Precipitation inversion method and system based on deep learning from satellite data

Abstract

The present invention discloses a deep learning-based precipitation inversion method and system for satellite data. The method comprises: using data acquired from satellite data in a preset area within a specified time as the data to be analyzed; determining a first precipitation area using the determined data, performing uncertainty estimation on the uncertain data to obtain a second precipitation area, performing rainfall scale prediction on the data to be analyzed within the first and second precipitation areas to obtain a precipitation scale probability; extracting precipitation scale features from the data to be analyzed to obtain precipitation scale characteristics, and performing feature fusion on the precipitation scale probability and the precipitation scale characteristics to obtain fused data; constructing a deep learning precipitation inversion model based on the attenuation rate using the fused data, outputting the data to be inverted to the deep learning precipitation inversion model, and outputting an inversion result. This method not only improves the accuracy of satellite data precipitation inversion but can also be directly applied to satellite data precipitation inversion systems.

Description

Technical Field

[0001] The present invention relates to the field of precipitation inversion, and in particular to a method and system for precipitation inversion using satellite data based on deep learning. Background Art

[0002] Precipitation is a core element of the global water cycle and climate change. High-precision, high-temporal and high-resolution precipitation monitoring is crucial for weather forecasting, disaster warning, agricultural planning, and water resources management. Traditional precipitation inversion methods rely primarily on ground-based meteorological stations and weather radar observations, but their coverage is limited, with significant data gaps particularly in oceans, plateaus, and remote areas. Satellite remote sensing technology, with its wide-area coverage and continuous observations, has become an important complement to ground-based observations. However, satellite precipitation inversion faces challenges such as insufficient multi-source data fusion, difficulty in quantifying uncertainty, and low accuracy in modeling complex precipitation processes.

[0003] Currently, mainstream satellite precipitation inversion methods can be divided into two categories: statistical inversion and physical models. Statistical methods rely on empirical relationships in historical data but are difficult to adapt to dynamic atmospheric conditions. While physical models have clear physical mechanisms, they are computationally complex and sensitive to initial fields and auxiliary data. In recent years, deep learning technology, with its powerful nonlinear modeling capabilities, has been gradually applied to the field of precipitation inversion. However, it still has the following limitations: it does not fully exploit the complementarity of multi-source satellite data; it lacks systematic uncertainty estimation for uncertain data in satellite observations, resulting in insufficient robustness of the model in complex scenarios; and existing deep learning models mostly adopt an end-to-end structure, ignoring the time decay characteristics of the precipitation process and the spatial correlation of multi-scale features, which affects the inversion accuracy of extreme precipitation events.

[0004] In addition, existing studies usually simply splice deterministic data and uncertain data into the model without designing a differentiated feature fusion mechanism based on their confidence differences, which may introduce noise interference. At the same time, the impact of environmental factors such as terrain undulation and surface type heterogeneity on the vertical structure of precipitation has not been effectively modeled, resulting in a significant increase in precipitation inversion errors in complex terrain areas. Therefore, there is an urgent need to develop a deep learning inversion method that integrates multi-source satellite data, integrates uncertainty estimation mechanisms, and adapts to the spatiotemporal evolution of precipitation, so as to improve the reliability and refinement of global-scale precipitation monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a precipitation inversion method from satellite data based on deep learning.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0007] The present invention comprises the following steps:

[0008] The data acquired from satellite data in a preset area within a specified time is used as the data to be analyzed; the satellite data includes satellite observation data, auxiliary data, and label data; the data to be analyzed includes certain data and uncertain data; the certainty of the certain data is higher than that of the uncertain data; the satellite observation data includes microwave radiation data, infrared data, multi-channel radiometer data, and radar joint observation data; the auxiliary data includes terrain elevation, surface type, and atmospheric reanalysis data;

[0009] Determine a first precipitation area using the determined data, perform uncertainty estimation on the uncertain data to obtain a second precipitation area, and perform rainfall scale prediction on the data to be analyzed within the first precipitation area and the second precipitation area to obtain a precipitation scale probability;

[0010] Extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features, and fusing the precipitation scale probability and the precipitation scale features to obtain fused data;

[0011] Based on the attenuation rate, a deep learning precipitation inversion model is constructed using the fused data, the data to be inverted is output to the deep learning precipitation inversion model, and an inversion result is output.

