CN-122021982-A - Method for predicting precision of remote sensing precipitation data of site-free area

CN122021982ACN 122021982 ACN122021982 ACN 122021982ACN-122021982-A

Abstract

The application belongs to the technical field of meteorological hydrology, and particularly discloses a method for predicting the precision of remote sensing precipitation data of a site-free area, which comprises the following steps of acquiring and preprocessing multi-source satellite remote sensing precipitation data and corresponding environment auxiliary variable data; the method comprises the steps of carrying out precision evaluation on multi-source satellite remote sensing precipitation data based on a multiplication triple collocation analysis method, an extension double-station variable method and a verification method based on actual measurement sites to obtain corresponding precision evaluation indexes, constructing a deep neural network model, carrying out model training and cross verification by taking the precision evaluation indexes and environment auxiliary variable data as input features, revealing feature contribution by utilizing an interpretability analysis method to obtain a trained deep neural network model, carrying out precision prediction on satellite remote sensing precipitation data of a site-free area by using the trained deep neural network model, and outputting a precision prediction result. The method can realize reliable prediction of the precision of remote sensing precipitation data in the site-free area.

Inventors

XU LEI
Hong Youting

Assignees

中国地质大学（武汉）

Dates

Publication Date: 20260512
Application Date: 20251023

Claims (10)

1. The method for predicting the precision of remote sensing precipitation data of the site-free area is characterized by comprising the following steps of: S10, acquiring multi-source satellite remote sensing precipitation data and corresponding environment auxiliary variable data, and preprocessing the multi-source satellite remote sensing precipitation data and the environment auxiliary variable data to ensure that all data have a uniform space-time range, space-time resolution and projection coordinate system; S20, performing precision evaluation on multi-source satellite remote sensing precipitation data based on a multiplication triple collocation analysis method, an extension double tool variable method and a verification method based on an actual measurement site to obtain corresponding precision evaluation indexes, wherein the precision evaluation indexes comprise correlation coefficients of the remote sensing precipitation data and real precipitation data; S30, constructing a deep neural network model, performing model training and cross verification by taking the precision evaluation index and the environment auxiliary variable data as input features, and revealing feature contribution by using an interpretability analysis method to obtain a trained deep neural network model; s40, accurately predicting satellite remote sensing precipitation data of the site-free area by using the trained deep neural network model, and outputting an accuracy prediction result.
2. The method for predicting the accuracy of remote sensing precipitation data in a site-free area according to claim 1, wherein in step S10, the step of acquiring multisource satellite remote sensing precipitation data and corresponding environment auxiliary variable data further comprises acquiring actual site observation precipitation data, analyzing precipitation data and precipitation data based on soil humidity inversion; the auxiliary environment variable data comprise a digital elevation model, a drought index, a leaf area index, soil texture, land utilization data and climate zone classification data; The preprocessing includes time alignment, spatial resampling and projection alignment.
3. The method for predicting precision of remote sensing precipitation data in no-site area as claimed in claim 1, wherein in step S20, correlation coefficients are obtained by a multiplicative triple collocation analysis method The calculation formula of (2) is as follows: In the formula, Three data sets are represented that satisfy the triple collocation analysis assumption; Represented as a dataset And true value Correlation coefficients between; representing covariance between the data.
4. The method for predicting the accuracy of remote sensing precipitation data in a site-free area according to claim 1, wherein in step S20, the correlation coefficient is calculated by the extended duplex variable method The calculation formula of (2) is as follows: In the formula, An error term representing the data set X, Representing variance, and P representing true precipitation value.
5. The method for predicting precision of remote sensing precipitation data in no-site area as claimed in claim 1, wherein in step S20, correlation coefficient is calculated in verification method based on measured site The calculation formula of (2) is as follows: In the formula, Representing the logarithm of the sample; indicating the measured precipitation; Indicating the remote sensing precipitation.
6. The method for predicting the precision of remote sensing precipitation data in a site-free area according to claim 1, wherein in step S30, the deep neural network model adopts a fully connected structure, and comprises an input layer, an intermediate layer and an output layer, wherein the input layer node corresponds to a multi-source feature, the intermediate layer comprises a hidden layer and adds a attention mechanism, and the output layer corresponds to a precision prediction result; And the training of the deep neural network model adopts an Adam optimizer, and reduces the risk of overfitting by utilizing five-fold cross validation.
7. The method for predicting the accuracy of site-free regional remote sensing precipitation data of claim 1, wherein a SHAP method is added to the deep neural network model to perform an interpretable analysis on the results, and Shapley values of each characteristic variable are calculated, wherein the Shapley values represent contributions of the characteristic variables to the model output result.
8. The method for predicting the accuracy of remote sensing precipitation data in a site-free area of claim 7, wherein the SHapley values are calculated according to the formula: In the formula, Representing characteristics Is a Shapley value of (2); representing the number of features; Is to remove A set of all features except; Representation of Is a feature set of the group; a function representing a predictive model; the model prediction output is expressed as the sum of a reference value and each characteristic contribution, and the calculation formula is as follows: In the formula, Wherein 1 indicates that the corresponding feature is selected in the subset; Average predicted values representing all observed values; Representing model predictors.
9. The method for predicting the precision of remote sensing precipitation data in a site-free area according to claim 1, wherein in step S40, a dense precipitation site area is selected, and a 0.25 ° large grid area thereof is divided into small grids of 1km×1km, and a deep neural network model is applied to obtain a small grid precision prediction result; And then, a spatial precision distribution map is generated by using a prediction result, so that visual analysis and visual evaluation of the precision of remote sensing precipitation data in the site-free area are realized.
10. A system for implementing the method of predicting site-free area remote sensing precipitation data accuracy of any one of claims 1-9, comprising: The data acquisition and preprocessing unit is used for acquiring multi-source satellite remote sensing precipitation data and corresponding environment auxiliary variable data, and preprocessing the multi-source satellite remote sensing precipitation data and the environment auxiliary variable data to ensure that all data have a uniform space-time range, space-time resolution and projection coordinate system; The precision evaluation unit is used for performing precision evaluation on the multi-source satellite remote sensing precipitation data based on a multiplication triple collocation analysis method, an extension duplex tool variable method and a verification method based on an actual measurement site to obtain corresponding precision evaluation indexes, wherein the precision evaluation indexes comprise correlation coefficients of the remote sensing precipitation data and real precipitation data; The model training unit is used for constructing a deep neural network model, taking the precision evaluation index and the environment auxiliary variable data as input features, performing model training and cross verification, and revealing feature contribution by using an interpretability analysis method so as to obtain a trained deep neural network model; the prediction unit is used for accurately predicting satellite remote sensing precipitation data of the site-free area by using the trained deep neural network model and outputting an accuracy prediction result.

