To ensure that weight conservation is satisfied during multi-model fusion, the basic weights of all participating sub-models need to be normalized. The obtained wet path delay inversion results are obtained by weighted summation of the outputs of each sub-model according to their corresponding normalized weights.

[0064] When abnormal wind speeds or invalid sub-model predictions are detected, the output of the global unified model is automatically used as the final inversion result.

[0065] In this embodiment, abnormal wind speed includes: no wind speed and wind speed outside the preset wind speed range. The preset wind speed range in this embodiment is 0-30m / s. In other embodiments, the preset wind speed range can be adjusted according to actual needs.

[0066] To evaluate the impact of sample equalization strategies and wind speed zoning models on WTC inversion performance, this study constructed three progressive comparative models: the Global Unweighted Model (GUM), the Global Weighted Model (GWM), and the Overlap Regime Weighted Model (ORWM). These three models maintain consistency in their basic network type, input / output feature spaces, and core training hyperparameters, differing only in their sample weighting methods and wind speed zoning models. GUM serves as the baseline model, without introducing sample weighting or wind speed zoning strategies. GWM, building upon GUM, introduces a global inverse frequency weighting mechanism based on power-law smoothing and dynamic hard truncation. ORWM, building upon GWM, combines wind speed-based overlap zoning modeling, local sample rebalancing, and a model soft fusion strategy. Tables 1 and 2 present the quantitative evaluation results of the three models and Jason-3AMR-2 wet path delay observations from June 2022 to May 2023.

[0067] Table 1: Quantitative Assessment Results at All Wind Speeds

[0068] Table 1 shows that the full wind speed assessment demonstrates that ORWM exhibits the best inversion performance, with its RMSE and STD reduced to 0.8823 cm and 0.8803 cm, respectively, Bias reduced to 0.0601 cm, and correlation coefficient increased to 0.9962. Compared to GUM and GWM, ORWM's RMSE decreased by 21.2% and 5.5%, respectively; STD decreased by 20.5% and 4.2%, respectively; and the absolute value of systematic bias decreased by 63.6% and 64.0%, respectively. This indicates that the ORWM model possesses extremely high inversion accuracy and stability in independent satellite cross-validation.

[0069] Table 2: Quantitative evaluation results under different wind speed ranges

[0070] Table 2 shows the robustness of ORWM further validated by the wind speed range assessment. Compared to GUM, ORWM reduced RMSE by 15.65%, 25.34%, and 18.09% in low, medium, and high wind speed ranges, respectively. Compared to GWM, ORWM reduced RMSE by 3.98%, 6.86%, and 4.89% in low, medium, and high wind speed ranges, respectively, achieving a comprehensive improvement in accuracy across all sea states. Regarding systematic bias, the bias decreased by 16.39% in the low wind speed range; the bias in the medium wind speed range was significantly optimized from 0.1184 cm to -0.0441 cm. The results indicate that ORWM has strong robustness and generalization ability across different wind speeds and sea states.

[0071] Example 4 The invention will now be described below using a three-band spaceborne microwave radiometer (MWR) as an example, with reference to the accompanying drawings, from both the methodological and system perspectives. It is assumed that the MWR operates in an orbit at an altitude of 970 km and an inclination of 66°, with observation frequencies of 18.7 GHz, 23.8 GHz, and 37 GHz. The MWR data products are released on a 10-day cycle.

[0072] This embodiment provides an adaptive wet tropospheric path delay inversion system that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion, to implement the wet tropospheric path delay inversion method of Embodiment 1 above. This system can be deployed in spaceborne data processing systems, ground-based operational processing systems, or offline scientific research processing platforms, etc.

[0073] I. System Overall Structure like Figure 2 As shown, the system includes: a data acquisition and preprocessing module, a wind speed zoning module, a multi-scale collaborative sample equalization module, an adaptive sub-model construction and training module, and a model output fusion module.

[0074] (1) Data acquisition and preprocessing module: used to acquire MWR observation data and auxiliary data, and to perform quality control, feature construction and normalization processing on them.

[0075] (2) Wind speed zoning module: Based on the preset physical threshold, the sea surface wind speed sample is divided into three sub-intervals with overlapping transition zones: low, medium and high.

[0076] (3) Multi-scale collaborative sample equalization module: used to calculate the global weight of the sample to be inverted and the local rebalancing weight in each wind speed sub-interval, and generate the final weight for model training.

[0077] (4) Adaptive sub-model construction and training module: used to construct and train the corresponding wet tropospheric path delay inversion sub-model for different wind speed sub-intervals.

[0078] (5) Model output fusion module: It is used to perform weighted fusion of the output structure of multiple sub-models according to the wind speed membership weight, and call the global model output result in abnormal situations.

