Method for retrieving concentration of nutrients in a gulf from multi-source optical remote sensing data fusion
By constructing a dynamic trigger-based remote sensing observation scheduling mechanism and fusing multi-source optical remote sensing data, combined with a hydrodynamic model, the problem of delayed emergency response of remote sensing monitoring systems in sudden pollution events was solved, and dynamic tracking and prediction of nutrient concentration with high spatiotemporal resolution was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MINJIANG UNIVERSITY
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-16
Smart Images
Figure CN122221667A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of temperature sensors in remote sensing technology, specifically involving a method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data. Background Technology
[0002] With the increasing demand for marine ecological environment monitoring, optical remote sensing technology has become an important means of inverting nutrient concentrations in bays due to its advantages such as wide coverage and short observation cycle. Multi-source remote sensing data fusion methods aim to improve inversion accuracy and spatial coverage by integrating spectral information acquired from different sensors. Currently, mainstream remote sensing monitoring systems generally adopt a data acquisition mode with a fixed revisit period. Their observation frequency and spatial resolution are limited by satellite orbit design and mission planning, making it difficult to dynamically adapt to the urgent need for high spatiotemporal resolution monitoring in the event of sudden pollution incidents.
[0003] Retrieval of nutrient concentrations in the Gulf relies on precise modeling of the water's optical properties, typically requiring the integration of multi-source data from polar-orbiting satellites, drones, or ground-based observation stations to compensate for the limitations of a single platform. While existing methods have achieved some success in routine water quality monitoring, their passive response mechanisms suffer from the following problems when dealing with sudden pollution events such as red tides, oil spills, or industrial leaks: geostationary satellites, while possessing high temporal resolution, lack sufficient spatial resolution, making it difficult to identify local anomalies; high-resolution satellites or drones, while providing detailed images, cannot capture the critical evolutionary processes in the early stages of pollution due to mission scheduling delays.
[0004] Existing technologies generally suffer from problems such as delayed emergency response, lack of multi-platform collaboration, and rigid inversion algorithms. Fixed-period data acquisition strategies cannot automatically trigger enhanced observations based on sudden changes in water quality, resulting in missing data during critical windows; there is a lack of linkage mechanisms between multi-source platforms, and detailed investigation resources cannot be scheduled as needed; at the same time, traditional inversion models are not optimized for sudden events, the calculation process is lengthy, and real-time correction capabilities are not embedded.
[0005] The aforementioned deficiencies are particularly prominent in sudden pollution scenarios, severely restricting the timeliness and accuracy of dynamic tracking of nutrient concentrations. There is an urgent need to build a space-air-ground collaborative emergency monitoring system with event perception, autonomous triggering, and rapid inversion capabilities. Summary of the Invention
[0006] This invention provides a method for inverting nutrient concentration in bay waters by fusing multi-source optical remote sensing data. By constructing a dynamic triggering remote sensing observation scheduling mechanism and combining collaborative mission planning of heterogeneous satellite constellations with a multi-scale spatiotemporal interpolation model, it achieves high spatiotemporal resolution and high precision dynamic inversion of nutrient concentration in bay waters under sudden pollution events.