[0012] Furthermore, the method for determining the first precipitation area by using the determination data includes:

[0013] A neural network algorithm is used to determine the satellite images and their corresponding rainfall data in the data through learning, automatically extract rainfall features and establish a mapping relationship, predict future rainfall conditions based on the mapping relationship, and output the rainy area as the first precipitation area.

[0014] Furthermore, the method of performing uncertainty estimation on the uncertain data to obtain a second precipitation area includes:

[0015] The Bayesian neural network is used to estimate the posterior distribution of the weights of each layer of the neural network by treating the weights of each layer of the neural network as random variables, and the uncertainty estimate about the uncertain data is predicted according to the posterior distribution;

[0016] Spatial analysis is performed based on the uncertain data, and the uncertainty assessment value of the predicted rainfall area is obtained by combining the uncertain information; the rainfall area with an uncertainty assessment value lower than 0.152 is output as the second precipitation area.

[0017] Furthermore, the method for obtaining the precipitation scale probability includes:

[0018] A feature-enhanced hybrid model combines statistics and machine learning algorithms, uses advanced feature engineering to improve prediction accuracy, analyzes the data to be analyzed in the precipitation area, estimates the probability distribution of multiple rainfall scales, and outputs the probability distribution of multiple rainfall scales as precipitation scale probabilities.

[0019] Furthermore, the method of extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features includes:

[0020] The data to be analyzed in the first and second precipitation areas are input into the precipitation scale feature extractor based on multi-head attention, and the expression is:

[0021]

[0022] The output feature of the i-th encoder layer is The intermediate features after the multi-head attention layer are The multi-layer attention layer is MSA, the multi-layer perceptron layer is MLP, and the output feature of the i-1th encoder layer is The layer normalization is LN(·), the position encoding function is PosEnc(·), and the feature after position encoding of the i-th layer is

[0023] The output features are output as rainfall scale features.

[0024] Furthermore, the method of performing feature fusion on the precipitation scale probability and the precipitation scale feature to obtain fused data includes:

[0025] The precipitation scale probability and precipitation scale features are input into the multi-scale feature fusion device, which divides the precipitation scale features into a global label that aggregates the entire area and an embedded representation of the image block sequence;

[0026] The global mark is converted into a form combined with the detailed feature K through the multi-layer perceptron layer, and the sampling feature is introduced. The expression of feature fusion is:

[0027]

[0028] The multi-scale feature data is The fusion feature is The feature embedding from the backbone network is K, and the feature after reorganizing the global mark is z tk , the feature obtained after sampling by the factor is u, the multi-layer perceptron layer is MLP, the fusion module is Fusion(·), the sampling feature is u, and the preliminary fusion result is The probability of precipitation scale is p, and the characteristics of precipitation scale are c;

[0029] The classification results of the scale classification module are used to constrain the fusion process of precipitation features of different scales. The expression is:

[0030]

[0031] The precipitation scale predictor is Γ(·), and the precipitation scale feature extractor is The feature map after weighted sum is Specify the number of precipitation scales as n, the normalized exponential function as softmax(·), and the predicted precipitation scale of multi-scale feature data as The i-th precipitation scale feature of the multi-scale feature data is The second precipitation scale feature of the multi-scale feature data is The nth precipitation scale feature of the multi-scale feature number is The probability distribution of precipitation scale is The fusion precipitation feature mapping of the precipitation scale characteristics and precipitation scale probability distribution in the ath row and jth column is: The index of the xth precipitation scale in the ath row and jth column is The index of the xth precipitation scale in the probability distribution of the precipitation scale in the ath row and the jth column

[0032]

[0033] The loss function is given based on the temperature coefficient:

[0034]

[0035] The loss function is The cross entropy loss function is The temperature coefficient is The mean square error loss function is Output the fused precipitation feature map as fused data.

