Marine oil spill multi-source data fusion monitoring method and system based on deep learning

Abstract

The invention discloses an offshore oil spill multi-source data fusion monitoring method and system based on deep learning, and the method comprises the steps: collecting remote sensing data and land-sea data of a preset region, and carrying out the preprocessing of the remote sensing data and the land-sea data; carrying out oil spill identification according to the synthetic aperture radar data to obtain a suspected oil spill area, and carrying out comparison and oil-like film removal on the suspected oil spill area according to the remote sensing data and the land-sea data to obtain an adjusted oil spill area; performing time sequence discrimination feature extraction on the remote sensing data and the land-sea data of the adjusted oil spill area to obtain an oil spill discrimination rule, and constructing a fusion monitoring function in the adjusted oil spill area according to an oil spill evaluation index; and according to the oil spill distinguishing rule and the fusion monitoring function, guiding to combine a wind field and the ocean current data to construct an offshore oil spill multi-source fusion monitoring model, inputting to-be-monitored data into the offshore oil spill multi-source fusion monitoring model, and outputting a monitoring result.

Description

Technical Field

[0001] This invention relates to the field of marine monitoring, and in particular to a method and system for monitoring marine oil spills based on deep learning and multi-source data fusion. Background Technology

[0002] With the increasing frequency of marine economic activities such as ocean shipping and oil and gas development, marine oil spills are occurring frequently, posing a serious threat to the marine ecological environment, fishery resources, and coastal economy. Rapid and accurate identification of oil spill areas and monitoring of their spread have become critical requirements for marine environmental protection.

[0003] Traditional marine oil spill monitoring often relies on a single data source. For example, although synthetic aperture radar data has all-weather monitoring capabilities, it is easily interfered with by oil film-like substances (such as biofilms and suspended sediments), leading to misjudgments. Optical remote sensing data has high resolution but is greatly affected by weather, making it impossible to achieve all-weather monitoring. Land-sea data (such as buoys and ship reports) can provide local real-time information, but they suffer from limited coverage and data fragmentation.

[0004] Meanwhile, traditional monitoring methods lack an effective multi-source data fusion mechanism, making it difficult to integrate the advantages of remote sensing and land-sea data. They are unable to accurately extract the temporal evolution characteristics of oil spills, and their prediction accuracy for oil spill drift trajectory and diffusion speed is insufficient. Furthermore, the deployment of monitoring stations lacks scientific basis, making it difficult to meet the needs of emergency response for real-time and comprehensive monitoring results.

[0005] Furthermore, existing technologies for distinguishing between oil spills and oil slicks largely rely on traditional machine learning algorithms, which exhibit poor robustness and low accuracy when faced with nonlinear characteristics in complex marine environments. Therefore, there is an urgent need for a deep learning-based multi-source data fusion monitoring method to overcome the limitations of single data sources, achieve accurate oil spill identification, dynamic monitoring, and scientific prediction, and provide technical support for marine oil spill emergency response. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for monitoring multi-source oil spill data fusion at sea.

[0007] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Remote sensing data and land-sea data of a preset area are collected, and the remote sensing data and land-sea data are preprocessed; the remote sensing data includes synthetic aperture radar data, optical remote sensing data, laser fluorescence remote sensing data and airborne remote sensing data; the land-sea data includes shore-based radar data, buoy data, ship reports, underwater robot monitoring data and ocean current data. Based on the synthetic aperture radar data, oil spill identification is performed to obtain suspected oil spill areas. Based on the remote sensing data and the land-sea data, the suspected oil spill areas are compared and oil film-like substances are removed to obtain the adjusted oil spill areas. Temporal discriminant feature extraction is performed on the remote sensing data and the land-sea data of the calibrated oil spill area to obtain oil spill discrimination rules, and a fusion monitoring function is constructed in the calibrated oil spill area based on the oil spill assessment index. Based on the oil spill discrimination rules and the fusion monitoring function, a multi-source fusion monitoring model for marine oil spills is constructed by combining wind field and ocean current data. The data to be monitored is input into the multi-source fusion monitoring model for marine oil spills, and the monitoring results are output. The monitoring results include drift range, drift trajectory, diffusion speed and monitoring station location.

[0008] Furthermore, the method for identifying suspected oil spill areas based on the synthetic aperture radar data includes: The gray information and pixel coordinates in the synthetic aperture radar data are combined into a three-dimensional feature vector. Cluster centers are initialized according to a preset number of superpixels. The gray-level distance and spatial distance between pixels and cluster centers are obtained. The distance from a pixel to a cluster center is calculated within a local region, expressed as: in The normalized grayscale distance. The normalized spatial distance The distance from the pixel to the cluster center. To control the weighting parameters of grayscale, To control the weighting parameters of the distance space; like If the distance between a pixel and its cluster center is less than the minimum value, the current pixel is assigned to the superpixel of the cluster center. After the search is completed, the cluster center of the superpixel is updated to the mean of all pixels within the superpixel. The residual between the two cluster centers is calculated, and this process is repeated until the residual is less than the threshold. The segmented superpixel set is then output. The pixel set includes similar gray levels, textures, and spatial locations. Construct a Hidden Markov Random Field (HMRF), where the observation field consists of image grayscale values ​​and the label field consists of pixel class labels. The joint probability of the label field is given by a Gibbs distribution, expressed as: in For pixel labels The energy function, the sum of group functions within the neighborhood system; The joint probability of pixel labels; The observation model assumes that the observed gray values ​​under a given label follow a Gaussian distribution. The expression for oil spill identification is: in The optimal label. The observed energy is given by the pixel grayscale y label x condition; The John-Canney edge detection operator is used to extract edge information from the segmented synthetic aperture radar image for edge detection, resulting in a binary edge map. An edge prior energy term is then constructed, expressed as follows: in The edge prior energy term for the pixel label. This is the edge identifier for the i-th pixel. Let j be the edge identifier of the j-th pixel in the group, where j is the pixel index in the neighboring group of pixel i. and Edge markers for adjacent pixels. Let i be a neighborhood group of the i-th pixel. Energy value; The suspected oil spill area is obtained by correcting the optimal label from the edge energy term, as expressed in the following expression: in This is the set of parameters corresponding to the minimum of the log-posterior probability. The optimized optimal label is then output as the suspected oil spill area.

