A rapid absolute radiometric calibration method for ocean color satellite sensors on-orbit

Abstract

The present invention discloses a method for rapid on-orbit absolute radiometric calibration of an ocean color satellite sensor, belonging to the technical field of radiometric calibration. The method comprises the following specific steps: I. analyzing the optical properties of the atmosphere and water at the station site using aerosol optical depth data at 865 nm and normalized water-leaving radiance data at 443 nm in historical AEROENT-OC observation data; II. establishing a machine learning model with on-site multispectral data as input and hyperspectral data as output; and utilizing on-site station data at different locations to calculate the sensor's on-orbit absolute radiometric calibration coefficient in an alternative calibration manner, and verifying the sensor's on-board calibration results. This method can evaluate the overall operating status of the sensor in a relatively short period of time during the initial launch phase, and can fully reflect the sensor's performance over a wide dynamic range.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation calibration, and in particular to an on-orbit rapid absolute radiation calibration method for an ocean color satellite sensor. Background Art

[0002] Accurate radiometric calibration of ocean color satellite sensors is a crucial prerequisite for supporting their diverse global applications, including marine ecosystems, ecological-dynamic interactions, fisheries, and marine pollution. To accurately characterize the sensor's radiometric characteristics at all wavelengths, the sensor undergoes absolute radiometric calibration on the ground using radiation with known characteristics before launch. However, after the sensor is in orbit, the applicability of this absolute radiometric calibration data becomes limited due to factors such as vibration during launch, changes in the operating environment, and component aging over time, resulting in a decline in sensor data quality. To accurately characterize the sensor's radiometric response characteristics after orbit, in-orbit radiometric calibration is typically performed using various methods, including onboard calibration, proxy calibration, and cross-calibration.

[0003] Onboard calibration primarily relies on solar calibration, a method successfully applied to multiple on-orbit sensors (MODIS, VIIRS, and OLCI). However, it uses the sun as the radiation source and can only capture the sensor's radiometric signature at a single response intensity. Alternative calibration uses in-situ measurements as a benchmark, such as those from the established MOBY and BOUSSOLE buoys. Cross-calibration, an extension of the alternative calibration method, utilizes observations from a second, precisely calibrated reference payload as a benchmark. This allows the use of observations from different regions around the globe to ensure coverage of the sensor's dynamic range. Due to the varying atmospheric and water composition across the globe, the top-of-atmosphere radiance received by ocean color satellite sensors exhibits relatively low magnitudes but also exhibits significant variability. However, the limited number of sites used does not fully exploit the advantages of using diverse sites in terms of time and dynamic range. Therefore, developing a rapid, absolute radiometric calibration method for on-orbit ocean color satellite sensors is crucial. Summary of the Invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose a method for rapid on-orbit absolute radiometric calibration of ocean color satellite sensors.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for rapid on-orbit absolute radiometric calibration of an ocean color satellite sensor, the specific steps of the calibration method are as follows:

[0007] I. Analyze the optical properties of the atmosphere and water at the site using the aerosol optical depth data at 865 nm and the normalized water-leaving radiance data at 443 nm from the historical AEROENT-OC observation data.

[0008] II. Build a machine learning model with field multispectral data as input and hyperspectral data as output;

[0009] III. Using the established machine learning model, a hyperspectral field dataset was constructed. Using the OSOAA radiative transfer model, the reconstructed field water-leaving radiance was used as the basis, combined with the field atmospheric and water environmental conditions as input parameters, to calculate the theoretical top-of-atmosphere radiance that matches the satellite sensor.

[0010] IV. Compare the calculated results with the sensor on-board calibration results and the alternative calibration results currently carried out using MOBY on a 6-month basis.