[0007] The present invention provides a method for inverting bay nutrient concentrations by fusing multi-source optical remote sensing data, comprising: Establish a threshold determination mechanism for the water quality anomaly index in the bay. When the water quality anomaly index calculated from real-time remote sensing data is greater than the preset threshold, the multi-source optical remote sensing data fusion and inversion process is triggered. In response to the trigger signal, emergency observation commands are sent to the low-Earth orbit high-resolution optical remote sensing satellite constellation, the medium-Earth orbit medium-resolution wide-swath optical remote sensing satellite constellation, and the geostationary orbit high-frequency imaging optical remote sensing satellite constellation to acquire multi-source optical remote sensing raw data covering the target bay area. The raw data from the multi-source optical remote sensing are preprocessed to obtain the water reflectance products corresponding to each source of remote sensing data. Construct a spatiotemporal fusion model of multi-source water-leaving reflectivity to form a multi-source fused water-leaving reflectivity cube under a unified spatiotemporal grid. Based on the multi-source fusion water-leaving reflectance cube, the input is fed into a pre-trained nutrient concentration inversion neural network model. The nutrient concentration inversion neural network model adopts a convolutional long short-term memory network architecture. The input layer receives the water-leaving reflectance sequence, the hidden layer contains a stacked structure of 3 convolutional long short-term memory units, and the output layer generates the concentration distribution maps of the three main nutrients: nitrate, phosphate and silicate. The nutrient concentration distribution map is coupled with the output of the hydrodynamic numerical model of the bay area. The hydrodynamic numerical model provides velocity field, flow direction field and mixing layer depth field as physical constraint priors. The nutrient concentration field is spatiotemporally smoothed and extrapolated by variational assimilation algorithm to generate dynamic evolution sequence of nutrient concentration.
[0008] Preferably, the water quality anomaly index is constructed based on the statistical regression relationship between the spectral characteristics of remote sensing reflectance during the same period in history and the on-site measured nutrient concentration.
[0009] Preferably, the multi-source optical remote sensing raw data is preprocessed, including: The preprocessing includes atmospheric correction, solar flare removal, geometric fine correction, and tidal level synchronization correction; Atmospheric correction was performed using a vector radiative transfer model. The input parameters included aerosol optical thickness, ozone column concentration, water vapor content, and wind speed. The aerosol optical thickness was obtained by inversion of the remote sensing reflectance of the adjacent clean sea area, the ozone column concentration and water vapor content were extracted from the numerical weather prediction model, and the wind speed data came from the sea surface scatterometer. Solar flare removal is performed based on a bidirectional reflectance distribution function model. The flare reflectance threshold is calculated under the geometric configuration consisting of the solar incidence angle, the observed zenith angle, and the relative azimuth angle. When the water-free reflectance of a pixel is greater than the flare reflectance threshold under the geometric configuration, it is marked as flare contamination and removed. A rational function model is used to perform geometric fine correction, and ground control points are used to perform sub-pixel level registration of the image, with the registration error controlled within 0.5 pixels; The instantaneous tide height at the moment of remote sensing imaging is calculated using a global tidal model, and the spatial coordinates of all remote sensing data are uniformly converted to the mean sea level datum, and tidal level synchronous correction is performed.
[0010] Preferably, a spatiotemporal fusion model of multi-source water-leaving reflectivity is constructed, including: A continuous reference timeline is constructed using timestamps provided every 15 minutes by geostationary satellites; The water reflectivity data of low-Earth orbit and medium-Earth orbit satellites were interpolated using a cubic spline method, aligned to the reference time axis in the time dimension, with the interpolation control points being the water reflectivity data acquired within two hours before and after the interpolation. In terms of spatial dimension, using the 10-meter grid of low-orbit satellite imagery as the baseline spatial grid, the 40-meter data of medium-orbit satellites and the 500-meter raw data of geostationary satellites are resampled to the 10-meter grid through bilinear interpolation. A unified 10m×10m spatial grid and a multi-source fusion water reflectivity cube with a 15-minute time step are formed.
[0011] Preferably, the training dataset of the nutrient concentration inversion neural network model consists of historically synchronized remotely sensed water reflectance and on-site measured nutrient concentrations from shipborne sampling. During training, a joint loss function of mean square error and correlation coefficient is used.
[0012] Preferably, coupling the nutrient concentration distribution map with the output of the hydrodynamic numerical model of the bay area includes: The hydrodynamic numerical model is a three-dimensional primitive equation ocean circulation model. The forcing field includes wind stress, heat flux, freshwater flux and open boundary tidal signal. The model output is updated once per hour. The variational assimilation algorithm is used to fuse the neural network inversion results with the initial nutrient concentration field output by the hydrodynamic model. The objective function is composed of a weighted sum of the background field error term and the observation residual term, and the weight coefficients are dynamically adjusted according to the observation error covariance matrix. After minimizing the objective function, the hydrodynamic-ecological coupled model is driven to perform short-term forecasting, generating a dynamic evolution sequence of nutrient concentrations with a time step of 30 minutes and a spatial resolution of 10 meters for the next 6 hours.
