Deep learning method for remote sensing retrieval of lake optical substances based on embedded radiative transfer equation
By embedding a radiative transfer model into a deep learning network, high-precision remote sensing inversion of optical materials in inland lakes was achieved. This solved the problems of poor generalization and overfitting of traditional models, improved the inversion accuracy and robustness, and is suitable for lake eutrophication monitoring and water environment management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING INST OF GEOGRAPHY & LIMNOLOGY
- Filing Date
- 2026-04-27
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies suffer from optical crosstalk in water quality inversion of inland lakes and nearshore seas, making it difficult to invert biogeochemical parameters. Traditional models have poor generalization in cross-regional and cross-seasonal scenarios, and pure data-driven deep learning models are prone to overfitting and lack physical constraints.
The analytical radiative transfer model is hard-coded into a differentiable layer of a deep learning network. Through a two-stage training strategy, a deep learning model is constructed that includes feature extraction, physical constraint reconstruction, and multi-task prediction, achieving optical closure and physical consistency.
It improves the physical consistency and interpretability of the inversion results, solves the problem of poor generalization of traditional models in cross-regional and cross-seasonal scenarios, realizes high-precision remote sensing inversion of lake optical materials, reduces the cost of on-site monitoring, and provides a basis for scientific decision-making.
Smart Images
Figure CN122493313A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of satellite remote sensing and data processing, specifically to a deep learning method for remote sensing inversion of optical matter in lakes with an embedded radiative transfer equation. Background Technology
[0002] Inland lakes and nearshore seas are important indicators of global carbon cycling and the health of terrestrial-aquatic ecosystems. However, these water bodies typically exhibit significant optical complexity, with their water radiance signal influenced by phytoplankton, suspended particulate matter, and other factors. SPM ) and colored soluble organic matter ( CDOM Inherent optical properties ( ) IOPs It is jointly controlled by highly nonlinear and dynamic interactions between different optically active components ( ). OACs Combinations of these may produce approximately the same remote sensing reflectance. The so-called optical crosstalk phenomenon (manifested as significant spectral aliasing and overlapping absorption features) is essentially an "illness inversion" problem in the inversion of biogeochemical parameters, which remains one of the core challenges in the fields of satellite oceanography and limnology.
[0003] From a methodological development perspective, water quality inversion algorithms have long faced a trade-off between physical interpretability and empirical flexibility. Traditional empirical models and semi-empirical models (such as band ratios or regression-based models) heavily rely on synchronously measured data to establish statistical relationships, thus exhibiting significant regional dependence and often failing to achieve effective generalization under different optical water conditions. In contrast, semi-analytical algorithms ( SAAs ), such as quasi-analytic algorithms ( QAA Based on the theory of radiative transfer, but relying on a pre-set and rigid parameterization scheme (such as a fixed spectral shape of phytoplankton absorption coefficient), it often fails in highly turbid or hypereutrophic waters.
[0004] In recent years, machine learning methods, including deep learning, have demonstrated powerful nonlinear mapping capabilities. However, purely data-driven neural networks are essentially "black box" models lacking explicit physical constraints, making them prone to overfitting, exhibiting significantly reduced generalization performance in out-of-distribution scenarios, and potentially producing non-physical intermediate results (such as negative values) that violate fundamental principles of biooptics. To bridge the gap between physical consistency and data-driven optimization, physically-guided deep learning (Physically Guided Deep Learning) has emerged as a solution. PGDL In recent years, water color remote sensing has gradually emerged in the field of Earth sciences. However, existing methods in water color remote sensing... PGDLMany methods incorporate physical knowledge as empirical regularization terms (i.e., soft constraints) into the loss function, without fundamentally altering the network structure. During gradient-based optimization, the model may still generate physically meaningless intermediate variables, and simply reducing output error through mathematical optimization makes it difficult to achieve true structural optical closure. Summary of the Invention
[0005] To address the aforementioned issues, this invention proposes a deep learning-based method for remote sensing inversion of optical materials in lakes with embedded radiative transfer equations. This invention aims to break the "black box" paradigm of traditional pure data-driven deep learning models by directly hard-coding the analytical radiative transfer model as a differentiable layer into the network topology, thereby achieving true structured optical closure.
[0006] This invention employs the following technical solution: a deep learning-based lake optical material remote sensing inversion method with embedded radiative transfer equations, comprising the following steps: Noise-free simulated optical data is generated in advance, and measured optical data is acquired; both the simulated and measured optical data include remote sensing reflectance data and corresponding water body optical data; based on the target sensor, differentiated preprocessing is performed on the remote sensing reflectance data and corresponding water body optical data to construct training and validation sets; A deep learning network model is constructed that includes a feature extraction module, a physical constraint reconstruction module, and a multi-task prediction module. The radiative transfer equation is hard-coded into a differentiable forward operator layer and embedded into the physical constraint reconstruction module. A two-stage training strategy of pre-training with simulated optical data and fine-tuning with measured optical data is adopted. Based on the training set and validation set, the deep learning network model is optimized using a multi-task loss function. The remote sensing reflectance of the remote sensing image to be inverted is input into the optimized deep learning network model, which outputs various optical material characteristics of the target lake.
