A method for remote sensing inversion of nutrient salt concentration in near-shore sea area based on satellite fusion

Abstract

The application discloses a kind of based on satellite fusion's near-shore sea area nutrient salt concentration remote sensing inversion method, comprising the following steps: S1, the original L1C data of the optical image of high spatial resolution satellite and low spatial resolution satellite are respectively carried out Rayleigh correction, obtain after Rayleigh correction remote sensing reflectance;S2, cloud and land pixel in after Rayleigh correction remote sensing reflectance data are carried out mask processing;S3, based on low space respectively satellite data constructs nutrient salt training dataset;S4, constructs cross-satellite fusion training dataset;S5, establish AutoGluon-DIN machine learning model and AutoGluon-DIP machine learning model based on low resolution satellite, and model training is carried out;S6, establish AutoGluon-transfer machine learning model of fusion high resolution and low resolution satellite, and model training is carried out;S7, sequentially apply AutoGluon-transfer machine learning, AutoGluon-DIN machine learning model, AutoGluon-DIP machine learning model after training to high resolution satellite, obtain high spatial resolution's near-shore sea area nutrient salt concentration remote sensing inversion product.

Description

Technical Field

[0001] This invention relates to the field of marine observation technology, specifically to a remote sensing inversion method for nutrient concentration in nearshore waters based on satellite fusion. Background Technology

[0002] Nutrients such as nitrogen and phosphorus promote phytoplankton growth, drive primary production, and maintain biodiversity. However, human activities, including agriculture, urban runoff, and industrial emissions, often result in excessive nutrient discharge, leading to eutrophication of nearshore waters. This process not only degrades water quality but also disrupts aquatic ecosystems and triggers harmful algal blooms. Algal blooms negatively impact water resources, marine life, and local economies reliant on tourism and fisheries; therefore, effective water quality monitoring and assessment are crucial. The concentrations of dissolved inorganic nitrogen (DIN) and dissolved inorganic phosphorus (DIP) are commonly used to characterize nutrient levels and are key parameters for assessing water quality.

[0003] Because the physical mechanism between optical satellite signals and the concentration of nutrients in water (including dissolved inorganic nitrogen and dissolved inorganic phosphorus) is not yet clear, most optical satellite nutrient concentration retrieval algorithms still rely on empirical and neural network algorithms. Input parameters include satellite products such as remote sensing reflectance (Rrs), sea surface temperature (SST), chlorophyll concentration (Chl), and sea surface salinity (SSS). However, for nearshore waters, environmental factor products such as SST, Chl, and SSS have low resolution and are difficult to use as effective inputs for algorithms. Furthermore, Rrs requires atmospheric correction, and atmospheric correction algorithms have significant uncertainties in nearshore shallow waters. If atmospheric correction fails, it can easily lead to invalid or erroneous Rrs data, thus affecting the retrieval accuracy of DIN and DIP, and making it impossible to obtain reliable remote sensing nutrient distribution. In particular, the spatiotemporal heterogeneity of nearshore waters is significant, and operational nutrient monitoring requires even higher spatiotemporal resolution. Furthermore, especially for small estuaries and bays like Xiamen Bay, nutrient retrieval algorithms require high spatial resolution satellite data input. However, due to the long revisit period of high spatial resolution satellites, the amount of measured data that can be matched is relatively small, thus limiting model development. 。 Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a remote sensing inversion method for nutrient concentration in nearshore waters based on satellite fusion. This method can accurately estimate the nutrient concentration distribution of DIN and DIP in nearshore waters using the AutoGluon automatic machine learning model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A remote sensing inversion method for nutrient concentration in nearshore waters based on satellite fusion includes the following steps:

[0007] S1. Rayleigh correction is performed on the raw L1C data of optical images from high spatial resolution satellites and low spatial resolution satellites respectively to obtain the Rayleigh-corrected remote sensing reflectance. The calculation formula is as follows:

[0008]