[0079] II. System Operation Process During system operation, the data acquisition and preprocessing module first processes the input data; then the wind speed partitioning module completes the sample partitioning; next, the multi-scale collaborative sample equalization module generates the corresponding training weights; then the sub-model construction and training module completes the training of the corresponding sub-models; finally, the output fusion module generates the final wet tropospheric path delay inversion result.

[0080] In summary, this application addresses the issue of insufficient WTC inversion accuracy under extreme wind speeds and sea states (extremely low / extremely high wind speeds). It proposes a physically constrained data-driven inversion scheme from three aspects: wind speed physical zoning modeling, sample distribution optimization, and model output fusion. Its advantages are: (1) It breaks through the limitations of the global single model and constructs a physical perception wind speed partition modeling mechanism. Based on the strong stage and high nonlinearity of the response of sea surface microwave radiation characteristics to wind speed, it innovatively establishes a hierarchical strategy that includes overlapping transition zones. By independently training sub-models in low, medium and high wind speed ranges, it enhances the model's ability to characterize the differences in radiation response under different wind speed conditions.

[0081] (2) To address the inherent uneven distribution of observed data in the wind speed dimension, a multi-scale collaborative sample equalization strategy is proposed. Based on the global weights, this strategy further implements sample rebalancing within each wind speed zone to improve the model's ability to learn features from sparse samples within the zone, especially samples in the extremely low and high wind speed ranges.

[0082] (3) To avoid the discontinuity of inversion results caused by traditional hard threshold partitioning at the wind speed boundary, a soft fusion mechanism based on membership function is introduced to weight and integrate the output results of each wind speed sub-model, thereby ensuring the physical continuity and numerical stability of the inversion results across the entire wind speed range.

[0083] This invention constructs a wind speed zoning modeling framework with overlapping transition zones based on the variation law of sea surface radiative response with wind speed, and independently trains sub-models to enhance the ability to characterize the differences in sea surface radiative response characteristics under different wind speeds. A multi-scale sample equalization mechanism combining global weighting and local rebalancing is designed to improve the model's feature learning ability for sparse samples. A membership function-based model soft fusion strategy is used to dynamically weight the outputs of each sub-model to achieve continuity and smoothness of the inversion results. This invention can significantly suppress systematic biases under low and high wind speed sea states while ensuring overall inversion accuracy and stability.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. An adaptive wet tropospheric path delay inversion method that integrates wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion, comprising: Acquire multi-source data and preprocess the multi-source data to obtain the sample to be inverted; The full-range wind speed of the sample to be inverted is divided into several wind speed sub-intervals with overlapping transition zones according to a preset threshold. The sample to be inverted is input into the trained inversion sub-model. The output results of each inversion sub-model are weighted and summed according to the final fusion weight of the sample in each wind speed range to obtain the final wet tropospheric path delay inversion result. When abnormal wind speeds are detected or the sub-model prediction output is invalid, the output of the pre-trained global unified model is used as the final wet tropospheric path delay inversion result.

2. The adaptive wet tropospheric path delay inversion method according to claim 1, characterized in that, The training process of the aforementioned inversion sub-model is as follows: Acquire multi-source data and preprocess the multi-source data to obtain preprocessed samples; The full-range wind speed of the preprocessed sample is divided into several wind speed sub-intervals with overlapping transition zones according to a preset threshold. By employing a multi-scale collaborative sample equalization strategy, the wind speed within the wind speed sub-interval is binned to obtain wind speed bins. The global weight and local rebalancing weight corresponding to the preprocessed samples are calculated and normalized. If the number of preprocessed samples in the wind speed sub-interval meets the preset threshold, the corresponding inversion sub-model is directly trained using the normalized local rebalancing weights. Conversely, if the rollback mechanism is triggered, the model will be trained directly using the global normalized weights corresponding to that sample. The samples located in the overlapping transition region of the preprocessed samples also participate in the training of adjacent inversion sub-models; The inversion sub-model is a neural network model, including: an error backpropagation neural network and a recurrent neural network.

3. The adaptive wet tropospheric path delay inversion method according to claim 2, characterized in that, The acquisition of multi-source data and the preprocessing of the multi-source data include: The multi-source data includes: microwave radiometer observation data and global reanalysis data; The observation data from the microwave radiometer include: observed brightness temperature, satellite latitude, time, rainfall markers, land-sea markers, and sea ice markers; The global reanalysis data includes: sea surface temperature, sea surface wind speed, land-sea markers, sea ice markers, and cloud liquid water content; The acquired observation data and auxiliary parameters are subjected to quality control and validity screening. Input feature vectors are constructed and the wet tropospheric path delay correction value is used as the reference ground value for supervised learning. The annual day accumulation and latitude features are encoded using sine and cosine functions. Finally, the input and output variables are normalized. The wet tropospheric path delay correction value includes: simulated wet path delay and wet tropospheric path delay measurements from other microwave radiometers.