[0009] Furthermore, the method for comparing and removing oil film-like areas to obtain the adjusted oil spill area includes: The suspected oil spill area is divided into grids, and high-resolution optical images that are in the same phase or close phase as the synthetic aperture radar data are acquired. The image features of the high-resolution optical images are extracted. If the image features are obvious non-oil film features, the corresponding suspected oil spill area is removed to obtain the first calibration area. Acquire ship navigation data, wind field and ocean current data. When there are ship tracks upstream or inside the suspected oil spill area, and the speed and heading are abnormal, enhance the correlation information of the corresponding first calibration area. Based on the wind field and ocean current data, reverse the deduction of the source of the first calibration area, obtain the operating trajectory, remove the first calibration area that does not match the operating trajectory, and obtain the second calibration area. Laser fluorescence data is acquired and compared with high-resolution optical images. The second calibration region, in which no characteristic fluorescence peaks of petroleum hydrocarbons are detected, is removed to obtain the calibrated oil spill region.

[0010] Furthermore, the method for extracting time-series discriminative features from the remote sensing data and the land-sea data of the calibrated oil spill area to obtain oil spill discrimination rules includes: Remote sensing data and land-sea data are correlated by region and constructed into a time series dataset according to time sequence. Oil spill correlation features are extracted from the time series dataset. The calibrated oil spill area is divided into regions to obtain oil spill zones. A spatiotemporal data cube is constructed according to the oil spill zones. The spatiotemporal data cube includes time dimension, spatial dimension and feature dimension. Oil spill correlation features include morphological evolution features, motion trajectory features, spectral features, thermal features and land-sea data correlation features. Random forest is used to rank the importance of oil spill-related features to obtain the feature importance of oil spill-related features. Oil spill-related features with feature importance greater than the importance threshold are combined to obtain a feature subset. When oil spills and oil-like films are linearly separable in the feature space, obtain the category labels of the oil spill partitions, construct the optimal hyperplane, and give constraints, where the expression for the constraints is: in It is the normal vector. For displacement terms, For transpose, For the a-th spatiotemporal data cube, The category label for the a-th spatiotemporal data cube is +1 for real oil spill and -1 for oil film-like substance. By introducing Lagrange multipliers, we obtain the linear optimal classification function, expressed as: in It is a linear optimal classification function. For Lagrange multipliers, The number of oil spill zones; The linear optimal classification function is output as the oil spill discrimination rule: when the linear optimal classification function is greater than zero, it is an oil spill; otherwise, it is an oil film-like substance. When oil spills and oil film exhibit a complex nonlinear relationship, the oil spill correlation features are mapped to a high-dimensional Hilbert space using a kernel function, as expressed by: in For kernel width parameter, For spatiotemporal data cubes, Spatiotemporal data cube and spatiotemporal data cube The kernel function; Based on the radial basis function kernel, the nonlinear optimal classification function is given by the following expression: in The nonlinear optimal classification function is used to output oil spill identification rules.

[0011] Furthermore, the method for constructing a fusion monitoring function based on oil spill assessment indicators within the calibrated oil spill area includes: Oil spill assessment indicators include the concentration of petroleum pollutants, pH value, dissolved oxygen, chemical oxygen demand, amount of petroleum pollutant deposition, amount of sulfide deposition, amount of organic carbon deposition, species and quantity of benthic organisms, degree of damage to the ecosystem, and amount of petroleum hydrocarbon residues in organisms. The monitoring area, monitoring station layout, and monitoring methods are obtained according to the oil spill area calibration and oil spill discrimination rules. Remote sensing inversion is used to obtain the oil spill diffusion trend and pollution diffusion path, as well as the current oil spill location and movement characteristics. The time-series discrimination features are normalized, and a cross-modal fusion monitoring function is constructed based on the time-series discrimination features. Where is the fusion monitoring function, Let c be the importance weight coefficient of the evaluation dimension. The score for the c-th evaluation dimension is... The number of evaluation dimensions; in This includes sub-functions for oil spill identification accuracy, monitoring area site selection rationality, and monitoring station deployment rationality, with the following expressions: in For the oil spill identification accuracy subfunction, To monitor the site selection rationality sub-function, To configure reasonable sub-functions for monitoring stations, This is the area of ​​the oil spill, as determined by remote sensing. This is the actual high concentration area. This represents the spatial overlap of the actual high-concentration areas within the remotely sensed oil spill region. The oil spill is spreading. This represents the actual concentration change trend. The correlation coefficient between the oil spill diffusion trend and the actual concentration change trend. For spatial overlap weights, For the correlation coefficient weight, For the number of sites, Let L be the shortest distance from the k-th station to the pollution diffusion path, and L be the distance scale of the time-series discriminant feature. To determine the strength of the correlation between site data and pollution diffusion paths, This is the current location of the oil spill. As a characteristic of motion, A risk map based on diffusion simulation of the current oil spill location and movement characteristics. This is an environmental sensitivity level map. For diffusion weights, Environmentally sensitive weights.

