Search

CN-121330503-B - Depth learning inversion method and system based on physical constraint reflectivity and electronic equipment

CN121330503BCN 121330503 BCN121330503 BCN 121330503BCN-121330503-B

Abstract

The invention discloses a deep learning inversion method, a system and electronic equipment based on physical constraint reflectivity, wherein the method comprises the steps of constructing a deep neural network model, inputting hyperspectral remote sensing sample data into the deep neural network model, respectively utilizing hyperspectral remote sensing sample data to carry out model training by a transmittance product combination parameter submodel, an atmospheric intrinsic reflectivity submodel and an atmospheric hemispherical albedo submodel, and obtaining hyperspectral remote sensing data of a research area, inputting the hyperspectral remote sensing data into the deep neural network model and predicting and outputting the top reflectivity of an atmosphere layer by a physical inversion model. The invention creatively builds a parameter model, respectively predicts the transmittance product combination parameter, the atmosphere intrinsic reflectivity and the atmosphere hemispherical albedo by means of the parameter model, realizes the high-efficiency and rapid prediction of the output atmosphere layer top reflectivity by utilizing three parameters based on the built physical radiation transmission equation, and has the advantages of high prediction efficiency, high precision and the like.

Inventors

  • CHEN WEI
  • ZHAO JING
  • Xiong Wenwu

Assignees

  • 中国矿业大学(北京)

Dates

Publication Date
20260512
Application Date
20251103

Claims (10)