[0007] In a further embodiment, the differential preprocessing includes the following steps: Obtain the relative spectral response function of the target sensor For remote sensing reflectance data, the relative spectral response function is used. Integral convolution processing is performed to convert the hyperspectral remote sensing reflectance... Convert to the corresponding target sensor reflectivity data for the band ; The optical data of the water body are classified into optical parameters of the absorption coefficient category. and optical component concentration parameters , The center wavelength of each band of the sensor; for the absorption coefficient-like optical parameters, the hyperspectral absorption coefficient... Take the reciprocal, and combine it with the relative spectral response function. Obtain the weighted average value of the corresponding band. Take the weighted average value. The reciprocal of the first digit gives the second digit. Average absorption coefficient of the band ; For optical component concentration parameters A base-10 logarithmic transformation was performed to obtain the transformed concentration data. .
[0008] In a further embodiment, the feature extraction module is configured to use reflectance data. As input, multi-scale spectral features are extracted through multi-layer convolution and attention mechanisms, and high-dimensional feature vectors related to the optical properties of water bodies are output. The physical constraint reconstruction module reconstructs the phytoplankton absorption coefficient based on the aforementioned high-dimensional feature vector. Non-algal particulate matter CDOM absorption coefficient and particulate backscattering coefficient The total absorption coefficient was calculated by combining the optical properties of pure water. and total backscattering coefficient Through hard-coded Gordon Theoretical reconstruction of remote sensing reflectance using an approximate analytical expression for radiative transfer This achieves optical closure within the model. The multi-task prediction module is configured to reconstruct remote sensing reflectance based on the theory. It can independently predict multiple optical material characteristics of the target water body.
[0009] In a further embodiment, the two-stage training strategy specifically includes the following training process: The first stage is the pre-training stage: the network is trained using noise-free simulated optical data to learn the physically consistent positive and negative mapping relationships between apparent optical properties and inherent optical properties, and to theoretically reconstruct remote sensing reflectance. Cross-validation of absorption coefficient and backscattering coefficient; The second stage is the fine-tuning stage: using the weights of the pre-trained model as initial values, the network parameters are fine-tuned and optimized using in-situ measured optical data of global nearshore and inland water bodies that contain real environmental variability.
[0010] In a further embodiment, the multi-task loss function is expressed as follows: ; in, This is the loss value. For physical reconstruction loss, For the loss in factor concentration inversion, Regularization terms are used to prevent overfitting. , and These are the corresponding weight hyperparameters.
[0011] In a further embodiment, the various optical material features include: chlorophyll a Concentration, suspended particulate matter concentration, and 443 nm The absorption coefficient of colored soluble organic matter at the location; The target sensor includes: Sentinel -2 series satellite sensors, Sentinel -3 series satellite sensors, S-NPP / NOAA-20 / NOAA-21 One or more of a series of satellite sensors.
[0012] In a further embodiment, the phytophyte absorption coefficient Dynamically construct the principal component weights by extracting them using empirical orthogonal functions: The mean phytoplankton absorption coefficient was obtained based on spectral data from a global standard water color dataset. First principal component weights Second principal component weights After saving, dynamic reconstruction is performed using a linear reconstruction formula to ensure differentiability of network transmission.
[0013] In a further embodiment, the non-algae particles / CDOM absorption coefficient The reconstruction formula is as follows: ; In the formula, 443 nm Non-algal particulate matter at the location / CDOM Absorption coefficient, This represents the spectral slope.
[0014] In a further embodiment, the particulate backscattering coefficient The reconstruction formula is as follows: ; In the formula, 443 nm The backscattering coefficient of particulate matter at that location, It is a power function sampling quantity.
[0015] In a further embodiment, the theoretically reconstructed remote sensing reflectance The reconstruction process is as follows: Will GordonThe radiative transfer approximation model is hard-coded into a differentiable tensor operation layer without training parameters. The total absorption coefficient is calculated using the following formulas. and total backscattering coefficient : ; ; in, The absorption coefficient of pure water. The backscattering coefficient of pure water; Based on the total absorption coefficient and total backscattering coefficient The backscattering ratio was calculated. : ; Using the backscatter ratio The remote sensing reflectance was calculated to be exactly below the water surface. : ;in, and All are preset constants; Based on the remote sensing reflectance that is just below the water surface The theoretical reconstructed remote sensing reflectance was calculated. : .
[0016] The beneficial effects of this invention are as follows: This invention directly embeds the analytical radiative transfer model into the deep learning network topology, and constrains the model training process with physically interpretable radiative transfer equations. This completely breaks through the technical bottleneck of traditional pure data-driven models that rely solely on data fitting and lack physical mechanism support. It realizes a paradigm upgrade from data fitting to physical constraints combined with data driving, constructs a true structured optical closure, and significantly improves the physical consistency and interpretability of the inversion results.
