Lake chlorophyll concentration adaptive remote sensing inversion method and system based on test

Abstract

The invention relates to the technical field of water environment remote sensing monitoring, and discloses a lake chlorophyll concentration self-adaptive remote sensing inversion method and system based on testing, and the method comprises the following steps: data preparation and preprocessing: obtaining in-situ spectral data for training and corresponding chlorophyll concentration data, obtaining remote sensing image data for testing, and carrying out data preprocessing; preprocessing the training data and the test data; and model construction and joint training: constructing a dual-task architecture model comprising a feature extractor, a main regression module and a self-adaptive module during testing, and performing end-to-end training on the model by adopting a joint loss function. A self-adaptive optimization mechanism is introduced in a test stage, and domain offset detection and a dynamic parameter adaptation process are combined, so that the model can perform self-supervised parameter fine adjustment by using unmarked test data, and lake water body data distribution differences in different time periods and different geographic areas are dynamically adapted; the problem that the generalization performance of an existing static remote sensing inversion model is reduced in a new monitoring scene is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of water environment remote sensing monitoring technology, specifically to a test-based adaptive remote sensing inversion method and system for lake chlorophyll concentration. Background Technology

[0002] Chlorophyll a concentration in lakes is a key indicator for assessing eutrophication and monitoring dynamic changes in water quality. Accurate monitoring of chlorophyll a concentration is crucial for watershed ecological protection, sustainable water resource utilization, and early warning of algal blooms. With the rapid development of remote sensing technology, large-scale, long-term dynamic inversion of lake chlorophyll a concentration using satellite remote sensing imagery has become a core technological direction in water environment monitoring. Compared to traditional in-situ sampling monitoring, remote sensing inversion can overcome spatial and temporal limitations, significantly improving monitoring efficiency and coverage, and providing data support for large-scale water environment assessment.

[0003] Currently, remote sensing inversion algorithms for chlorophyll a concentration in lakes can be divided into three main categories based on their technical principles: empirical models, semi-analytical models, and data-driven machine learning and deep learning models. Each type of model has developed its own technical characteristics and application boundaries during its development, while also revealing limitations to varying degrees.

[0004] Among these, empirical models were the earliest developed and applied inversion methods. Their core idea is to establish a statistical regression relationship between chlorophyll a concentration and specific band reflectance or band combinations (such as the blue-green band ratio, normalized difference index, etc.) of remote sensing images based on a large amount of in-situ observation data. Typical blue-green band ratio models, due to their simple principle and low computational cost, have been widely used in monitoring open ocean waters with relatively simple optical characteristics. However, in inland lakes with complex optical characteristics, the inversion accuracy of such models decreases significantly. This is because inland lakes contain not only chlorophyll a but also a large amount of chromatic dissolved organic matter, total suspended solids, and other substances. These substances, along with chlorophyll a, affect the surface optical properties of the water, causing the statistical relationship relied upon by the empirical model to lose stability and making it impossible to accurately separate the contributions of various substances to spectral reflectance, thus resulting in significant deviations in the inversion results.

[0005] Semi-analytical models, building upon empirical models, incorporate water optics theory. They construct a physical model connecting inherent optical properties (such as the absorption and scattering coefficients of various water components) with apparent optical properties (such as surface reflectance obtained from remote sensing images) to retrieve chlorophyll a concentration. Compared to empirical models, these models have stronger physical support and are theoretically adaptable to variations in the optical properties of different water bodies. However, in practical applications, semi-analytical models are highly dependent on the accuracy of inherent optical property parameters. The inherent optical properties of various components in inland lakes exhibit significant spatiotemporal heterogeneity, making accurate description through a unified parameterization scheme difficult. When the algal community structure changes, the source of chromatic dissolved organic matter alters, or the total suspended solids concentration fluctuates, the preset parameters of the semi-analytical model do not match the actual water characteristics, easily leading to the invalidation of model assumptions and insufficient retrieval accuracy to meet long-term monitoring requirements.