[0013] Preferably, the cubic spline interpolation method avoids the oscillation problem of higher-order polynomial interpolation while maintaining data smoothness.
[0014] Preferably, the ground control points are derived from a high-precision digital elevation model and coastline vector data.
[0015] Preferably, the observation error covariance matrix is dynamically constructed based on the uncertainty quantification results of neural network inversion.
[0016] Preferably, the low-orbit high-resolution optical remote sensing satellite constellation consists of no fewer than three sun-synchronous orbit satellites, and the multispectral imager it carries includes four bands: blue, green, red, and near-infrared.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention breaks through the limitation of traditional remote sensing monitoring relying on fixed-period observation by establishing a dynamic triggering mechanism based on water quality anomaly index. It can automatically start multi-source satellite collaborative observation in the early stage of sudden pollution events, thereby improving the timeliness of emergency response.
[0018] 2. By fusing three types of optical remote sensing data—low-Earth orbit, medium-Earth orbit, and geostationary orbit—complementary enhancement of spatial and temporal resolution was achieved, solving the technical bottleneck that a single data source cannot simultaneously meet the requirements of high spatial detail and high-frequency updates.
[0019] 3. The spatiotemporal adaptive fusion algorithm based on physical constraints effectively eliminates radiation and geometric differences between multi-source data, ensuring the physical consistency of the fused product.
[0020] 4. This invention uses variational assimilation of neural network inversion results with hydrodynamic models, which not only improves inversion accuracy but also enables dynamic prediction of nutrient concentration evolution in short-term periods, providing data support for pollution diffusion early warning, emergency response decision-making, and ecological risk assessment. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the dynamic triggering remote sensing observation scheduling mechanism based on water quality anomaly index in this invention. Figure 3 This is a logical flowchart of the multi-source optical remote sensing data preprocessing and spatiotemporal fusion in this invention; Figure 4 This is a schematic diagram of the logical framework for coupling the nutrient concentration inversion neural network model and the marine water color bio-optical model in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of multi-source satellite constellation (low Earth orbit, medium Earth orbit, geostationary orbit) collaborative observation and data acquisition in this invention; Figure 6 This is a logical flowchart of the spatiotemporal evolution prediction module that couples the nutrient concentration field with the hydrodynamic numerical model in this invention. Detailed Implementation
[0022] refer to Figures 1 to 6 This invention provides a method for inverting nutrient concentration in bay waters by fusing multi-source optical remote sensing data. Its core lies in constructing a dynamic trigger-based remote sensing observation scheduling mechanism and achieving high spatiotemporal resolution and high precision dynamic inversion of nutrient concentration in bay waters under sudden pollution events through heterogeneous satellite constellation collaborative mission planning and multi-scale spatiotemporal interpolation models.
[0023] This method first establishes the triggering conditions for remote sensing data acquisition based on water quality anomaly index; When abnormal water quality signals are detected in the bay area, the emergency access process for multi-source optical remote sensing data is automatically activated. Subsequently, optical remote sensing satellite observation data with different orbital altitudes, different revisit periods, and different spatial resolutions are integrated to form a multi-level spatiotemporal coverage dataset; Based on this, a spatiotemporal adaptive fusion algorithm based on physical constraints is adopted to perform radiometric consistency correction and spatial registration on multi-source remote sensing reflectance data, and an ocean color bio-optical model is introduced to convert the fused remote sensing reflectance data into nutrient concentration distribution. Finally, by embedding a spatiotemporal evolution prediction module with prior information in fluid dynamics, the nutrient concentration field is continuously reconstructed over time to generate dynamic inversion products of nutrient concentration with an update frequency of minutes to hours and a spatial resolution of 10 meters that meet the needs of emergency response.