  1. 1. A depth learning inversion method based on physical constraint reflectivity is characterized by comprising the following steps: S1, constructing a deep neural network model combining physical constraint and radiation transmission, wherein the deep neural network model comprises a parameter model and a physical inversion model, and the parameter model comprises a transmittance product combination parameter sub-model, an atmosphere intrinsic reflectivity sub-model and an atmosphere hemispherical albedo sub-model; S2, constructing hyperspectral remote sensing sample data of a research area, wherein the hyperspectral remote sensing sample data comprises apparent reflectivity received by corresponding satellites The hyperspectral remote sensing sample data comprises transmittance product combination parameter label data, atmosphere intrinsic reflectivity label data, atmosphere hemispherical albedo label data and atmosphere layer top reflectivity label data, and the transmittance product combination parameter comprises a data set of data sets Inputting hyperspectral remote sensing sample data into a deep neural network model, respectively utilizing hyperspectral remote sensing sample data to carry out model training on the transmittance product combination parameter submodel, the atmospheric intrinsic reflectivity submodel and the atmospheric hemispherical albedo submodel, and utilizing the transmittance product combination parameter output by the deep neural network model by the physical inversion model Intrinsic reflectance of atmosphere Albedo of atmospheric hemisphere Model training is carried out according to the following physical constraint formula; , 、 、 As a parameter of the model, it is possible to provide, Is the top reflectivity of the atmosphere; S3, acquiring hyperspectral remote sensing data of the research area, inputting the hyperspectral remote sensing data into a trained deep neural network model, respectively outputting transmittance product combination parameters, atmosphere intrinsic reflectivity and atmosphere hemispherical albedo by the deep neural network model, and outputting atmospheric layer top reflectivity by a physical inversion model through the transmittance product combination parameters, the atmosphere intrinsic reflectivity and the atmosphere hemispherical albedo.
  2. 2. The depth learning inversion method based on physical constraint reflectivity is characterized in that the depth neural network model further comprises an encoder, a spatial self-attention module, a decoder and a residual optimization module, hyperspectral remote sensing sample data or hyperspectral remote sensing data are used as input images of the encoder, the encoder extracts multi-scale spatial features and high-dimensional semantic information from the input images and obtains multi-scale features, the spatial self-attention module is used for obtaining spatial attention weights of the multi-scale features, the decoder is used for decoding the features layer by layer, the residual optimization module adopts a residual connection structure to conduct residual analysis on the decoded features and the input images and improve the capability of capturing local features, and the residual optimization module inputs the features into a parameter model, and the transmittance product combination parameter sub-model, the atmospheric intrinsic reflectivity sub-model and the atmospheric hemispherical albedo sub-model of the parameter model respectively output transmittance product combination parameters, atmospheric intrinsic reflectivity and parameter results of atmospheric hemispherical albedo.
  3. 3. The method for deep learning inversion of reflectivity based on physical constraints of claim 2, wherein said transmittance product combination parameter submodel further employs the following physical constraint formula: wherein , For the sensor to observe the zenith angle, In order to achieve an optical thickness of the atmosphere, For the total diffuse scattering transmittance of the ascending atmosphere, In order to be able to diffuse the transmission power, 、 、 、 Derived from hyperspectral remote sensing data or hyperspectral remote sensing sample data and used as a necessary characteristic item of a deep neural network model, Is a model parameter.
  4. 4. The method for deep learning inversion based on physical constraint reflectivity of 2 or 3 further comprising a global information optimization module, wherein the global information optimization module introduces scaling factors and offset factors to perform scaling and offset processing of features including spatial features and high-dimensional semantic information.
  5. 5. The method for deep learning inversion based on physical constraint reflectivity according to claim 1, wherein the atmospheric intrinsic reflectivity label data of the hyperspectral remote sensing sample data are obtained through inversion of hyperspectral remote sensing sample data with clear and cloudless research areas, and the atmospheric hemispherical albedo label data of the hyperspectral remote sensing sample data are obtained through radiation measurement and calculation by four-component radiometers under clear and cloudless research areas.
  6. 6. The deep learning inversion method based on the physical constraint reflectivity is characterized in that the hyperspectral remote sensing sample data and the hyperspectral remote sensing data comprise satellite hyperspectral remote sensing images, in the method S2, a transmittance product combination parameter sub-model, an atmospheric intrinsic reflectivity sub-model and an atmospheric hemispherical albedo sub-model of the deep neural network model respectively conduct pixel-by-pixel and space topology associated model training on the satellite hyperspectral remote sensing images, the physical inversion model conducts pixel-by-pixel and space topology associated model training on the satellite hyperspectral remote sensing images, and in the method S3, the deep neural network model outputs the atmospheric hemispherical albedo corresponding to each pixel of the satellite hyperspectral remote sensing images.
  7. 7. The method for deep learning inversion based on physical constraint reflectivity is characterized in that the hyperspectral remote sensing sample data comprises satellite hyperspectral remote sensing images acquired by satellites, and the satellite hyperspectral remote sensing images are preprocessed as follows: A1, performing radiation calibration and atmosphere correction on a satellite hyperspectral remote sensing image; A2, removing a water vapor absorption wave band in the satellite hyperspectral remote sensing image, wherein the wavelength range of the water vapor absorption wave band comprises 1363.481 nm-1447.714 nm, 1801.491 nm-1953.111 nm and 2475.353 nm-2509.043 nm; and A3, intercepting window band data in the satellite hyperspectral remote sensing image by adopting a sliding window, calculating an average value and a standard deviation, and eliminating data exceeding the average value plus or minus 3 standard deviations in the sliding window.
  8. 8. The method S2 is characterized in that a transmittance product combination parameter sub-model, an atmospheric intrinsic reflectance sub-model and a prediction error threshold of an atmospheric hemispherical albedo sub-model are respectively set, difference calculation is conducted on the transmittance product combination parameter predicted value output by the transmittance product combination parameter sub-model and the true value of transmittance product combination parameter tag data, the difference value is smaller than the prediction error threshold and is used as a model training constraint, difference calculation is conducted on the atmospheric intrinsic reflectance predicted value output by the atmospheric intrinsic reflectance sub-model and the true value of the atmospheric intrinsic reflectance tag data, the difference value is smaller than the prediction error threshold and is used as a model training constraint, difference calculation is conducted on the atmospheric hemispherical albedo predicted value output by the atmospheric hemispherical albedo sub-model and the true value of the atmospheric hemispherical albedo tag data, and the difference value is smaller than the prediction error threshold and is used as a model training constraint.
  9. 9. A depth learning inversion system based on physical constraint reflectivity is characterized by comprising a depth neural network model, a hyperspectral remote sensing sample database and a hyperspectral remote sensing data acquisition module, wherein the depth neural network model is constructed by combining physical constraint and radiation transmission, the depth neural network model comprises a parameter model and a physical inversion model, the parameter model comprises a transmittance product combination parameter sub-model, an atmospheric intrinsic reflectivity sub-model and an atmospheric hemispherical reflectivity sub-model, and hyperspectral remote sensing sample data of a research area and apparent reflectivities received by corresponding satellites are stored in the hyperspectral remote sensing sample database The hyperspectral remote sensing sample data comprises transmittance product combination parameter label data, atmosphere intrinsic reflectivity label data, atmosphere hemispherical albedo label data and atmosphere layer top reflectivity label data, and the transmittance product combination parameter comprises a data set of data sets The model training is carried out by respectively utilizing hyperspectral remote sensing sample data in a hyperspectral remote sensing sample database by the transmittance product combination parameter submodel, the atmospheric intrinsic reflectivity submodel and the atmospheric hemispherical albedo submodel of the deep neural network model, and the physical inversion model utilizes the transmittance product combination parameter output by the deep neural network model Intrinsic reflectance of atmosphere Albedo of atmospheric hemisphere Model training is carried out according to the following physical constraint formula; , 、 、 As a parameter of the model, it is possible to provide, The hyperspectral remote sensing data acquisition module is used for acquiring hyperspectral remote sensing data of a research area and inputting the hyperspectral remote sensing data into a trained deep neural network model, the deep neural network model respectively outputs a transmittance product combination parameter, an atmospheric intrinsic reflectivity and an atmospheric hemispherical albedo, and then the physical inversion model outputs the atmospheric top reflectivity by using the transmittance product combination parameter, the atmospheric intrinsic reflectivity and the atmospheric hemispherical albedo.
  10. 10. An electronic device comprising at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor to cause the at least one processor to perform the steps of the method of any of claims 1-8.