[0012] Furthermore, the method for constructing a multi-source fusion monitoring model for marine oil spills by combining wind field and ocean current data includes: The loss function is obtained based on the actual monitoring results and the predicted monitoring results. The loss function, the fusion monitoring function and the regression loss are objectively weighted and output as the objective function of the marine oil spill multi-source fusion monitoring model. The oil spill discrimination rule is used to construct the positive and negative sample set for model training and serves as an auxiliary classifier. The system acquires the current-time calibration binary map of the oil spill area, multi-band remote sensing data stacking, oil spill morphology vector composed of historical time series oil spill area morphology features, current and future wind field data, and current and future ocean current data as input data; the current and future wind field data are resampled to the same spatial grid as the remote sensing data. The input data is normalized to the range of 0 to 1. A multi-source fusion monitoring model for marine oil spills is constructed based on the spatial feature coding branch, the temporal dynamic coding branch, the cross-modal spatiotemporal feature fusion module, and the monitoring result decoder. The spatial feature coding branch includes a remote sensing data encoder and an environmental field encoder. The remote sensing data encoder uses a convolutional neural network as the backbone to extract multi-scale spatial features of the oil spill area. The environmental field encoder processes stacked wind field and ocean current data to extract the spatial scale of environmental driving forces. The temporal dynamic coding branch inputs the feature vectors of past oil spill morphology into the long short-term memory network to capture the dynamic laws of oil spill morphology evolution and outputs the temporal context vector. The cross-modal spatiotemporal feature fusion module includes spatial fusion, spatiotemporal injection, and attention guidance. Spatial fusion, at the jump connection of the multi-source fusion monitoring model for marine oil spills, stitches together remote sensing feature maps and environmental feature maps, and then fuses them through convolutional layers to generate a fused feature map. Spatiotemporal injection expands and reshapes the temporal context vector into a temporal feature map through a fully connected layer, and then adds and fuses it with the deepest temporal feature map in the decoder through spatial broadcasting, injecting temporal dynamic information into spatial prediction. Attention guidance uses the amplitude of wind field and ocean current data to generate a spatial attention map, guiding the model to pay attention to the influence of the environmental field in areas with strong winds and currents, and to rely more on remote sensing image features in calm areas. The monitoring result decoder gradually restores the spatial resolution through upsampling and skip connections. The decoding layer connects convolution and sigmoid activation functions to output an oil spill probability map. By setting a threshold, the future drift range is obtained, and the drift trajectory is obtained by connecting the oil spill centroids at consecutive time steps. On another parallel branch, the average expansion distance of the oil spill range boundary in the next adjacent time step is calculated, and the average diffusion velocity value is regressed. The fused high-level feature map is input into the regressor, and the predicted pixel position is used as the priority score of the candidate monitoring station. The score map is post-processed in combination with the fusion objective function, and the candidate monitoring stations with priority scores greater than the priority threshold are selected as monitoring stations.

[0013] Secondly, a deep learning-based multi-source data fusion monitoring system for marine oil spills includes: Data acquisition module: used to collect remote sensing data and land-sea data of a preset area, and to preprocess the remote sensing data and land-sea data; the remote sensing data includes synthetic aperture radar data, optical remote sensing data, laser fluorescence remote sensing data and airborne remote sensing data; the land-sea data includes shore-based radar data, buoy data, ship reports, underwater robot monitoring data and ocean current data; Monitoring area locking module: used to identify suspected oil spill areas based on the remote sensing data and the land-sea data, and to compare and remove oil film-like areas to obtain the adjusted oil spill area; Rule extraction and evaluation module: used to extract time-series discriminant features from the remote sensing data domain of the calibrated oil spill area to obtain oil spill discrimination rules, and to construct a fusion monitoring function based on oil spill evaluation indicators within the calibrated oil spill area; Modeling output module: It is used to guide the construction of a marine oil spill multi-source fusion monitoring model by combining wind field and ocean current data according to the oil spill discrimination rules and the fusion monitoring function. The monitoring data to be monitored is input into the marine oil spill multi-source fusion monitoring model, and the monitoring results are output. The monitoring results include drift range, drift trajectory, diffusion speed and monitoring station location.

[0014] The beneficial effects of this invention are: This invention relates to a deep learning-based method and system for multi-source data fusion monitoring of marine oil spills. Compared with existing technologies, this invention has the following technical advantages: This invention integrates remote sensing and multi-source (land and sea) data through preprocessing, oil spill identification, comparison and oil film removal, temporal discriminative feature extraction, construction of a fusion monitoring function, and model building. This overcomes the limitations of single data sources, reduces oil film interference, and improves the accuracy of oil spill identification. By combining deep learning to extract temporal features, a fusion monitoring function and a multi-source fusion model are constructed to accurately predict the drift range, trajectory, and diffusion speed of oil spills. Monitoring stations are scientifically generated to meet emergency response needs, enabling dynamic and comprehensive monitoring of oil spills and improving the efficiency and scientific nature of marine oil spill emergency response. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of the deep learning-based multi-source data fusion monitoring method for marine oil spills according to the present invention. Detailed Implementation