Description

Method for predicting precision of remote sensing precipitation data of site-free area Technical Field The application belongs to the technical field of meteorological hydrology, and particularly relates to a method for predicting the precision of remote sensing precipitation data of a site-free area. Background Precipitation is a key element in meteorological and hydrologic processes as a key element of water circulation. Precipitation has a high degree of variability and complexity in its spatial-temporal distribution, making its precise capture a significant challenge. The accuracy of accurately estimating satellite remote sensing precipitation is important for knowing regional precipitation, preventing flood drought disasters and supporting ecological hydrologic applications. At present, the method for evaluating the satellite remote sensing precipitation precision is mainly divided into a point scale method and a grid scale, wherein the point scale method is used for calculating a statistical index by taking site observation as a reference, but is affected by site space distribution unevenness and is easy to generate a representative error, the grid scale method is used for evaluating the precision under the condition of no real value by integrating a plurality of precipitation data sets, however, the consistency of evaluation results of different methods under different geographic areas and environmental conditions is poor, and the system comparison and fusion of multisource evaluation results are not available, so that the comprehensiveness and reliability of precision evaluation are limited. In recent years, machine learning techniques have been widely used for satellite precipitation error modeling by introducing environmental variables such as terrain, vegetation index, land utilization, etc. as input features, and predicting the error spatial distribution using various algorithms. However, the existing method still has the obvious defects that the evaluation coverage of global multisource precipitation data is insufficient, the precision difference analysis between the point scale and grid scale method is insufficient, various precision evaluation results and environmental factors cannot be effectively integrated into machine learning input, and the model has weak interpretation, so that the prediction accuracy and the practicability in a site-free area are low. Therefore, how to integrate multiple precision evaluation methods and integrate multisource environment data, and construct an interpretable machine learning model to realize reliable prediction of the precision of remote sensing precipitation data in a site-free area is a current urgent problem to be solved. Disclosure of Invention Aiming at the defects of the prior art, the application aims to provide a method for predicting the precision of remote sensing precipitation data of a non-site area, which can realize the reliable prediction of the precision of remote sensing precipitation data of the non-site area. To achieve the above object, in a first aspect, the present application provides a method for predicting precision of remote sensing precipitation data of a site-free area, comprising the steps of: S10, acquiring multi-source satellite remote sensing precipitation data and corresponding environment auxiliary variable data, and preprocessing the multi-source satellite remote sensing precipitation data and the environment auxiliary variable data to ensure that all data have a uniform space-time range, space-time resolution and projection coordinate system; S20, performing precision evaluation on multi-source satellite remote sensing precipitation data based on a multiplication triple collocation analysis method, an extension double tool variable method and a verification method based on an actual measurement site to obtain corresponding precision evaluation indexes, wherein the precision evaluation indexes comprise correlation coefficients of the remote sensing precipitation data and real precipitation data; S30, constructing a deep neural network model, performing model training and cross verification by taking the precision evaluation index and the environment auxiliary variable data as input features, and revealing feature contribution by using an interpretability analysis method to obtain a trained deep neural network model; s40, accurately predicting satellite remote sensing precipitation data of the site-free area by using the trained deep neural network model, and outputting an accuracy prediction result. The method for predicting the precision of the remote sensing precipitation data in the website-free area has the advantages that comprehensive precision evaluation is carried out by combining a multiplication triple collocation analysis method, an extension duplex tool variable method and a verification method based on an actual measurement website, deviation of a single method can be effectively overcome, comprehensive and