CN-121389071-B - Small sample water quality parameter inversion method, storage medium and computer equipment

CN121389071BCN 121389071 BCN121389071 BCN 121389071BCN-121389071-B

Abstract

The invention discloses a small sample water quality parameter inversion method, a storage medium and computer equipment, which relate to the technical field of remote sensing water quality monitoring, and mainly comprise the following steps: data acquisition and preprocessing, source domain Inherent Optical Parameter (IOPs) inversion model training, cross-domain generation countermeasure enhancement, target domain IOPs prediction and water quality parameter inversion. By implementing the small sample water quality parameter inversion method, the storage medium and the computer equipment provided by the invention, the accuracy of river water quality inversion under the condition of the small sample can be improved.

Inventors

LV YUFENG
CHEN FANGFANG
XU HONGGEN
TANG YUANYUAN
ZOU PU
Peng Nengli
JIANG WEIXIA
ZHANG YISHU
Qi Mengru
WANG KAIHUA

Assignees

中国地质调查局长沙自然资源综合调查中心

Dates

Publication Date: 20260508
Application Date: 20251224

Claims (7)

1. The small sample water quality parameter inversion method is characterized by comprising the following steps of: s1, acquiring spectral data of a source domain water body, corresponding inherent optical characteristic data, spectral data of a target domain water body and corresponding water quality parameter data, and preprocessing and data enhancement processing are carried out on the spectral data of the source domain water body and the target domain water body to obtain processed spectral data; s2, constructing an inherent optical characteristic inversion model, training the inherent optical characteristic inversion model by utilizing the processed spectral data of the source domain water body and corresponding inherent optical characteristic data to obtain a trained inherent optical characteristic inversion model, wherein the inherent optical characteristic inversion model comprises a multi-scale convolution characteristic extraction module, a characteristic fusion module and a full-connection regression module, the multi-scale convolution characteristic extraction module comprises a plurality of parallel convolution kernels for carrying out parallel convolution extraction on spectral characteristics in different wave band ranges, the convolution kernels are 1×1 convolution kernels, 3×3 convolution kernels and 5×5 convolution kernels, the characteristic fusion module is a splicing operation unit or a weighted fusion unit for splicing or weighting fusion of different scale convolution characteristics, and the weighted fusion unit is a compression-excitation module or a convolution block attention module; S3, inputting the spectral data of the target domain water body to generate an countermeasure network, and performing countermeasure training by utilizing the spectral data of the source domain water body to generate target domain enhanced spectral data consistent with the spectral data distribution of the source domain water body; S4, inputting the target domain enhancement spectrum data into the trained inherent optical characteristic inversion model to obtain a target domain inherent optical characteristic predicted value; s5, obtaining a target domain water quality parameter prediction result by utilizing a water quality parameter inversion model according to the target domain inherent optical characteristic prediction value.
2. The method of claim 1, wherein the spectral data of the source domain water and the target domain water comprise image, blue, green, red, near infrared and short wave infrared band reflectivity data of corresponding dates and positions acquired from multispectral satellite data, the intrinsic optical characteristic data comprises absorption coefficients and backscattering coefficients, and the water quality parameter data comprises chlorophyll concentration, total suspended matter concentration and colored soluble organic matter absorption coefficients.
3. The method of small sample water quality parameter inversion of claim 1 wherein said preprocessing comprises radiometric scaling and atmospheric correction and said data enhancement processing comprises random noise perturbation, band substitution and spectral smoothing.
4. The small sample water quality parameter inversion method of claim 1 wherein said generating an countermeasure network generates a countermeasure network for cyclical consistency, said countermeasure training loss function being as follows: , , , Wherein, the 、 And The contrast loss, the cyclic consistency loss and the spectrum reconstruction loss are respectively; 、、、 respectively representing a generator, a discriminator, source domain data distribution and target domain data distribution; 、 respectively representing source domain data and target domain data; Representing the classification confidence expectation of the discriminator on the real target domain sample y; Means performing a logarithmic process; A classification function for the discriminator for discriminating whether the input data is authentic or counterfeit; Representing the classification confidence expectations of the arbiter for the generated samples G (x); A mapping function of the generator, which represents a conversion sample obtained after the source domain sample x is input into the generator; respectively representing the desire for source/destination domain data; Representing a target domain to source domain generator; Representing a source domain to target domain generator; Represents an L1 norm; representing the prediction/generation spectrum, generator output or downstream decoder result; is a real spectrum label; Representing the L2 norm.
5. The small sample water quality parameter inversion method according to claim 1, wherein step S5 specifically comprises: performing initial training on the water quality parameter inversion model by utilizing water quality parameter samples of a source domain and a target domain to obtain an initially trained water quality parameter inversion model; According to the water quality parameter data of the target domain water body, utilizing an advanced stopping and learning rate attenuation strategy to finely adjust the parameters of the initially trained water quality parameter inversion model to obtain an optimized water quality parameter inversion model; and inputting the target domain inherent optical characteristic predicted value into the optimized water quality parameter inversion model to obtain a target domain water quality parameter predicted result.
6. A computer readable storage medium having stored thereon a computer program, which when executed by a processor performs the steps of the small sample water quality parameter inversion method according to any of claims 1-5.
7. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor performs the steps of the small sample water quality parameter inversion method of any one of claims 1-5 when the computer program is executed.