[0016] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0017] The present invention provides a deep learning-based method and system for monitoring multi-source oil spills at sea, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Remote sensing data and land-sea data of a preset area are collected, and the remote sensing data and land-sea data are preprocessed; the remote sensing data includes synthetic aperture radar data, optical remote sensing data, laser fluorescence remote sensing data and airborne remote sensing data; the land-sea data includes shore-based radar data, buoy data, ship reports, underwater robot monitoring data and ocean current data. In the actual assessment, an operating area of ​​an oilfield in the Bohai Sea (38.5°N-39.5°N, 118.5°E-119.5°E) was selected as the preset area, and data were collected from May 10th to 15th, 2024. Remote sensing data included: synthetic aperture radar data (10m resolution, VV polarization), optical remote sensing data (Landsat-9 satellite, 30m resolution), and laser fluorescence remote sensing data (petroleum hydrocarbon detection wavelength 250-350nm). Land and sea data included: shore-based radar data (monitoring radius 20km), buoy data (water temperature 18℃, salinity 32‰), ship reports (trajectories of 3 operating vessels), underwater robot monitoring data (petroleum hydrocarbon concentration at 5 locations), and ocean current data (current velocity 0.8m / s, flowing northeast). Preprocessing includes: removing radar data noise, unifying the optical image coordinate system, completing missing buoy data, and normalizing the spatial resolution of all data to 10m. Based on the synthetic aperture radar data, oil spill identification is performed to obtain suspected oil spill areas. Based on the remote sensing data and the land-sea data, the suspected oil spill areas are compared and oil film-like substances are removed to obtain the adjusted oil spill areas. In the actual assessment, the grayscale values ​​of the synthetic aperture radar data and pixel coordinates were used to construct a three-dimensional vector. With 500 superpixels, M=0.6, and S=0.4, the distance from the pixel to the cluster center was calculated. The process was iterated until the residual was <0.01, and the superpixel set was output. A hidden Markov random field was constructed with an energy value of 0.8. After John Canney edge detection, three suspected oil spill areas (total area of ​​approximately 8 km²) were identified. Using concurrent 30m high-resolution optical images, one strip-shaped non-oil film area (identified as a ship's wake) was removed, resulting in two first calibration areas. Combining the ship's trajectory (a working vessel's speed suddenly dropped to 1 knot upstream of the area on May 12), the ocean current trajectory was reverse-engineered, and one area inconsistent with the trajectory was removed, resulting in one second calibration area. Laser fluorescence data detected a 430nm characteristic fluorescence peak of petroleum hydrocarbons in this area, ultimately confirming the calibration oil spill area (area 3.2 km²). Temporal discriminant feature extraction is performed on the remote sensing data and the land-sea data of the calibrated oil spill area to obtain oil spill discrimination rules, and a fusion monitoring function is constructed in the calibrated oil spill area based on the oil spill assessment index. In the actual assessment, a dataset was constructed based on time series data. Features of oil spill morphology (irregular patches), movement (daily drift of 5 km), and spectral density (reflectance 0.12-0.18) were extracted. After random forest sorting, the top 8 high-contribution features were selected. Because oil spills and oil-like films exhibit a non-linear relationship, a non-linear classification function based on radial basis function was constructed with a kernel width parameter of 0.3. It was determined to be an oil spill; Eight evaluation indicators were selected, including petroleum hydrocarbon concentration and dissolved oxygen. Take values ​​between 0.1 and 0.15, and calculate. The value is 0.85 (spatial overlap is 0.8, correlation coefficient is 0.9). 0.5 (0.5) The value is 0.82 (diffusion risk is 0.8, environmental sensitivity is 0.84). 0.4 (0.6) The value is 0.78 (10 stations, average distance from the diffusion path 1.2km), resulting in a fusion monitoring function value of 0.81; Based on the oil spill discrimination rules and the fusion monitoring function, a multi-source fusion monitoring model for marine oil spills is constructed by combining wind field and ocean current data. The data to be monitored is input into the multi-source fusion monitoring model for marine oil spills, and the monitoring results are output. The monitoring results include drift range, drift trajectory, diffusion speed, and monitoring station location. In the actual assessment, the inputs were a binary map of the calibration area, multi-band remote sensing data stacking, 5-day oil spill morphology vector, wind field (wind speed 5 m / s, wind direction northeast), and ocean current data, all normalized to [0,1]. The spatial coding branch used ResNet50 to extract remote sensing features, and the environmental field encoder extracted wind and ocean current features. The temporal branch used LSTM (128-dimensional hidden layer) to obtain the context vector. After cross-modal fusion, the decoder output an oil spill probability map. The monitoring results are as follows: drift range (reached 4.5 km² on May 16), drift trajectory (northeast direction, average 5 km per day), diffusion speed (0.2 km² / h), and monitoring station locations (3 high-priority locations, less than 1 km from the diffusion path).

[0018] In this embodiment, the method for identifying suspected oil spill areas based on the synthetic aperture radar data includes: The gray information and pixel coordinates in the synthetic aperture radar data are combined into a three-dimensional feature vector. Cluster centers are initialized according to a preset number of superpixels. The gray-level distance and spatial distance between pixels and cluster centers are obtained. The distance from a pixel to a cluster center is calculated within a local region, expressed as: in The normalized grayscale distance. The normalized spatial distance The distance from the pixel to the cluster center. To control the weighting parameters of grayscale, To control the weighting parameters of the distance space; like If the distance between a pixel and its cluster center is less than the minimum value, the current pixel is assigned to the superpixel of the cluster center. After the search is completed, the cluster center of the superpixel is updated to the mean of all pixels within the superpixel. The residual between the two cluster centers is calculated, and this process is repeated until the residual is less than the threshold. The segmented superpixel set is then output. The pixel set includes similar gray levels, textures, and spatial locations. Construct a Hidden Markov Random Field (HMRF), where the observation field consists of image grayscale values ​​and the label field consists of pixel class labels. The joint probability of the label field is given by a Gibbs distribution, expressed as: in For pixel labels The energy function, the sum of group functions within the neighborhood system; The joint probability of pixel labels; The observation model assumes that the observed gray values ​​under a given label follow a Gaussian distribution. The expression for oil spill identification is: in The optimal label. The observed energy is given by the pixel grayscale y label x condition; The John-Canney edge detection operator is used to extract edge information from the segmented synthetic aperture radar image for edge detection, resulting in a binary edge map. An edge prior energy term is then constructed, expressed as follows: in The edge prior energy term for the pixel label. This is the edge identifier for the i-th pixel. Let j be the edge identifier of the j-th pixel in the group, where j is the pixel index in the neighboring group of pixel i. and Edge markers for adjacent pixels. Let i be a neighborhood group of the i-th pixel. Energy value; The suspected oil spill area is obtained by correcting the optimal label from the edge energy term, as expressed in the following expression: in This is the set of parameters corresponding to the minimum of the log-posterior probability. The optimized optimal label is then output as the suspected oil spill area.