[0009] In the formula, ρ rc ρ is the Rayleigh-corrected remote sensing reflectance; λ is the wavelength; t The reflectivity observed by satellite sensors at the top of the atmosphere; t g ρ is the gas transmittance in the atmosphere. r Rayleigh reflectance due to multiple molecular scattering under aerosol-free conditions; T is direct transmittance; ρ g This is a scattered signal caused by a solar flare;

[0010] S2. Mask the cloud and land pixels in the Rayleigh-corrected remote sensing reflectance data;

[0011] S3, The ρ values ​​of the low spatial resolution satellites selected in step S2 rc Water body pixels are matched with the measured nutrient concentrations of the corresponding pixels to construct a nutrient training dataset;

[0012] S4. The ρ values ​​of the low spatial resolution satellites selected in step 2 are... rc Water body pixels and high spatial resolution satellites over the same study area on the same day. rc Water body pixels are matched to construct a satellite fusion training dataset;

[0013] S5. Establish an AutoGluon-DIN machine learning model for estimating dissolved inorganic nitrogen concentration and an AutoGluon-DIP machine learning model for estimating dissolved inorganic phosphorus concentration, and train the AutoGluon-DIN machine learning model and the AutoGluon-DIP machine learning model respectively using the nutrient training dataset.

[0014] S6. Establish 8 AutoGluon-transfer machine learning models to map the band signals of high spatial resolution satellites to bands within the common wavelength range of low spatial resolution satellites; train the 8 AutoGluon-transfer machine learning models using the satellite fusion training dataset.

[0015] S7. The trained AutoGluon-DIN machine learning model, AutoGluon-DIP machine learning model, and AutoGluon-transfer machine learning model were applied to the remote sensing inversion of nutrient concentration in nearshore waters by satellite fusion.

[0016] Preferably, the specific process of step S2 is as follows: a threshold is set according to the characteristics of the satellite sensor, and pixels that are higher than the threshold are determined to be cloud or land pixels and are directly removed, retaining only the bands within the wavelength range shared by high spatial resolution satellites and low spatial resolution satellites.

[0017] Preferably, the matching process in step S3 covers measured nutrient concentration data under different seasons and atmospheric conditions.

[0018] Preferably, during model training in step S5, the input to the AutoGluon-DIN machine learning model is the ρ of a low spatial resolution satellite. rc (λ) and related satellite metadata, outputting dissolved inorganic nitrogen concentration; the AutoGluon-DIP machine learning model inputs ρ from low spatial resolution satellites. rc (λ) and related satellite metadata, output as dissolved inorganic phosphorus concentration.

[0019] Preferably, during model training in step S6, the inputs to the eight AutoGluon-transfer machine learning models are the ρ values ​​of the high spatial resolution satellites. rc (λ), the output is the ρ of low spatial resolution satellites. rc (λ).

[0020] Preferably, the specific process of step S7 is as follows: the trained AutoGluon-DIN machine learning model and AutoGluon-DIP machine learning model are applied to the low spatial resolution satellite optical images in the satellite fusion verification dataset in step S4 to generate a low spatial resolution nutrient concentration distribution map of the nearshore sea area on the same day; the trained AutoGluon-DIN machine learning model, AutoGluon-DIP machine learning model and AutoGluon-transfer machine learning model are applied to the high spatial resolution satellite optical images in the satellite fusion verification dataset in step S4 to generate a high spatial resolution nutrient concentration distribution map of the nearshore sea area on the same day; the high spatial resolution nutrient concentration distribution is resampled to low spatial resolution and compared with the low spatial resolution nutrient concentration distribution map to draw a density scatter plot comparison map.

[0021] After adopting the above technical solution, the present invention has the following beneficial effects: The present invention

[0022] 1. This invention utilizes the Rayleigh-corrected remote sensing reflectance ρ rc As input to the algorithm, it does not require atmospheric correction, thus avoiding the problems of invalid and erroneous Rrs data. Based on this, multiple neural network models are integrated to significantly improve the accuracy of the model in retrieving nutrient concentrations in nearshore waters such as Xiamen Bay.