[0017] The hard-coded constraints of the radiative transfer model provide the model with universal water optical physics priors, enabling the model to maintain stable inversion accuracy in lake scenarios with different hydrological conditions and different optical types (shallow / deep water, eutrophic / oligotrophic) without relying on large-scale, full-coverage measured training data. This effectively solves the problems of poor generalization and severe overfitting of traditional models in cross-regional and cross-seasonal scenarios, and significantly improves the robustness and engineering applicability of the model.
[0018] This invention hardcodes the radiative transfer model into a differentiable tensor operation layer without training parameters, which preserves the accuracy and interpretability of the physical model while avoiding the training burden and overfitting risk brought about by the introduction of additional parameters. At the same time, the differentiable structure supports end-to-end gradient backpropagation without the need to split the physical model and deep learning module, which greatly simplifies the model training process, improves inversion efficiency, and achieves triple optimization of accuracy, efficiency and interpretability.
[0019] This invention can achieve the reduction of total nitrogen, total phosphorus, and chlorophyll in lakes. a High-precision remote sensing inversion of key water quality parameters such as suspended solids provides accurate and efficient technical means for monitoring lake eutrophication, water environment management, and ecological protection. It can significantly reduce the manpower and material costs of traditional field monitoring, provide scientific decision-making basis for watershed water environment management and water ecological restoration, and has significant social and economic benefits and promotion value. Attached Figure Description
[0020] Figure 1 This is a spatial and statistical distribution map of the training points in this embodiment; wherein, Figure ( a ), ( b )and( c ) are chlorophyll a concentration( Chl - a ), suspended solids concentration ( SPM )and a g (443) frequency distribution diagram.
[0021] Figure 2 This is the embodiment. OCRT - Net A schematic diagram of the overall model.
[0022] Figure 3 This is the embodiment. OCRT - Net The model reconstructs the relevant spectral representation on the simulated dataset; where, ( a A comparison diagram of the reconstructed remote sensing reflectance spectrum and the true spectrum. b ( ) is the reconstructed spectrum of the phytoplankton absorption coefficient. c (This refers to non-algae particulate matter) CDOM Reconstructed spectrum of absorption coefficient, ( d ( ) is the reconstructed spectrum of the backscattering coefficient of particulate matter.
[0023] Figure 4 This is the embodiment. OCRT - Net The model's performance inverted on the measured datasets from three sensors.
[0024] Figure 5This is the embodiment. OCRT - Net The model covers 12 lakes in China Sentinle -3 OLCI Verification results diagram at satellite matching points.
[0025] Figure 6 This is the embodiment. OCRT - Net The model shows the lake groups of the Qinghai-Tibet Plateau and the lake groups of the eastern plains at a certain moment. Sentinel -3 OLCI The inversion results are shown in the satellite image. Detailed Implementation
[0026] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.
[0027] Example 1 This embodiment discloses a deep learning-based remote sensing inversion method for lake optical matter with embedded radiative transfer equations, characterized by the following steps: Noise-free simulated optical data is generated in advance, and measured optical data is acquired; both the simulated and measured optical data include remote sensing reflectance data and corresponding water body optical data; based on the target sensor, differentiated preprocessing is performed on the remote sensing reflectance data and corresponding water body optical data to construct training and validation sets; it should be noted that the simulated optical data described in this embodiment is generated using... HydroLight The model simulation generates remote sensing reflectance data and corresponding water body optical data covering the typical variation range of inland lakes. Furthermore, the water body optical data described in this embodiment includes at least: synchronized chlorophyll... a Suspended matter and CDOM Concentration data of optical substances.
[0028] A deep learning network model is constructed that includes a feature extraction module, a physical constraint reconstruction module, and a multi-task prediction module. The radiative transfer equation is hard-coded into a differentiable forward operator layer and embedded into the physical constraint reconstruction module. A two-stage training strategy of pre-training with simulated optical data and fine-tuning with measured optical data is adopted. Based on the training set and validation set, the deep learning network model is optimized using a multi-task loss function. The remote sensing reflectance of the remote sensing image to be inverted is input into the optimized deep learning network model, which outputs various optical material characteristics of the target lake.
[0029] To address the technical issues of mismatch between hyperspectral remote sensing reflectance data and the multispectral bands of the target sensor, and the tendency for direct input into deep learning models to lead to band misalignment and radiance distortion; and to address the problems of information loss in absorption coefficient-type optical parameters under limited sensor bands and difficulties in training convergence due to large dynamic range differences in optical component concentration parameters, the following differentiated preprocessing is performed: Acquire target sensor (400-800) nm The relative spectral response function of ) Using the relative spectral response function Integral convolution processing is performed to convert the hyperspectral remote sensing reflectance... Convert to the corresponding target sensor reflectivity data for the band , ; ; The optical data of the water body are classified into optical parameters of the absorption coefficient category. and optical component concentration parameters , The center wavelength of each band of the sensor; for the absorption coefficient-like optical parameters, first, the hyperspectral absorption coefficient... Take the reciprocal, which is Combined with the aforementioned relative spectral response function Obtain the weighted average value of the corresponding band. This can be understood by referring to the following formula: ; For the weighted average Taking the reciprocal, we get the first... Average absorption coefficient of the band The formula is expressed as: ; For optical component concentration parameters A base-10 logarithmic transformation was performed to obtain the transformed concentration data. The conversion formula is: .