[0019] In this embodiment, the method for comparing and removing suspected oil spill areas to obtain adjusted oil spill areas includes: The suspected oil spill area is divided into grids, and high-resolution optical images that are in the same phase or close phase as the synthetic aperture radar data are acquired. The image features of the high-resolution optical images are extracted. If the image features are obvious non-oil film features, the corresponding suspected oil spill area is removed to obtain the first calibration area. Acquire ship navigation data, wind field and ocean current data. When there are ship tracks upstream or inside the suspected oil spill area, and the speed and heading are abnormal, enhance the correlation information of the corresponding first calibration area. Based on the wind field and ocean current data, reverse the deduction of the source of the first calibration area, obtain the operating trajectory, remove the first calibration area that does not match the operating trajectory, and obtain the second calibration area. Laser fluorescence data is acquired and compared with high-resolution optical images. The second calibration region, in which no characteristic fluorescence peaks of petroleum hydrocarbons are detected, is removed to obtain the calibrated oil spill region.

[0020] In this embodiment, the method for extracting time-series discriminative features from the remote sensing data and the land-sea data of the calibrated oil spill area to obtain oil spill discrimination rules includes: Remote sensing data and land-sea data are correlated by region and constructed into a time series dataset according to time sequence. Oil spill correlation features are extracted from the time series dataset. The calibrated oil spill area is divided into regions to obtain oil spill zones. A spatiotemporal data cube is constructed according to the oil spill zones. The spatiotemporal data cube includes time dimension, spatial dimension and feature dimension. Oil spill correlation features include morphological evolution features, motion trajectory features, spectral features, thermal features and land-sea data correlation features. Random forest is used to rank the importance of oil spill-related features to obtain the feature importance of oil spill-related features. Oil spill-related features with feature importance greater than the importance threshold are combined to obtain a feature subset. When oil spills and oil-like films are linearly separable in the feature space, obtain the category labels of the oil spill partitions, construct the optimal hyperplane, and give constraints, where the expression for the constraints is: in It is the normal vector. For displacement terms, For transpose, For the a-th spatiotemporal data cube, The category label for the a-th spatiotemporal data cube is +1 for real oil spill and -1 for oil film-like substance. By introducing Lagrange multipliers, we obtain the linear optimal classification function, expressed as: in It is a linear optimal classification function. For Lagrange multipliers, The number of oil spill zones; The linear optimal classification function is output as the oil spill discrimination rule: when the linear optimal classification function is greater than zero, it is an oil spill; otherwise, it is an oil film-like substance. When oil spills and oil film exhibit a complex nonlinear relationship, the oil spill correlation features are mapped to a high-dimensional Hilbert space using a kernel function, as expressed by: in For kernel width parameter, For spatiotemporal data cubes, Spatiotemporal data cube and spatiotemporal data cube The kernel function; Based on the radial basis function kernel, the nonlinear optimal classification function is given by the following expression: in The nonlinear optimal classification function is used to output oil spill identification rules.

[0021] In this embodiment, the method for constructing a fusion monitoring function based on oil spill assessment indicators within the calibrated oil spill area includes: Oil spill assessment indicators include the concentration of petroleum pollutants, pH value, dissolved oxygen, chemical oxygen demand, amount of petroleum pollutant deposition, amount of sulfide deposition, amount of organic carbon deposition, species and quantity of benthic organisms, degree of damage to the ecosystem, and amount of petroleum hydrocarbon residues in organisms. The monitoring area, monitoring station layout, and monitoring methods are obtained according to the oil spill area calibration and oil spill discrimination rules. Remote sensing inversion is used to obtain the oil spill diffusion trend and pollution diffusion path, as well as the current oil spill location and movement characteristics. The time-series discrimination features are normalized, and a cross-modal fusion monitoring function is constructed based on the time-series discrimination features. Where is the fusion monitoring function, Let c be the importance weight coefficient of the evaluation dimension. The score for the c-th evaluation dimension is... The number of evaluation dimensions; in This includes sub-functions for oil spill identification accuracy, monitoring area site selection rationality, and monitoring station deployment rationality, with the following expressions: in For the oil spill identification accuracy subfunction, To monitor the site selection rationality sub-function, To configure reasonable sub-functions for monitoring stations, This is the area of ​​the oil spill, as determined by remote sensing. This is the actual high concentration area. This represents the spatial overlap of the actual high-concentration areas within the remotely sensed oil spill region. The oil spill is spreading. This represents the actual concentration change trend. The correlation coefficient between the oil spill diffusion trend and the actual concentration change trend. For spatial overlap weights, For the correlation coefficient weight, For the number of sites, Let L be the shortest distance from the k-th station to the pollution diffusion path, and L be the distance scale of the time-series discriminant feature. To determine the strength of the correlation between site data and pollution diffusion paths, This is the current location of the oil spill. As a characteristic of motion, A risk map based on diffusion simulation of the current oil spill location and movement characteristics. This is an environmental sensitivity level map. For diffusion weights, Environmentally sensitive weights.