Description

Small sample water quality parameter inversion method, storage medium and computer equipment Technical Field The invention relates to the technical field of remote sensing water quality monitoring, in particular to a small sample water quality parameter inversion method, a storage medium and computer equipment. Background With the rapid development of remote sensing technology, water quality monitoring by utilizing multispectral or hyperspectral remote sensing images has become an important means in water environment protection and treatment. Existing water quality parameter inversion methods can be generally divided into three types, i.e. experience/statistics model, semi-experience semi-analysis model and deep learning-based method. The empirical/statistical model usually adopts an empirical relationship between measured water quality parameters and remote sensing reflectivity (Rrs), such as linear regression, exponential regression, random forest, support vector machine and the like. The method is simple to realize and easy to popularize and apply. However, the method is highly dependent on the number and distribution of samples, and when the number of samples is limited, the over-fitting is easy to occur, the model generalization capability is insufficient, and the applicability under different water bodies or cross-region conditions is limited. The semi-empirical semi-analytical model is based on the radiation transmission relationship between the intrinsic optical properties (Inherent Optical Properties, IOPs) and the remote sensing reflectivity of the body of water. Such as IOCCG recommended algorithms, the QAA algorithm proposed by Lee, etc. The method has definite physical meaning, and can link the spectrum and the water quality parameters through IOPs (such as absorption coefficient a and backscattering coefficient bb). However, the model parameters have large differences among different water body types, and a large amount of measured data is required to be used as constraint in application, so that the application under the condition of diversified water bodies is limited. In recent years, convolutional Neural Networks (CNNs), transfer learning, ensemble learning and other methods are introduced into water quality parameter inversion. The complex nonlinear mapping relation between the reflectivity and the water quality parameters is automatically learned through training of large-scale actually measured sample data, and a good effect is obtained in partial research. However, all three methods mentioned above imply the premise of 'training domain and target domain optical properties are equally distributed'. When the target area is a river high turbidity water body and the training area is a clean ocean water body, the precondition is invalid, so that the model deviation is uncontrollable. In actual monitoring of river basins, it is often difficult to obtain a sufficient number of measured water quality samples. On one hand, the field sampling and experimental analysis are high in cost, and on the other hand, the river basin has strong spatial heterogeneity and time dynamic property, so that the number of samples in a single basin is limited. Meanwhile, the large-scale water quality sample set which can be obtained is mostly derived from ocean water, the optical characteristics of the large-scale water quality sample set are obviously different from those of river water, and the model accuracy is reduced and the applicability is insufficient due to direct migration and use. Disclosure of Invention The invention aims to provide a small sample water quality parameter inversion method, a storage medium and computer equipment, which can improve the accuracy of river water quality inversion under the condition of small samples. The invention provides a water quality parameter inversion method for a small sample, which comprises the following steps: S1, acquiring spectral data of a source domain water body, corresponding inherent optical characteristic data, spectral data of a target domain water body and corresponding water quality parameter data, and preprocessing and data enhancement processing are carried out on the spectral data of the source domain water body and the target domain water body to obtain processed spectral data; S2, constructing an inherent optical characteristic inversion model, and training the inherent optical characteristic inversion model by utilizing the processed spectral data of the source domain water body and the corresponding inherent optical characteristic data to obtain a trained inherent optical characteristic inversion model; S3, inputting the spectrum data of the target domain water body into a countermeasure network, and performing countermeasure training by utilizing the spectrum data of the source domain water body to generate target domain enhanced spectrum data consistent with the spectrum data distribution of the source domain water body; S4, inputting the targ