[0030] Based on the preprocessed data obtained through the aforementioned differential processing, during model training, the preprocessed data is randomly divided into a training set and a validation set at a ratio of 80-20%. The target sensors mentioned in this embodiment include: Sentinel -2 series satellite sensors, Sentinel -3 series satellite sensors and S-NPP / NOAA-20 / NOAA-21 One or more of a series of satellite sensors.
[0031] In a further embodiment, the feature extraction module is configured to use reflectance data. As input, multi-scale spectral features are extracted through multi-layer convolution and attention mechanisms, and high-dimensional feature vectors related to the optical properties of water bodies are output. The physical constraint reconstruction module reconstructs the phytoplankton absorption coefficient based on the aforementioned high-dimensional feature vector. Non-algal particulate matter [[ID=4 absorption coefficient and particulate backscattering coefficient The total absorption coefficient was calculated by combining the optical properties of pure water. and total backscattering coefficient Through hard-coded Theoretical reconstruction of remote sensing reflectance using an approximate analytical expression for radiative transfer This achieves optical closure within the model. The multi-task prediction module is configured to reconstruct remote sensing reflectance based on the theory. It independently predicts multiple optical material characteristics of the target water body. It should be noted that the multiple optical material characteristics described in this embodiment include: chlorophyll. a Concentration, suspended particulate matter concentration, and 443 The absorption coefficient of colored soluble organic matter at the location.
[0032] In a further embodiment, the phytophyte absorption coefficient Dynamically construct the principal component weights by extracting them using empirical orthogonal functions: The mean phytoplankton absorption coefficient was obtained based on spectral data from a global standard water color dataset. First principal component weights Second principal component weights After saving, dynamic reconstruction is performed using the linear reconstruction formula to ensure differentiability of network transmission. For ease of understanding, the linear reconstruction formula described in this embodiment is as follows: ; In the formula, and These are the dynamic weight coefficients obtained through network learning.
[0033] To address the technical problems of high spectral dimensionality of phytoplankton absorption coefficients, easy network overfitting and poor physical consistency caused by direct mapping, this embodiment uses an empirical orthogonal function ( Principal component linear reconstruction enables low-dimensional, physically consistent dynamic reconstruction of phytoplankton absorption coefficients while ensuring network transmission differentiability, thereby improving model generalization ability and inversion accuracy.
[0034] In a further embodiment, the non-algae particles described in this embodiment / absorption coefficient The reconstruction formula is as follows: ; In the formula, 443 Non-algal particulate matter at the location / Absorption coefficient, The spectral slope has a value between 0.008 and 0.026.
[0035] The above reconstruction formula is used to address the issue of non-algal particulate matter / Given the physical property that the absorption coefficient decays exponentially with wavelength, directly inputting the entire wavelength range into a neural network can easily lead to redundant parameters, distortion of radiation, and technical problems that cannot guarantee the differentiability of the radiative transfer equation.
[0036] In a further embodiment, the particulate backscattering coefficient The reconstruction formula is as follows: ; In the formula, 443 The backscattering coefficient of particulate matter at that location, It is a power function sample, which follows a normal distribution with a mean of -1 and a standard deviation of 0.6.
[0037] This embodiment uses a power-law model to reconstruct the physical constraints of the particulate backscattering coefficient. Under the premise of ensuring differentiability of network transmission, it solves the problems that the spectral morphology of the particulate backscattering coefficient is greatly affected by the particle size distribution, direct full-band mapping easily leads to the network learning non-physical mapping relationships, and the high parameter dimensionality.
[0038] Based on this, the theoretical reconstruction of remote sensing reflectance described in this embodiment The reconstruction process is as follows: Will The radiative transfer approximation model is hard-coded into a differentiable tensor operation layer without training parameters. The total absorption coefficient is calculated using the following formulas. and total backscattering coefficient : ; ; in, The absorption coefficient of pure water. This is the backscattering coefficient of pure water.
[0039] Based on the total absorption coefficient and total backscattering coefficient The backscattering ratio was calculated. : ; Using the backscatter ratio The remote sensing reflectance was calculated to be exactly below the water surface. : ;in, and These are all preset constants, typically taking values of 0.0895 and 0.1247 respectively; Based on the remote sensing reflectance that is just below the water surface The theoretical reconstructed remote sensing reflectance was calculated. : .