[0036] Furthermore, a method for constructing a deep learning precipitation inversion model based on the attenuation rate using the fused data includes:

[0037] Calculate the attenuation coefficient:

[0038]

[0039] The attenuation cross section is σ e , wavelength is ρ, particle diameter is D, and mass-weighted average diameter is D m , the normalized intercept is N w , the particle number concentration is N, the particle number concentration of the particle diameter is N(D), the gamma distribution is β(·), the temperature is T, the particle diameter D and the mass-weighted average diameter D m The particle size distribution function is f(D,Dm ), the shape parameter of the precipitation particle size distribution is γ;

[0040] The attenuation coefficient is corrected based on the on-site monitoring precipitation data, and the expression is:

[0041]

[0042] The corrected normalized intercept is The vertical height from the curvature of the earth is h, and the correction factor for the height h is The actual precipitation is y, and the predicted precipitation is The temperature difference at the beginning of rainfall is ΔT, the humidity difference at the beginning of rainfall is ΔS, the constants are r1, r2, r3, the adjustment factor is ω, and the corrected attenuation coefficient is

[0043] Calculate rainfall:

[0044]

[0045] The rate of precipitation decrease is V(D), and the amount of precipitation is Q;

[0046] Based on the deep learning precipitation inversion model of attenuation coefficient, mass-weighted mean diameter and rainfall, the adjustment factor and mass-weighted mean diameter are obtained to invert the corrected normalized intercept, and the rainfall is inverted according to the inverted normalized intercept to output the precipitation inversion results.

[0047] The second aspect is the deep learning-based satellite data precipitation inversion system, which includes:

[0048] Data acquisition module: used to use the data acquired from satellite data in a preset area within a specified time as data to be analyzed; the satellite data includes satellite observation data, auxiliary data and label data; the data to be analyzed includes certain data and uncertain data; the certainty of the certain data is higher than that of the uncertain data;

[0049] Regional probability module: used to determine a first precipitation area based on the determined data, perform uncertainty estimation on the uncertain data to obtain a second precipitation area, and perform rainfall scale prediction on the data to be analyzed in the first precipitation area and the second precipitation area to obtain a precipitation scale probability;

[0050] Feature fusion module: used for extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features, and performing feature fusion on the precipitation scale probability and the precipitation scale features to obtain fused data;

[0051] Modeling output module: used to construct a deep learning precipitation inversion model based on the attenuation rate using the fusion data, output the data to be inverted to the deep learning precipitation inversion model, and output the inversion result.

[0052] The beneficial effects of the present invention are:

[0053] The present invention is a method and system for inverting precipitation from satellite data based on deep learning. Compared with the existing technology, the present invention has the following technical effects:

[0054] The present invention can improve the accuracy of satellite data precipitation inversion by determining a first precipitation area, obtaining a second precipitation area, obtaining precipitation scale probability, obtaining precipitation scale characteristics, obtaining fusion data and model construction steps, thereby improving the precision of satellite data precipitation inversion. The satellite data precipitation inversion is optimized, which can greatly save resources and improve work efficiency. It can realize intelligent precipitation inversion of satellite data, partition and fuse data of satellite data precipitation inversion in real time, which is of great significance to satellite data precipitation inversion, can adapt to satellite data precipitation inversion of different standards and different satellite data precipitation inversion requirements, and has certain universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flowchart of the steps of the precipitation inversion method based on satellite data based on deep learning of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below through specific examples. The illustrative examples and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0057] The satellite data precipitation inversion method and system based on deep learning of the present invention include the following steps:

[0058] like Figure 1 As shown, in this embodiment, the following steps are included:

[0059] The data acquired from satellite data in a preset area within a specified time is used as the data to be analyzed; the satellite data includes satellite observation data, auxiliary data, and label data; the data to be analyzed includes certain data and uncertain data; the certainty of the certain data is higher than that of the uncertain data; the satellite observation data includes microwave radiation data, infrared data, multi-channel radiometer data, and radar joint observation data; the auxiliary data includes terrain elevation, surface type, and atmospheric reanalysis data;

[0060] Determine a first precipitation area using the determined data, perform uncertainty estimation on the uncertain data to obtain a second precipitation area, and perform rainfall scale prediction on the data to be analyzed within the first precipitation area and the second precipitation area to obtain a precipitation scale probability;

[0061] Extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features, and fusing the precipitation scale probability and the precipitation scale features to obtain fused data;

[0062] Based on the attenuation rate, a deep learning precipitation inversion model is constructed using the fused data, the data to be inverted is output to the deep learning precipitation inversion model, and an inversion result is output.

[0063] In this embodiment, the method for determining the first precipitation area using the determined data includes:

[0064] A neural network algorithm is used to determine the satellite images and their corresponding rainfall data in the data through learning, automatically extract rainfall features and establish a mapping relationship, predict future rainfall conditions based on the mapping relationship, and output the rainy area as the first precipitation area.