[0022] In this embodiment, the method for guiding the construction of a multi-source fusion monitoring model for marine oil spills by combining wind field and ocean current data includes: The loss function is obtained based on the actual monitoring results and the predicted monitoring results. The loss function, the fusion monitoring function and the regression loss are objectively weighted and output as the objective function of the marine oil spill multi-source fusion monitoring model. The oil spill discrimination rule is used to construct the positive and negative sample set for model training and serves as an auxiliary classifier. The system acquires the current-time calibration binary map of the oil spill area, multi-band remote sensing data stacking, oil spill morphology vector composed of historical time series oil spill area morphology features, current and future wind field data, and current and future ocean current data as input data; the current and future wind field data are resampled to the same spatial grid as the remote sensing data. The input data is normalized to the range of 0 to 1. A multi-source fusion monitoring model for marine oil spills is constructed based on the spatial feature coding branch, the temporal dynamic coding branch, the cross-modal spatiotemporal feature fusion module, and the monitoring result decoder. The spatial feature coding branch includes a remote sensing data encoder and an environmental field encoder. The remote sensing data encoder uses a convolutional neural network as the backbone to extract multi-scale spatial features of the oil spill area. The environmental field encoder processes stacked wind field and ocean current data to extract the spatial scale of environmental driving forces. The temporal dynamic coding branch inputs the feature vectors of past oil spill morphology into the long short-term memory network to capture the dynamic laws of oil spill morphology evolution and outputs the temporal context vector. The cross-modal spatiotemporal feature fusion module includes spatial fusion, spatiotemporal injection, and attention guidance. Spatial fusion, at the jump connection of the multi-source fusion monitoring model for marine oil spills, stitches together remote sensing feature maps and environmental feature maps, and then fuses them through convolutional layers to generate a fused feature map. Spatiotemporal injection expands and reshapes the temporal context vector into a temporal feature map through a fully connected layer, and then adds and fuses it with the deepest temporal feature map in the decoder through spatial broadcasting, injecting temporal dynamic information into spatial prediction. Attention guidance uses the amplitude of wind field and ocean current data to generate a spatial attention map, guiding the model to pay attention to the influence of the environmental field in areas with strong winds and currents, and to rely more on remote sensing image features in calm areas. The monitoring result decoder gradually restores the spatial resolution through upsampling and skip connections. The decoding layer connects convolution and sigmoid activation functions to output an oil spill probability map. By setting a threshold, the future drift range is obtained, and the drift trajectory is obtained by connecting the oil spill centroids at consecutive time steps. On another parallel branch, the average expansion distance of the oil spill range boundary in the next adjacent time step is calculated, and the average diffusion velocity value is regressed. The fused high-level feature map is input into the regressor, and the predicted pixel position is used as the priority score of the candidate monitoring station. The score map is post-processed in combination with the fusion objective function, and the candidate monitoring stations with priority scores greater than the priority threshold are selected as monitoring stations. In actual assessment, the learned oil spill discrimination rules are used to automatically label the historical dataset and filter out high-confidence real oil spill and oil film samples. The post-processing of the score map in conjunction with the fusion objective function is as follows: the calculation result of the monitoring area location rationality sub-function in the fusion monitoring function is used as prior knowledge and multiplied pixel by pixel with the priority score map predicted by the model. In areas with high environmental sensitivity, the monitoring priority is further improved to generate monitoring stations.

[0023] Secondly, a deep learning-based multi-source data fusion monitoring system for marine oil spills includes: Data acquisition module: used to collect remote sensing data and land-sea data of a preset area, and to preprocess the remote sensing data and land-sea data; the remote sensing data includes synthetic aperture radar data, optical remote sensing data, laser fluorescence remote sensing data and airborne remote sensing data; the land-sea data includes shore-based radar data, buoy data, ship reports, underwater robot monitoring data and ocean current data; Monitoring area locking module: used to identify suspected oil spill areas based on the remote sensing data and the land-sea data, and to compare and remove oil film-like areas to obtain the adjusted oil spill area; Rule extraction and evaluation module: used to extract time-series discriminant features from the remote sensing data domain of the calibrated oil spill area to obtain oil spill discrimination rules, and to construct a fusion monitoring function based on oil spill evaluation indicators within the calibrated oil spill area; Modeling output module: It is used to guide the construction of a marine oil spill multi-source fusion monitoring model by combining wind field and ocean current data according to the oil spill discrimination rules and the fusion monitoring function. The monitoring data to be monitored is input into the marine oil spill multi-source fusion monitoring model, and the monitoring results are output. The monitoring results include drift range, drift trajectory, diffusion speed and monitoring station location.

[0024] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based multi-source data fusion monitoring method for marine oil spills, characterized in that, Includes the following steps: Remote sensing data and land-sea data of a preset area are collected, and the remote sensing data and land-sea data are preprocessed; the remote sensing data includes synthetic aperture radar data, optical remote sensing data, laser fluorescence remote sensing data and airborne remote sensing data; the land-sea data includes shore-based radar data, buoy data, ship reports, underwater robot monitoring data and ocean current data. Based on the synthetic aperture radar data, oil spill identification is performed to obtain suspected oil spill areas. Based on the remote sensing data and the land-sea data, the suspected oil spill areas are compared and oil film-like substances are removed to obtain the adjusted oil spill areas. Temporal discriminant feature extraction is performed on the remote sensing data and the land-sea data of the calibrated oil spill area to obtain oil spill discrimination rules, and a fusion monitoring function is constructed in the calibrated oil spill area based on the oil spill assessment index. Based on the oil spill discrimination rules and the fusion monitoring function, a multi-source fusion monitoring model for marine oil spills is constructed by combining wind field and ocean current data. The data to be monitored is input into the multi-source fusion monitoring model for marine oil spills, and the monitoring results are output. The monitoring results include drift range, drift trajectory, diffusion speed and monitoring station location.