[0023] 2. This invention first constructs a nutrient inversion algorithm based on low spatial resolution satellite matching data, then develops a data fusion algorithm for satellites of different resolutions, and then the constructed algorithm can be directly applied to high spatial resolution satellite data, effectively solving the problem of insufficient training data when matching high spatial resolution satellites with measured data due to long revisit times.

[0024] 3. This invention uses an open-source automatic machine learning algorithm that integrates multiple neural network models. Compared with traditional empirical algorithms and single neural network models, it significantly improves the accuracy of the nutrient concentration inversion model in the Xiamen Bay area, enabling high spatiotemporal resolution remote sensing observation of nutrients in small bays. Attached Figure Description

[0025] Figure 1 This is a diagram of the AutoGluon neural network structure used for nutrient estimation in an embodiment of the present invention;

[0026] Figure 2 This is the verification result of the nutrient concentration algorithm in the embodiments of the present invention;

[0027] Figure 3 This is a diagram of the AutoGluon neural network structure used for band migration of Sentinel-2 and Sentinel-3 satellites in an embodiment of the present invention.

[0028] Figure 4 This is a low-spatial-resolution and high-spatial-resolution nutrient concentration distribution map of the Xiamen Bay area on the same day, obtained by inverting Sentinel-3 and Sentinel-2 satellite image data in an embodiment of the present invention.

[0029] Figure 5 This is a comparison of normalized density scatter plots of nutrient concentrations at low and high spatial resolutions in the Xiamen Bay area on the same day in an embodiment of the present invention.

[0030] Figure 6 This is a flowchart of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] like Figures 1 to 6 As shown, a remote sensing inversion method for nutrient concentration in nearshore waters based on satellite fusion includes the following steps:

[0033] S1. Rayleigh correction is performed on the raw L1C data of optical images from high spatial resolution satellites and low spatial resolution satellites respectively to obtain the Rayleigh-corrected remote sensing reflectance. The calculation formula is as follows:

[0034]

[0035] In the formula, ρ rc ρ is the Rayleigh-corrected remote sensing reflectance; λ is the wavelength; t The reflectivity observed by satellite sensors at the top of the atmosphere; t g ρ is the gas transmittance in the atmosphere. r Rayleigh reflectance due to multiple molecular scattering under aerosol-free conditions; T is direct transmittance; ρ g This is a scattered signal caused by a solar flare;

[0036] S2. Mask the cloud and land pixels in the Rayleigh-corrected remote sensing reflectance data;

[0037] The specific process of step S2 is as follows: set a threshold according to the characteristics of the satellite sensor. Pixels that are higher than the threshold are judged as cloud or land pixels and are directly removed. Only the bands in the wavelength range shared by high spatial resolution satellites and low spatial resolution satellites are retained.

[0038] S3, The ρ values ​​of the low spatial resolution satellites selected in step S2 rc Water body pixels are matched with the measured nutrient concentrations of the corresponding pixels to construct a nutrient training dataset;

[0039] The matching process in step S3 covers measured nutrient concentration data under different seasons and atmospheric conditions;

[0040] S4. The ρ values ​​of the low spatial resolution satellites selected in step 2 are... rc Water body pixels and high spatial resolution satellites over the same study area on the same day. rc Water body pixels are matched to construct a satellite fusion training dataset;

[0041] S5. Establish an AutoGluon-DIN machine learning model for estimating dissolved inorganic nitrogen concentration and an AutoGluon-DIP machine learning model for estimating dissolved inorganic phosphorus concentration, and train the AutoGluon-DIN machine learning model and the AutoGluon-DIP machine learning model respectively using the nutrient training dataset.

[0042] During model training in step S5, the input to the AutoGluon-DIN machine learning model is the ρ of a low spatial resolution satellite. rc(λ) and related satellite metadata, outputting dissolved inorganic nitrogen concentration; the AutoGluon-DIP machine learning model inputs ρ from low spatial resolution satellites. rc (λ) and related satellite metadata, output as dissolved inorganic phosphorus concentration;

[0043] S6. Establish 8 AutoGluon-transfer machine learning models to map the band signals of high spatial resolution satellites to bands within the common wavelength range of low spatial resolution satellites; train the 8 AutoGluon-transfer machine learning models using the satellite fusion training dataset.

