Marine oil spill monitoring method and system based on multiband fluorescence spectrum neural network

Abstract

The invention relates to the technical field of marine oil spill monitoring, in particular to a marine oil spill monitoring method and system based on a multiband fluorescence spectrum neural network. The method comprises the following steps: acquiring fluorescence spectrums and fluorescence image data of spilled oil of different oil types and different oil film thicknesses; respectively preprocessing the fluorescence spectrum and the fluorescence image data of the marine oil spill; constructing three-dimensional fluorescence spectrum data and a thickness difference matrix by using the preprocessed fluorescence spectrum; utilizing a deep neural network to carry out waveband feature and thickness feature extraction on the three-dimensional fluorescence spectrum data and the thickness difference matrix; according to the invention, the marine oil spill fluorescence spectrum and fluorescence image data are effectively integrated, and a reliable basis is provided for accurate identification and early warning of oil spill.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine oil spill monitoring, in particular to a marine oil spill monitoring method and system based on a multi-band fluorescence spectrum neural network. BACKGROUND

[0002] The sources of marine oil spills usually include natural seepage, ship accidents and leaks from offshore engineering (drilling platforms, oil pipelines, etc.). Oil is a complex mixture containing a variety of compounds, and the hydrocarbon organic matter contains many highly toxic and persistent harmful substances. Not only does the change in the physical properties of the ocean water cause irreversible damage, but it also destroys marine biodiversity and even causes biological extinction. Therefore, rapid detection and monitoring of marine oil spills are of great significance to the ecological environment of the deep sea and coastal areas.

[0003] Traditional remote sensing technology is a common marine oil spill monitoring and early warning technology, and the related satellite, airborne and shipborne instruments are relatively mature. The principle is based on the large reflectivity difference between oil film and seawater in different wave bands (microwave, visible light, infrared wave band). Static images or dynamic video information formed by reflected light spectrum are used for marine oil spill monitoring and early warning. However, the reflected light spectrum information of marine oil spills and seawater only has overall differences, and remote sensing technology still has certain difficulties in identifying oil types and inverting oil film thickness. Moreover, traditional remote sensing technology relies on sunlight as the light source, which is affected by fog, clouds and other weather conditions, and cannot be used at night. Ultraviolet-induced fluorescence technology is an active remote sensing detection technology. According to the laser excitation of some organic matter contained in oil at the ultraviolet wave band, the oil can emit fluorescence with a longer wavelength than the excitation wavelength. The sensing and detection method is based on the analysis of the characteristic peaks and intensity information of the fluorescence spectrum. It has strong detection capability for different oil types and thin oil films. Through the analysis and processing of fluorescence information, real-time monitoring and rapid early warning of marine oil spills can be achieved.

[0004] In the existing marine oil spill monitoring using fluorescence information, the fluorescence information in a limited range is usually obtained at a fixed vertical angle, ignoring the fluorescence directional pattern characteristics of marine oil spills under the action of wind and waves. Secondly, the fluorescence spectrum can only collect point data in a small area, and the fluorescence spectrum characteristics of different oil types are obvious. However, although the fluorescence image can collect data in a larger area, the feature difference in the image is small and cannot effectively distinguish oil types and oil film thickness. By using the characteristic information of the fluorescence spectrum and fusing it into the analysis and processing of the fluorescence image, the problem of breaking through the fluorescence spectrum point detection to the fluorescence image large-range accurate detection is solved, and the problem of heterogeneous data fusion of fluorescence spectrum and image is solved, thereby improving the accuracy of the monitoring and early warning model. SUMMARY

[0005] To address the aforementioned problems, this invention provides a method and system for monitoring marine oil spills based on a multi-band fluorescence spectral neural network. This invention integrates fluorescence spectral features and fluorescence image data, fusing heterogeneous spectral and image data features to provide a novel method for monitoring and early warning of marine oil spills.