[0011] As a further embodiment of the present invention, the specific steps for analyzing the optical properties of the atmosphere and water at the site using the aerosol optical depth data at 865 nm and the normalized water-leaving radiance data at 443 nm described in step I are as follows:

[0012] S1.1: Extract the aerosol optical depth data at 865 nm and the normalized water-leaving radiance data at 443 nm, i.e., AOD_865 and Lwn_443, from the selected AEROENT-OC historical observation data, and select records with complete data and high spatial coverage consistency within the time range according to user requirements;

[0013] S1.2: Perform data cleaning on the extracted AOD_865 data and Lwn_443 data, unify the data of different temporal or spatial resolutions into the target format, and then align AOD_865 and Lwn_443 by timestamp and spatial position to form analysis data pairs;

[0014] S1.3: Based on the spatial distribution of AOD_865, identify areas of high aerosol concentration and determine the temporal variation of aerosols. Use Lwn_443 to plot spectral characteristic curves for different water types. Analyze the differences in Lwn_443 across different water environments through multi-point comparisons, and record the influencing factors.

[0015] As a further solution of the present invention, the specific steps of establishing the machine learning model in step II are as follows:

[0016] S2.1: Normalize the multispectral data collected on-site to the range [0, 1] and divide the processed multispectral data into a training set and a test set. Then, improve the MLP model to obtain the MRH model, use ReLU as the activation function, and perform a weighted combination of the MSE and MAPE to construct the combined loss function of the model to complete the construction of the machine learning model.

[0017] S2.2: With a training batch size of 64 and a training epoch of 100, the training data is fed into the machine learning model in batches. Each batch of training data passes through the machine learning model's forward propagation, outputting a predicted value. The combined loss function is used to calculate the loss between the predicted value and the actual value. The Adam optimizer adjusts the model weights based on the generated loss value.

[0018] S2.3: After each round of training, the test set is input into the machine learning model, and the loss value is calculated by combining the loss function to evaluate the accuracy of the model in reconstructing the hyperspectral data. The spectral curve comparison chart is used to check the degree of match between the predicted hyperspectral data and the actual value in the entire wavelength range. Training and testing are repeated until the preset training rounds are reached or the model is overfitted. In this case, training is stopped and the final machine learning model is output.

[0019] As a further solution of the present invention, the machine learning model described in S2.1 includes two hidden layers, the first layer has 64 neurons, and the second layer has 32 neurons. The number of neurons in each layer is optimized according to the complexity of the input and output, and the number of neurons in the output layer of the machine learning model is consistent with the number of wavelengths of the target hyperspectral data. Each neuron corresponds to the radiance value of a hyperspectral band. At the same time, the output layer does not use an activation function and directly outputs a linear result to match the continuous spectral characteristics.

[0020] As a further solution of the present invention, the specific steps for calculating the theoretical value of the top-of-atmosphere radiance in step IV are as follows:

[0021] S3.1: Use the hyperspectral data generated by the machine learning model and combine it with the atmospheric and water environmental conditions recorded by AEROENT-OC to construct a hyperspectral field dataset. Then, the hyperspectral water-leaving radiance, field atmospheric conditions, and water optical properties data from the hyperspectral field dataset are input into the OSOAA radiative transfer model.

[0022] S3.2: The OSOAA radiative transfer model simulates the radiative transfer process from the ocean surface to the top of the atmosphere based on the input data and multiple environmental factors, and calculates the radiance at the top of the atmosphere. The simulated TOA radiance is then compared with the actual radiance measured on site for spectral correction.

[0023] As a further solution of the present invention, the environmental factors described in S3.2 specifically include solar and ground conditions, seasonal changes, sunshine cycles, wind speed and ocean fluctuations, vertical structure of the atmosphere, precipitation and clouds, and terrain factors.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] This method is a fast absolute radiometric calibration method for ocean color satellite sensors on-orbit. It uses 865nm aerosol optical depth data and 443nm normalized water-free radiance data to analyze the optical properties of the atmosphere and water at the site. Then, the MLP model is improved to obtain the MRH model, and ReLU is used as the activation function of the model. The MSE and MAPE are weightedly combined to construct the combined loss function of the model to complete the construction of the machine learning model. The machine learning model is trained and tested with a training batch size of 64 and a training round of 100 times. A hyperspectral field dataset is constructed using the established machine learning model, and then the OSOA A radiation transfer model uses the reconstructed on-site water-leaving radiance as its basis and combines on-site atmospheric and water environmental conditions as input parameters to calculate the theoretical value of the top-of-atmosphere radiance that matches the satellite sensor. The calculated results are compared with the sensor's on-board calibration results and the current alternative calibration results using MOBY on a six-month cycle. Using on-site station data at different locations, the sensor's on-orbit absolute radiation calibration coefficient is calculated using an alternative calibration method, and the sensor's on-board calibration results are verified. This allows the evaluation of the sensor's overall operating status to be completed in a relatively short period of time during the initial launch phase, and can fully reflect the sensor's performance over a wide dynamic range. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0027] Figure 1 This is a flowchart of a method for rapid on-orbit absolute radiometric calibration of an ocean color satellite sensor proposed by the present invention;