[0006] In recent years, with the development of computer technology and data science, data-driven machine learning models (such as support vector machines, extreme gradient boosting, and random forests) and deep learning models (such as convolutional neural networks, recurrent neural networks, and Transformers) have gradually become the mainstream technologies for remote sensing inversion of chlorophyll a concentration in lakes. These models autonomously learn the complex nonlinear relationship between chlorophyll a concentration and remote sensing spectral characteristics from a large amount of observational data, without relying on explicit physical model assumptions. This significantly improves the fitting ability to complex water body optical properties and demonstrates superior inversion accuracy compared to traditional empirical and semi-analytical models in short-term monitoring tasks of some lakes.

[0007] However, existing data-driven models generally suffer from a fundamental technical flaw: a contradiction exists between the static nature of the model and the dynamic nature of the lake water environment. Specifically, after the training phase, the internal parameters of such models are fixed, forming a static model. However, in actual long-term, large-scale lake chlorophyll a concentration monitoring tasks, both the water environment and data conditions change significantly, leading to a deviation in the distribution of model training data and actual test data, i.e., the domain shift problem. On the one hand, the inherent optical properties of lake water are affected by factors such as algal community succession, changes in chromatic dissolved organic matter input, and fluctuations in total suspended matter concentration, exhibiting strong spatiotemporal heterogeneity. The spectral characteristics of lake water differ significantly across seasons and geographical regions. On the other hand, satellite sensors may experience performance degradation during long-term operation, and the variability of atmospheric conditions (such as cloud cover, aerosols, and water vapor content) can also lead to differences in the radiometric correction results of remote sensing images, further exacerbating the data distribution shift.

[0008] The aforementioned domain offset problem causes a sharp decline in the generalization performance of static data-driven models when applied to new time periods or new geographical areas, resulting in a significant decrease in inversion accuracy and even systematic biases. This situation makes it difficult for existing models to meet the needs of long-term, stable, and accurate lake water environment monitoring, and to provide continuous and reliable data support for watershed water quality management and eutrophication risk assessment. This has become a core technical problem that urgently needs to be solved in the development of current remote sensing inversion technology for lake chlorophyll a concentration. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a test-based adaptive remote sensing inversion method and system for lake chlorophyll concentration, which solves the problems mentioned in the background section.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a test-based adaptive remote sensing inversion method for lake chlorophyll concentration, comprising the following steps: Data preparation and preprocessing: Acquire in-situ spectral data and corresponding chlorophyll concentration data for training, acquire remote sensing image data for testing, and preprocess the training and testing data; Model building and joint training: A dual-task architecture model including a feature extractor, a main regression module and a test-time adaptive module is built, and the model is trained end-to-end using a joint loss function; Adaptive inversion during testing: The preprocessed test data is input into the trained model, and the domain offset detection, dynamic parameter adaptation, and test-time training optimization loop are performed sequentially to obtain model parameters that are adapted to the current test data distribution. Final concentration inversion: Based on the adapted model parameters, the lake chlorophyll concentration inversion results are output through the main regression module.

[0011] Preferably, the data preprocessing includes normalizing the spectral data and performing a logarithmic transformation on the chlorophyll concentration data.

[0012] Preferably, the feature extractor is a multi-scale one-dimensional convolutional neural network that extracts features from the input spectral data in parallel using convolutional kernels of different sizes, and generates feature vectors after processing by the feature fusion module.

[0013] Preferably, the feature fusion module includes an activation operation, a max pooling operation, and an average pooling operation, wherein the activation operation is GeLU activation.

[0014] Preferably, the joint loss function is composed of a weighted sum of principal regression loss and spectral reconstruction loss, wherein the principal regression loss is mean squared error loss or smoothing L1 loss, and the spectral reconstruction loss is mean squared error loss.

[0015] Preferably, the domain offset detection quantifies the degree of domain offset by calculating the difference between the statistical features of the test data and the reference statistical features of the training data.

[0016] Preferably, the dynamic parameter adaptation selects the corresponding learning rate, adaptation steps, and stability weights according to the degree of domain offset, and the degree of domain offset is divided into three levels: slight domain offset, moderate domain offset, and significant domain offset.