Description

Depth learning inversion method and system based on physical constraint reflectivity and electronic equipment Technical Field The invention relates to the field of atmospheric correction inversion by utilizing hyperspectral images, in particular to a depth learning inversion method, a system and electronic equipment based on physical constraint reflectivity. Background With the rapid development of remote sensing technology, hyperspectral remote sensing data plays an increasingly important role in the fields of resource investigation, environmental monitoring, agricultural evaluation, atmospheric research and the like due to rich spectral information and fine wave band division. However, the radiation information recorded by the hyperspectral sensor needs to pass through the atmosphere before reaching the sensor, so that the radiation information is inevitably influenced by the processes of atmospheric scattering, absorption and the like, so that a significant difference exists between the original observed data and the actual reflectivity of the ground object, and thus the atmospheric correction processing is required, the traditional atmospheric correction processing method mainly solves the atmospheric radiation transmission equation to eliminate the atmospheric influence, such as models of MODTAN, 6S, FLAASH and the like, but has certain limitations in practical application, depends on accurate atmospheric parameters (such as water vapor content, aerosol optical thickness and the like) and often needs to be estimated through auxiliary data or an empirical model, so that additional errors are introduced, the calculation process is complex, a large number of iterative operations and lookup table interpolation are involved, the calculation efficiency is low, and the requirement of large-scale data processing is difficult to meet, and the model assumes that the atmospheric condition is uniform, and the actual atmosphere has space-time heterogeneity, so that the correction accuracy under the complex atmospheric condition is reduced. In recent years, the deep learning method has strong potential in the field of atmosphere correction, and compared with the traditional atmosphere correction processing method, the deep learning method has the advantages of being capable of automatically learning a complex nonlinear mapping relation between atmospheric influence and earth surface reflectivity from data, reducing dependence on artificial feature design, being high in calculation efficiency, being capable of realizing rapid reasoning after model training is completed, being suitable for real-time processing of massive remote sensing data, being high in adaptability, and being capable of learning different atmospheric conditions and change rules of sensor characteristics through training data. Thus, how to take advantage of the apparent reflectivity of satellite receptionAnd the deep learning of hyperspectral remote sensing data and the correction of the true reflectivity of the inversion ground object (namely the top reflectivity of the atmosphere) are technical problems to be solved in the present technology. Disclosure of Invention The invention aims to provide a depth learning inversion method, a system and electronic equipment based on physical constraint reflectivity, wherein a depth neural network model is used for extracting characteristic data of hyperspectral remote sensing data through processing such as coding, spatial attention, decoding and residual error, and the like, and inputting the characteristic data into a parameter model, wherein the parameter model is combined with physical constraint prediction to obtain three parameters of transmittance product combination parameters, atmosphere intrinsic reflectivity and atmosphere hemispherical albedo, and then the three parameters are used by the physical inversion model to output the predicted atmosphere layer top reflectivity based on a physical radiation transmission equation, so that the high-precision atmosphere layer top reflectivity is obtained through the hyperspectral remote sensing data, and the depth neural network model has the advantages of high efficiency, strong generalization capability and the like. The aim of the invention is achieved by the following technical scheme: A depth learning inversion method based on physical constraint reflectivity comprises the following steps: S1, constructing a deep neural network model combining physical constraint and radiation transmission, wherein the deep neural network model comprises a parameter model and a physical inversion model, and the parameter model comprises a transmittance product combination parameter sub-model, an atmosphere intrinsic reflectivity sub-model and an atmosphere hemispherical albedo sub-model; S2, constructing hyperspectral remote sensing sample data of a research area, wherein the hyperspectral remote sensing sample data comprises apparent reflectivity received by correspon