[0028] Figure 2 This is an algorithm block diagram of an on-orbit rapid absolute radiometric calibration method for ocean color satellite sensors proposed by the present invention;

[0029] Figure 3 This is an error comparison chart of the machine learning model and the BOM model for the on-orbit rapid absolute radiometric calibration method for ocean color satellite sensors proposed in this invention. DETAILED DESCRIPTION

[0030] Reference Figure 1-3A method for rapid on-orbit absolute radiometric calibration of an ocean color satellite sensor is provided. The specific steps of the calibration method are as follows:

[0031] The optical properties of the atmosphere and water at the site were analyzed using the aerosol optical depth data at 865 nm and the normalized water-leaving radiance data at 443 nm in the historical AEROENT-OC observation data.

[0032] Specifically, the aerosol optical depth data at 865nm and the normalized water-offset radiance data at 443nm, namely AOD_865 and Lwn_443, were extracted from the historical AEROENT-OC observation data. According to user needs, records with complete data and high spatial coverage consistency within the time range were screened. The extracted AOD_865 data and Lwn_443 data were cleaned, and data with different time or spatial resolutions were unified into the target format. AOD_865 and Lwn_443 were then aligned according to timestamps and spatial positions to form analysis data pairs. According to the spatial distribution of AOD_865, high aerosol concentration areas were identified, and the temporal variation pattern of aerosols was determined. The spectral characteristic curves of different water types were then drawn using Lwn_443. The differences in Lwn_443 in different water environments were analyzed through multi-point comparison, and the influencing factors were recorded.

[0033] Build a machine learning model with field multispectral data as input and hyperspectral data as output.

[0034] Specifically, the multispectral data collected on site are normalized and unified into the range of [0, 1], and the processed multispectral data are divided into a training set and a test set. Then, the MLP model is improved to obtain the MRH model, and ReLU is used as the activation function of the model. The MSE and MAPE are weightedly combined to construct the combined loss function of the model to complete the construction of the machine learning model. According to the training batch size of 64 and the training rounds of 100 times, the training set data are input into the machine learning model in batches. The training set data of each batch outputs the predicted value after forward propagation of the machine learning model, and the loss value between the predicted value and the actual value is calculated by the combined loss function. The Adam optimizer adjusts the model weight according to the generated loss value. After each round of training, the test set is input into the machine learning model, and the loss value is calculated by the combined loss function to evaluate the accuracy of the model in reconstructing the hyperspectral data. The spectral curve comparison chart is used to check the matching degree between the predicted hyperspectral data and the actual value in the entire wavelength range. The training and testing are repeated until the preset training rounds are reached or the model is overfitted. The training is stopped and the final machine learning model is output.

[0035] It should be further explained that the machine learning model contains two hidden layers. The first layer has 64 neurons and the second layer has 32 neurons. The number of neurons in each layer is optimized according to the complexity of the input and output, and the number of neurons in the output layer of the machine learning model is consistent with the number of wavelengths of the target hyperspectral data. Each neuron corresponds to the radiance value of a hyperspectral band. At the same time, the output layer does not use an activation function and directly outputs a linear result to match the continuous spectral characteristics.