[0017] Preferably, the test-time training optimization loop freezes the feature extractor parameters, optimizing only the reconstruction head parameters in the main regression module and the test-time adaptive module. The optimization objective is a composite loss that includes spectral reconstruction loss, stability loss, and prediction range constraint loss.

[0018] A test-based adaptive remote sensing inversion system for lake chlorophyll concentration includes: Data acquisition and preprocessing module: used to acquire in-situ spectral data for training, corresponding chlorophyll concentration data and remote sensing image data for testing, and to preprocess the data; Model training module: used to build a dual-task architecture model, and uses a joint loss function to train the model end-to-end; Adaptive inversion module: used to perform domain offset detection, dynamic parameter adaptation, and test-time training optimization loop to achieve online adaptation of model parameters; Results output module: Used to output the lake chlorophyll concentration inversion results based on the adapted model parameters.

[0019] Preferably, the adaptive inversion module includes a domain offset detection unit, a parameter selection unit, and an optimization unit. The domain offset detection unit is used to calculate the distribution difference between the test data and the training data. The parameter selection unit is used to select the adaptation parameters according to the distribution difference. The optimization unit is used to perform parameter fine-tuning optimization.

[0020] This invention provides a test-based adaptive remote sensing inversion method and system for chlorophyll concentration retrieval in lakes. It has the following beneficial effects: 1. This invention introduces an adaptive optimization mechanism during the testing phase, combined with domain offset detection and dynamic parameter adaptation processes, enabling the model to perform self-supervised parameter fine-tuning using unlabeled test data. This allows the model to dynamically adapt to the differences in lake water data distribution across different time periods and geographical regions, effectively improving the generalization performance degradation of existing static remote sensing inversion models in new monitoring scenarios and enhancing the applicability of the model in diverse lake water monitoring tasks.

[0021] 2. This invention relies on a dual-task architecture for joint training, and uses a multi-scale one-dimensional convolutional neural network to extract spectral features. It is optimized through a composite loss that includes stability loss and prediction range constraints. This not only enhances the model's ability to capture the optical properties of complex water bodies, but also reduces the prediction bias that is prone to occur in traditional models. This improves the accuracy of the lake chlorophyll concentration inversion results. Its key indicators such as the coefficient of determination and root mean square error are improved compared with traditional CNN, XGBoost and other models.

[0022] 3. This invention possesses excellent long-term monitoring robustness and practical deployment convenience. It can stably handle long-term, large-scale remote sensing monitoring tasks of lake chlorophyll concentration, clearly reveal the spatiotemporal distribution pattern and trend of lake chlorophyll concentration, and accurately identify high-risk areas of eutrophication. At the same time, the dynamic parameter adaptation mechanism can automatically match hyperparameters without manual intervention. The optimization process only requires fine-tuning the parameters of the master regression module and reconstruction head to reduce computational costs, making it easy to achieve large-scale rapid water body inversion on conventional computing equipment, and providing reliable data support for watershed water quality management. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the enhanced TTT model architecture of the present invention; Figure 2 This is a scatter plot comparing the prediction performance of different models in this invention. Figure 3 This is a time series diagram of the Chl-a spatial distribution of representative lakes in five major lake regions of this invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see the appendix Figure 1 - Appendix Figure 3 This invention provides a test-based adaptive remote sensing inversion method for lake chlorophyll concentration, comprising the following steps: Data preparation and preprocessing: In-situ spectral data and corresponding chlorophyll concentration data for training were acquired, as well as remote sensing image data for testing. Preprocessing of both training and testing data was performed, including normalization of the spectral data and logarithmic transformation of the chlorophyll concentration data. The feature extractor is a multi-scale one-dimensional convolutional neural network that extracts features from the input spectral data in parallel using convolutional kernels of different sizes. These features are then processed by the feature fusion module to generate feature vectors.