[0024] The method includes the following steps: S1, Establish a threshold determination mechanism for abnormal water quality index in the bay; S2, in response to the trigger signal, sends emergency observation commands to the low-Earth high-resolution optical remote sensing satellite constellation, the medium-Earth medium-resolution wide-swath optical remote sensing satellite constellation, and the geostationary orbit high-frequency imaging optical remote sensing satellite constellation respectively, to acquire multi-source optical remote sensing raw data covering the target bay area; S3, preprocess the raw multi-source optical remote sensing data; S4, construct a spatiotemporal fusion model of multi-source water reflectivity; S5. Input the fused water-free reflectance cube into the pre-trained nutrient concentration inversion neural network model to generate a nutrient concentration distribution map. S6. The nutrient concentration distribution map is coupled with the output of the hydrodynamic numerical model of the bay area, and a dynamic evolution sequence of nutrient concentration with a time step of 30 minutes and a spatial resolution of 10 meters is generated in the next 6 hours through a variational assimilation algorithm.
[0025] In step S1, a threshold determination mechanism for the bay's water quality anomaly index is established. This mechanism is constructed based on the statistical regression relationship between the spectral characteristics of historical remote sensing reflectance and the field-measured nutrient concentrations. Specifically, the water quality anomaly index is defined as the standard deviation multiple of the ratio of the current remote sensing reflectance in the 555 nm and 490 nm bands relative to the historical average for the same period. Historical data refers to remote sensing reflectance data and corresponding field-measured nutrient concentration data acquired under the same season, time period, and meteorological conditions within the past 5 years.
[0026] A mapping function was established between the reflectance ratio of the 555 nm and 490 nm bands and the total concentrations of nitrates, phosphates, and silicates by performing linear regression analysis on the historical dataset. The residual standard deviation of this mapping function was used to set the anomaly detection threshold. When the water quality anomaly index calculated from real-time remote sensing data exceeds the preset threshold of 2.5, it is determined to be a water quality anomaly and subsequent procedures are triggered. The mapping function is set based on a 99% confidence level critical value under a Gaussian distribution to ensure a false alarm rate of less than 1%.
[0027] The calculation of the water quality anomaly index is performed every 15 minutes, using staring imaging data provided by geostationary satellites as the input source to ensure the timeliness of anomaly detection.
[0028] In step S2, in response to the trigger signal, emergency observation commands are sent to the three types of heterogeneous optical remote sensing satellite constellations respectively.
[0029] The low-Earth orbit high-resolution optical remote sensing satellite constellation consists of no fewer than three sun-synchronous orbit satellites with an orbital altitude of 600 to 800 kilometers and a revisit period of no more than 4 hours. The onboard multispectral imager includes four bands: blue, green, red, and near-infrared, with center wavelengths of 490 nm, 555 nm, 670 nm, and 850 nm, respectively. The ground sampling distance is 10 meters, and the signal-to-noise ratio is greater than 300.
[0030] The medium-resolution wide-swath optical remote sensing satellite constellation operates in an inclined circular orbit at an altitude of 1,000 kilometers, passing over the same area no less than twice a day. Its multispectral sensor contains eight spectral bands, covering the range of 400 nanometers to 900 nanometers, with a ground sampling distance of 40 meters and a swath width of no less than 600 kilometers.
[0031] The geostationary orbit high-frequency imaging optical remote sensing satellite is located above the equator at 105 degrees east longitude. Its staring imager acquires a full-disk image every 15 minutes, and the imaging time resolution for the target bay area can reach 5 minutes, with a raw spatial resolution of 500 meters.
[0032] Emergency observation commands are sent through the standard interface of the National Space Mission Scheduling Center. The commands include the latitude and longitude range of the target area, priority identifier, data format requirements, and return time limit.
[0033] After receiving the instructions, the three types of satellites will prioritize imaging the target bay area within the next transit window and transmit the raw data to the ground receiving station in real time via a high-speed data transmission link.