[0065] In this embodiment, the method for performing uncertainty estimation on the uncertain data to obtain the second precipitation area includes:

[0066] The Bayesian neural network is used to estimate the posterior distribution of the weights of each layer of the neural network by treating the weights of each layer of the neural network as random variables, and the uncertainty estimate about the uncertain data is predicted according to the posterior distribution;

[0067] Spatial analysis is performed based on the uncertain data, and the uncertainty assessment value of the predicted rainfall area is obtained by combining the uncertain information; the rainfall area with an uncertainty assessment value lower than 0.152 is output as the second precipitation area.

[0068] In this embodiment, the method for obtaining the precipitation scale probability includes:

[0069] A feature-enhanced hybrid model combines statistics and machine learning algorithms, uses advanced feature engineering to improve prediction accuracy, analyzes the data to be analyzed in the precipitation area, estimates the probability distribution of multiple rainfall scales, and outputs the probability distribution of multiple rainfall scales as precipitation scale probabilities.

[0070] In this embodiment, the method of extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features includes:

[0071] The data to be analyzed in the first and second precipitation areas are input into the precipitation scale feature extractor based on multi-head attention, and the expression is:

[0072]

[0073] The output feature of the i-th encoder layer is The intermediate features after the multi-head attention layer are The multi-layer attention layer is MSA, the multi-layer perceptron layer is MLP, and the output feature of the i-1th encoder layer is The layer normalization is LN(·), the position encoding function is PosEnc(·), and the feature after position encoding of the i-th layer is

[0074] The output features are output as rainfall scale features.

[0075] In this embodiment, the method of performing feature fusion on the precipitation scale probability and the precipitation scale feature to obtain fused data includes:

[0076] The precipitation scale probability and precipitation scale features are input into the multi-scale feature fusion device, which divides the precipitation scale features into a global label that aggregates the entire area and an embedded representation of the image block sequence;

[0077] The global mark is converted into a form combined with the detailed feature K through the multi-layer perceptron layer, and the sampling feature is introduced. The expression of feature fusion is:

[0078]

[0079] The multi-scale feature data is The fusion feature is The feature embedding from the backbone network is K, and the feature after reorganizing the global mark is z tk , the feature obtained after sampling by the factor is u, the multi-layer perceptron layer is MLP, the fusion module is Fusion(·), the sampling feature is u, and the preliminary fusion result is The probability of precipitation scale is p, and the characteristics of precipitation scale are c;

[0080] The classification results of the scale classification module are used to constrain the fusion process of precipitation features of different scales. The expression is:

[0081]

[0082]

[0083] The precipitation scale predictor is Γ(·), and the precipitation scale feature extractor is The feature map after weighted sum is Specify the number of precipitation scales as n, the normalized exponential function as softmax(·), and the predicted precipitation scale of multi-scale feature data as The i-th precipitation scale feature of the multi-scale feature data is The second precipitation scale feature of the multi-scale feature data is The nth precipitation scale feature of the multi-scale feature number is The probability distribution of precipitation scale is The fusion precipitation feature mapping of the precipitation scale characteristics and precipitation scale probability distribution in the ath row and jth column is: The index of the xth precipitation scale in the ath row and jth column is The index of the xth precipitation scale in the probability distribution of the precipitation scale in the ath row and the jth column

[0084]

[0085] The loss function is given by the temperature coefficient:

[0086]

[0087] The loss function is The cross entropy loss function is The temperature coefficient is The mean square error loss function is Output the fused precipitation feature map as fused data.

[0088] In this embodiment, the method for constructing a deep learning precipitation inversion model based on the attenuation rate using the fused data includes:

[0089] Calculate the attenuation coefficient:

[0090]

[0091] The attenuation cross section is σ e , wavelength is ρ, particle diameter is D, and mass-weighted average diameter is D m , the normalized intercept is N w , the particle number concentration is N, the particle number concentration of the particle diameter is N(D), the gamma distribution is β(·), the temperature is T, the particle diameter D and the mass-weighted average diameter D m The particle size distribution function is f(D,D m ), the shape parameter of the precipitation particle size distribution is γ;

[0092] The attenuation coefficient is corrected based on the on-site monitoring precipitation data, and the expression is:

[0093]