[0006] In a first aspect, the present invention provides a marine oil spill monitoring method based on a multi-band fluorescence spectrum neural network, which adopts the following technical solution: A method for monitoring marine oil spills based on multi-band fluorescence spectral neural networks, comprising: Obtain fluorescence spectra and fluorescence image data of oil spills of different oil types and with different oil film thicknesses; The fluorescence spectra and fluorescence image data of marine oil spills were preprocessed respectively; Three-dimensional fluorescence spectral data and thickness difference matrix were constructed using the preprocessed fluorescence spectra. Deep neural networks were used to extract band features and thickness features from three-dimensional fluorescence spectral data and thickness difference matrices. Feature extraction from fluorescence image data based on a joint convolutional neural network; The extracted features are analyzed and predicted based on the Lasso regression model after feature fusion; Output the warning results.

[0007] Furthermore, the preprocessing of the fluorescence spectral and fluorescence image data of marine oil spills includes performing moving average smoothing and logistic normalization on the fluorescence spectral data of marine oil spills with different oil types and oil film thicknesses to ensure data consistency and integrity. The average smoothing is expressed as follows: , in, Indicates in x i The result after smoothing the fluorescence spectrum intensity at the wavelength, f(x) i-1 ), f(x i ) and f(x i+1 ) respectively represent x i-1 x i and x i+1 The fluorescence spectral intensity measured at a given wavelength is expressed by the Logisti function as follows: The Logistic normalization process is represented as: , Where e represents the base of the natural logarithm, g(x) i ) indicates that in x i The fluorescence spectral intensity after Logistic function calculation at the specified wavelength, f(x) i ) indicates that in x iFluorescence spectral intensity measured at a specific wavelength Indicates in x i The result of normalizing the fluorescence spectrum intensity at the specified wavelength, g(x). max and g(x) min These represent the maximum and minimum fluorescence spectral intensities after the Logistic function operation, respectively.

[0008] Furthermore, the preprocessing of the fluorescence spectrum and fluorescence image data of the marine oil spill also includes preprocessing of the fluorescence image by using Sobel operators in the horizontal and vertical directions for edge detection to detect the edges of the glass container containing the marine oil spill sample, and cropping the fluorescence image based on the edge detection results to retain only the image of the simulated marine oil spill area.

[0009] Furthermore, the construction of three-dimensional fluorescence spectral data and a thickness difference matrix using preprocessed fluorescence spectra includes constructing a mathematical model between fluorescence spectral intensity and oil film thickness using the preprocessed fluorescence spectra. The fluorescence spectral intensity and oil film thickness follow an exponential decay model. Based on the model and combined with fluorescence spectral data of different oil types and oil film thicknesses, three-dimensional fluorescence spectral data is constructed. A thickness difference matrix is ​​constructed using the band characteristics in the fluorescence spectral data and the mathematical relationship between oil film thickness. Euclidean distance is used to calculate the fluorescence spectral intensity difference between adjacent oil film thicknesses at different wavelengths. The expression for the exponential decay model is: , Where I(d) is the fluorescence intensity when the oil film thickness is d, and I0 is the initial fluorescence intensity when the oil film thickness approaches zero. The intensity attenuation coefficient, The oil film thickness (mm) represents the rate at which the fluorescence signal decays with increasing thickness.

[0010] Furthermore, the construction of three-dimensional fluorescence spectral data and thickness difference matrix using preprocessed fluorescence spectra also includes constructing three-dimensional fluorescence spectral data for each oil type with different oil film thicknesses, taking each oil type as a unit and different oil film thicknesses as distinctions; and constructing a thickness difference matrix based on Euclidean distance, where the thickness difference matrix A... oil Each element in Represented as: , Among them, f t (x i ) and f t+1 (x i ) represent a certain type of oil film at x i The fluorescence spectral intensity between adjacent thicknesses t and t+1 at the wavelength, f j (x i) indicates that a certain type of oil film exists at x i The fluorescence spectral intensity at wavelength j for an oil film thickness of j; Fourier transform is used to perform pooling operations on the three-dimensional fluorescence spectral data in the frequency domain to extract the band features of the fluorescence spectrum, including using discrete two-dimensional Fourier transform to transform the fluorescence spectral curves of a certain oil with different oil film thicknesses. The transformed frequency domain curves are as follows: , Where M and N represent the data lengths of the wavelength range and oil film thickness range of the fluorescence spectrum, respectively; k and l represent the frequency indices of the fluorescence spectrum in the wavelength and oil film thickness directions, respectively.

