CN-120997696-B - Training method for neglecting cloud interference in remote sensing deep learning parameter inversion

CN120997696BCN 120997696 BCN120997696 BCN 120997696BCN-120997696-B

Abstract

The invention discloses a training method for neglecting cloud interference in remote sensing deep learning parameter inversion, and relates to the field of geospatial artificial intelligence, comprising the steps of carrying out cloud identification on an original remote sensing image, generating cloud mask data, and rasterizing ground actual measurement data into tag data matched with the remote sensing image space; the cloud mask data is utilized to mark cloud shielding areas in the tag data to obtain cloud mask tag data, a remote sensing deep learning parameter inversion model is constructed, an original remote sensing image is input to be predicted, a parameter prediction graph is output, a loss function is ignored based on a cloud mask, the ignored cloud interference loss between the parameter prediction graph and the cloud mask tag data is calculated, and model parameters are optimized through gradient back propagation. The method solves the technical problem that cloud interference is difficult to effectively process and inversion accuracy is limited in the existing training method, and achieves the technical effect of effectively processing interference of cloud cover on model training and improving inversion accuracy.

Inventors

LUAN WENBO
LIU PENG
YI ZHILI
LI XIAOLIANG
WANG KAISONG
Ke Jichang
LIU GUANG
YAN NING
LIU YANSONG
LI LEI
CUI YANG
LI HAN

Assignees

北京市测绘设计研究院

Dates

Publication Date: 20260508
Application Date: 20250807

Claims (5)

1. A training method for neglecting cloud interference in remote sensing deep learning parameter inversion is characterized by comprising the following steps: Performing cloud identification on an original remote sensing image to generate cloud mask data, and rasterizing ground actual measurement data into tag data matched with the remote sensing image space; Marking the cloud shielding area in the tag data by utilizing the cloud mask data to obtain cloud mask tag data; improving a pre-training image segmentation network, constructing an image-based remote sensing deep learning parameter inversion model, inputting the original remote sensing image into the remote sensing deep learning parameter inversion model for prediction, and outputting a parameter prediction graph; improving the mean square error loss function to obtain a cloud mask neglected loss function, and calculating neglected cloud interference loss between the parameter prediction graph and the cloud mask tag data based on the cloud mask neglected loss function; based on the neglect of cloud interference loss, the training of the remote sensing deep learning parameter inversion model is completed through gradient back propagation optimization model parameters; Improving the mean square error loss function to obtain a cloud mask neglected loss function, comprising: Obtaining a mean square error loss function; modifying the mean square error loss function into a piecewise function to obtain a cloud mask neglected loss function; The expression of the cloud mask neglecting the loss function is as follows: CM- ; Wherein CM-MSE represents a cloud mask neglecting loss function, H represents the total line number of the original remote sensing image, W represents the total column number of the original remote sensing image, H represents the line number of the current pixel, W represents the column number of the current pixel, Representing parameter tag values at locations (h, w) in the cloud mask tag data, Representing model prediction parameter values at positions (h, w) in the parameter prediction graph, wherein ignore_index represents values marked as cloud mask areas in the cloud mask tag data; Based on the neglecting cloud interference loss, completing training of the remote sensing deep learning parameter inversion model by gradient back propagation optimization model parameters, comprising: calculating the gradient of the neglected cloud interference loss to the remote sensing deep learning parameter inversion model parameter through a back propagation algorithm; and updating parameters of the remote sensing deep learning parameter inversion model by using the gradient through an optimizer, and performing iterative training until the remote sensing deep learning parameter inversion model converges.
2. The method of claim 1, wherein the performing cloud recognition on the original remote sensing image to generate cloud mask data comprises: Judging whether the original remote sensing image has a data quality evaluation wave band or not; if yes, identifying cloud pixels based on cloud information in the data quality evaluation wave band to obtain cloud mask data; and if not, carrying out cloud identification on the original remote sensing image through the segmentation model to obtain cloud mask data.
3. The method of claim 1, wherein the step of rasterizing ground measured data into tag data spatially matched with the remote sensing image comprises: Converting the ground actual measurement data into vector space data through coordinate analysis and data format conversion; Performing geospatial interpolation processing on the vector space data to obtain continuous field data; Determining grid meshes by taking the space range of the original remote sensing image as a constraint, and determining grid units by taking the resolution of the original remote sensing image as a constraint; and rasterizing the continuous field data through the grid mesh and the grid unit to obtain tag data.
4. The method of claim 1, wherein the training method for ignoring cloud interference in remote sensing deep learning parameter inversion is characterized by improving a pre-training image segmentation network to construct an image-based remote sensing deep learning parameter inversion model, and comprises the following steps: acquiring a pre-training image segmentation network; deleting the terminal network by taking the pre-training image segmentation network as a basic framework to obtain a truncated image segmentation network; and presetting an output convolution layer, wherein the output convolution layer is connected with the truncated image segmentation network to obtain a remote sensing deep learning parameter inversion model.
5. A remote sensing deep learning system as defined in claim 1 the training method of ignoring cloud interference in the parameter inversion, the method is characterized by further comprising the following steps after the training of the remote sensing deep learning parameter inversion model is completed: and deploying the trained remote sensing deep learning parameter inversion model to a remote sensing data processing system, inputting a remote sensing image to be inverted, and outputting a parameter inversion result neglecting cloud interference through the remote sensing deep learning parameter inversion model, wherein the parameter inversion result comprises surface temperature, leaf area index, soil humidity or water quality parameters.

