An end-to-end cross-calibration method and system for ocean color satellite sensors

Abstract

The present application belongs to the technical field of satellite remote sensing data processing and radiation calibration, and discloses an end-to-end cross-radiation calibration method and system for marine color satellite sensors. The present application obtains digital count images of a satellite sensor to be calibrated and radiance images of a reference satellite sensor; inputs the digital count images and the radiance images into a double-branch air-spectrum collaborative neural network model, which includes a spatial local fine coding branch and a spectral global dependence coding branch, and is respectively used for extracting multi-scale local spatial features and global spectral dependence between full wave bands; fuses the extracted multi-scale local spatial features and global spectral dependence to obtain air-spectrum fusion features; and inputs the air-spectrum fusion features into a feature reconstruction module for up-sampling reconstruction to generate radiance images after radiation calibration. The present application has good calibration accuracy, provides a new idea for cross-radiation calibration, and can be widely applied to cross-radiation calibration of satellite data.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing data processing and radiometric calibration technology, and particularly relates to an end-to-end cross-radiometric calibration method and system for ocean color satellite sensors. Background Technology

[0002] Ocean color satellites, with their advantages of wide-area and long-term observation, have become a major technical means for quantitative monitoring of the marine environment. Radiometric calibration is crucial for the quantitative application of ocean color satellites, directly affecting the reliability of subsequent remote sensing quantitative products. Radiometric calibration methods for ocean color satellite sensors mainly include on-board calibration, field calibration, and cross-calibration. On-board calibration offers advantages such as high precision and immunity to surface and atmospheric influences, but it requires onboard calibration equipment, resulting in high costs, and its accuracy decreases as the onboard calibration equipment degrades. Field calibration has strict requirements for the calibration field environment and observation conditions, relying on high-precision ground measurements.

[0003] Cross-calibration breaks through the spatiotemporal limitations of traditional calibration methods. It achieves radiometric calibration through radiometric transfer between multiple sensors, significantly improving the consistency between multi-source satellite sensor data. It is an important supplement to on-board calibration and field calibration. Researchers have achieved accurate radiometric calibration of multiple satellite sensors through cross-radiometric calibration. Although multiple factors such as sensor observation geometric differences and spectral differences are considered, the traditional cross-radiometric calibration method still has the following shortcomings: (1) The traditional radiometric calibration mode based on pixels is difficult to effectively mine and characterize spatial context information; (2) The traditional method adopts a single-band calibration strategy, ignoring the correlation between bands.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) Traditional cross-radiometric calibration methods use pixels as the basic processing unit and treat each pixel as an independent object for calibration. This makes it difficult to effectively model the spatial context of the image and to use neighboring pixel information to suppress the interference of local abnormal pixels. As a result, spatial anomalies such as strip noise are retained or even amplified during the radiometric calibration process.

[0005] (2) Traditional cross-radiative calibration methods adopt a strategy of independent calibration for each band, which ignores the inherent physical correlation and spectral correlation between different bands. It is impossible to effectively constrain the spectral shape of the calibration results and it is difficult to guarantee the spectral integrity of radiation consistency between each band.

[0006] (3) Although existing deep learning-based radiation calibration methods have introduced neural network technology, they still follow the traditional approach of fitting linear calibration coefficients. That is, they only use deep learning to simulate linear calibration coefficients and then complete the calibration through a linear model. This fails to achieve end-to-end nonlinear mapping from digital counts to radiance, thus limiting the full utilization of the nonlinear representation capabilities of deep learning. Summary of the Invention

[0007] To overcome the problems existing in related technologies, the present invention discloses an end-to-end cross-radiometric calibration method and system for ocean color satellite sensors, the technical solution of which is as follows: This invention is implemented as follows: an end-to-end cross-radiometric calibration method for ocean color satellite sensors, comprising the following steps: S1. Acquire digital counting images of the satellite sensor to be calibrated and radiance images of the spatiotemporally synchronized reference satellite sensor, and preprocess the digital counting images and radiance images. S2. Input the preprocessed digital counting image and radiance image into a pre-constructed multi-layer spatial-spectral collaborative neural network, which contains multiple cascaded feature extraction layers. In each feature extraction layer, multi-scale local spatial features of the current layer are extracted through a spatial local fine coding branch, and global spectral dependencies between all bands of the current layer are extracted through a spectral global dependency coding branch. S3. In each feature extraction layer, the multi-scale local spatial features extracted by the layer are fused with the global spectral dependence by the corresponding feature aggregation module to obtain the spatial-spectral fusion feature of the current layer. Except for the last feature extraction layer, the spatial-spectral fusion feature of the current layer is used as the input of the next feature extraction layer and passed to the next layer. In this way, feature extraction and fusion are carried out layer by layer until the last feature extraction layer outputs the final deep spatial-spectral fusion feature. S4. Input the obtained final deep spatial spectrum fusion features into the feature reconstruction module for upsampling reconstruction to generate a radiometrically calibrated radiance image.

[0008] Another objective of this invention is to provide an end-to-end cross-radiometry calibration system for ocean color satellite sensors. This system is used to execute the aforementioned end-to-end cross-radiometry calibration method for ocean color satellite sensors. The system is implemented based on the PyTorch deep learning framework and runs on a graphics processor, comprising: The data acquisition module is used to acquire digital counting images from the satellite sensor to be calibrated and radiance images from the spatiotemporally synchronized reference satellite sensor. The dual-branch spatial-spectral feature extraction module includes a spatial local fine-grained coding branch and a spectral global dependency coding branch. The spatial local fine-grained coding branch is used to extract multi-scale local spatial features, and the spectral global dependency coding branch is used to extract global spectral dependencies across all bands. The spatial-spectral fusion feature module is used to fuse the multi-scale local spatial features and the global spectral dependence to output spatial-spectral fusion features; The feature reconstruction module is used to upsample and reconstruct the spatial-spectral fusion features to generate a radiometrically calibrated radiance image.

[0009] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, ocean color satellites, with their wide-area and periodic observation capabilities, play a crucial role in quantitative monitoring of the marine environment. Radiometric calibration is key to quantitative remote sensing, directly impacting the reliability of subsequent quantitative remote sensing products. While on-board and field calibration offer high accuracy, they are technically and costly. Cross-radiometric calibration utilizes precisely radiometrically calibrated sensors to calibrate target sensors, offering advantages in high frequency and low cost. However, existing cross-radiometric calibration methods often employ linear calibration models, neglecting spatial and spectral correlations between pixels. To address these issues, this invention proposes a novel end-to-end cross-radiometric calibration method based on deep learning, aiming to uncover nonlinear relationships between satellites and achieve end-to-end cross-radiometric calibration. For the spatial-spectral three-dimensional features of satellite imagery, this method designs a dual-branch structure to mine local spatial features and global spectral dependencies; for maintaining the shape of the radiometric calibration spectrum, a full-band collaborative learning mechanism is employed to preserve the accuracy of the radiometric calibration spectrum. Based on this, taking the Chinese Ocean Color and Temperature Scanner (COCTS) aboard the Haiyang-1D satellite (HY-1D) as an example, a sample library covering different types of water bodies was constructed, and model training and application were carried out. Experimental results show that the radiometric calibration results of each band of HY-1D COCTS obtained based on the method of this invention have good consistency with the standard radiance product of the AQUA Moderate-resolution Imaging Spectroradiometer (MODIS). The average mean squared error (MSE) of the eight bands is 0.2349, and the average relative error (RE) is 1.14%. Compared with other bands, the blue light band has the highest accuracy (RE = 0.7%, MSE = 0.4353), while the near-infrared band has the worst accuracy (RE = 1.83%, MSE = 0.0236). The HY-1D COCTS radiometric product generated by this method has better accuracy than the radiance product, with an average reduction of 84% in RE and 96% in MSE. Meanwhile, this invention exhibits good spatial and spectral preservation capabilities, with a Structural Similarity Index Measure (SSIM) of 0.9942 and a Spectral Angle (SA) of 0.29°, effectively suppressing stripe noise and local distortion, and improving the quality of radiometric calibration products. Furthermore, this method demonstrates good applicability in various water environments, suitable for both clear ocean waters and turbid nearshore waters. This invention provides a new approach to cross-radiometric calibration and can be widely applied to the cross-radiometric calibration of satellite data.