[0036] refer to Figure 3 It can be seen that a machine learning model is established to realize the reconstruction of multispectral field observation data, and the model performance is verified by comparing with the results obtained by the existing empirical relationship method. The errors of the hyperspectral reconstruction results of the visible light domain field data by the MRH and BOM models are as follows: except for the 17.42% error in the 670 nm band, the error of the MRH model in the 400-580 nm band is less than 5%, and the error in the 580-670 nm band is less than 9%. The error gradually increases after the 670 nm band, which shows that in the top-of-atmosphere radiance it reflects, the uncertainty in the blue-green band is 0.5%, and the uncertainty in the red light band is between 1% and 2%. In comparison, the error generated by the BOM model is always higher than that of the MRH model, with the maximum reconstruction error reaching 34.19%.

[0037] Using the established machine learning model, a hyperspectral field dataset was constructed. Then, with the help of the OSOAA radiation transfer model, based on the reconstructed field water-leaving radiance, combined with the field atmospheric and water environmental conditions as input parameters, the theoretical value of the top of the atmosphere radiance that matches the satellite sensor was calculated.

[0038] Specifically, a hyperspectral field dataset was constructed using the hyperspectral data generated by the machine learning model, combined with the atmospheric and water environmental conditions recorded by AEROENT-OC. The hyperspectral off-water radiance, on-site atmospheric conditions, and water optical properties data in the hyperspectral field dataset were then input into the OSOAA radiation transfer model. The OSOAA radiation transfer model simulates the radiation transfer process from the ocean surface to the top of the atmosphere based on each set of input data and multiple sets of environmental factors, and calculates the top-of-atmosphere radiance. The simulated TOA radiance is then compared with the actual radiance measured on-site for spectral correction.

[0039] It should be further explained that environmental factors specifically include solar and ground conditions, seasonal changes, sunlight cycles, wind speed and ocean fluctuations, the vertical structure of the atmosphere, precipitation and clouds, and topographic factors.

[0040] The calculated results are compared with the sensor on-board calibration results and the current alternative calibration results using MOBY on a 6-month basis.

[0041] Furthermore, it should be noted that the calculated results are compared with actual satellite observations and with the existing MOBY results, highlighting the advantages of the present invention in terms of the number of matches and sensor dynamic range. A hyperspectral in situ dataset reconstructed using a machine learning model was used to perform an alternative calibration of the MODIS-Aqua sensor and compared with the top-of-atmosphere radiance obtained from its onboard calibration at red (678nm), green (547nm), and blue (412nm) wavelengths. A comparison of the top-of-atmosphere radiance coverage as measured by AERONET-OC (red dots) and MOBY (blue dots) is also presented over a six-month period. From the perspective of alternative calibration and onboard calibration, all comparison results show correlation coefficients greater than 0.97, with absolute mean percentage errors less than 3.70%. From the perspective of the top-of-atmosphere radiance coverage as measured by AERONET-OC and MOBY, AERONET-OC matched MODIS-Aqua 56 times within the set period, while MOBY matched only three times. At 412 nm, AERONET-OC reports a radiance range of 4-12 mW / cm2 / μm / sr at the top of the atmosphere, while MOBY covers a range of only 9-10.5 mW / cm2 / μm / sr. At 555 nm, AERONET-OC reports a radiance range of 1.5-5 mW / cm2 / μm / sr at the top of the atmosphere, while MOBY covers a range of only 3-3.5 mW / cm2 / μm / sr. At 678 nm, AERONET-OC reports a radiance range of 0.7-2.2 mW / cm2 / μm / sr at the top of the atmosphere, while MOBY covers a range of only 1.1-1.4 mW / cm2 / μm / sr.