[0026] Specifically, the GLORIA global in-situ dataset was used for model training. This dataset contains 7,572 hyperspectral in-situ measurement data points (Rrs(λ)) and corresponding water quality parameters from 450 water bodies worldwide. This invention uses 4,839 chlorophyll a (Chl-a) matching data points (concentration range 0.03-2,317 mgm⁻³) as the basis for training and validation.

[0027] Test / Application Data: Sentinel-2 MSI imagery was used as the application data source. L2A-level surface reflectance products were used, with 12 core bands (B1 to B12) selected as model input.

[0028] Data preprocessing: Normalization was performed on the training and testing data. To fit the right-skewed distribution of Chl-a concentration, a logarithmic transformation was performed on the measured Chl-a values. The model also predicts logarithmic values, and finally uses an exponential function ( The final concentration is obtained through inversion.

[0029] The feature fusion module includes activation, max pooling, and average pooling operations, with the activation operation being GeLU activation.

[0030] The feature fusion module includes activation, max pooling, and average pooling operations, with the activation operation being GeLU activation.

[0031] Model Construction and Joint Training: A dual-task architecture model was built, comprising a feature extractor, a main regression module, and a test-time adaptive module. End-to-end training was performed using a joint loss function. The joint loss function consists of a weighted average of the main regression loss and the spectral reconstruction loss. The main regression loss is either the mean squared error loss or the smoothing L1 loss, while the spectral reconstruction loss is also the mean squared error loss. Domain Shift Detection: The degree of domain shift is quantified by calculating the difference between the statistical characteristics of the test data and the reference statistical characteristics of the training data.

[0032] Specifically, the CNN feature extractor receives spectral data across 12 bands from its input layer. A multi-scale 1D-CNN (1x1, 5x1, 7x1 kernels) is used to extract local and adjacent band features in parallel. The final feature vector is then generated through a feature fusion module (GeLU activation, max / average pooling, Sigmoid normalization). Main return header: Uses an MLP to receive... Output ( (Predicted value).

[0033] TTT Reconstruction Header: Employs another MLP, receives f, outputs... (Reconstructed 12-band spectrum).

[0034] Joint training: Using the GLORIA training set, the joint loss function described in step (1) is employed. The entire model (CNN extractor and two MLP heads) is trained end-to-end. The root mean square error (RMSE) is used. The root mean square error of spectral reconstruction.

[0035] Adaptive inversion during testing: The preprocessed test data is input into the trained model, and the domain offset detection, dynamic parameter adaptation, and test-time training optimization loop are performed in sequence to obtain model parameters that are adapted to the current test data distribution. Dynamic parameter adaptation selects the corresponding learning rate, adaptation steps and stability weights according to the degree of domain offset. The degree of domain offset is divided into three levels: slight domain offset, moderate domain offset and significant domain offset.

[0036] Specifically, the adaptive inversion during the test was carried out using Sentinel-2 images of Taihu Lake in China during the summer of 2020 as an example: Image Acquisition: Acquire Sentinel-2L2A images covering the Taihu Lake area during the summer of 2020, perform cloud and land masking, and extract the spectral reflectance of pure water pixels. .

[0037] Forward propagation (initial prediction): The input is fed into the trained model. The CNN extractor generates features. . The data is fed into the main regression head to generate an initial (unadapted) predicted value. .

[0038] Domain offset detection: Calculate the batch Spectral statistical characteristics ( , and statistical characteristics during training (GLORIA) Compare by offset The formula is used for calculation. Assume the calculation yields... .

[0039] Select hyperparameters: based on Querying Table 1, it is determined to be "moderate domain offset". Select the adaptation parameter: learning rate. Adapt to step count Stability weights .

[0040] TTT optimization loop: Freeze the parameters of the CNN feature extractor.

[0041] Parameters for the master regression head and the TTT reconstruction head only ,conduct The gradient descent optimization is performed.

[0042] The goal of optimization is to minimize the self-supervised loss. .

[0043] Calculations based on TTT head reconstruction Reconstructed spectrum and minimize its relationship with The difference between them (MSE).

[0044] Minimize the prediction in the current step Compared with the initial prediction obtained in step 2 The differences between them.