[0034] In step S3, the raw multi-source optical remote sensing data is preprocessed. Preprocessing includes four sub-steps: atmospheric correction, solar flare removal, geometric fine correction, and tidal level synchronization correction. Atmospheric correction employs a vector radiative transfer model, which considers the effects of multiple scattering, polarization effects, and sea surface roughness on radiative transfer.
[0035] Input parameters include aerosol optical thickness, ozone column concentration, water vapor content, and wind speed. Aerosol optical thickness was obtained by inverting the reflectance of a nearby clean sea area using remote sensing. Specifically, an open sea area 50 km upwind of the target area was selected as a clean reference area, and the aerosol type and optical thickness were inverted using the blue-green band reflectance ratio. Ozone column concentration and water vapor content were extracted from numerical weather prediction models provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), with a spatial resolution of 25 km and a temporal resolution of 6 hours. Wind speed data were obtained from a sea surface scatterometer, with a spatial resolution of 12.5 km and a temporal resolution of 30 minutes.
[0036] Solar flare removal is based on a two-way reflectance distribution function model, which calculates the flare reflectance threshold under a geometric configuration consisting of the solar incidence angle, the observed zenith angle, and the relative azimuth angle. Pixels with water reflectance exceeding this threshold are marked as flare contamination and removed. Geometric fine correction employs a rational function model, using ground control points for sub-pixel-level image registration. These control points are derived from a high-precision digital elevation model and coastline vector data, with registration errors controlled within 0.5 pixels.
[0037] The tidal level synchronization correction uses the global tidal model FES2014 to calculate the instantaneous tidal height at the time of remote sensing imaging and unifies the spatial coordinates of all remote sensing data to the mean sea level datum. This correction is achieved through three-dimensional coordinate transformation, eliminating the influence of shoreline position shifts caused by tidal fluctuations on nutrient concentration inversion, with particularly significant effects in estuaries and shallow bay areas.
[0038] In step S4, a spatiotemporal fusion model of water-leaving reflectivity from multiple sources is constructed. This model uses the time dimension of geostationary orbit satellite data as the reference time axis. Geostationary orbit satellites provide a timestamp every 15 minutes, forming a continuous time series. Water-leaving reflectivity data from low-Earth orbit and medium-Earth orbit satellites are discrete sampling points in time and need to be aligned to this reference time axis. Time alignment employs cubic spline interpolation, which maintains data smoothness while avoiding the oscillation problem of high-order polynomial interpolation.
[0039] During the interpolation process, the time point of the geostationary orbit data is used as the node, and water reflectance data acquired by low-Earth orbit and medium-Earth orbit satellites within two hours before and after the time point are used as control points to generate a continuous time function. Spatially, a 10-meter grid of low-Earth orbit satellite imagery is used as the baseline spatial grid. The 40-meter data from medium-Earth orbit satellites and the original 500-meter data from geostationary orbit satellites are both resampled to a 10-meter grid using bilinear interpolation. The bilinear interpolation performs a weighted average among the four nearest neighbor pixels, with the weights determined by the inverse square distance.
[0040] After resampling, the three types of data were combined into a three-dimensional data cube, namely the multi-source fused water reflectance cube, using a unified 10m × 10m spatial grid and a 15-minute time step. Each voxel of this cube contains water reflectance values in multiple bands within the range of 400 nm to 900 nm, with a time dimension of 6 hours and a spatial dimension covering the entire target bay area.
[0041] In step S5, the multi-source fused water-leaving reflectance cube is input into a pre-trained nutrient concentration inversion neural network model. This neural network model employs a convolutional long short-term memory (LSTM) network architecture. The input layer receives water-leaving reflectance sequences in the 400 nm to 700 nm wavelength range, specifically including five key wavelength bands: 490 nm, 510 nm, 555 nm, 620 nm, and 670 nm. The hidden layer contains a stacked structure of three convolutional LTM units, each with a 3×3 kernel size and 63, 128, and 256 channels respectively.