[0010] Secondly, this invention proposes a dual-branch cross-radiometric calibration method, DB-SSJN, based on a spatial-spectral synergy mechanism. Taking the HY-1D COCTS sensor as an example, it achieves end-to-end conversion from DN to apparent radiance. Compared to traditional calibration methods based on single pixels and independent bands, this method introduces convolution and attention mechanisms to construct a dual-branch architecture, extracting multi-scale spatial features and long-range spectral correlations respectively, and achieving complementarity and synergy of spatial-spectral information through a feature fusion module. Addressing the significant differences in dynamic range across different bands of satellite sensors, this invention designs a band-weighted loss function, significantly improving the calibration effect in the infrared band. Validation results with synchronous AQUA / MODIS radiance products show that the DB-SSJN method exhibits excellent calibration accuracy and effectively maintains spectral shape. Compared to the COCTS standard product, the results obtained by this method show higher consistency with MODIS reference data, demonstrating superior calibration performance. Furthermore, this method exhibits good adaptability and robustness to different aquatic environments (including nearshore turbid waters and open, clean ocean waters). Thanks to its inherent spatial-spectral collaborative processing mechanism, the model constructed in this invention has a significant ability to suppress local abnormal pixels such as strip noise that may exist in the image.

[0011] Third, based on a dual-branch spatial-spectral collaborative architecture and a full-band collaborative learning mechanism, this invention can effectively suppress satellite image stripe noise, correct local pixel distortion, accurately preserve the spectral integrity of each band, and stably output high-precision satellite apparent radiance images with high spatial and spectral consistency. Simultaneously, this invention abandons the traditional linear fitting mode and eliminates band response differences and nonlinear response deviations between different satellite sensors through an end-to-end deep nonlinear mapping method, achieving high-precision radiometric alignment between ocean color satellites such as HY-1D COCTS and satellites such as AQUA MODIS.

[0012] Fourth, this invention addresses the shortcomings of existing cross-radiometric calibration techniques, such as shallow spatial-spectral feature fusion, simplistic modeling methods, and unbalanced multi-band optimization. It solves the technical problems of multi-level, vertically progressive spatial-spectral collaborative cross-radiometric calibration mechanisms from three core dimensions: network architecture, modeling approach, and loss function design. Existing technologies can only perform simple single-scale, single-step superposition of spatial and spectral features. This invention innovatively proposes a dual-branch, progressively fused DB-SSJN feature extraction network, achieving deep cascaded fusion of multi-scale local spatial context features and long-range dependencies of global spectral data across all bands, breaking through the shallow bottleneck of traditional spatial-spectral feature fusion. This invention also solves the technical problem of end-to-end deep nonlinear cross-calibration from remote sensing DN values ​​to L1-level radiance, constructing and verifying an end-to-end deep nonlinear network model that directly maps raw remote sensing digital counts to physical radiance images, thus improving the nonlinear feature representation technology system in the field of remote sensing radiometric calibration. This invention solves the problem of designing a band-weighted calibration loss function for sensors with dynamic range differences. It customizes the band-weighted loss function to address the physical characteristics of significant differences in the dynamic range of visible and infrared bands in ocean color sensors, effectively solving the industry technical problem of imbalance in high and low dynamic band optimization during multi-band joint calibration.

[0013] Fifth, this invention, relying on a spatial-spectral collaborative processing mechanism and multi-scale feature adaptive learning capabilities, overcomes two major technical challenges that have long plagued the field of marine remote sensing radiometric calibration. On the one hand, it solves the industry problem of the incompatibility between spatial denoising and striping and the preservation of spectral features. Traditional calibration techniques, in the process of removing image stripe noise and correcting local abnormal pixels, are prone to destroying the spectral morphology and band correlation features of the original pixels. This invention, by fusing spatial deformation perception and composite channel attention mechanisms, relies on neighborhood spatial context information to achieve accurate denoising and striping, while simultaneously locking spectral features through global spectral dependency constraints, achieving a two-way unity of spatial optimization and spectral fidelity. On the other hand, it solves the technical problem of poor cross-environmental calibration adaptability between nearshore turbid water bodies and open ocean water bodies. Due to the great difference in the physical characteristics of reflectivity between nearshore water bodies and open ocean water bodies, traditional fixed parameter calibration models cannot be adapted to the two types of water environments, resulting in problems such as unbalanced calibration accuracy and insufficient applicability. This invention, through end-to-end nonlinear mapping and multi-scale adaptive feature reconstruction technology, has excellent generalization ability and robustness for both turbid nearshore water bodies and clear ocean water bodies, thus broadening the applicable scenarios for marine satellite radiometric calibration.