[0045] Obtaining the final prediction: After 15 optimization steps, a set of MLP parameters adapted to the characteristics of Taihu Lake was obtained. Use this and original features Then, forward propagation is performed again to obtain the final, adapted prediction value. .

[0046] Inversion output: for Perform exponential transformation The final Chl-a concentration value for that pixel is obtained.

[0047] Final concentration inversion: Based on the adapted model parameters, the lake chlorophyll concentration inversion results are output through the main regression module. During testing, the feature extractor parameters are frozen in a loop during training optimization, and only the reconstruction head parameters in the main regression module and the test-time adaptive module are optimized. The optimization objective is a composite loss that includes spectral reconstruction loss, stability loss, and prediction range constraint loss.

[0048] The technical effectiveness verification compares the method of this invention (TTT) with seven static models that do not use the TTT mechanism, such as CNN, XGBoost, and SVR. The verification dataset consists of 936 independent global samples.

[0049] Accuracy Comparison: As shown in Table 2, the TTT model significantly outperforms all the comparison models in both R² (0.797) and RMSE (0.584 μg / L). For example, its R² is 42% higher than that of the standard CNN model (0.561).

[0050] Deviation comparison: such as Figure 2 As shown in (a), the scatter plot of predicted vs. measured values ​​for the TTT model is closely distributed around the 1:1 line, with a mean residual of 0.033, indicating no significant systematic bias. In contrast, as... Figure 2The XGBoost model shown in (d) exhibits a clear trend of prediction range compression and underestimation (mean residual -0.117).

[0051] Application Results: Applying this method to nationwide Sentinel-2 imagery from 2018 to 2025, over 170,000 lake observation data points were retrieved. Retrieval results (e.g.) Figure 3 The study successfully revealed the spatiotemporal dynamics of eutrophication in Chinese lakes, including the "extremely high risk" status in the eastern plain area (44.7% algal bloom rate) and the significant deterioration trend in the northeastern plain area (Sen's slope = +1.871 μg / L / yr), demonstrating the robustness and practicality of the invention in long-term, large-scale monitoring.

[0052] Microscale application (specific examples): To further verify the application effect of the present invention on specific targets, five representative lakes (Chaohu Lake, Dianchi Lake, Lianhuan Lake, Bosten Lake, and Eling Lake) were selected, and the spatiotemporal evolution of Chl-a concentration from summer 2019 to 2024 was inverted using the TTT model of this invention.

[0053] A test-based adaptive remote sensing inversion system for lake chlorophyll concentration includes: Data acquisition and preprocessing module: used to acquire in-situ spectral data for training, corresponding chlorophyll concentration data and remote sensing image data for testing, and to preprocess the data; Model training module: used to build a dual-task architecture model, and uses a joint loss function to train the model end-to-end; Specifically, the data acquisition and preprocessing module works by acquiring the in-situ spectral data required for training and the corresponding measured chlorophyll concentration data, as well as the remote sensing image data required for testing. These raw data are then standardized to eliminate format differences and noise interference from different sources, ensuring the consistency and reliability of the input data and providing a high-quality data foundation for subsequent model training. The model training module, based on the preprocessed dataset, constructs a dual-task architecture comprising a feature extraction section, a main regression task module, and an auxiliary task module. A joint loss function combines the prediction error of the main task with the feature reconstruction error of the auxiliary task, guiding the model through end-to-end training. This allows the model to accurately learn the mapping relationship between chlorophyll concentration and spectral features while simultaneously enhancing its ability to characterize the complex optical properties of water bodies through the auxiliary task. Ultimately, this results in an initial model with basic predictive capabilities and a stable structure, laying the foundation for adaptive optimization in the subsequent testing phase.

[0054] The adaptive inversion module performs domain offset detection, dynamic parameter fitting, and test-time training optimization loops to achieve online adaptation of model parameters. It includes a domain offset detection unit, a parameter selection unit, and an optimization unit. The domain offset detection unit calculates the distribution difference between the test and training data, the parameter selection unit selects suitable parameters based on this difference, and the optimization unit performs fine-tuning of the parameters.