[0044] During model training in step S5, the input to the AutoGluon-DIN machine learning model is the ρ of a low spatial resolution satellite. rc (λ) and related satellite metadata, outputting dissolved inorganic nitrogen concentration; the AutoGluon-DIP machine learning model inputs ρ from low spatial resolution satellites. rc (λ) and related satellite metadata, output as dissolved inorganic phosphorus concentration;

[0045] S7. The trained AutoGluon-DIN machine learning model, AutoGluon-DIP machine learning model and AutoGluon-transfer machine learning model are applied to the satellite-fused remote sensing inversion of nutrient concentration in nearshore waters.

[0046] The specific process of step S7 is as follows: The trained AutoGluon-DIN machine learning model and AutoGluon-DIP machine learning model are applied to the low spatial resolution satellite optical images in the satellite fusion verification dataset in step S4 to generate a low spatial resolution nutrient concentration distribution map of the nearshore sea area on the same day; The trained AutoGluon-DIN machine learning model, AutoGluon-DIP machine learning model and AutoGluon-transfer machine learning model are applied to the high spatial resolution satellite optical images in the satellite fusion verification dataset in step S4 to generate a high spatial resolution nutrient concentration distribution map of the nearshore sea area on the same day; The high spatial resolution nutrient concentration distribution is resampled to low spatial resolution and compared with the low spatial resolution nutrient concentration distribution map to draw a density scatter plot comparison map.

[0047] Example of using the technical solution of this invention to invert nutrient concentration in the Xiamen Bay area:

[0048] Download optical satellite imagery of the Xiamen Bay area. In this embodiment, Sentinel-2 is used as a high spatial resolution optical satellite and Sentinel-3 as a low spatial resolution satellite. Rayleigh correction is performed on the raw L1C satellite imagery using the open-source software Acolite, and the ρ value of each pixel is calculated. rc The calculation formula is:

[0049]

[0050] In the formula, ρ rc ρ is the Rayleigh-corrected remote sensing reflectance; λ is the wavelength; t The reflectivity observed by satellite sensors at the top of the atmosphere; t g ρ is the gas transmittance in the atmosphere. r Rayleigh reflectance due to multiple molecular scattering under aerosol-free conditions; T is direct transmittance; ρ g This is a scattered signal caused by a solar flare;

[0051] With the ρ of the pixel rc The (865) value is used as the criterion for determining whether a pixel is a cloud or a land cell. If the value is greater than 0.2, it is marked as a land or cloud cell and masked; if the value is less than 0.2, it is marked as a water cell and retained. All bands of Sentinel-3 have a spatial resolution of 300m, while the spatial resolution of different bands of Sentinel-2 is inconsistent, ranging from 10m to 60m. To facilitate the construction of the nutrient inversion algorithm, linear interpolation was used to unify the resampling of different bands of Sentinel-2 satellites to a spatial resolution of 10m. In order to construct the satellite fusion algorithm, only the common or similar bands of Sentinel-2 and Sentinel-3 were retained, and finally 8 bands were selected; the bands of Sentinel-2 were marked as λ. S2 The wavelengths include 443, 490, 560, 620, 709, 754, 779, and 865 nm; the band designation for Sentinel-3 is λ. S3 The corresponding nm wavelengths are 443, 492, 560, 665, 704, 740, 783 and 865 nm.

[0052] Sentinel-3 satellite ρ rc (λ S3This study matched water body pixels with measured nutrient data to construct a nutrient dataset. The measured data used in this example included monitoring data, survey data, and data from four buoys located in the Baozhuyu (BZY), Yefengzhai (YFZ), Tong'an Bay (TAW), and Hulishan (HLS) areas of Xiamen Bay. The study area was Xiamen Bay, and to improve data representativeness, the coverage of the monitoring data was expanded to include the Fujian coast. Sentinel-3 satellite data from 2018 to 2022 under different seasons and atmospheric conditions were matched with the measured nutrient data. The matched data from the Hulishan buoy was used to construct a nutrient validation dataset with 390 data sets; the matched data from the other measured data were used to construct a nutrient training dataset with 2117 data sets.