[0011] Furthermore, the extraction of band and thickness features from the three-dimensional fluorescence spectral data and thickness difference matrix using a deep neural network includes using a differentiable pooling deep neural network (DP-DNN) to extract band and thickness features from the three-dimensional fluorescence spectral data and thickness difference matrix. Specifically, Fourier transform is used to perform pooling operations on the three-dimensional fluorescence spectral data in the frequency domain to extract the band features of the fluorescence spectrum; a globally learnable pooling method is used to perform pooling operations on the thickness difference matrix to extract the oil film thickness features of the fluorescence spectrum; the globally learnable pooling method generates spatial weights through an attention mechanism and performs a weighted summation, as shown below: , Where, x cij It is a thickness difference matrix The elements, output The attention weight a ij Through a small learning network of 1x1 convolution and ReLU Generate, represented as: .

[0012] Furthermore, the feature extraction of fluorescence image data based on a joint convolutional neural network includes extracting texture features of marine oil spill fluorescence images using a convolutional neural network (CNN), a deep residual network (ResNet), and a densely connected network (DenseNet). Key features of marine oil spills are captured by performing convolution operations on multi-band fluorescence images. Specifically, when performing convolution operations on the filtered multi-band fluorescence images, a k×k convolution kernel is gradually slid across the image from the top left corner, covering the entire image. After the convolution operation, each pixel in the image is weighted according to the band characteristics of the fluorescence spectrum, and the weighted sum of each pixel is represented as its pixel value. , Among them, w cThis represents the weight values ​​extracted based on the fluorescence spectral band features; I(x, y) is the pixel value at (x, y) of a certain band of the input fluorescence image; K(i, j) is the weight of the convolution kernel at position (i, j). P(x, y) represents the pixel value of the convolved feature map at position (x, y); P(x, y) represents the pixel value of the fluorescence image at position (x, y) of a certain band after fusing fluorescence spectral features.

[0013] Furthermore, the Lasso regression model is used to analyze and predict the extracted features after feature fusion. This includes fusing the band features of the fluorescence spectrum, oil film thickness features, and image features extracted from the fluorescence image data extracted from the differentiable pooling deep neural network, and then inputting these features into the Lasso regression model for analysis and prediction. A Bayesian deep learning framework is integrated into the Lasso regression model to further quantify the uncertainty of the prediction, thereby achieving accurate prediction and early warning of marine oil spills. First, the three types of features are encoded separately to obtain vector representations: Fs: fluorescence spectrum band feature vector; Ft: oil film thickness feature vector; Fi: fluorescence image texture feature vector. Then, the three types of features are concatenated along the feature dimension to form a unified high-dimensional feature representation. ,in This represents a vector concatenation operation; the concatenated and fused features are input into a Lasso regression model for analysis and prediction. Lasso regression achieves feature selection and sparse modeling by introducing L1 regularization into the loss function. Its optimization objective function is: , in, This represents the actual oil spill monitoring value for the j-th sample. For bias terms, The regression coefficients corresponding to the i-th feature are... This is a regularization parameter used to control the sparsity of the model.

[0014] Furthermore, the analysis and prediction of the extracted features based on the Lasso regression model after feature fusion also includes the joint optimization of the deep neural network structure parameters and the regularization hyperparameter λ of Lasso regression using the genetic algorithm GA and particle swarm optimization PSO. After multiple rounds of iterative optimization, the trained model uses the fused feature vector F_concat for monitoring and early warning, outputting the prediction results of marine oil spills and the corresponding uncertainty estimates, providing a reliable basis for accurate identification and early warning.

[0015] Secondly, a marine oil spill monitoring system based on a multi-band fluorescence spectral neural network includes: The data acquisition module is configured to acquire fluorescence spectrum and fluorescence image data of oil spills of different oil types and different oil film thicknesses; The preprocessing module is configured to preprocess the fluorescence spectral and fluorescence image data of marine oil spills, respectively. The difference module is configured to construct three-dimensional fluorescence spectral data and a thickness difference matrix using preprocessed fluorescence spectra; The spectral feature module is configured to extract band features and thickness features from three-dimensional fluorescence spectral data and thickness difference matrix using a deep neural network. The image feature module is configured to extract features from fluorescence image data based on a joint convolutional neural network. The prediction module is configured to analyze and predict the extracted features after feature fusion based on the Lasso regression model. The early warning module is configured to output early warning results.