[0014] Sixth, existing satellite radiometric calibration is merely an independent measurement process of the sensor's physical response, requiring individual pixels and single bands to be treated as independent calculation units, completely ignoring the spatial correlation between pixels and the spectral coupling between bands. This invention, by introducing neighborhood spatial context information and cross-band physical correlation features, constructs a spatial-spectral collaborative calibration model. Experimental verification shows that this method not only does not interfere with the accuracy of radiometric calibration but also effectively resists error interference from local abnormal pixels, significantly improving the radiometric consistency and overall calibration accuracy of the entire remote sensing image, thus reconstructing the technical concept and implementation system of cross-radiometric calibration for marine satellites. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of the disclosure of this invention. Figure 1 This is a general framework diagram of the end-to-end cross-radiometric calibration method for ocean color satellite sensors provided in this embodiment of the invention; wherein, (a) is a diagram of the globally dependent spectral encoder, (b) is a diagram of the spatially local fine encoder, and (c) is a diagram of the spatial-spectral feature aggregation module; Figure 2 This is a comparison chart of MODIS radiance and HY-1D COCTS radiometric calibration results provided in the embodiments of the present invention; wherein, (a1) to (a8) are MODIS radiance images corresponding to the 1st to 8th study bands, and (b1) to (b8) are HY-1D COCTS radiometric calibration results corresponding to the 1st to 8th study bands; Figure 3 This is a scatter plot comparing the statistical results of HY-1D COCTS cross-radiometric calibration with MODIS radiance provided in this embodiment of the invention; wherein, the 8 subplots (a) to (h) correspond to the 1st to 8th study bands respectively; Figure 4 These are the HY-1D radiometric calibration results, MODIS radiance products, and NPP VIIRS radiance provided in the embodiments of the present invention; wherein, (a1) to (a5) are sub-figures comparing the HY-1D radiometric calibration results and VIIRS radiance corresponding to the 1st to 5th study bands, and (b1) to (b5) are sub-figures comparing the MODIS radiance products and VIIRS radiance corresponding to the 1st to 5th study bands; Figure 5 These are the HY-1D radiometric calibration results, MODIS radiance products, and NPP VIIRS spectrum comparison charts provided in this embodiment of the invention. Figure 6 The scatter plot comparing the HY-1D COCTS calibration results with the standard radiance product provided in the embodiments of the present invention; wherein, the 8 subplots (a) to (h) correspond to the 1st to 8th study bands in turn; Figure 7 The image provided in this embodiment of the invention is a 412nm band grayscale image after maximum and minimum linear stretching; wherein, (a) is an uncalibrated DN image, (b) is an L1B radiance image, and (c) is a calibrated radiance image. Figure 8 This is a comparison of the cross-sections at 412nm of the uncalibrated DN image, L1B radiance image, and calibrated radiance image provided in the embodiments of the present invention; wherein, (a) is the uncalibrated DN image, (b) is the L1B radiance image, and (c) is the calibrated radiance image. Figure 9 This is an RGB color composite image of the L1B radiance image, the calibrated radiance image and the MODIS radiance image provided in the embodiments of the present invention, which is composed using 670nm, 490nm and 412nm; wherein, (a) is the L1B radiance image, (b) is the calibrated radiance image and (c) is the MODIS radiance image. Figure 10 This is a comparison of the spectral curves of the COCTS L1B radiance image, the DB-SSJN calibrated radiance image, and the MODIS L1B radiance image provided in the embodiments of the present invention. Figure 11 These are comparison images of the calibration results in the near-infrared band (825nm) provided in the embodiments of the present invention; where (a) is the calibration result without weighted loss function, (b) is the calibration result with weighted loss function, and (c) is the MODIS reference image; Figure 12 A comparison of AQUA MODIS and calibrated COCTS radiance of clean water bodies provided in this embodiment of the invention; wherein, (a1) to (a8) are AQUA MODIS radiance images corresponding to the 1st to 8th study bands, and (b1) to (b8) are calibrated COCTS radiance images corresponding to the 1st to 8th study bands. Figure 13 This is a comparison of the spectral curves of AQUA MODIS and HY-1D COCTS under different water bodies provided in the embodiments of the present invention; wherein, (a) is the spectral curve of a clean marine water sample, and (b) is the spectral curve of a turbid water sample along the Yangtze River estuary. Figure 14 The above is a comparison of the AQUA MODIS and calibrated COCTS radiance of turbid water bodies provided in the embodiments of the present invention; wherein (a1) to (a8) are the AQUA MODIS radiance images corresponding to the 1st to 8th study bands, and (b1) to (b8) are the calibrated COCTS radiance images corresponding to the 1st to 8th study bands. Figure 15The image shows a comparison of COCTS / HY-1D suspended solids concentration and Sentinel-3 OLCI concentration provided in the embodiments of the present invention; wherein, (a) is the suspended solids concentration result of Sentinel-3 OLCI, and (b) is the suspended solids concentration result of COCTS / HY-1D obtained by the method of the present invention. Figure 16 This is a graph showing the changing trends of training loss and validation loss provided in an embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0017] The innovation of this invention lies in: (1) End-to-end radiometric calibration: abandoning the traditional linear regression model, end-to-end cross-radiative calibration is achieved.

[0018] (2) Spatial-spectral collaborative network architecture: A dual-branch network architecture was designed to extract local spatial features and global spectral features, ensuring the consistency between the spatial structure and spectral shape of the calibration results.

[0019] (3) Band weighted loss function: A weighted loss is designed to balance the optimization effect of each band and improve the calibration accuracy of the infrared band.

[0020] An end-to-end cross-radiometric calibration method for ocean color satellite sensors includes the following steps: S1. Acquire digital counting images of the satellite sensor to be calibrated and radiance images of the spatiotemporally synchronized reference satellite sensor, and preprocess the digital counting images and radiance images. S2. Input the preprocessed digital counting image and radiance image into a pre-constructed multi-layer spatial-spectral collaborative neural network, which contains multiple cascaded feature extraction layers. In each feature extraction layer, multi-scale local spatial features of the current layer are extracted through a spatial local fine coding branch, and global spectral dependencies between all bands of the current layer are extracted through a spectral global dependency coding branch. S3. In each feature extraction layer, the multi-scale local spatial features extracted by the layer are fused with the global spectral dependence by the corresponding feature aggregation module to obtain the spatial-spectral fusion feature of the current layer. Except for the last feature extraction layer, the spatial-spectral fusion feature of the current layer is used as the input of the next feature extraction layer and passed to the next layer. In this way, feature extraction and fusion are carried out layer by layer until the last feature extraction layer outputs the final deep spatial-spectral fusion feature. S4. Input the obtained final deep spatial spectrum fusion features into the feature reconstruction module for upsampling reconstruction to generate a radiometrically calibrated radiance image.

[0021] Example 1: Implementation environment and hardware / software configuration; The end-to-end cross-radiometric calibration method for ocean color satellite sensors based on deep learning, as described in this invention, is implemented using the PyTorch deep learning framework and trained and executed on an NVIDIA RTX A6000 GPU. It should be noted that the aforementioned hardware platform and software framework are merely one optional implementation of this invention, and the invention is not limited thereto. Other hardware platforms with equivalent computing power (such as NVIDIA V100, A100, etc.) and deep learning frameworks (such as TensorFlow, MindSpore, etc.) can also implement the technical solution of this invention.

[0022] Example 2: Construction of the training sample library; 1. Satellite data; In this embodiment, to train and verify the cross-radiometric calibration model constructed in this invention, the Chinese Ocean Color and Temperature Scanner (COCTS) aboard the HY-1D satellite was selected as the sensor to be calibrated. The COCTS sensor has 10 bands, with bands 1-8 being visible and near-infrared bands, and bands 9 and 10 being thermal infrared bands. It has an observation swath of 2900 km and a spatial resolution of 1100 m, enabling large-area, precise monitoring of environmental parameters such as ocean color and temperature. The COCTS satellite data used in this embodiment comes from the China Ocean Satellite Data Service System.

[0023] The AQUA MODIS sensor, with its stable radiometric calibration accuracy and high spectral resolution, is shown in Table 1 to be largely consistent with HY-1D COCTS in the visible-near-infrared band. Both are afternoon satellites with overpass times close to 1:30 PM, exhibiting a small time difference that reduces errors caused by variations in satellite imaging environments. The AQUA MODIS data used in this embodiment was acquired from the NASA Earth Data platform. Additionally, to verify and compare the radiometric calibration results, NPP VIIRS satellite data was also acquired in this embodiment.

[0024] Table 1 Sensor Parameter Configuration

[0025] 2. Dataset construction; To ensure the quality of the samples, this embodiment performs strict quality control on the spatiotemporally matched MODIS data and COCTS data.

[0026] First, based on the MODIS L2 Ocean Color Flags, data contaminated by solar flares (HIGLINT, MODGLINT), clouds (CLDICE), and sea ice (SEAICE) are removed. On this basis, residual cloud shadows, flare contamination, and anomalous pixels are further removed through manual visual interpretation to obtain the final effective pixels.