[0094] The corrected normalized intercept is The vertical height from the curvature of the earth is h, and the correction factor for the height h is The actual precipitation is y, and the predicted precipitation is The temperature difference at the beginning of rainfall is ΔT, the humidity difference at the beginning of rainfall is ΔS, the constants are r1, r2, r3, the adjustment factor is ω, and the corrected attenuation coefficient is

[0095] Calculate rainfall:

[0096]

[0097] The rate of precipitation decrease is V(D), and the amount of precipitation is Q;

[0098] Based on the deep learning precipitation inversion model of attenuation coefficient, mass-weighted mean diameter and rainfall, the adjustment factor and mass-weighted mean diameter are obtained to invert the corrected normalized intercept, and the rainfall is inverted according to the inverted normalized intercept to output the precipitation inversion results.

[0099] The second aspect is the deep learning-based satellite data precipitation inversion system, which includes:

[0100] Data acquisition module: used to use the data acquired from satellite data in a preset area within a specified time as data to be analyzed; the satellite data includes satellite observation data, auxiliary data and label data; the data to be analyzed includes certain data and uncertain data; the certainty of the certain data is higher than that of the uncertain data;

[0101] Regional probability module: used to determine a first precipitation area based on the determined data, perform uncertainty estimation on the uncertain data to obtain a second precipitation area, and perform rainfall scale prediction on the data to be analyzed in the first precipitation area and the second precipitation area to obtain a precipitation scale probability;

[0102] Feature fusion module: used for extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features, and performing feature fusion on the precipitation scale probability and the precipitation scale features to obtain fused data;

[0103] Modeling output module: used to construct a deep learning precipitation inversion model based on the attenuation rate using the fusion data, output the data to be inverted to the deep learning precipitation inversion model, and output the inversion result.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based precipitation inversion method based on satellite data, characterized by: The following steps are involved: The data acquired from satellite data in a preset area within a specified time is used as the data to be analyzed; the satellite data includes satellite observation data, auxiliary data, and label data; the data to be analyzed includes certain data and uncertain data; the certainty of the certain data is higher than that of the uncertain data; the satellite observation data includes microwave radiation data, infrared data, multi-channel radiometer data, and radar joint observation data; the auxiliary data includes terrain elevation, surface type, and atmospheric reanalysis data; Determine a first precipitation area using the determined data, perform uncertainty estimation on the uncertain data to obtain a second precipitation area, and perform rainfall scale prediction on the data to be analyzed within the first precipitation area and the second precipitation area to obtain a precipitation scale probability; Extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features, and fusing the precipitation scale probability and the precipitation scale features to obtain fused data; Based on the attenuation rate, the fused data is used to construct a deep learning precipitation inversion model, the data to be inverted is output to the deep learning precipitation inversion model, and the inversion result is output; including: Calculate the attenuation coefficient: The attenuation cross section is , the wavelength is , the particle diameter is D, and the mass-weighted average diameter is , the normalized intercept is , the particle number concentration is N, and the particle number concentration of the particle diameter is , the gamma distribution is , temperature T, particle diameter D and mass-weighted average diameter The particle size distribution function is , the shape parameter of the precipitation particle size distribution is ; The attenuation coefficient is corrected based on the on-site monitoring precipitation data, and the expression is: The corrected normalized intercept is , the vertical height from the earth's curvature is h, and the correction factor for height h is , the actual precipitation is y, and the predicted precipitation is , the temperature difference of the ground at the beginning of rainfall is , the humidity difference of the ground at the beginning of rainfall is , the constants are 、 、 , the adjustment factor is , the corrected attenuation coefficient is ; Calculate rainfall: The rate of precipitation decrease is , precipitation is Q; Based on the deep learning precipitation inversion model of attenuation coefficient, mass-weighted mean diameter and rainfall, the adjustment factor and mass-weighted mean diameter are obtained to invert the corrected normalized intercept, and the rainfall is inverted according to the inverted normalized intercept to output the precipitation inversion results.

2. The method for inverting precipitation from satellite data based on deep learning according to claim 1, characterized in that: The method for determining the first precipitation area by using the determination data includes: A neural network algorithm is used to determine the satellite images and their corresponding rainfall data in the data through learning, automatically extract rainfall features and establish a mapping relationship, predict future rainfall conditions based on the mapping relationship, and output the rainy area as the first precipitation area.