[0040] In another embodiment, the two-stage training strategy specifically includes the following training process: The first stage is the pre-training stage: the network is trained using noiseless simulated optical data, with an adaptive learning rate optimizer (such as...). Pre-training is performed under these conditions. The physically consistent positive and negative mapping relationships between apparent optical properties and inherent optical properties are learned, and the theoretical reconstruction of remote sensing reflectance is achieved. Cross-validation of absorption coefficient and backscattering coefficient; The second stage is the fine-tuning stage: using the weights of the pre-trained model as initial values, in-situ measured optical data of global nearshore and inland water bodies that include real-world environmental variability (such as... The network parameters were fine-tuned and optimized using the dataset to improve the model's generalization ability to real lake water bodies, completing the two-stage training.
[0041] To address the technical issues of pure data-driven deep learning models being prone to learning non-physical mapping relationships and having poor generalization ability, a two-stage training strategy is adopted. First, a physical consistency weight baseline is established using noise-free simulated data, and then fine-tuning and optimization are performed using measured data. Under the premise of ensuring the physical consistency of the model, the generalization ability and inversion accuracy of the model to actual lake water bodies are significantly improved.
[0042] In a further embodiment, the multi-task loss function used in this embodiment is expressed as follows: ; in, This is the loss value. For physical reconstruction loss, This represents the loss from the element concentration inversion, specifically the root mean square error between the measured and predicted values. To prevent overfitting, an L1 regularization term from the neural network is used. , and These are the corresponding weight hyperparameters.
[0043] It should be noted that the physical reconstruction loss Mean square error (MSE) ) and spectral angle mapping ( The mixed metrics are specifically: ; in, This is the root mean square error value, used to constrain absolute differences in energy levels. It is a spectral angle mapping used to penalize the difference in geometry (angle) between the real spectrum and the reconstructed spectrum in multidimensional space, so as to highlight absorption features and suppress systematic bias caused by atmospheric correction; For balance coefficient, This refers to the input measured optical data / simulated optical data.
[0044] Based on the above description, the deep learning-based lake optical material remote sensing inversion method with embedded radiative transfer equations presented in this embodiment can be applied for verification and practical inversion according to the following process: First, model validation was conducted at the satellite-to-ground synchronization matching point to evaluate the model's robustness and sensitivity to atmospheric correction uncertainties. Secondly, we will conduct inversion applications of the model on satellite imagery to test the model's generalization ability and inversion stability on image spatiotemporal texture. Specifically, typical lake areas in the Qinghai-Tibet Plateau and the eastern plains of my country... -3 Satellite imagery, After completing preprocessing such as atmospheric correction and radiometric calibration, the system performs quality control on the image pixels; the spectra of the quality-controlled pixels are then input into the trained image. - The model inverts and outputs the spatial distribution of optical material concentration in lakes, enabling rapid and accurate remote sensing inversion of optical elements in large-scale lake water bodies.
[0045] Multi-element lake optical materials (chlorophyll) measured in the study area a Suspended particulate matter a g (443) Data and synchronization -3 Based on satellite remote sensing reflectance data, high-quality reflectance spectra are obtained through rigorous atmospheric correction; then... 410-778 Using the 12 water color feature bands as input, a deep learning network with an embedded radiative transfer equation is used ( - ), developed applicable -3 A joint inversion model of multiple optical elements for inland lakes was developed. Then, the algorithm was applied to typical representative lake cases to evaluate its inversion accuracy and physical consistency performance. Finally, the algorithm was comprehensively applied to the lake groups of the Qinghai-Tibet Plateau in China (such as Qinghai Lake, Siling Co, Nam Co and other high-altitude clear lakes) and the lake groups of the eastern plains (such as Taihu Lake, Chaohu Lake, Hongze Lake and other shallow eutrophic lakes) to obtain high-precision spatiotemporal distribution characteristics of multiple optical substances in complex optical water bodies.
[0046] As an exemplary description, the implementation of the aforementioned method will be specifically explained below with reference to the accompanying drawings.
[0047] Step 1: Acquisition and preprocessing of datasets for specific sensors.
[0048] 1) Utilize The model simulation generates remote sensing reflectance data and corresponding water body optical data covering the typical variation range of inland lakes. (A total of approximately 120,000 groups); and simultaneously sorting A global in-situ measured matching dataset, totaling 2097 samples. To make the model applicable... -3 Images were used to extract 12 main water color bands (e.g., 412) from the corresponding sensors in the visible to near-infrared range. 442 490 510 560 620 665 673 681 708 754 and 779 The relative spectral response function of ) ).
[0049] 2) Compare the above hyperspectral data with the corresponding relative spectral response function ( Integral convolution calculations are performed to generate discrete 12 bands. Simulated reflectance was used as an input baseline for the deep learning network model; concentration-labeled chlorophyll was analyzed. a Suspended particulate matter concentration ( ), a g (443) These three types of concentration labels are used for The transformation converts the long-tailed distribution of the original data into an approximately normal distribution, improving the convergence stability of model training.