[0027] Subsequently, geometric registration was performed on the spatiotemporally quasi-synchronous MODIS and COCTS images to ensure spatial consistency between the two images. To ensure radiometric consistency of the matched image pairs, further quality control was performed on the samples: the Pearson correlation coefficients of corresponding bands of MODIS and COCTS were calculated, and sample pairs with correlation coefficients below 0.5 were removed to reduce deviations caused by changes in the imaging environment.

[0028] Considering the difference in center wavelength between COCTS and MODIS sensors (see Table 1), this embodiment uses bicubic interpolation to perform spectral resampling on the MODIS radiance data to unify it to the center wavelength of each COCTS band.

[0029] Based on the quality-controlled sample data, a COCTS-MODIS matching dataset containing 357 pairs of 32×32 pixels was constructed, covering different seasons and different aquatic environments. To meet the model training requirements, the dataset was randomly divided into training and validation sets in an 8:2 ratio.

[0030] Example 3: An end-to-end cross-radiation calibration method based on deep learning; 1. Overall framework; This embodiment abandons the traditional approach of fitting linear radiometric calibration coefficients and designs an end-to-end cross-radiometric calibration method based on dual-branch spatial-spectral joint network (DB-SSJN), such as... Figure 1As shown, to address the inherent spatial-spectral 3D features of satellite imagery, two parallel branches were designed to model the spatial-spectral features of the imagery. The spatial local fine-grained coding branch employs a pre-activated ResNet structure, comprising four feature extraction stages. Each stage stacks 3, 6, 2, and 2 residual units respectively. The base channel width is set to 64, and a width factor of 0.5 is introduced for scaling. Therefore, the actual number of input / output channels in each stage is as follows: 64 channels in the first stage, 128 channels in the second stage, 256 channels in the third stage, and 512 channels in the fourth stage. Each residual unit internally performs channel transformation and spatial feature extraction through 1×1 and 3×3 convolutions, gradually abstracting from fine-grained texture information to a high-level semantic structure. The spectral global dependency coding branch is based on an improved Swing Transformer architecture with an initial embedding dimension of 64. It consists of four stages, each stacked with 2, 2, 2, and 2 Transformer modules respectively. The number of attention heads in each stage is set to 4, 2, 2, and 8 respectively, and the window size is fixed at 8×8. The long-range dependencies between all bands are captured through multi-head self-attention and moving window mechanism. In each stage of the encoder (a total of four stages), both branches perform multi-scale feature fusion through layer-by-layer embedded Spatial-Spectral Feature Aggregation Module (SFAM), thereby making full use of spatial details and global semantic information at different resolutions. The specific fusion operation is as follows: For the current stage, the feature map output by the spatial branch first undergoes a deformable convolution to enhance the model's adaptability to geometric deformations in the image. This convolution dynamically adjusts the sampling position by learning the offset and modulation parameters. Simultaneously, the feature map output by the Transformer branch first undergoes a 1×1 convolution to adjust the number of channels to match the current number of channels in the spatial branch. Then, it passes through a channel attention module, which performs global average pooling and global max pooling in parallel on the adjusted features. The pooling results are fed into a shared multilayer perceptron (containing two linear layers) to generate channel weights. The weights are then normalized using the Sigmoid function and multiplied channel by channel with the original features to achieve adaptive recalibration of the Transformer branch features. Finally, the spatial features after deformable convolution and the recalibrated Transformer features are added element by element to complete the spatial-spectral fusion of the current stage. The fusion process is performed at each of the four levels of the encoder. The number of output channels after fusion at each level is consistent with the number of channels in the spatial branch at that stage, which are 64, 128, 256 and 512 respectively. The feature space size is also the same as the feature map size of the corresponding spatial branch. Therefore, after fusion, the feature dimension does not need to be changed before it can be directly sent to the next stage.The SFAM module itself contains a large number of trainable parameters, mainly including the offset and modulation parameters of deformable convolutions, the kernel parameters of 1×1 convolutional layers, and the weights and biases of the multilayer perceptron in channel attention. The number of parameters increases with the number of channels in the stage. Taking a stage with 128 channels as an example, the total number of trainable parameters in this module is approximately in the hundreds of thousands. These parameters are optimized together with the rest of the network during end-to-end training to ensure that the fusion strategy is adapted to the specific calibration task. Through the above-mentioned step-by-step fusion, the local texture details of the spatial branch and the global spectral dependence of the Transformer branch are fully complementary, providing an informative joint spatial-spectral representation for subsequent feature reconstruction. In the feature reconstruction stage, this method uses a strategy of cascading convolution and upsampling operations instead of transposed convolution to avoid the checkerboard artifacts caused by kernel overlap and non-uniform mapping during the reconstruction process, thereby ensuring the spatial continuity of the radiometric calibration results. In addition, to address the issue of large differences in radiance values ​​across different bands in multispectral images, a band weighting mechanism was designed. In the loss function, bands with smaller radiance values ​​in the near-infrared band are given higher weights to enhance their gradient contribution in backpropagation, thereby improving the overall accuracy and stability of multi-band joint calibration.

[0031] 2. Dual-branch spatial spectrum feature extraction module; To simultaneously preserve local spatial details and global spectral dependencies in remote sensing images, this paper constructs a dual-branch spatial-spectral feature extraction framework consisting of a Spatial Local Refinement Encoder and a Spectral Global Dependency Encoder. This architecture extracts multi-scale local spatial features and long-range spectral dependencies from the images through two parallel branches, achieving complementary and collaborative representation of spatial-spectral features.

[0032] The spatial local fine-grained coding branch adopts a multi-layer stacked residual network structure. Through a bottleneck structure composed of sequential convolutional operations (1×1, 3×3, 1×1), it achieves efficient extraction of local spatial features while reducing computational complexity. Its basic residual unit can be represented as: ; ; ; ; In the formula, These are the first, second, and third intermediate feature tensors, respectively. It is a ReLU activation function with non-linear mapping capabilities. For batch normalization operations, The kernel size is Two-dimensional convolution operation; The input feature tensor of the basic residual unit, Input the number of channels. For feature map height, The width of the feature map; The final spatial feature tensor output after fusion and reconstruction of the basic residual units is given. for Convolutional projection transformation.

[0033] This branch leverages the local receptive field and weight sharing characteristics of convolutional operations to gradually expand the receptive field during layer-by-layer downsampling, thereby effectively capturing multi-scale spatial features from low-level texture to high-level semantics. Residual connections effectively alleviate the vanishing gradient problem, ensuring stable training of deep networks.

[0034] However, the inherent locality of convolutional operations limits their ability to model long-range spectral dependencies. To address this, the spectral feature extraction branch introduces a self-attention-based global spectral dependency encoding mechanism to model global correlations between bands. This branch is based on an improved Swing Transformer architecture, such as... Figure 1 As shown in Figure (a), cross-window information interaction is achieved through the Shifted Window Multi-head Self-Attention (SW-MSA) mechanism. Window Multi-head Self-Attention (W-MSA) and SW-MSA are executed alternately in consecutive SwingTransformer blocks, establishing long-range dependencies while maintaining O(N) computational complexity. Specifically: ; ; ; ; In the formula, and The first Layer and First The intermediate output feature tensor of the layer , and The first Layer, First Layer and First The final output feature tensor of the layer, The length of the spectral feature sequence; For layer normalization operation, This is a multi-head self-attention mechanism for windows, used to compute local self-attention between bands within a divided local window; This is a multi-head self-attention mechanism for shifted windows, used to achieve cross-window feature interaction through window shifting; It is a multilayer perceptron used to perform nonlinear mapping and feature reconstruction along the channel dimension.

