CN-122021231-A - Lake chlorophyll concentration self-adaptive remote sensing inversion method and system based on test

CN122021231ACN 122021231 ACN122021231 ACN 122021231ACN-122021231-A

Abstract

The invention relates to the technical field of remote sensing monitoring of water environment and discloses a lake chlorophyll concentration self-adaptive remote sensing inversion method and system based on testing, comprising the following steps of data preparation and preprocessing, namely acquiring training in-situ spectrum data and corresponding chlorophyll concentration data, acquiring testing remote sensing image data, and preprocessing the training data and the testing data; model construction and joint training, namely constructing a dual-task architecture model comprising a feature extractor, a main regression module and a self-adaptive module in test, and performing end-to-end training on the model by adopting a joint loss function. By introducing a self-adaptive optimization mechanism in a test stage and combining domain offset detection and a dynamic parameter adaptation flow, the model can utilize unlabeled test data to carry out self-supervision parameter fine adjustment, dynamically adapt to lake water body data distribution differences of different time periods and different geographic areas, and effectively solve the problem that the generalization performance of the existing static remote sensing inversion model is reduced in a new monitoring scene.

Inventors

Yang Cuilin
ZHANG ZHEN
ZHANG BINGRU
Qi Yemao
LIU CHENXI

Assignees

昆明理工大学

Dates

Publication Date: 20260512
Application Date: 20251114

Claims (10)

1. A lake chlorophyll concentration self-adaptive remote sensing inversion method based on a test is characterized by comprising the following steps of: the data preparation and preprocessing, namely acquiring in-situ spectrum data for training and corresponding chlorophyll concentration data, acquiring remote sensing image data for testing, and preprocessing the training data and the testing data; Constructing a dual-task architecture model comprising a feature extractor, a main regression module and a self-adaptive module in test, and performing end-to-end training on the model by adopting a joint loss function; the self-adaptive inversion during testing is that the preprocessed test data is input into a model with completed training, and domain offset detection, dynamic parameter adaptation and training optimization circulation during testing are sequentially carried out, so that model parameters adapting to current test data distribution are obtained; And (3) final concentration inversion, namely outputting a lake chlorophyll concentration inversion result through a main regression module based on the adapted model parameters.
2. The test-based lake chlorophyll concentration adaptive remote sensing inversion method of claim 1, wherein the data preprocessing includes normalizing spectral data and logarithmically converting chlorophyll concentration data.
3. The lake chlorophyll concentration self-adaptive remote sensing inversion method based on the test of claim 1, wherein the feature extractor is a multi-scale one-dimensional convolutional neural network, features are extracted in parallel by different-size convolutional check input spectrum data, and feature vectors are generated after the feature fusion module processes the feature data.
4. A method of adaptive remote sensing inversion of chlorophyll concentration in a lake based on testing according to claim 3, wherein said feature fusion module comprises an activation operation, a max pooling operation, and an average pooling operation, said activation operation being GeLU activation.
5. The method for adaptive remote sensing inversion of chlorophyll concentration in a lake according to claim 1, wherein the joint loss function is formed by weighting a primary regression loss and a spectral reconstruction loss, the primary regression loss is a mean square error loss or a smooth L1 loss, and the spectral reconstruction loss is a mean square error loss.
6. The method for adaptive remote sensing inversion of chlorophyll concentration in a lake according to claim 1, wherein the domain shift detection is quantified by calculating the difference between the statistical characteristics of the test data and the reference statistical characteristics of the training data.
7. The method for adaptive remote sensing inversion of chlorophyll concentration in lakes based on the test of claim 1, wherein the dynamic parameter adaptation selects the corresponding learning rate, adaptation step number and stability weight according to the degree of domain shift, and the degree of domain shift is divided into three classes of slight domain shift, moderate domain shift and significant domain shift.
8. The lake chlorophyll concentration adaptive remote sensing inversion method based on the test of claim 1, wherein the test-time training optimizes the circulation freezing feature extractor parameters, only optimizes the reconstruction head parameters in the main regression module and the test-time adaptive module, and optimizes the composite loss including the spectral reconstruction loss, the stability loss and the prediction range constraint loss.
9. A test-based lake chlorophyll concentration adaptive remote sensing inversion system for a test-based lake chlorophyll concentration adaptive remote sensing inversion method as set forth in any one of claims 1 to 8, comprising: the data acquisition and preprocessing module is used for acquiring in-situ spectrum data for training, corresponding chlorophyll concentration data and remote sensing image data for testing, and preprocessing the data; the model training module is used for building a dual-task architecture model and carrying out end-to-end training on the model by adopting a joint loss function; the self-adaptive inversion module is used for executing domain offset detection, dynamic parameter adaptation and training optimization circulation during test, and realizing on-line adaptation of model parameters; And the result output module is used for outputting the inversion result of the chlorophyll concentration of the lake based on the adapted model parameters.
10. The adaptive remote sensing inversion system for lake chlorophyll concentration based on test of claim 9, wherein said adaptive inversion module comprises a domain offset detection unit for calculating a distribution difference of test data and training data, a parameter selection unit for selecting an adaptation parameter according to the distribution difference, and an optimization unit for performing fine-tuning optimization of the parameter.