3. The method for inverting precipitation from satellite data based on deep learning according to claim 1, characterized in that: The method for performing uncertainty estimation on the uncertain data to obtain a second precipitation area includes: The Bayesian neural network is used to estimate the posterior distribution of the weights of each layer of the neural network by treating the weights of each layer of the neural network as random variables, and the uncertainty estimate about the uncertain data is predicted according to the posterior distribution; Spatial analysis is performed based on the uncertain data, and the uncertainty assessment value of the predicted rainfall area is obtained by combining the uncertain information; the rainfall area with an uncertainty assessment value lower than 0.152 is output as the second precipitation area.

4. The method for inverting precipitation from satellite data based on deep learning according to claim 1, characterized in that: The method for obtaining the precipitation scale probability includes: A feature-enhanced hybrid model combines statistics and machine learning algorithms, uses advanced feature engineering to improve prediction accuracy, analyzes the data to be analyzed in the precipitation area, estimates the probability distribution of multiple rainfall scales, and outputs the probability distribution of multiple rainfall scales as precipitation scale probabilities.

5. The method for inverting precipitation from satellite data based on deep learning according to claim 1, characterized in that: The method for extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features includes: The data to be analyzed in the first and second precipitation areas are input into the precipitation scale feature extractor based on multi-head attention, and the expression is: The output feature of the i-th encoder layer is , the intermediate features after the multi-head attention layer are , the multi-layer attention layer is , the multi-layer perceptron layer is , the output feature of the i-1th encoder layer is , the layer is normalized to , the position encoding function is , the feature after position encoding at layer i is ; The output features are output as rainfall scale features.

6. The method for inverting precipitation from satellite data based on deep learning according to claim 1, characterized in that: The method for performing feature fusion on the precipitation scale probability and the precipitation scale feature to obtain fused data includes: The precipitation scale probability and precipitation scale features are input into the multi-scale feature fusion device, which divides the precipitation scale features into a global label that aggregates the entire area and an embedded representation of the image block sequence; The global mark is converted into a form combined with the detailed feature K through the multi-layer perceptron layer, and the sampling feature is introduced. The expression of feature fusion is: The multi-scale feature data is , the feature embedding from the backbone network is K, and the feature after reorganizing the global mark is , the feature obtained after sampling by the factor is u, and the multi-layer perceptron layer is , the fusion module is , the sampling characteristics are The preliminary fusion result is , the probability of precipitation scale is p, and the characteristics of precipitation scale is c; The classification results of the scale classification module are used to constrain the fusion process of precipitation features of different scales. The expression is: The precipitation scale predictor is , the precipitation scale feature extractor is , the feature map after weighted sum is , specify the number of precipitation scales as n, and the normalized exponential function is , the predicted precipitation scale of multi-scale feature data is , the i-th precipitation scale feature of the multi-scale feature data is , the second precipitation scale feature of the multi-scale feature data is , the nth precipitation scale feature of the multi-scale feature number is , the probability distribution of precipitation scale is The fusion precipitation feature mapping of the precipitation scale characteristics in the ath row and the jth column and the precipitation scale probability distribution is: , the index of the xth precipitation scale in the ath row and jth column is , the index of the xth precipitation scale in the probability distribution of precipitation scale in row a and column j ; The loss function is given by the temperature coefficient: The loss function is , the cross entropy loss function is , the temperature coefficient is , the mean square error loss function is , and output the fused precipitation feature map as fused data.

7. A deep learning-based precipitation inversion system based on satellite data, used to execute the method according to any one of claims 1 to 6, characterized in that: include: Data acquisition module: used to use the data acquired from satellite data in a preset area within a specified time as data to be analyzed; the satellite data includes satellite observation data, auxiliary data and label data; the data to be analyzed includes certain data and uncertain data; the certainty of the certain data is higher than that of the uncertain data; Regional probability module: used to determine a first precipitation area based on the determined data, perform uncertainty estimation on the uncertain data to obtain a second precipitation area, and perform rainfall scale prediction on the data to be analyzed in the first precipitation area and the second precipitation area to obtain a precipitation scale probability; Feature fusion module: used for extracting precipitation scale features from the data to be analyzed to obtain precipitation scale features, and performing feature fusion on the precipitation scale probability and the precipitation scale features to obtain fused data; Modeling output module: used to construct a deep learning precipitation inversion model based on the attenuation rate using the fusion data, output the data to be inverted to the deep learning precipitation inversion model, and output the inversion result.

Images