4. The adaptive wet tropospheric path delay inversion method according to claim 3, characterized in that, The aforementioned wind speed sub-intervals with overlapping transition zones include three corresponding inversion sub-models constructed based on the wind speed sub-intervals, including: Medium wind speed baseline model; A low-wind-speed lightweight model, wherein the number of hidden layers and / or the number of hidden layer nodes in the low-wind-speed lightweight model is less than the number of hidden layers or the number of hidden layer nodes in the medium-wind-speed baseline model. A high-wind-speed enhancement model, wherein the number of hidden layers and / or the number of hidden layer nodes in the high-wind-speed enhancement model is greater than the number of hidden layers or the number of hidden layer nodes in the medium-wind-speed baseline model.

5. The adaptive wet tropospheric path delay inversion method according to claim 2, characterized in that, The multi-scale collaborative sample equalization strategy is as follows: calculate the global weight of the preprocessed sample across the entire wind speed range, and calculate the corresponding local rebalancing weight within each wind speed sub-interval for the allocation of sample weights.

6. The adaptive wet tropospheric path delay inversion method according to claim 5, characterized in that, The calculation of the global weights is as follows: Within the entire wind speed range, box-based statistics are performed according to a preset physical step size to construct a full wind speed histogram. The histogram is then calculated using a power-law smoothing inverse frequency-weighted method with a global smoothing constant and a dynamic hard truncation mechanism. Subsequently, the global weights are normalized to a mean of 1. The calculation formula is: ; In the formula, Indicates sample Statistical frequency within the corresponding wind speed compartment; The average density of sample frequencies within the non-air separation chamber across the entire wind speed range; This is the global scaling index. It is the global smoothing constant. This is the global dynamic hard truncation threshold.

7. The adaptive wet tropospheric path delay inversion method according to claim 5, characterized in that, The calculation of the local rebalancing weights involves: re-statistically distributing the samples within each wind speed interval, and then calculating the weights using a power-law smoothing inverse frequency-weighted method with a local smoothing constant and a dynamic hard truncation mechanism. The local rebalancing weights are then normalized. The calculation formula is: ; In the formula, Indicates sample The sample frequency of each box within the corresponding local wind speed zone; The average frequency of effective bins; This is the local scaling index; It is a local smoothing constant; This is the local hard truncation threshold.

8. The adaptive wet tropospheric path delay inversion method according to claim 5, characterized in that, The sample weight allocation method is as follows: when the number of samples in the wind speed range is greater than or equal to the preset threshold, the local rebalancing weight is used as the sample weight; when the number of samples in the wind speed range is less than the preset threshold, the global weight is used as the sample weight.

9. The adaptive wet tropospheric path delay inversion method according to claim 1, characterized in that, The fusion weights of the samples to be inverted in each wind speed range include: The basic weights of each inversion sub-model are calculated using the trapezoidal membership function. The basic weights of each inversion sub-model are normalized to obtain the final fusion weights. These final fusion weights are used to fuse the outputs of each inversion sub-model according to the final fusion weights, generating the final wet tropospheric path delay inversion result. The calculation formula is: ; in, The total number of inversion sub-models fused according to the final fusion weights. For the inversion sample in the th The basic weights within each wind speed sub-interval are calculated using the following formula: ; in, The sea surface wind speed of the sample to be inverted. and Respectively characterize the first The lower and upper limits of the wind speed sub-intervals corresponding to each inversion sub-model. It represents the total bandwidth of the overlapping transition zone between adjacent wind speed sub-intervals.

10. An adaptive wet tropospheric path delay inversion system integrating wind speed overlap partitioning modeling, multi-scale collaborative sample equalization, and model soft fusion, used to execute the method described in any one of claims 1-9, characterized in that, include: Data acquisition and preprocessing module: used to acquire multi-source data and preprocess the multi-source data to obtain preprocessed samples; Wind speed partitioning module: The preprocessed sample is divided into several wind speed sub-intervals with overlapping transition zones according to preset parameters / thresholds. Multi-scale collaborative sample equalization module: used to divide the wind speed into bins within the wind speed sub-interval, obtain wind speed bins, calculate the global weight of the preprocessed sample in the full wind speed range and the local rebalancing weight in the corresponding wind speed sub-interval, and perform normalization processing to generate the final sample weights for training the inversion sub-model. Inversion sub-model construction and training module: Construct corresponding inversion sub-models based on the wind speed sub-intervals and train each sub-model. When the preprocessed sample is in the overlapping transition region, the sample will participate in the training of adjacent inversion sub-models at the same time. and Model output fusion module: Input the sample to be inverted into each trained inversion sub-model, calculate the basic weight of each sub-model corresponding to the sample based on the wind speed information of the sample to be inverted, and dynamically weight and sum the output results of each sub-model based on the final fusion weight obtained after normalization of the basic weight of each sub-model to obtain the final wet tropospheric path delay inversion result; when abnormal wind speed is detected or the sub-model prediction output is invalid, the output result of the pre-trained global unified model is used as the final wet tropospheric path delay inversion result.