[0050] Will Simulated data and The measured data were randomly divided into training and validation sets, with a ratio of 80%-20%.
[0051] Step 2 -3 Image Atmospheric Correction and Satellite-Ground Synchronization Matching Dataset Generation Download from the European Space Agency -3 of L 1 B Level satellite imagery; combined with auxiliary meteorological data and Rayleigh scattering lookup tables, in (National Oceanic and Atmospheric Administration) 12 software for L 1 B Atmospheric correction processing is performed on the high-resolution imagery to obtain remote sensing reflectance data.
[0052] Based on existing lake sampling data, such as As shown, the following criteria were used to screen satellite-to-ground synchronization matching data: ① Samples covered by clouds, solar flares, and algal blooms were removed by combining measured records and satellite imagery; ② ±3 h The time window was used to achieve time matching between satellite imagery and in-situ measured data; ③ Within the 3×3 window corresponding to each sample point, the spectral variation coefficient was controlled within 10%. Based on the above standards, 89 satellite-to-ground synchronous matching datasets were finally obtained, distributed across 12 lakes in my country.
[0053] Step 3, Adaptation Deep learning networks for sensors ( - Architecture construction and parameter setting The deep learning network shown in this embodiment is an example. - The basic framework of the network is as follows. The input layer dimension is set to 12, corresponding to... -3 The sensor's 12 water color bands; through including (Batch normalization) and Multilayer perceptron with activation function ( The following parameters will be output sequentially: 1) Non-algal particulate matter / Absorption coefficient related parameters: 443 Non-algal particulate matter at the location / absorption coefficient and spectral slope ; Parameters related to particulate backscattering coefficient: power function sampling ; 2) Phytophyte absorption coefficient To address the complex spectral morphology of phytoplankton absorption coefficients in lakes, instead of using a single fixed model, empirical orthogonal functions are employed. Extracting the weights of the first principal component from real water bodies Second principal component weights The network outputs the weight coefficients of these two principal components, and dynamically generates the corresponding adapters for 11 bands through linear combination. Phytoplankton absorption coefficient of the sensor; 3) Physical constraint reconstruction: Through tensor concatenation operations, all inherent optical property parameters output above are summarized and substituted into... The radiative transfer approximation equation is used for remote sensing reflectance reconstruction; since the entire calculation process only involves addition, subtraction, multiplication, division, and exponentiation, Framework (version) v 2.2) This layer can achieve full differentiability, and ultimately outputs the theoretical reconstructed remote sensing reflectance for 12 bands. ; 4) Multi-task prediction: Construct three parallel fully connected subnetworks at the end of the network, each outputting the transformed data. 10 ( - a ), 10 ( )and [ a g (443)
[0054] Step 4: Loss Function Design and Hyperparameter Setting In this embodiment, the total loss function In this paper, a physical reconstruction loss is introduced to constrain the physical consistency of the model, and its expression is as follows: ; in, This is the loss value. For physical reconstruction loss This represents the loss from the element concentration inversion, specifically the root mean square error between the measured and predicted values. To prevent overfitting, the L1 regularization term of the neural network is used. , and These are the corresponding weight hyperparameters. Set to 2.0. and Setting it to 1.0 guides the network to strictly preserve the absorption by phytoplankton (e.g., 673) while fitting the absolute energy of the spectrum. Chlorophyll absorption valley at 681 The spectral geometry is controlled by the fluorescence peak at the location.
[0055] It should be noted that the physical reconstruction loss Mean square error (MSE) ) and spectral angle mapping ( The mixed metrics are specifically: ; Step 5: Two-stage fine-tuning training of model parameters 1) Pre-training based on simulated data: using The optimizer has an initial learning rate set to 0.001. _ The value is set to 64, and the above 12-band input model is pre-trained. This stage mainly relies on physical reconstruction loss. The parameter updates are guided to establish positive optical closure constraints, ensuring that the model learns a physically consistent positive and negative mapping relationship between apparent optical properties and inherent optical properties.
[0056] The model reconstruction shows the remote sensing reflectance, phytoplankton absorption coefficient, and non-algal particulate matter / Case studies include absorption and particulate backscattering coefficients. Overall, the remote sensing reflectance spectrum reconstructed from the differentiable radiative transport layer embedded in the model closely matches the actual input spectrum. Furthermore, the network can operate without directly inherent optical properties (…). Under label supervision, the underlying optical parameters with clear physical meaning were successfully decoupled from the latent space. Among them, the predicted phytoplankton absorption coefficient accurately reproduced the complex characteristic absorption peaks, and the non-algal absorption and particulate matter backscattering coefficients also perfectly matched the physical exponential decay and power-law spectral shape, which fully confirms that the model truly achieves structured optical closure.
[0057] 2) Fine-tuning based on measured data: Load the model weights obtained in the pre-training stage, reduce the learning rate to 0.0001, freeze some low-level feature extraction parameters, and utilize data that includes real atmospheric interference and instrument noise. In-situ measured dataset, 500 Fine-tuning training improves the model's noise resistance and real-world adaptability; after fine-tuning, model performance is evaluated on an independent validation set.