[0035] To further enhance feature representation capabilities and alleviate the problem of high-frequency information attenuation in deep encoding, an Attention Scale Module was designed, such as... Figure 1 As shown in Figure (a), this module decomposes the attention matrix into low-frequency components representing smooth trends and high-frequency components representing local details, and dynamically adjusts the ratio of the two through learnable weight coefficients, thereby optimizing the feature weight distribution. Specifically: ; ; ; ; In the formula, To output the attention result, For querying the matrix, The key matrix, The value matrices are all obtained by mapping the input spectral features through a transformation matrix. The length of the spectral feature sequence. For feature dimension, For query vector dimensions, The dimension of the value vector; To do for each row Obtain attention weights; This is the transpose of the key matrix. and From respectively The low-frequency and high-frequency components extracted from them By calculating the global average of the attention matrix, smooth background information of the attention distribution is captured; The residual calculation captures local detail fluctuations relative to the average level; For a matrix where all elements are 1, Original attention matrix The Middle line, number The element values ​​of the column, The original attention matrix, For the reweighted output attention matrix, The weights are learnable and satisfy the constraints. .

[0036] 3. Feature Reconstruction Module; In the feature reconstruction stage, to reconstruct the low-resolution feature map to the original image size while preserving and restoring spectral spatial structure information to the maximum extent, this method employs a reconstruction architecture based on progressive upsampling and feature transformation cascades, such as... Figure 1 As shown, high-fidelity spatial-spectral information reconstruction is achieved by smoothly transitioning spatial frequency and deeply fusing multi-scale features. This module abandons the transposed convolution operation that easily leads to checkerboard effects, and instead adopts an upsampling strategy that combines bilinear interpolation with convolution. Specifically, the reconstruction process is shown in Equation (13). Each reconstruction layer first performs a nonlinear transformation on the input low-resolution features through 3×3 convolution and ReLU activation function to enhance their representation ability; then, it performs a 2x upsampling through bilinear interpolation to smoothly expand the spatial scale of the feature map; then, the upsampled features are spliced ​​with the corresponding scale features in the encoder through skip connections to introduce multi-scale spatial and spectral features obtained by the dual-branch encoder, thereby enhancing the detail preservation ability of the reconstruction results. The fused features are further aligned and integrated through 3×3 convolution to finally output the enhanced high-resolution features.

[0037] ; In the formula, For the first The final feature tensor output by the layer decoder after reconstruction. For the first The number of feature channels in the layer output. For the first Layer output feature map height, No. Layer output feature map width; For the first The feature tensor output by the layer decoder The number of feature channels, and These are the height and width of the feature map, respectively. It is a linear rectified activation function. For use in fusing cascaded features and achieving a 50% reduction in the number of channels Two-dimensional convolution operation, This refers to a feature concatenation operation that concatenates two feature tensors along the channel dimension. This is a spatial dimension upsampling operation achieved through bilinear interpolation. Used to keep the number of feature channels constant Two-dimensional convolution operation; For the encoder's number The skip connection feature tensor of the layer The number of feature channels of the encoder. and These are the height and width of the target feature map in the current layer, respectively, and satisfy the following conditions: , ; This is the layer index for the decoder and encoder.

[0038] 4. Band-weighted loss function; Due to the large differences in radiance values ​​across bands in ocean color satellite data, the backpropagation gradient contribution is insufficient for some bands with smaller value ranges (such as the near-infrared band). To address this issue, this embodiment proposes an improved MSE loss function based on band weighting, aiming to optimize the spectral balance during propagation. This function effectively balances the contribution weights of each band during training by introducing band weight parameters, as shown in the following equation: ; In the formula, This is the global band weighted mean square error loss function. The total number of bands contained in the image. These represent the height and width of the image's spatial location, respectively. These are the height and width indices of the spatial location, respectively. , ; For the model in the band Spatial location The predicted value at that location; To correspond to the band Spatial location Reference value at that location; With band The band weight coefficients are determined by the reciprocal of the radiance value range of the corresponding dataset. Specifically, based on the eight bands of the HY-1D COCTS sensor (center wavelengths of 412nm, 443nm, 490nm, 520nm, 560nm, 670nm, 750nm, and 865nm respectively), this embodiment first statistically analyzes the radiance value range (i.e., the difference between the maximum and minimum values) of each band using a large number of training samples. Statistical results show that the radiance value range of the visible light band (412nm~670nm) is relatively close, while the radiance value range of the near-infrared band (750nm and 865nm) is significantly smaller, approximately 1 / 3 to 1 / 6 of that of the visible light band. To balance the contribution of each band to the loss function, this embodiment uses the principle of inverse proportionality to the value range to determine the weight coefficients. Using the visible light band as a benchmark (weight set to 1), the weight of the near-infrared band is proportionally amplified. After multiple experimental verifications, the final weight coefficients for the eight bands are set. The possible values ​​are: =1、 , =1、 , , , , This weighting configuration effectively enhances the gradient in the near-infrared band during backpropagation, compensating for insufficient gradient contribution due to its small value range and improving the overall accuracy and stability of multi-band joint calibration. It is worth noting that the aforementioned weighting coefficients can be fine-tuned according to the radiation response characteristics of different sensors. The values ​​given in this embodiment are typical values ​​based on statistical analysis and experimental optimization using HY-1D COCTS data, and their effectiveness has been fully verified.

[0039] The loss function first calculates the pixel-wise mean square error for all spatial locations in each band, then performs a weighted summation of the loss values ​​for each band, and finally obtains the global loss value through an arithmetic mean to ensure the stability of gradient calculation. By using differentiated weighting along the spectral dimension, the optimization insufficiency caused by the small radiance range in the near-infrared band is effectively mitigated, significantly improving the radiation accuracy in the long-wavelength region while ensuring the reconstruction accuracy of the visible light band.

[0040] Example 4: Model training method and hyperparameter settings; In this embodiment, stochastic gradient descent (SGD) is used as the optimizer, with an initial learning rate of 0.01 and L2 regularization applied (weight decay coefficient 1×10-4).

[0041] The learning rate adjustment strategy employs polynomial decay, specifically... In the formula Represents the current iteration step. The total number of iterations is calculated by multiplying the maximum training epochs of 3000 by the number of iterations per epoch. No learning rate warm-up is used during the entire training process. The batch size is set to 16 during training, and the maximum training epochs are 3000. For dataset partitioning, a random partitioning method is used, with a random seed dividing all samples into training and validation sets in an 8:2 ratio. The input data preprocessing workflow is as follows: each band of the image is independently processed using minimum... Maximum normalization linearly maps pixel values ​​to the [0,1] interval. No data augmentation was used during training to maintain the authenticity of the original radiometric information of the images. Model weight initialization used the default Kaiming uniform initialization (He initialization) of the PyTorch framework, which is compatible with the ReLU activation function in the network. During training, an early stopping mechanism was implemented to determine the model's convergence state. When the validation set loss did not decrease for 50 consecutive epochs, early stopping was triggered, and the current optimal model weights were saved, effectively preventing overfitting. In actual training, the loss function decreased rapidly in the early stages of training, then gradually slowed down. The validation loss triggered the early stopping mechanism at 2360 epochs, at which point training terminated and the optimal model parameters were saved. The loss change curves throughout the entire training process (including training loss and validation loss) are shown below. Figure 16 As shown, both training loss and validation loss show a decreasing trend and there is no significant overfitting, indicating that the model has good generalization ability.