[0055] Specifically, a functional closed loop is formed through the coordinated operation of three internal units to achieve online dynamic adaptation of model parameters, addressing potential distribution differences between lake water data and training data during the testing phase: The domain offset detection unit first quantifies the distribution differences between the test and training data by comparing key statistical features (such as central tendency and dispersion) to provide an objective basis for subsequent parameter adaptation; the parameter selection unit selects suitable adaptation parameters (such as learning rate, optimization steps, and stability control parameters) from a preset set of reasonable parameter configurations based on the aforementioned distribution difference degree, ensuring that the parameter settings conform to the current data distribution characteristics; the optimization unit, based on the selected adaptation parameters, extracts some parameters from the core features in the model that are not sensitive to changes in data distribution, and only fine-tunes the parameters of the main regression module and the auxiliary self-supervised task module directly related to chlorophyll a concentration inversion. Through iterative optimization, the model parameters are gradually adapted to the distribution characteristics of the test data, while avoiding the decrease in model prediction stability caused by excessive parameter adjustment. Ultimately, the model is accurately adapted to the current lake water data during the testing phase, ensuring the reliability of the inversion results.

[0056] Results output module: Used to output the lake chlorophyll concentration inversion results based on the adapted model parameters.

[0057] Figure 1 A schematic diagram of the enhanced TTT model architecture proposed in this invention is shown. This architecture consists of three core modules: CNN spectral feature extractor: This is used as a shared backbone network to process input remote sensing spectral data (such as the 12 bands of Sentinel-2). Figure 1 As shown, in order to capture multi-scale spectral features, the input data... The feature vector is passed in parallel through three 1D convolutional (1D-CNN) layers with different kernels (e.g., 1x1, 5x1, 7x1). The multi-scale outputs are concatenated and fed into a deeper CNN module for feature fusion and refinement (including GeLU activation, max pooling, average pooling, etc.), ultimately generating a robust feature vector f.

[0058] The MainRegressionHead is a multilayer perceptron (MLP) responsible for the core Chl-a prediction task. It receives the feature vector f and predicts the log-transformed chlorophyll a concentration. .

[0059] Test-Time Adaptive Module (TTTOptimization): The core of this invention. It includes an auxiliary self-supervised TTT header (for input spectrum reconstruction) and an online parameter adaptation loop. The technical solution of this invention is implemented in stages: The joint training phase involves jointly optimizing the model to simultaneously handle the primary task (regression) and the auxiliary task (reconstruction). The total loss function... It is the main regression loss and spectral reconstruction loss Weighted sum: ; where L main It can be mean squared error (MSE) or Smooth L1 loss. It is the MSE loss from spectral reconstruction. This involves balancing the auxiliary task weights of the two. This stage trains the entire network, including the CNN extractor and two MLP heads.

[0060] The Inference Phase during testing occurs after the model has been deployed and is adapted to new test samples. At that time, the following steps are performed: Domain offset detection (Biasdetection): The model first quantizes the test samples. The distribution difference between the training and training domains. The model maintains reference statistics (mean) from the training domain. and standard deviation ).calculate Batch mean and standard deviation Offset It can be quantified using the following formula: ; in It is a tiny value used to maintain numerical stability.

[0061] Dynamic parameter adaptation: based on the calculated offset. The model selects a set of hyperparameters (learning rate at test time) from a predefined logical table (as shown in Table 1). Adapt to step count Stability weights ).

[0062] Table 1. Examples of TTT Adaptive Logic Hyperparameter Configuration

[0063] TTT Feedback Optimization Loop: In this stage, the parameters of the CNN feature extractor are frozen, and only the MLP parameters (main regression module and TTT reconstruction module) most relevant to the domain shift are optimized. The optimization objective is to minimize a composite self-supervised loss. : in: This is the loss for spectral reconstruction (the core of self-supervised tasks). For stability loss. Penalty for prediction during the adaptation process. Compared with the original prediction before adaptation The deviation between them ensures that the model does not drift excessively. To constrain the prediction range and prevent predictions from "drifting" into unrealistic ranges.