[0058] If this method needs to be applied , -2 For other satellite sensors, only the relative spectral response function extracted in step 1 needs to be used. Replace the band response function with the corresponding sensor's band response function, and repeat the above pre-training and fine-tuning process. This method has good sensor adaptability.
[0059] This demonstrates the widespread global distribution of the model of this invention. Validation results on the in-situ measured dataset. -3 , -2 and With various sensor band configurations, the reconstructed physical reflectance spectra within the network show high agreement with the measured spectra, effectively overcoming the extreme value bias problem of traditional purely empirical algorithms in high / low concentration regions. Within a wide dynamic range spanning three to four orders of magnitude, the inversion results for the optical substances in the three lakes maintain extremely high consistency with the measured concentrations; Taking sensors as an example, chlorophyll a Inversion determination coefficients of suspended particulate matter The values were as high as 0.86 and 0.77 respectively, and the systematic bias ( All deviations are strictly controlled within ±5%. This fully demonstrates that the embedded radiative transfer physical layer can not only achieve structured optical closure, but also ensure that the model maintains robust inversion performance with high accuracy and low bias when facing real and extremely complex natural water bodies.
[0060] Step 6, the model is in -3 Image-based verification and practical applications The model was applied to the 89 satellite-to-ground matching points generated in step 2 to test its sensitivity to atmospheric corrections. Based on -3 Based on atmospheric correction matching point data from 12 typical lakes, the model constructed in this invention demonstrates robust performance for practical applications. ), among which, chlorophyll a and The inversion results are highly consistent with the measured values. (0.63 and 0.66 respectively), and chlorophyll a Systematic bias in inversion ( The figure is only 0.3%, proving that embedded physical constraints can effectively suppress the propagation of atmospheric correction noise; for suspended particulate matter ( Despite the slight underestimation of the retrieved values due to systematic errors in the satellite input signals themselves, the model can still accurately capture their relative spatial variability characteristics. This demonstrates a powerful ability to resolve element gradients in complex optical water bodies.
[0061] Given that the model's accuracy and uncertainty are well-defined, it can be applied to the target lake. -3 Image inversion, the specific process is as follows: will 12 processed image bands are read in and a panoramic pixel matrix is constructed (denoted as ). M Line × After flattening the matrix, input the result of the fine-tuning. - Model; network output corresponding The transformed concentration value is then subjected to a base-10 exponential operation to restore the linear true concentration; the restored concentration data is then reconstructed... M × N The spatial matrix is used to generate high-precision lake chlorophyll through pseudo-color mapping. a Concentration, suspended particulate matter concentration Thematic map of spatial distribution.
[0062] It shows - The models are located in the lake clusters of the Qinghai-Tibet Plateau (left column, photographed on July 31, 2021) and the lake clusters of the eastern plains (right column, photographed on August 3, 2020). -3 The inversion results from the satellite image, with the first row showing chlorophyll. a Concentration spatial distribution, second row is suspended particulate matter ( Concentration spatial distribution, third line Spatial distribution. As shown in the figure, in chlorophyll... a In terms of suspended particulate matter inversion, the model can accurately capture the highly heterogeneous aggregation characteristics of summer cyanobacterial blooms in shallow lakes in eastern China and the dynamic resuspension process driven by wind and waves. It can stably reproduce the extremely low oligotrophic background of lakes on the Qinghai-Tibet Plateau and the local concentration gradient reduction phenomenon caused by glacial meltwater entering the lakes. At the same time, the model effectively reveals the high concentration of cyanobacterial blooms in lakes in the eastern plains affected by intensive human activities and land-based inputs. The spatial pattern is in stark contrast to the significantly low-value distribution of alpine lakes on the Qinghai-Tibet Plateau.
[0063] Thanks to the effective decoupling of strong optical crosstalk in complex water bodies by the internal physically differentiable layer of the network, this model completely avoids the common problems of high-value saturation underestimation and low-value negative distortion in wide-area concentration inversion spanning multiple orders of magnitude. Overall implementation results show that the deep learning method with embedded radiative transfer equations described in this invention successfully breaks through the generalization bottleneck of traditional "black box" models when dealing with huge geographical altitude differences and extreme water quality changes, achieving robust joint remote sensing monitoring of complex inland water bodies with multiple optical materials and high physical consistency.
[0064] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention.
Claims
1. A method for remote sensing inversion of lake optical substances by embedding a depth learning radiation transfer equation, characterized in that, Includes the following steps: Noise-free simulated optical data is generated in advance, and measured optical data is acquired; both the simulated and measured optical data include remote sensing reflectance data and corresponding water body optical data; based on the target sensor, differentiated preprocessing is performed on the remote sensing reflectance data and corresponding water body optical data to construct training and validation sets; A deep learning network model is constructed that includes a feature extraction module, a physical constraint reconstruction module, and a multi-task prediction module. The radiative transfer equation is hard-coded into a differentiable forward operator layer and embedded into the physical constraint reconstruction module. A two-stage training strategy of pre-training with simulated optical data and fine-tuning with measured optical data is adopted. Based on the training set and validation set, the deep learning network model is optimized using a multi-task loss function. The remote sensing reflectance of the remote sensing image to be inverted is input into the optimized deep learning network model, which outputs various optical material characteristics of the target lake.