[0042] Example 5: Accuracy Evaluation Indicators The proposed method is evaluated using mean squared error (MSE), relative error (RE), structural similarity index (SSIM), and spectral angle (SA). The formulas for calculating these metrics are as follows: ; ; ; ; in, , , , These represent mean square error, relative error, structural similarity, and spectral angle, respectively. The total number of samples involved in the calculation; For sample index; and The first The true observed value (or true spectral vector) and the model predicted value (or predicted spectral vector) of each sample. Operations for calculating the absolute value or the magnitude of a vector; and These are the distribution mean of all true values ​​and predicted values, respectively; and These are the standard deviations of the distributions of all true values ​​and predicted values, respectively. The covariance between all true values ​​and predicted values; and These are all preset minimum constants used to prevent the denominator from being zero in order to maintain computational stability; It is the inverse cosine function; Representing vectors and The inner product between them. (Note: The spectral angle is...) The angle between high-dimensional spectral vectors is calculated to measure the shape similarity between the true and predicted spectra after eliminating the influence of brightness. Example 6: Radiation calibration results and accuracy verification; 1. Comparison and verification with AQUA MODIS reference products; To verify the performance of the cross-radiometric calibration method of this invention, a comparative analysis was conducted using synchronized AQUA MODIS L1B radiance products. The results show that the radiometric calibration results of each COCTS band obtained based on the method of this invention (DB-SSJN) have good consistency with the MODIS radiance products. The radiance value range and distribution of each band are consistent, and the local detail information is well recovered, such as... Figure 2 As shown. Statistical results of radiance distribution show that the calibrated HY-1D COCTS data exhibits a high linear correlation with MODIS. The radiance distribution trend of the eight COCTS bands after calibration is basically consistent with MODIS, with most data points distributed near the 1:1 line, such as... Figure 3 As shown.

[0043] Meanwhile, the COCTS radiometric calibration results were quantitatively evaluated using MODIS radiance products, and the evaluation results are shown in Table 2. The results show that the COCTS radiometric calibration results have good accuracy, with an average MSE of 0.2349 and an RE of 1.14% across the eight bands. Among the bands, the blue light band has the highest accuracy, with an MSE of 0.4353 and an RE of 0.7%. The near-infrared band has slightly lower accuracy, with an MSE of 0.0236 and an RE of 1.83%. Furthermore, the cross-radiometrically calibrated images exhibit good spatial structure and spectral consistency, with an SSIM value of 0.9942 and an SA of 0.29°. These experimental results demonstrate that this method can achieve accurate radiometric calibration of the HY-1D COCTS across all eight bands, while also possessing good spatial structure and spectral shape preservation capabilities.

[0044] Table 2. Accuracy Evaluation Results Based on AQUA MODIS

[0045] 2. Comparative verification with NPP VIIRS radiation calibration products; To further verify the algorithm of this invention, a comparative analysis was conducted using NPP VIIRS radiance data. Considering the differences in band settings between the COCTS and VIIRS sensors, five bands—412nm, 443nm, 490nm, 670nm, and 865nm—were selected for comparison with the VIIRS bands—412nm, 445nm, 488nm, 672nm, and 865nm. The comparison results show that the COCTS calibration results and VIIRS radiance have a good trend consistency, but the calibration results exhibit a systematic underestimation. Figure 4 As shown, the specific relative errors for the five bands are: 9.74% for the 412nm band, 5.39% for the 443nm band, 11.35% for the 490nm band, 9.61% for the 670nm band, and 11.49% for the 865nm band. The overall average relative error is less than 11.5%, and the spectral angle (SA) is 1.55°. A comparison of simultaneous MODIS observations with VIIRS reveals that the MODIS data also exhibits the same bias characteristics. Existing research indicates that there is approximately a 4% difference in radiation between MODIS and VIIRS. Therefore, the bias in the current comparison results likely stems from systematic bias between the sensors. From the spectral shape, as... Figure 5 As shown, the calibrated average radiance spectrum curves exhibit high similarity to those of NPP VIIRS and AQUA MODIS.

[0046] 3. Comparison with COCTS standard radiation products; Meanwhile, the HY-1D COCTS radiance product generated in this embodiment was compared with the radiance product, and the results are as follows: Figure 6 As shown in the figure. The comparison results show that although the two are basically consistent in overall distribution trends, the calibration results of this invention are closer to the MODIS radiance values ​​in each band range. Quantitative evaluation results show that the calibration results of this invention are more consistent with the MODIS product, with relative errors of less than 1.3% in all eight bands, an average relative error of 0.65%, and an SA value of 0.23°. In contrast, the COCTS standard radiance product differs significantly from the MODIS radiance product, with a minimum relative error of 1.99%, a maximum relative error of 5.16%, an average relative error of 4.26%, and an SA value of 1.22° in the eight bands.

[0047] Example 7: Technical Effects and Application Verification; 1. Space preservation effect; Traditional radiometric calibration methods typically use linear calibration coefficients for each band, treating each pixel as an independent unit for calibration. This method fails to model spatial context relationships and struggles to utilize neighboring pixel information to suppress local anomalous pixel interference, resulting in the retention or even amplification of anomalous pixels such as stripe noise during radiometric calibration. In the spectral dimension, traditional radiometric calibration methods employ a band-by-band calibration strategy, neglecting the dependencies between different bands and failing to constrain the spectral shape. In contrast, the DB-SSJN method proposed in this invention achieves preservation of both image spatial and spectral shape through a spatial-spectral co-modeling mechanism. This mechanism enhances the perception of contextual structure through local feature extraction in the spatial dimension, and ensures spectral shape consistency by constructing global dependencies in the spectral dimension.

[0048] Calibration results show that DB-SSJN, based on the spatial-spectral coordination mechanism, exhibits significant advantages in radiometric calibration. By modeling the spatial-spectral context, the consistency of calibration results in both spatial and spectral dimensions is improved. In the spatial dimension, DB-SSJN effectively suppresses interference from local anomalous pixels such as stripe noise and random noise, enhancing the robustness of the calibration results. Figure 6 As shown in the image, both the image and the cross-sectional view reveal significant noise in the original uncalibrated COCTS DN value image, as indicated by... Figure 8 As shown in Figure (a). This strip noise is preserved in the COCTS L1B standard radiance product using the conventional linear radiometric calibration model, as... Figure 8 As shown in Figure (b). After applying DB-SSJN calibration, strip noise was effectively suppressed, and the cross-sectional data tended to be smoother, as shown in Figure (b). Figure 8 As shown in Figure (c).

[0049] 2. Spectral preservation effect; In the spectral dimension, DB-SSJN exhibits excellent spectral preservation capabilities. For example... Figure 9 As shown in Figure (b), the calibration results based on DB-SSJN in RGB composite images, compared to... Figure 9 The COCTS L1B standard product shown in Figure (a) is closer to... Figure 9 The MODIS reference image shown in Figure (c) is selected. Figure 9 The region within the black box in the image shown in Figure (a) is taken as the region of interest. Its average spectral curve is extracted and plotted, as follows: Figure 10 As shown in the figure. The results show that the spectral curve of the L1B product deviates significantly from the MODIS reference curve, while the spectrum of the DB-SSJN calibration result is more consistent with the MODIS spectral shape.