[0042] The output layer generates concentration distribution maps of three major nutrients: nitrates, phosphates, and silicates, with a spatial resolution of 10 meters and a temporal resolution of 15 minutes. The training dataset for this model consists of historically synchronized remotely sensed water reflectance and nutrient concentrations measured by on-site shipborne sampling. On-site sampling was conducted at no fewer than 50 fixed stations within the target bay area, with a sampling frequency of once per hour for 30 days, covering different tidal phases and meteorological conditions.
[0043] During training, a joint loss function of mean squared error and correlation coefficient was used, with a weight ratio of 1:0.8. The optimizer employed the Adam algorithm, with an initial learning rate of 0.001, a batch size of 31, and 200 training epochs. After training, the model's root mean square error for nitrate concentration retrieval on the independent validation set was less than 0.8 μmol / L, and the correlation coefficient was greater than 0.92.
[0044] In step S6, the nutrient concentration distribution map is coupled with the output of the hydrodynamic numerical model of the bay area. The hydrodynamic numerical model is a three-dimensional primitive equation ocean circulation model with a horizontal grid resolution of 500 meters, a vertical grid of 20 layers, and a time step of 10 seconds. Its forcing field includes wind stress, heat flux, freshwater flux, and open-boundary tidal signals.
[0045] Wind stress is driven by reanalysis wind field data, heat flux and freshwater flux are provided by satellite remote sensing sea surface temperature and precipitation products, and open-boundary tidal signals are specified by a global tidal model. The model output is updated hourly, providing velocity, direction, and mixing layer depth fields as physical constraint priors. The coupling process is implemented using a variational assimilation algorithm. The objective function of this algorithm is... It is composed of a weighted sum of the background field error term and the observation residual term: ; For the nutrient concentration field to be estimated, This is the initial field (background field) of nutrient concentration output by the hydrodynamic model. The results of neural network inversion (observations). For observation operators (mapping the model mesh to the observation location). The background error covariance matrix, The observation error covariance matrix, This is a transpose. Estimated using the ensemble Kalman filter method The objective function is dynamically constructed based on the uncertainty quantification results obtained from neural network inversion. Minimize using the conjugate gradient method, with the iterative convergence threshold set to... .
[0046] The optimal analytical field obtained after assimilation is used as initial conditions to drive the hydrodynamic-ecological coupled model for short-term forecasting, generating a dynamic evolution sequence of nutrient concentrations over the next 6 hours with a time step of 30 minutes and a spatial resolution of 10 meters. This sequence not only smooths out the random noise in the neural network inversion results, but also extrapolates the concentration changes during periods without observation through physical laws, effectively supporting pollution diffusion path simulation and emergency response decision-making.
[0047] The implementation of the above method relies on a complete hardware and software system. This system includes a satellite mission scheduling interface, a high-speed data receiving station, a high-performance computing cluster, an ocean hydrodynamic model server, and a visualization terminal. The satellite mission scheduling interface interfaces with the national aerospace tracking and control network, enabling the issuance of emergency commands within seconds. The high-speed data receiving station is equipped with dual-frequency X-band and Ka-band receiving antennas, with a maximum downlink rate of 600 megabits per second.
[0048] The high-performance computing cluster consists of no fewer than 128 computing nodes, each configured with a dual-processor CPU and four graphics processing units (GPUs) to perform computationally intensive tasks such as atmospheric correction, spatiotemporal fusion, and neural network inference. The ocean hydrodynamic model server runs the ROMS open-source ocean model, automatically initializing and continuously integrating daily. The visualization terminal, developed based on WebGL technology, supports a 3D interactive display of the dynamic evolution sequence of nutrient concentrations, with a time slider accuracy of up to one minute.
[0049] Throughout the methodology, data flows exhibit strict temporal dependencies. Geostationary orbit satellite data serves as the source for anomaly detection and time reference, driving the system's startup and timing alignment. Low Earth Orbit (LEO) and Medium Earth Orbit (MEO) satellite data provide high spatial detail, compensating for the insufficient resolution of GEO data. The preprocessing module ensures the physical consistency of all data, eliminating atmospheric, geometric, and tidal interferences. The spatiotemporal fusion model is the core of multi-source data collaboration, enabling seamless integration of heterogeneous observations within a unified spatiotemporal framework.