2. The deep learning-based multi-source data fusion monitoring method for marine oil spills according to claim 1, characterized in that, A method for identifying suspected oil spill areas based on the synthetic aperture radar data includes: The gray information and pixel coordinates in the synthetic aperture radar data are combined into a three-dimensional feature vector. Cluster centers are initialized according to a preset number of superpixels. The gray-level distance and spatial distance between pixels and cluster centers are obtained. The distance from a pixel to a cluster center is calculated within a local region, expressed as: in The normalized grayscale distance. The normalized spatial distance The distance from the pixel to the cluster center. To control the weighting parameters of grayscale, To control the weighting parameters of the distance space; like If the distance between a pixel and its cluster center is less than the minimum value, the current pixel is assigned to the superpixel of the cluster center. After the search is completed, the cluster center of the superpixel is updated to the mean of all pixels within the superpixel. The residual between the two cluster centers is calculated, and this process is repeated until the residual is less than the threshold. The segmented superpixel set is then output. The pixel set includes similar gray levels, textures, and spatial locations. Construct a Hidden Markov Random Field (HMRF), where the observation field consists of image grayscale values ​​and the label field consists of pixel class labels. The joint probability of the label field is given by a Gibbs distribution, expressed as: in For pixel labels The energy function, the sum of group functions within the neighborhood system; The joint probability of pixel labels; The observation model assumes that the observed gray values ​​under a given label follow a Gaussian distribution. The expression for oil spill identification is: in The optimal label. The observed energy is given by the pixel grayscale y label x condition; The John-Canney edge detection operator is used to extract edge information from the segmented synthetic aperture radar image for edge detection, resulting in a binary edge map. An edge prior energy term is then constructed, expressed as follows: in The edge prior energy term for the pixel label. This is the edge identifier for the i-th pixel. Let j be the edge identifier of the j-th pixel in the group, where j is the pixel index in the neighboring group of pixel i. and Edge markers for adjacent pixels. Let i be a neighborhood group of the i-th pixel. Energy value; The suspected oil spill area is obtained by correcting the optimal label from the edge energy term, as expressed in the following expression: in This is the set of parameters corresponding to the minimum of the log-posterior probability. The optimized optimal label is then output as the suspected oil spill area.

3. The deep learning-based multi-source data fusion monitoring method for marine oil spills according to claim 1, characterized in that, A method for comparing and removing suspected oil spill areas to obtain adjusted oil spill areas includes: The suspected oil spill area is divided into grids, and high-resolution optical images that are in the same phase or close phase as the synthetic aperture radar data are acquired. The image features of the high-resolution optical images are extracted. If the image features are obvious non-oil film features, the corresponding suspected oil spill area is removed to obtain the first calibration area. Acquire ship navigation data, wind field and ocean current data. When there are ship tracks upstream or inside the suspected oil spill area, and the speed and heading are abnormal, enhance the correlation information of the corresponding first calibration area. Based on the wind field and ocean current data, reverse the deduction of the source of the first calibration area, obtain the operating trajectory, remove the first calibration area that does not match the operating trajectory, and obtain the second calibration area. Laser fluorescence data is acquired and compared with high-resolution optical images. The second calibration region, in which no characteristic fluorescence peaks of petroleum hydrocarbons are detected, is removed to obtain the calibrated oil spill region.

4. The deep learning-based multi-source data fusion monitoring method for marine oil spills according to claim 1, characterized in that, A method for extracting oil spill discrimination rules from the remote sensing data and the land-sea data of the calibrated oil spill area using time-series discriminative feature extraction includes: Remote sensing data and land-sea data are correlated by region and constructed into a time series dataset according to time sequence. Oil spill correlation features are extracted from the time series dataset. The calibrated oil spill area is divided into regions to obtain oil spill zones. A spatiotemporal data cube is constructed according to the oil spill zones. The spatiotemporal data cube includes time dimension, spatial dimension and feature dimension. Oil spill correlation features include morphological evolution features, motion trajectory features, spectral features, thermal features and land-sea data correlation features. Random forest is used to rank the importance of oil spill-related features to obtain the feature importance of oil spill-related features. Oil spill-related features with feature importance greater than the importance threshold are combined to obtain a feature subset. When oil spills and oil-like films are linearly separable in the feature space, obtain the category labels of the oil spill partitions, construct the optimal hyperplane, and give constraints, where the expression for the constraints is: in It is the normal vector. For displacement terms, For transpose, For the a-th spatiotemporal data cube, The category label for the a-th spatiotemporal data cube is +1 for real oil spill and -1 for oil film-like substance. By introducing Lagrange multipliers, we obtain the linear optimal classification function, expressed as: in It is a linear optimal classification function. For Lagrange multipliers, The number of oil spill zones; The linear optimal classification function is output as the oil spill discrimination rule: when the linear optimal classification function is greater than zero, it is an oil spill; otherwise, it is an oil film-like substance. When oil spills and oil film exhibit a complex nonlinear relationship, the oil spill correlation features are mapped to a high-dimensional Hilbert space using a kernel function, as expressed by: in For kernel width parameter, For spatiotemporal data cubes, Spatiotemporal data cube and spatiotemporal data cube The kernel function; Based on the radial basis function kernel, the nonlinear optimal classification function is given by the following expression: in The nonlinear optimal classification function is used to output oil spill identification rules.