[0050] 3. The effect of the band-weighted strategy; The band weighting strategy proposed in this invention plays a crucial role in balancing the various bands. By assigning higher weight coefficients to the near-infrared band, which has a smaller value range, the contribution of the loss term of this band to the overall loss is increased, thereby obtaining a larger gradient during backpropagation and ensuring that each band receives relatively balanced attention during training. Calibration results show that, using the traditional root mean square error loss function, the calibration results for the 825nm band exhibit significant spatial blurring, resulting in severe loss of image spatial information. Figure 11 As shown in Figure (a). In contrast, the calibration quality of this band is significantly improved after the introduction of the band weighting mechanism, as shown in Figure (a). Figure 11 As shown in Figure (b), the image details are richer and the spatial consistency with the reference image is significantly improved.

[0051] 4. Applicability to different aquatic environments; To assess the applicability of this method in different water body types, two typical areas were selected for applicability assessment: open ocean (clean water) and turbid water along the Yangtze River estuary.

[0052] In open ocean areas, the calibrated COCTS radiance shows good consistency with MODIS in terms of spatial distribution and value range, such as... Figure 12 As shown. Quantitative evaluation results show that the average relative error (RE) of its eight bands is 1.49%, the structural similarity index (SSIM) reaches 0.9954, and the spectral angle (SA) is 0.26°, indicating that the method has good calibration accuracy in clean water areas. Furthermore, Figure 12 A comparison of the average spectral curves within the black-framed region shows that the radiance curve obtained by the method of this invention highly matches the shape of the MODIS spectrum. Figure 13 As shown in Figure (a).

[0053] In turbid water areas, the calibrated COCTS radiance and MODIS also showed good consistency in spatial distribution and value range, such as... Figure 14 As shown, the average relative error of its eight bands is 1.33%, the structure similarity index (SSIM) reaches 0.9692, and the spectral angle (SAM) is 0.38°. These results indicate that the method also possesses good radiometric calibration accuracy in nearshore turbid water areas. The comparison of average spectral curves further demonstrates that, even under turbid water conditions with complex optical properties, this method can effectively maintain the spectral morphology of the radiance, such as... Figure 13 As shown in Figure (b).

[0054] The above experimental results show that the radiometric calibration method proposed in this invention has good applicability and stability in different water body types (including clean water and turbid water), and can provide high-precision radiance products for quantitative remote sensing applications of different types of water bodies.

[0055] 5. Application Demonstration—Suspended Sediment Concentration Inversion; First, the HY-1D COCTS data were radiometrically calibrated using the developed DB-SSJN method, and then atmospherically corrected using the Seadas-c atmospheric correction method. Based on the remote sensing reflectance obtained from atmospheric correction, and referring to the suspended matter concentration inversion model proposed by Doxaran et al. and the TSM sensitive band, a regional inversion model suitable for the HY-1D COCTS sensor was constructed. Based on this model, TSM was inverted, and the results were compared with the TSM products of the Sentinel-3 Ocean and Land Colour Instrument (OLCI). The results are as follows: Figure 15 As shown in the figure, the comparison results show that the spatial distribution trends of suspended sediment concentration obtained from the two data sources are consistent, with high concentration areas mainly concentrated in estuaries and nearshore areas, and the suspended solids concentration gradually decreasing towards the open sea. This result indicates that the COCTS data after cross-radiological calibration has good application potential in aquatic environmental monitoring.

[0056] This invention proposes a dual-branch cross-radiometric calibration method, DB-SSJN, based on a spatial-spectral synergy mechanism. Using the HY-1D COCTS sensor as an example, it achieves end-to-end conversion from DN (Digital Density) to apparent radiance. Compared to traditional calibration methods based on single pixels and independent bands, this method introduces convolution and attention mechanisms to construct a dual-branch architecture, extracting multi-scale spatial features and long-range spectral correlations respectively, and achieving complementarity and synergy of spatial-spectral information through a feature fusion module. Addressing the significant differences in dynamic range across different bands of satellite sensors, a band-weighted loss function is designed, significantly improving the calibration effect in the infrared band. Validation results with synchronous AQUA / MODIS radiance products show that the DB-SSJN method exhibits excellent calibration accuracy and effectively maintains spectral shape. Compared to the COCTS standard product, the results obtained by this method show higher consistency with MODIS reference data, demonstrating superior calibration performance. Furthermore, this method demonstrates good adaptability and robustness to different aquatic environments (including nearshore turbid waters and open, clean ocean waters). Thanks to its inherent spatial-spectral collaborative processing mechanism, the model has a significant ability to suppress local abnormal pixels such as strip noise that may exist in the image.

[0057] Comparative analysis with different calibration benchmark models; To further quantify and verify the performance advantage of the DB-SSJN model in cross-radiometric calibration tasks, two typical methods were introduced as benchmarks for comparison: the Huber robust multiple linear regression model based on classical statistics, and the data-driven pixel-level multilayer perceptron model. From the global evaluation metrics in Table 3 and the band-by-band evaluation metrics in Tables 4 and 5, the three models exhibited significant performance differences. The Huber robust regression model showed relatively high global RMSE and RE, indicating that the purely linear model could not fully fit the nonlinear radiometric residuals caused by differences in sensor spectral response functions and the complex marine atmospheric environment. The MLP model, which introduced nonlinear mapping, reduced various errors to some extent. In contrast, the proposed DB-SSJN model performed better, with its global RMSE and RE reduced to 0.5157 W·m⁻¹. - ²·μm - ¹·sr - ¹ and 1.18%, and the SA, which characterizes spectral morphology fidelity, is as low as 0.28°.

[0058] Table 3 Comparison of global metrics for different calibration models

[0059] Table 4 Comparison of RMSE for different calibration models

[0060] Table 5 Comparison of REs for different calibration models

[0061] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An end-to-end cross-radiometric calibration method for an ocean color satellite sensor, characterized in that, The method includes the following steps: S1. Acquire digital counting images of the satellite sensor to be calibrated and radiance images of the spatiotemporally synchronized reference satellite sensor, and preprocess the digital counting images and radiance images. S2. Input the preprocessed digital counting image and radiance image into a pre-constructed multi-layer spatial-spectral collaborative neural network, which contains multiple cascaded feature extraction layers. In each feature extraction layer, multi-scale local spatial features of the current layer are extracted through a spatial local fine coding branch, and global spectral dependencies between all bands of the current layer are extracted through a spectral global dependency coding branch. S3. In each feature extraction layer, the multi-scale local spatial features extracted by the layer are fused with the global spectral dependence by the corresponding feature aggregation module to obtain the spatial-spectral fusion feature of the current layer. Except for the last feature extraction layer, the spatial-spectral fusion feature of the current layer is used as the input of the next feature extraction layer and passed to the next layer. In this way, feature extraction and fusion are carried out layer by layer until the last feature extraction layer outputs the final deep spatial-spectral fusion feature. S4. Input the obtained final deep spatial spectrum fusion features into the feature reconstruction module for upsampling reconstruction to generate a radiometrically calibrated radiance image.

2. The end-to-end cross-radiometric calibration method for ocean color satellite sensors according to claim 1, characterized in that, In step S1, the preprocessing of the digital count image and the radiance image includes: Pixels contaminated by solar flares, clouds, and sea ice are removed based on the quality indicators of reference satellite sensors. Geometric registration is performed between the digital counting image and the radiance image to ensure spatial consistency; Calculate the Pearson correlation coefficient between the reference satellite sensor and the satellite sensor to be calibrated in the corresponding bands, and remove image pairs with a correlation coefficient lower than 0.5; The radiance data of the reference satellite sensor is spectrally resampled and unified to the center wavelength of each band of the satellite sensor to be calibrated.