Description

Lake chlorophyll concentration self-adaptive remote sensing inversion method and system based on test Technical Field The invention relates to the technical field of remote sensing monitoring of water environment, in particular to a lake chlorophyll concentration self-adaptive remote sensing inversion method and system based on testing. Background The chlorophyll a concentration in the lake is a key index for evaluating the eutrophication degree of the water body and monitoring the dynamic change of the water quality, and the accurate monitoring has important significance for the ecological environment protection of the river basin, the sustainable utilization of water resources and the early warning of water bloom disasters. With the rapid development of remote sensing technology, the satellite remote sensing image is utilized to realize large-range and long-term dynamic inversion of the chlorophyll a concentration in the lake, which becomes the core technical direction of the water environment monitoring field, and compared with the traditional in-situ sampling monitoring, the remote sensing inversion can break through space-time limitation, greatly improve the monitoring efficiency and coverage range and provide data support for large-scale water environment evaluation. At present, remote sensing inversion algorithms for the chlorophyll a concentration in lakes can be divided into three major categories according to technical principles, namely an empirical model, a semi-analytical model and a data-driven machine learning and deep learning model, wherein each model forms own technical characteristics and application boundaries in the development process, and meanwhile, limitations of different degrees are also exposed. The empirical model is an inversion method which is developed and applied at the earliest time, and the core thought is to establish a statistical regression relation between chlorophyll a concentration and specific wave band reflectivity or wave band combination (such as blue-green wave band ratio, normalized difference index and the like) of the remote sensing image based on a large amount of in-situ observation data. The typical blue-green wave band ratio model is widely applied to open ocean water body monitoring with relatively single optical characteristics due to simple principle and low calculation cost, but the inversion accuracy of the model is obviously reduced in inland lake water bodies with complex optical characteristics. The method is characterized in that besides chlorophyll a, a large amount of substances such as chromaticity dissolved organic matters, total suspended matters and the like exist in the inland lake water body, and the substances and the chlorophyll a affect the apparent optical characteristics of the water body together, so that the statistical relationship relied on by an empirical model is unstable, the contribution of various substances to the spectral reflectivity cannot be accurately separated, and further inversion result deviation is larger. Based on an empirical model, the semi-analytical model introduces a water body optical theory, and inversion of chlorophyll a concentration is realized by constructing a physical model between inherent optical characteristics (such as absorption coefficient and scattering coefficient of each component in the water body) and apparent optical characteristics (such as surface reflectivity obtained by remote sensing images). Compared with an empirical model, the model has stronger physical significance support, can be theoretically adapted to the change of optical characteristics of different water bodies, but in practical application, the semi-analysis model has extremely high dependence on the accuracy of inherent optical characteristic parameters, and the inherent optical characteristics of each component in the inland lake water body have obvious space-time heterogeneity, so that the accurate description is difficult to carry out through a unified parameterization scheme. When the algae community structure in the water body changes, the chromaticity dissolved organic matter sources change or the total suspended matter concentration fluctuates, the preset parameters of the semi-analysis model are not matched with the actual water body characteristics, the model assumption condition is easy to lose efficacy, and the inversion accuracy cannot meet the long-term monitoring requirement. In recent years, with the development of computer technology and data science, data-driven machine learning models (such as support vector machines, extreme gradient lifting, random forests, etc.) and deep learning models (such as convolutional neural networks, cyclic neural networks, transformers, etc.) are becoming mainstream technologies for remote sensing inversion of chlorophyll a concentration in lakes. The model can be used for remarkably improving the fitting capacity of the optical characteristics of the complex water body by