5. The deep learning-based multi-source data fusion monitoring method for marine oil spills according to claim 1, characterized in that, The method for constructing a fusion monitoring function based on oil spill assessment indicators within the calibrated oil spill area includes: Oil spill assessment indicators include the concentration of petroleum pollutants, pH value, dissolved oxygen, chemical oxygen demand, amount of petroleum pollutant deposition, amount of sulfide deposition, amount of organic carbon deposition, species and quantity of benthic organisms, degree of damage to the ecosystem, and amount of petroleum hydrocarbon residues in organisms. The monitoring area, monitoring station layout, and monitoring methods are obtained according to the oil spill area calibration and oil spill discrimination rules. Remote sensing inversion is used to obtain the oil spill diffusion trend and pollution diffusion path, as well as the current oil spill location and movement characteristics. The time-series discrimination features are normalized, and a cross-modal fusion monitoring function is constructed based on the time-series discrimination features. Where is the fusion monitoring function, Let c be the importance weight coefficient of the evaluation dimension. The score for the c-th evaluation dimension is... The number of evaluation dimensions; in This includes sub-functions for oil spill identification accuracy, monitoring area site selection rationality, and monitoring station deployment rationality, with the following expressions: in For the oil spill identification accuracy subfunction, To monitor the site selection rationality sub-function, To configure reasonable sub-functions for monitoring stations, This is the area of ​​the oil spill, as determined by remote sensing. This is the actual high concentration area. This represents the spatial overlap of the actual high-concentration areas within the remotely sensed oil spill region. The oil spill is spreading. This represents the actual concentration change trend. The correlation coefficient between the oil spill diffusion trend and the actual concentration change trend. For spatial overlap weights, For the correlation coefficient weight, For the number of sites, Let L be the shortest distance from the k-th station to the pollution diffusion path, and L be the distance scale of the time-series discriminant feature. To determine the strength of the correlation between site data and pollution diffusion paths, This is the current location of the oil spill. As a characteristic of motion, A risk map based on diffusion simulation of the current oil spill location and movement characteristics. This is an environmental sensitivity level map. For diffusion weights, Environmentally sensitive weights.

6. The deep learning-based multi-source data fusion monitoring method for marine oil spills according to claim 1, characterized in that, The method for constructing a multi-source fusion monitoring model for marine oil spills by combining wind field and ocean current data includes: The loss function is obtained based on the actual monitoring results and the predicted monitoring results. The loss function, the fusion monitoring function and the regression loss are objectively weighted and output as the objective function of the marine oil spill multi-source fusion monitoring model. The oil spill discrimination rule is used to construct the positive and negative sample set for model training and serves as an auxiliary classifier. The system acquires the current-time calibration binary map of the oil spill area, multi-band remote sensing data stacking, oil spill morphology vector composed of historical time series oil spill area morphology features, current and future wind field data, and current and future ocean current data as input data; the current and future wind field data are resampled to the same spatial grid as the remote sensing data. The input data is normalized to the range of 0 to 1. A multi-source fusion monitoring model for marine oil spills is constructed based on the spatial feature coding branch, the temporal dynamic coding branch, the cross-modal spatiotemporal feature fusion module, and the monitoring result decoder. The spatial feature coding branch includes a remote sensing data encoder and an environmental field encoder. The remote sensing data encoder uses a convolutional neural network as the backbone to extract multi-scale spatial features of the oil spill area. The environmental field encoder processes stacked wind field and ocean current data to extract the spatial scale of environmental driving forces. The temporal dynamic coding branch inputs the feature vectors of past oil spill morphology into the long short-term memory network to capture the dynamic laws of oil spill morphology evolution and outputs the temporal context vector. The cross-modal spatiotemporal feature fusion module includes spatial fusion, spatiotemporal injection, and attention guidance. Spatial fusion, at the jump connection of the multi-source fusion monitoring model for marine oil spills, stitches together remote sensing feature maps and environmental feature maps, and then fuses them through convolutional layers to generate a fused feature map. Spatiotemporal injection expands and reshapes the temporal context vector into a temporal feature map through a fully connected layer, and then adds and fuses it with the deepest temporal feature map in the decoder through spatial broadcasting, injecting temporal dynamic information into spatial prediction. Attention guidance uses the amplitude of wind field and ocean current data to generate a spatial attention map, guiding the model to pay attention to the influence of the environmental field in areas with strong winds and currents, and to rely more on remote sensing image features in calm areas. The monitoring result decoder gradually restores the spatial resolution through upsampling and skip connections. The decoding layer connects convolution and sigmoid activation functions to output an oil spill probability map. By setting a threshold, the future drift range is obtained, and the drift trajectory is obtained by connecting the oil spill centroids at consecutive time steps. On another parallel branch, the average expansion distance of the oil spill range boundary in the next adjacent time step is calculated, and the average diffusion velocity value is regressed. The fused high-level feature map is input into the regressor, and the predicted pixel position is used as the priority score of the candidate monitoring station. The score map is post-processed in combination with the fusion objective function, and the candidate monitoring stations with priority scores greater than the priority threshold are selected as monitoring stations.

7. A deep learning-based multi-source oil spill data fusion monitoring system for performing the method described in any one of claims 1-6, characterized in that, include: Data acquisition module: used to collect remote sensing data and land-sea data of a preset area, and to preprocess the remote sensing data and land-sea data; the remote sensing data includes synthetic aperture radar data, optical remote sensing data, laser fluorescence remote sensing data and airborne remote sensing data; the land-sea data includes shore-based radar data, buoy data, ship reports, underwater robot monitoring data and ocean current data; Monitoring area locking module: used to identify suspected oil spill areas based on the remote sensing data and the land-sea data, and to compare and remove oil film-like areas to obtain the adjusted oil spill area; Rule extraction and evaluation module: used to extract time-series discriminant features from the remote sensing data domain of the calibrated oil spill area to obtain oil spill discrimination rules, and to construct a fusion monitoring function based on oil spill evaluation indicators within the calibrated oil spill area; Modeling output module: It is used to guide the construction of a marine oil spill multi-source fusion monitoring model by combining wind field and ocean current data according to the oil spill discrimination rules and the fusion monitoring function. The monitoring data to be monitored is input into the marine oil spill multi-source fusion monitoring model, and the monitoring results are output. The monitoring results include drift range, drift trajectory, diffusion speed and monitoring station location.

Images