[0053] Sentinel-3 satellite ρ rc (λ S3 Water body pixels and the Sentinel-2 satellite ρ on the same day rc (λ S2 Water body pixels were matched to construct a satellite fusion training dataset. Sentinel-3 and Sentinel-2 satellite images that passed over Xiamen Bay on the same day throughout 2022 were matched, covering different seasons and atmospheric conditions, resulting in 254,233 matching sets. In addition, Sentinel-2 and Sentinel-3 images from October 21, 2019, were matched separately to construct a satellite fusion validation dataset.

[0054] Two AutoGluon machine learning models were established for concentration estimation of DIN and DIP, respectively labeled AutoGluon-DIN and AutoGluon-DIP (see [link to model]). Figure 1 In this embodiment, the nutrient training dataset from step 3 is used to train the model. The model's input is the Sentinel-3 satellite ρ... rc (λ), ρ rc (λ) / ρ rc (560) and image date, outputting ln(DIN) or ln(DIP) respectively. After testing, the input variable ρ was increased. rc (λ) / ρ rc (560) and image date can effectively improve the accuracy of the model; taking the logarithm of DIN / DIP as the output can effectively stabilize the variance and reduce data skew, significantly improving the robustness of the model. The number of stacking layers of AutoGluon is set to 1, and the number of cross-validation folds is set to 5.

[0055] The accuracy of the AutoGluon-DIN / DIP model was validated using a nutrient validation dataset. Time-series plots of measured DIN / DIP (DINin situ / DINin situ) and retrieved DIN / DIP (DINest / DINest) for 2022 were generated, and the absolute correlation coefficient (R²) was calculated. 2 ), root mean square error (RMSD) and mean absolute correlation coefficient (MAPD) ​​(see Figure 2 Overall, R 2 The values ​​close to 0.6 and MAPD less than 30% indicate that the nutrient concentration estimated by the satellite is in good agreement with the concentration measured by the buoy, proving that the nutrient model of this invention has high accuracy.

[0056] Eight AutoGluon machine learning models (AutoGluon-transfer) were established to convert band data from the Sentinel-2 satellite into corresponding band signals from the Sentinel-3 satellite (see...). Figure 3 The model, through training, will apply the Sentinel-2 band ρ... rc (λ S2 ) and Sentinel-3 band ρ rc (λ S3 The relationships between the data are mapped to enable the application of the AutoGluon-DIN / DIP model to Sentinel-2 data at high spatial resolution. In this example, the model is trained using a satellite fusion training dataset. The inputs to the eight models are the ρ values ​​from the Sentinel-2 satellites. rc (λ S2 The output is the ρ of the Sentinel-3 satellite. rc (λ S3 ).

[0057] The AutoGluon-DIN / DIP model was applied to Sentinel-3 satellite imagery in the satellite fusion validation dataset to generate a low spatial resolution (300m) nutrient concentration distribution product map of Xiamen Bay. Simultaneously, the AutoGluon-DIN / DIP and AutoGluon-transfer models were applied to Sentinel-2 satellite imagery in the same dataset to generate a high spatial resolution (10m) nutrient concentration distribution product map of Xiamen Bay for the same day (see...). Figure 4 The high spatial resolution nutrient concentration distribution product was resampled to a low spatial resolution and compared with the low spatial resolution nutrient concentration distribution map to create a density scatter plot (see...). Figure 5 Overall, most points are located near the 1:1 line, R 2The values ​​close to 0.6 and MAPD less than 20% indicate that the inversion results at the two resolutions have good consistency.