[0050] The neural network inversion module converts optical signals into biochemical parameters, and its performance depends on high-quality training data and a reasonable network architecture. The hydrodynamic coupling assimilation module endows the inversion results with physical interpretability and temporal continuity, upgrading them from static snapshots to dynamic predictive products. All modules communicate through standardized data interfaces, using the NetCDF4 specification for data format and adhering to the CFConvention standard for metadata, ensuring the system's scalability and interoperability.
[0051] An anomaly handling mechanism is implemented throughout the entire process. If a certain type of satellite fails to acquire valid data due to malfunction or cloud cover, the system automatically activates a backup strategy: for missing low-Earth orbit (LEO) satellites, the system uses valid data from the previous time interval combined with a hydrodynamic model for short-term extrapolation; for missing medium-Earth orbit (MEO) satellites, the system skips their data and only fuses LEO and geostationary orbit (GEO) data; for interrupted geostationary orbit (GEO) data, the system switches to a fusion strategy based on LEO satellite timestamps and extends the anomaly detection judgment window.
[0052] All abnormal events are logged to the system log for post-event analysis and process optimization. In addition, the neural network inversion results are subject to a reasonableness check: if the output concentration exceeds the historical extreme value range (e.g., nitrate is greater than 50 micromoles per liter), a manual review process is triggered to prevent false warnings caused by model drift.
[0053] The method described in this embodiment was field-verified in a typical bay in the East China Sea. In a simulated red tide outbreak event, the system completed the entire process from satellite scheduling to concentration evolution sequence generation within 3 hours after the abnormal trigger.
[0054] The final product achieved a spatial resolution of 10 meters and a temporal resolution of 30 minutes, successfully capturing the movement trajectory and intensity changes of nutrient concentration fronts. The correlation coefficient with on-site buoy observations reached 0.89, outperforming the traditional single-satellite inversion method (correlation coefficient 0.73). This result demonstrates that the present invention effectively solves the core challenge of missing high spatiotemporal resolution nutrient concentration information in sudden pollution events, providing reliable technical support for marine environmental emergency response.
[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data, characterized in that, include: Establish a threshold determination mechanism for the water quality anomaly index in the bay. When the water quality anomaly index calculated from real-time remote sensing data is greater than the preset threshold, the multi-source optical remote sensing data fusion and inversion process is triggered. In response to the trigger signal, emergency observation commands are sent to the low-Earth orbit high-resolution optical remote sensing satellite constellation, the medium-Earth orbit medium-resolution wide-swath optical remote sensing satellite constellation, and the geostationary orbit high-frequency imaging optical remote sensing satellite constellation to acquire multi-source optical remote sensing raw data covering the target bay area. The raw data from the multi-source optical remote sensing are preprocessed to obtain the water reflectance products corresponding to each source of remote sensing data. Construct a spatiotemporal fusion model of multi-source water-leaving reflectivity to form a multi-source fused water-leaving reflectivity cube under a unified spatiotemporal grid. Based on the multi-source fusion water-leaving reflectance cube, the input is fed into a pre-trained nutrient concentration inversion neural network model. The nutrient concentration inversion neural network model adopts a convolutional long short-term memory network architecture. The input layer receives the water-leaving reflectance sequence, the hidden layer contains a stacked structure of 3 convolutional long short-term memory units, and the output layer generates the concentration distribution maps of the three main nutrients: nitrate, phosphate and silicate. The nutrient concentration distribution map is coupled with the output of the hydrodynamic numerical model of the bay area. The hydrodynamic numerical model provides velocity field, flow direction field and mixing layer depth field as physical constraint priors. The nutrient concentration field is spatiotemporally smoothed and extrapolated by variational assimilation algorithm to generate dynamic evolution sequence of nutrient concentration.
2. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 1, characterized in that, The water quality anomaly index is constructed based on the statistical regression relationship between the historical remote sensing reflectance spectral characteristics and the field-measured nutrient concentrations.
3. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 2, characterized in that, Preprocessing of the raw multi-source optical remote sensing data includes: The preprocessing includes atmospheric correction, solar flare removal, geometric fine correction, and tidal level synchronization correction; Atmospheric correction was performed using a vector radiative transfer model. The input parameters included aerosol optical thickness, ozone column concentration, water vapor content, and wind speed. Aerosol optical thickness was obtained by inversion of remote sensing reflectance from adjacent clean sea areas. Ozone column concentration and water vapor content were extracted from numerical weather prediction models. Wind speed data were obtained from sea surface scatterometers. Solar flare removal is performed based on a bidirectional reflectance distribution function model. The flare reflectance threshold is calculated under the geometric configuration consisting of the solar incidence angle, the observed zenith angle, and the relative azimuth angle. When the water-free reflectance of a pixel is greater than the flare reflectance threshold under the geometric configuration, it is marked as flare contamination and removed. A rational function model is used to perform geometric fine correction, and ground control points are used to perform sub-pixel level registration of the image, with the registration error controlled within 0.5 pixels; The instantaneous tide height at the moment of remote sensing imaging is calculated using a global tidal model, and the spatial coordinates of all remote sensing data are uniformly converted to the mean sea level datum, and tidal level synchronous correction is performed.
4. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 3, characterized in that, Constructing a spatiotemporal fusion model of multi-source water-leaving reflectivity, including: A continuous reference timeline is constructed using timestamps provided every 15 minutes by geostationary satellites; The water reflectivity data of low-Earth orbit and medium-Earth orbit satellites were interpolated using a cubic spline method, aligned to the reference time axis in the time dimension, with the interpolation control points being the water reflectivity data acquired within two hours before and after the interpolation. In terms of spatial dimension, the 10-meter grid of low-orbit satellite imagery is used as the baseline spatial grid. The 40-meter data of medium-orbit satellites and the 500-meter raw data of geostationary satellites are resampled to the 10-meter grid through bilinear interpolation. A unified 10m×10m spatial grid and a multi-source fusion water reflectivity cube with a 15-minute time step are formed.
5. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 4, characterized in that, The training dataset for the nutrient concentration inversion neural network model consists of historically synchronized remotely sensed water reflectance and on-site measured nutrient concentrations from shipborne sampling. During training, a joint loss function of mean square error and correlation coefficient is used.
6. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 5, characterized in that, The nutrient concentration distribution map is coupled with the output of the hydrodynamic numerical model of the bay area, including: The hydrodynamic numerical model is a three-dimensional primitive equation ocean circulation model. The forcing field includes wind stress, heat flux, freshwater flux and open boundary tidal signal. The model output is updated once per hour. The variational assimilation algorithm is used to fuse the neural network inversion results with the initial nutrient concentration field output by the hydrodynamic model. The objective function is composed of a weighted sum of the background field error term and the observation residual term, and the weight coefficients are dynamically adjusted according to the observation error covariance matrix. After minimizing the objective function, the hydrodynamic-ecological coupled model is driven to perform short-term forecasting, generating a dynamic evolution sequence of nutrient concentrations with a time step of 30 minutes and a spatial resolution of 10 meters for the next 6 hours.
7. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 6, characterized in that, The cubic spline interpolation method maintains data smoothness while avoiding the oscillation problem of higher-order polynomial interpolation.
8. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 7, characterized in that, The ground control points are derived from a high-precision digital elevation model and coastline vector data.
9. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 8, characterized in that, The observation error covariance matrix is dynamically constructed based on the uncertainty quantification results of neural network inversion.
10. The method for inverting bay nutrient concentration by fusing multi-source optical remote sensing data according to claim 9, characterized in that, The low-orbit high-resolution optical remote sensing satellite constellation consists of no fewer than three sun-synchronous orbit satellites, and its multispectral imager includes four bands: blue, green, red, and near-infrared.