Description

Training method for neglecting cloud interference in remote sensing deep learning parameter inversion Technical Field The application relates to the field of geospatial artificial intelligence (Geospatial AI, geoAI) related technology, in particular to a training method for ignoring cloud interference in remote sensing deep learning parameter inversion. Background The remote sensing deep learning parameter inversion is a process of reversely deducing physical, chemical or biological parameters in the earth surface or the atmosphere from remote sensing data by using a deep learning technology, is one of important research directions of geospatial artificial intelligence (Geospatial AI, geoAI), and has important application values in the aspects of environment monitoring, agricultural monitoring, hydrologic monitoring and the like. At present, the main method for solving the problem of cloud interference in remote sensing deep learning parameter inversion is to convert image data into table data and reject cloud-containing data, or convert a regression task of continuous value prediction into an image segmentation task of discrete value prediction. However, the existing method cannot effectively solve the problem of cloud interference due to limited feature extraction caused by destroying an image data structure or numerical precision loss caused by converting continuous numerical values into discrete data values, and limits the precision and efficiency of remote sensing deep learning parameter inversion. In the related art at the present stage, the training method for ignoring cloud interference in remote sensing deep learning parameter inversion has the technical problem that the cloud interference is difficult to effectively process, so that the inversion accuracy is limited. Disclosure of Invention According to the training method for neglecting cloud interference in remote sensing deep learning parameter inversion, cloud identification is carried out on remote sensing images to generate cloud masks, ground measured data are rasterized into matched tag data, cloud shielding areas in the tag data are marked by the cloud masks, an image segmentation network is improved to construct a parameter inversion model, a remote sensing image is input to predict and output a parameter prediction graph, a mean square error loss function is improved to be a cloud mask neglect loss function, neglect cloud interference loss between the prediction graph and the tag data is calculated, model parameters are optimized through gradient back propagation based on the losses, model training and other technical means are completed, the technical problem that cloud interference is difficult to effectively process in the existing training method for neglecting cloud interference in remote sensing deep learning parameter inversion, inversion accuracy is limited is solved, the interference problem of cloud layer shielding on model training is effectively processed, and the technical effect of inversion accuracy is improved. The application provides a training method for neglecting cloud interference in remote sensing deep learning parameter inversion, which comprises the steps of carrying out cloud identification on an original remote sensing image, generating cloud mask data, rasterizing ground actual measurement data into tag data matched with remote sensing image space, marking cloud shielding areas in the tag data by utilizing the cloud mask data to obtain cloud mask tag data, improving a pre-training image segmentation network, constructing an image-based remote sensing deep learning parameter inversion model, inputting the original remote sensing image into the remote sensing deep learning parameter inversion model to predict, outputting a parameter prediction graph, improving a mean square error loss function, obtaining a cloud mask neglect loss function, calculating neglect cloud interference loss between the parameter prediction graph and the cloud mask tag data based on the cloud mask neglect loss function, and optimizing model parameters through gradient reverse propagation based on the neglect cloud interference loss to finish training of the remote sensing deep learning parameter inversion model. In a possible implementation mode, cloud identification is carried out on an original remote sensing image to generate cloud mask data, and the following processing is carried out, wherein whether the original remote sensing image has a data quality evaluation wave band or not is judged, if so, cloud pixels are identified based on cloud information in the data quality evaluation wave band to obtain cloud mask data, and if not, cloud identification is carried out on the original remote sensing image through a segmentation model to obtain cloud mask data. In a possible implementation mode, the ground actual measurement data are rasterized into tag data matched with a remote sensing image space, the following processing is carried