[0058] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A remote sensing inversion method for nutrient concentration in nearshore waters based on satellite fusion, characterized in that, Includes the following steps: S1. Rayleigh correction is performed on the raw L1C data of optical images from high spatial resolution satellites and low spatial resolution satellites respectively to obtain the Rayleigh-corrected remote sensing reflectance at wavelength λ. The calculation formula is as follows: In the formula, Rayleigh-corrected remote sensing reflectance; λ Wavelength; The reflectivity observed by satellite sensors at the top of the atmosphere; The transmittance of gases in the atmosphere; Rayleigh reflectance is generated by multiple molecular scattering under aerosol-free conditions; T is direct transmittance. This is a scattered signal caused by a solar flare; S2. Mask the cloud and land pixels in the Rayleigh-corrected remote sensing reflectance data; S3. Select the low spatial resolution satellites from step S2. Water body pixels are matched with the measured nutrient concentrations of the corresponding pixels to construct a nutrient training dataset; S4. Select the low spatial resolution satellites from step 2. Water body pixels and high spatial resolution satellites passing over the study area on the same day Water body pixels are matched to construct a satellite fusion training dataset; S5. Establish an AutoGluon-DIN machine learning model for estimating dissolved inorganic nitrogen concentration and an AutoGluon-DIP machine learning model for estimating dissolved inorganic phosphorus concentration, and train the AutoGluon-DIN machine learning model and the AutoGluon-DIP machine learning model respectively using the nutrient training dataset. S6. Establish 8 AutoGluon-transfer machine learning models to map the band signals of high spatial resolution satellites to bands within the common wavelength range of low spatial resolution satellites; train the 8 AutoGluon-transfer machine learning models using the satellite fusion training dataset. S7. The trained AutoGluon-DIN machine learning model, AutoGluon-DIP machine learning model and AutoGluon-transfer machine learning model are applied to the satellite-fused remote sensing inversion of nutrient concentration in nearshore waters. The specific process of step S7 is as follows: The trained AutoGluon-DIN machine learning model and AutoGluon-DIP machine learning model are applied to the low spatial resolution satellite optical images in the satellite fusion verification dataset in step S4 to generate a low spatial resolution nutrient concentration distribution map of the nearshore sea area on the same day; The trained AutoGluon-DIN machine learning model, AutoGluon-DIP machine learning model and AutoGluon-transfer machine learning model are applied to the high spatial resolution satellite optical images in the satellite fusion verification dataset in step S4 to generate a high spatial resolution nutrient concentration distribution map of the nearshore sea area on the same day; The high spatial resolution nutrient concentration distribution is resampled to low spatial resolution and compared with the low spatial resolution nutrient concentration distribution map to draw a density scatter plot comparison map.

2. The method for remote sensing inversion of nutrient concentration in nearshore waters based on satellite fusion as described in claim 1, characterized in that, The specific process of step S2 is as follows: set a threshold according to the characteristics of the satellite sensor. Pixels that are higher than the threshold are judged as cloud or land pixels and are directly removed. Only the bands within the wavelength range shared by high spatial resolution satellites and low spatial resolution satellites are retained.

3. The method for remote sensing inversion of nutrient concentration in nearshore waters based on satellite fusion as described in claim 1, characterized in that, The matching process in step S3 covers measured nutrient concentration data under different seasons and atmospheric conditions.

4. The method for remote sensing inversion of nutrient concentration in nearshore waters based on satellite fusion as described in claim 1, characterized in that: During model training in step S5, the input to the AutoGluon-DIN machine learning model is low spatial resolution satellite data. ( λ The AutoGluon-DIP machine learning model takes low spatial resolution satellite data as input and related satellite metadata as input, and outputs dissolved inorganic nitrogen concentration. ( λ The data includes satellite metadata and outputs the dissolved inorganic phosphorus concentration.

5. The method for remote sensing inversion of nutrient concentration in nearshore waters based on satellite fusion as described in claim 1, characterized in that: During model training in step S6, the inputs to the eight AutoGluon-transfer machine learning models are high spatial resolution satellite data. ( λ The output is low spatial resolution satellite data. ( λ ).

Images