3. The end-to-end cross-radiometric calibration method for ocean color satellite sensors according to claim 1, characterized in that, In step S2, the spatial local fine coding branch adopts a multi-layer stacked residual network structure, extracts local spatial features through a bottleneck structure composed of sequentially performed 1×1 convolution, 3×3 convolution and 1×1 convolution, and uses residual connections to alleviate gradient vanishing. The basic residual unit of the spatial local fine coding branch is represented as follows: ； ； ； ； In the formula, These are the first, second, and third intermediate feature tensors, respectively. It is a ReLU activation function with nonlinear mapping capabilities. For batch normalization operations, The kernel size is Two-dimensional convolution operation; The input feature tensor of the basic residual unit, Input the number of channels. For feature map height, The width of the feature map; The final spatial feature tensor output after being fused and reconstructed by the basic residual units is... for Convolutional projection transformation.

4. The end-to-end cross-radiometric calibration method for ocean color satellite sensors according to claim 3, characterized in that, The spectral global dependency encoding branch is based on the Swing Transformer architecture. It establishes long-range dependencies between bands by alternately executing window multi-head self-attention mechanism and shift window multi-head self-attention mechanism in consecutive Swing Transformer blocks, as shown below: ； ； ； ； In the formula, and The first Layer and First The intermediate output feature tensor of the layer , and The first Layer, First Layer and First The final output feature tensor of the layer, The length of the spectral feature sequence; For layer normalization operation, This is a multi-head self-attention mechanism for windows, used to calculate local self-attention between bands within a divided local window; This is a multi-head self-attention mechanism for shifted windows, used to achieve cross-window feature interaction through window shifting; It is a multilayer perceptron used to perform nonlinear mapping and feature reconstruction along the channel dimension.

5. The end-to-end cross-radiometric calibration method for ocean color satellite sensors according to claim 4, characterized in that, The spectral global dependency encoding branch also includes an attention scaling module. This module decomposes the attention matrix into low-frequency components representing smooth trends and high-frequency components representing local details. It dynamically adjusts the ratio of the low-frequency and high-frequency components using learnable weight coefficients to optimize the feature weight distribution. The attention scaling module is represented as follows: ； ； ； ； In the formula, To output the attention result, For querying the matrix, The key matrix, The value matrices are all obtained by mapping the input spectral features through a transformation matrix. The length of the spectral feature sequence. For feature dimension, For query vector dimensions, The dimension of the value vector; To do for each row Obtain attention weights; This is the transpose of the key matrix. and From respectively The low-frequency and high-frequency components extracted from them By calculating the global average of the attention matrix, smooth background information of the attention distribution is captured; The residual calculation captures local detail fluctuations relative to the average level; For a matrix where all elements are 1, Original attention matrix The Middle line, number The element values ​​of the column, The original attention matrix, For the reweighted output attention matrix, The weights are learnable and satisfy the constraints. .

6. The end-to-end cross-radiometric calibration method for ocean color satellite sensors according to claim 5, characterized in that, In step S3, the spatial-spectral fusion feature is used to fuse the multi-scale local spatial features output by the spatial local fine coding branch with the global spectral dependence relationship across all bands output by the spectral global dependence coding branch, thereby achieving complementarity and synergy of spatial-spectral information.

7. The end-to-end cross-radiometric calibration method for ocean color satellite sensors according to claim 6, characterized in that, In step S4, the feature reconstruction module adopts an upsampling strategy combining bilinear interpolation and convolution. Each reconstruction layer performs a nonlinear transformation on the input low-resolution features through 3×3 convolution and ReLU activation function; bilinear interpolation is used to perform upsampling by 2 times to smoothly expand the spatial scale of the feature map; the upsampled features are concatenated with the corresponding scale features in the encoder through skip connections to introduce multi-scale spatial and spectral features obtained by the dual-branch encoder; the fused features are aligned and integrated through 3×3 convolution to finally output the enhanced high-resolution features. The reconstruction process of the feature reconstruction module is represented as follows: ； In the formula, For the first The final feature tensor output by the layer decoder after reconstruction. For the first The number of feature channels in the layer output. For the first Layer output feature map height, No. Layer output feature map width; For the first The feature tensor output by the layer decoder The number of feature channels, and These are the height and width of the feature map, respectively. It is a linear rectified activation function. For use in fusing cascaded features and achieving a 50% reduction in the number of channels Two-dimensional convolution operation, This refers to a feature concatenation operation that concatenates two feature tensors along the channel dimension. This is a spatial dimension upsampling operation achieved through bilinear interpolation. Used to keep the number of feature channels constant Two-dimensional convolution operation; For the encoder's number The skip connection feature tensor of the layer The number of feature channels of the encoder. and These are the height and width of the target feature map in the current layer, respectively, and satisfy the following conditions: , ; This is the layer index for the decoder and encoder.

8. The end-to-end cross-radiometric calibration method for ocean color satellite sensors according to claim 1, characterized in that, The training of the multi-layer spatial-spectral collaborative neural network employs a band-weighted improved mean square error loss function. This loss function predefines band weight coefficients based on the differences in radiance ranges across bands, assigning higher weight coefficients to the near-infrared band than to the visible light band. The band-weighted mean square error loss function is expressed as follows: ； In the formula, This is the global band weighted mean square error loss function. The total number of bands contained in the image. With band The band weight coefficients are determined by the reciprocal of the radiance value range of the corresponding dataset. These represent the height and width of the image's spatial location, respectively. These are the height and width indices of the spatial location, respectively. , ; For the model in the band Spatial location The predicted value at that location; To correspond to the band Spatial location Reference value at that location; Global band weighted mean square error loss function By calculating the pixel-wise mean square error of all spatial locations in each band, then weighting and summing the loss values ​​of each band, and finally obtaining the global loss value through arithmetic mean.

9. An end-to-end cross-radiometric calibration system for an ocean color satellite sensor, characterized in that, This system is used to execute the end-to-end cross-radiometric calibration method for ocean color satellite sensors as described in any one of claims 1-8. The system is implemented based on the PyTorch deep learning framework and runs on a graphics processor, comprising: The data acquisition module is used to acquire digital counting images from the satellite sensor to be calibrated and radiance images from the spatiotemporally synchronized reference satellite sensor. The dual-branch spatial-spectral feature extraction module includes a spatial local fine-grained coding branch and a spectral global dependency coding branch. The spatial local fine-grained coding branch is used to extract multi-scale local spatial features, and the spectral global dependency coding branch is used to extract global spectral dependencies across all bands. The spatial-spectral fusion feature module is used to fuse the multi-scale local spatial features and the global spectral dependence to output spatial-spectral fusion features; The feature reconstruction module is used to upsample and reconstruct the spatial-spectral fusion features to generate a radiometrically calibrated radiance image.

10. The end-to-end cross-radiometric calibration system for ocean color satellite sensors according to claim 9, characterized in that, The spatial local fine coding branch utilizes the local receptive field and weight sharing characteristics of convolution operations to gradually expand the receptive field during the downsampling process to extract multi-scale spatial features from low-level texture to high-level semantics. The spectral global dependency encoding branch is based on a self-attention mechanism, which captures long-range dependencies between all bands through query-key-value interactions.

Images