Claims

1. A method for rapid absolute radiometric calibration of an ocean color satellite sensor on-orbit, characterized in that: The specific steps of this calibration method are as follows: I. Analyze the optical properties of the atmosphere and water at the site using the aerosol optical depth data at 865 nm and the normalized water-leaving radiance data at 443 nm from the historical AEROENT-OC observation data. II. Build a machine learning model with field multispectral data as input and hyperspectral data as output; III. Using the established machine learning model, a hyperspectral field dataset was constructed. Using the OSOAA radiative transfer model, the reconstructed field water-leaving radiance was used as the basis, combined with the field atmospheric and water environmental conditions as input parameters, to calculate the theoretical top-of-atmosphere radiance that matches the satellite sensor. IV. Compare the calculated results with the sensor onboard calibration results and the current alternative calibration results using MOBY on a 6-month basis; The specific steps for building a machine learning model in Step II are as follows: S2.1: Normalize the multispectral data collected on-site to the range [0, 1] and divide the processed multispectral data into a training set and a test set. Then, improve the MLP model to obtain the MRH model, use ReLU as the activation function, and perform a weighted combination of the MSE and MAPE to construct the combined loss function of the model to complete the construction of the machine learning model. S2.2: With a training batch size of 64 and a training epoch of 100, the training data is fed into the machine learning model in batches. Each batch of training data passes through the machine learning model's forward propagation, outputting a predicted value. The combined loss function is used to calculate the loss between the predicted value and the actual value. The Adam optimizer adjusts the model weights based on the generated loss value. S2.3: After each round of training, the test set is input into the machine learning model, and the loss value is calculated using the combined loss function to evaluate the accuracy of the model in reconstructing the hyperspectral data. The spectral curve comparison chart is used to check the degree of match between the predicted hyperspectral data and the actual value across the entire wavelength range. Training and testing are repeated until the preset training rounds are reached or the model is overfitted. In this case, training is stopped and the final machine learning model is output. The machine learning model described in S2.1 contains two hidden layers. The first layer has 64 neurons and the second layer has 32 neurons. The number of neurons in each layer is optimized according to the complexity of the input and output. The number of neurons in the output layer of the machine learning model is consistent with the number of wavelengths of the target hyperspectral data. Each neuron corresponds to the radiance value of a hyperspectral band. At the same time, the output layer does not use an activation function and directly outputs a linear result to match the continuous spectral characteristics.

2. The method for rapid absolute radiometric calibration of an ocean color satellite sensor on-orbit according to claim 1, characterized in that: The specific steps for analyzing the optical properties of the atmosphere and water at the site using the 865nm aerosol optical depth data and the 443nm normalized water-leaving radiance data described in Step I are as follows: S1.1: Extract the aerosol optical depth data at 865 nm and the normalized water-leaving radiance data at 443 nm, i.e., AOD_865 and Lwn_443, from the selected AEROENT-OC historical observation data, and select records with complete data and high spatial coverage consistency within the time range according to user requirements; S1.2: Perform data cleaning on the extracted AOD_865 data and Lwn_443 data, unify the data of different temporal or spatial resolutions into the target format, and then align AOD_865 and Lwn_443 by timestamp and spatial position to form analysis data pairs; S1.3: Based on the spatial distribution of AOD_865, identify areas of high aerosol concentration and determine the temporal variation of aerosols. Use Lwn_443 to plot spectral characteristic curves for different water types. Analyze the differences in Lwn_443 across different water environments through multi-point comparisons, and record the influencing factors.

3. The method for rapid absolute radiometric calibration of an ocean color satellite sensor on-orbit according to claim 1, characterized in that: The specific steps for calculating the theoretical value of top-of-atmosphere radiance in step III are as follows: S3.1: Use the hyperspectral data generated by the machine learning model and combine it with the atmospheric and water environmental conditions recorded by AEROENT-OC to construct a hyperspectral field dataset. Then, the hyperspectral water-leaving radiance, field atmospheric conditions, and water optical properties data from the hyperspectral field dataset are input into the OSOAA radiative transfer model. S3.2: The OSOAA radiative transfer model simulates the radiative transfer process from the ocean surface to the top of the atmosphere based on the input data and multiple environmental factors, and calculates the radiance at the top of the atmosphere. The simulated TOA radiance is then compared with the actual radiance measured on site for spectral correction.

4. The method for rapid absolute radiometric calibration of an ocean color satellite sensor on-orbit according to claim 3, characterized in that: The environmental factors described in S3.2 specifically include solar and ground conditions, seasonal variations, insolation cycles, wind speeds and ocean fluctuations, the vertical structure of the atmosphere, precipitation and clouds, and topographic factors.

Images