[0064] Final prediction: After the adaptation loop is complete, the optimized master regression module will be used. Inference is performed on feature f to generate the final Chl-a concentration prediction value adapted to the current domain. .

[0065] As shown in Table 2, the proposed TTT model was comprehensively compared with seven mainstream machine learning algorithms (including traditional MLP, CNN, XGBoost, SVR, etc.) on a globally representative independent validation set containing 936 samples. The results show that the TTT model performs best on all evaluation metrics, with a determination coefficient (R²) of 0.797, significantly outperforming traditional MLP (R²=0.604) and CNN (R²=0.561). Regarding error control, the TTT model's root mean square error (RMSE) and mean absolute error (MAE) are 0.584 μg / L and 0.414 μg / L, respectively, representing reductions of 11%–21% and 18%–26% compared to traditional methods.

[0066] Table 2 Comparison of the accuracy of different models in retrieving chlorophyll a concentration

[0067] Overcoming systematic biases: such as Figure 2 The scatter plot showing the predicted values ​​versus the measured values ​​illustrates the TTT model of this invention. Figure 2a) It exhibits the best predictive consistency and the smallest systematic bias (mean residual is close to 0.033). Other traditional models (such as...) Figure 2 b-2f) exhibit varying degrees of "prediction range compression" (i.e., underestimating high concentration values ​​and overestimating low concentration values) or significant systematic negative bias (such as XGBoost). This demonstrates that TTT's online adaptive mechanism effectively overcomes the systematic bias of static models when facing different concentration ranges.

[0068] Spatiotemporal application effect: This invention successfully inverted the spatiotemporal distribution of Chl-a concentration in 2,621 lakes in China from 2018 to 2025 by applying the TTT model. Figure 3 This paper presents a spatiotemporal distribution map of Chl-a concentration in Chinese lakes based on the TTT model inversion. The results clearly reveal the macroscopic spatial pattern of "high in the east and low in the west" in lakes across the country (for example, the average concentration in the eastern plains is 31.29 μg / L, while it is 13.83 μg / L in the Qinghai-Tibet Plateau region), and successfully identify "extremely high-risk" areas (such as the eastern plains, where the algal bloom rate is 44.7%), verifying the stability and effectiveness of this invention in large-scale, long-term remote sensing monitoring applications.

[0069] Figure 4 visually illustrates the results of the inversion. For example, Figure 4(a) clearly shows that the Chl-a concentration (red high concentration area) in Chaohu Lake (eastern plain) has decreased significantly year by year; while Figure 4(c) shows that the Chl-a concentration (red area) in Lianhuan Lake (northern plain) has a clear trend of expansion and deterioration between 2019 and 2024.

[0070] Table 3 quantifies this effect. The Chl-a concentration in Chaohu Lake decreased by 58.89% over 6 years (annual average change of -4.36 μg / L / yr), while that in Lianhuanhu Lake increased by 25.19% (annual average change of +0.75 μg / L / yr).

[0071] This embodiment demonstrates that the TTT method of the present invention is not only accurate at the macroscopic level, but also highly robust and practical at the microscopic level, and can serve as a reliable tool to track the treatment effect or deterioration trend of specific water bodies.

[0072] Table 3. Statistics on Chl-a concentration inversion trends of representative lakes (2019-2024)