2. The deep learning lake water constituent remote sensing retrieval method embedded with the radiative transfer equation according to claim 1, characterized in that, The differential preprocessing includes the following steps: Acquiring relative spectral response function of target sensor , performing integral convolution processing on the remote sensing reflectance data by using the relative spectral response function , and converting hyperspectral remote sensing reflectance into reflectance data corresponding to the first band of the target sensor ; dividing the water body optical data into absorption coefficient-like optical parameters and optical component concentration parameters , λ is the center wavelength of each waveband of the sensor. For the absorption coefficient-like optical parameter, the hyperspectral absorption coefficient Taking the inverse, combined with the relative spectral response function gives the weighted average value for the corresponding waveband ; taking the inverse of the weighted average value gives the average absorption coefficient for the first waveband ; Concentration parameter of optical component Logarithmic conversion processing with base 10 is performed to obtain converted concentration data .
3. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 1, characterized in that, The feature extraction module is configured to take reflectance data as input, extract multi-scale spectral features through multi-layer convolution and attention mechanism, and output a high-dimensional feature vector related to the optical properties of the water body. The physical constraint reconstruction module reconstructs the phytoplankton absorption coefficient based on the aforementioned high-dimensional feature vector. Non-algal particulate matter CDOM absorption coefficient and particulate backscattering coefficient The total absorption coefficient was calculated by combining the optical properties of pure water. and total backscattering coefficient Through hard-coded Gordon Theoretical reconstruction of remote sensing reflectance using an approximate analytical expression for radiative transfer This achieves optical closure within the model. The multi-task prediction module is configured to reconstruct remote sensing reflectance based on the theory. It can independently predict multiple optical material characteristics of the target water body.
4. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 1, characterized in that, The two-stage training strategy specifically follows the following training process: The first stage is the pre-training stage: the network is trained using noise-free simulated optical data to learn the physically consistent positive and negative mapping relationships between apparent optical properties and inherent optical properties, and to theoretically reconstruct remote sensing reflectance. Cross-validation of absorption coefficient and backscattering coefficient; The second stage is the fine-tuning stage: using the weights of the pre-trained model as initial values, the network parameters are fine-tuned and optimized using in-situ measured optical data of global nearshore and inland water bodies that contain real environmental variability.
5. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 1, characterized in that, The multi-task loss function is expressed in the following form: ； in, The loss value. For physical reconstruction loss, For the loss in factor concentration inversion, Regularization terms are used to prevent overfitting. , and These are the corresponding weight hyperparameters.
6. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 1, characterized in that, The various optical material characteristics include: chlorophyll a Concentration, suspended particulate matter concentration, and 443 nm The absorption coefficient of colored soluble organic matter at the location; The target sensor includes: Sentinel -2 series satellite sensors, Sentinel -3 series satellite sensors, S- NPP / NOAA-20 / NOAA-21 One or more of a series of satellite sensors.
7. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 3, characterized in that, The absorption coefficient of the phytophyte Dynamically construct the principal component weights by extracting them using empirical orthogonal functions: The mean phytoplankton absorption coefficient was obtained based on spectral data from a global standard water color dataset. First principal component weights Second principal component weights After saving, dynamic reconstruction is performed using a linear reconstruction formula to ensure differentiability of network transmission.
8. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 3, characterized in that, The non-algae particles / CDOM absorption coefficient The reconstruction formula is as follows: ； In the formula, 443 nm Non-algal particulate matter at the location / CDOM Absorption coefficient, This represents the spectral slope.
9. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 3, characterized in that, The particulate matter backscattering coefficient The reconstruction formula is as follows: ； In the formula, 443 nm The backscattering coefficient of particulate matter at that location, It is a power function sampling quantity.
10. The deep learning lake optical material remote sensing inversion method with embedded radiative transfer equations according to claim 3, characterized in that, The theoretical reconstruction of remote sensing reflectance The reconstruction process is as follows: Will Gordon The radiative transfer approximation model is hard-coded into a differentiable tensor operation layer without training parameters. The total absorption coefficient is calculated using the following formulas. and total backscattering coefficient : ； ； in, The absorption coefficient of pure water is... The backscattering coefficient of pure water; Based on the total absorption coefficient and total backscattering coefficient The backscattering ratio was calculated. : ； Using the backscatter ratio The remote sensing reflectance was calculated to be exactly below the water surface. : ;in, and All are preset constants; Based on the remote sensing reflectance that is just below the water surface The theoretical reconstructed remote sensing reflectance was calculated. : 。