[0073] In summary, this invention introduces an adaptive optimization mechanism during the testing phase, combined with domain offset detection and dynamic parameter adaptation processes. This enables the model to perform self-supervised parameter fine-tuning using unlabeled test data, dynamically adapting to the differences in lake water data distribution across different time periods and geographical regions. This effectively improves the problem of declining generalization performance of existing static remote sensing inversion models in new monitoring scenarios, enhancing the model's applicability in diverse lake water monitoring tasks. Furthermore, relying on a dual-task architecture for joint training, combined with multi-scale one-dimensional convolutional neural networks to extract spectral features, and optimized through a composite loss including stability loss and prediction range constraints, this not only strengthens the model's ability to capture complex water optical characteristics but also reduces prediction biases that are prone to occur in traditional models, thereby improving the accuracy of lake chlorophyll concentration inversion results.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A test-based adaptive remote sensing inversion method for lake chlorophyll concentration, characterized in that, Includes the following steps: Data preparation and preprocessing: Acquire in-situ spectral data and corresponding chlorophyll concentration data for training, acquire remote sensing image data for testing, and preprocess the training and testing data; Model building and joint training: A dual-task architecture model including a feature extractor, a main regression module and a test-time adaptive module is built, and the model is trained end-to-end using a joint loss function; Adaptive inversion during testing: The preprocessed test data is input into the trained model, and the domain offset detection, dynamic parameter adaptation, and test-time training optimization loop are performed sequentially to obtain model parameters that are adapted to the current test data distribution. Final concentration inversion: Based on the adapted model parameters, the lake chlorophyll concentration inversion results are output through the main regression module.

2. The test-based adaptive remote sensing inversion method for lake chlorophyll concentration according to claim 1, characterized in that, The data preprocessing includes normalizing the spectral data and performing logarithmic transformation on the chlorophyll concentration data.

3. The test-based adaptive remote sensing inversion method for lake chlorophyll concentration according to claim 1, characterized in that, The feature extractor is a multi-scale one-dimensional convolutional neural network that extracts features from the input spectral data in parallel using convolutional kernels of different sizes, and generates feature vectors after processing by the feature fusion module.

4. The test-based adaptive remote sensing inversion method for lake chlorophyll concentration according to claim 3, characterized in that, The feature fusion module includes activation operation, max pooling operation and average pooling operation, and the activation operation is GeLU activation.

5. The test-based adaptive remote sensing inversion method for lake chlorophyll concentration according to claim 1, characterized in that, The joint loss function is composed of a weighted sum of the principal regression loss and the spectral reconstruction loss. The principal regression loss is either the mean squared error loss or the smoothing L1 loss, and the spectral reconstruction loss is the mean squared error loss.

6. The test-based adaptive remote sensing inversion method for lake chlorophyll concentration according to claim 1, characterized in that, The domain offset detection quantifies the degree of domain offset by calculating the difference between the statistical features of the test data and the reference statistical features of the training data.

7. The test-based adaptive remote sensing inversion method for lake chlorophyll concentration according to claim 1, characterized in that, The dynamic parameter adaptation selects the corresponding learning rate, adaptation steps, and stability weights based on the degree of domain offset. The degree of domain offset is divided into three levels: slight domain offset, moderate domain offset, and significant domain offset.

8. The test-based adaptive remote sensing inversion method for lake chlorophyll concentration according to claim 1, characterized in that, The test-time training optimization loop freezes the feature extractor parameters, optimizing only the reconstruction head parameters in the main regression module and the test-time adaptive module. The optimization objective is a composite loss that includes spectral reconstruction loss, stability loss, and prediction range constraint loss.

9. A test-based adaptive remote sensing inversion system for lake chlorophyll concentration, used in the test-based adaptive remote sensing inversion method for lake chlorophyll concentration as described in any one of claims 1-8, characterized in that, include: Data acquisition and preprocessing module: used to acquire in-situ spectral data for training, corresponding chlorophyll concentration data and remote sensing image data for testing, and to preprocess the data; Model training module: used to build a dual-task architecture model, and uses a joint loss function to train the model end-to-end; Adaptive inversion module: used to perform domain offset detection, dynamic parameter adaptation, and test-time training optimization loop to achieve online adaptation of model parameters; Results output module: Used to output the lake chlorophyll concentration inversion results based on the adapted model parameters.

10. A test-based adaptive remote sensing inversion system for lake chlorophyll concentration according to claim 9, characterized in that, The adaptive inversion module includes a domain offset detection unit, a parameter selection unit, and an optimization unit. The domain offset detection unit is used to calculate the distribution difference between the test data and the training data. The parameter selection unit is used to select the adaptation parameters based on the distribution difference. The optimization unit is used to perform parameter fine-tuning optimization.

Images