[0016] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for monitoring marine oil spills based on a multi-band fluorescence spectral neural network.

[0017] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the marine oil spill monitoring method based on a multi-band fluorescence spectrum neural network.

[0018] In summary, the present invention has the following beneficial technical effects: (1) This invention proposes a marine oil spill monitoring and early warning system based on a multi-band fluorescence spectrum neural network. It combines the feature extraction capabilities of differentiable pooling deep neural networks (DP-DNN), convolutional neural networks (CNN), deep residual networks (ResNet), and densely connected networks (DenseNet). Through pooling and multi-band convolution operations, it achieves accurate prediction and early warning of marine oil spills. This method effectively integrates marine oil spill fluorescence spectral and fluorescence image data, providing a reliable basis for accurate identification and early warning of oil spills.

[0019] (2) This invention innovatively integrates a graph neural network with a dual-channel multi-band fluorescence spectrum, enabling the system to comprehensively consider spectral band features and oil film thickness features, while simultaneously extracting texture features from the fluorescence image. This dual-channel design enhances the model's ability to monitor and warn of complex marine oil spills, especially in the process of processing marine oil spill fluorescence spectrum data, greatly improving the accuracy and efficiency of monitoring.

[0020] (3) By employing a differentiable pooling deep neural network DP-DNN to perform band feature and thickness feature analysis on three-dimensional fluorescence spectral data and thickness difference matrix, this system resolves the contradiction between spatial resolution and spectral resolution in current fluorescence monitoring methods for directional feature extraction and oil type thickness identification. Simultaneously, based on fluorescence spectral image data, the system fuses spectral and image heterogeneous data features and uses genetic algorithm (GA) and particle swarm optimization (PSO) methods to optimize the neural network structure and hyperparameters, enabling the system to possess higher robustness and accuracy when facing complex marine oil spill monitoring and early warning tasks. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a marine oil spill monitoring method based on a multi-band fluorescence spectrum neural network according to Embodiment 1 of the present invention.

[0022] Figure 2 This is a schematic diagram of fluorescence spectra of different oil film thicknesses in multiple bands according to Embodiment 1 of the present invention.

[0023] Figure 3 This is a schematic diagram of fluorescence images of different oil film thicknesses in multiple bands according to Embodiment 1 of the present invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings.

[0025] Example 1 Reference Figure 1 This embodiment of a marine oil spill monitoring method based on a multi-band fluorescence spectral neural network includes: Obtain fluorescence spectra and fluorescence image data of oil spills of different oil types and with different oil film thicknesses; Preprocessing was performed on the fluorescence spectra and fluorescence image data of marine oil spills, respectively; A mathematical model was constructed using preprocessed oil spill fluorescence spectra to establish the relationship between fluorescence intensity and oil film thickness, which follows an exponential decay model. Based on this mathematical model, and combined with fluorescence spectral data for different oil types and oil film thicknesses, three-dimensional fluorescence spectral data was constructed. A thickness difference matrix was then built using the mathematical relationship between the band characteristics of the fluorescence spectral data and oil film thickness. Euclidean distance was used to calculate the fluorescence intensity differences between adjacent oil film thicknesses at different wavelengths. Differentiable pooling deep neural network DP-DNN is used to extract band features and thickness features from three-dimensional fluorescence spectral data and thickness difference matrix. Specifically, Fourier transform is used to perform pooling operation on the three-dimensional fluorescence spectral data in the frequency domain to extract the band features of the fluorescence spectrum; and a globally learnable pooling method is used to perform pooling operation on the thickness difference matrix to extract the oil film thickness features of the fluorescence spectrum. Texture features of marine oil spill fluorescence images are extracted using convolutional neural networks (CNN), deep residual networks (ResNet), and densely connected networks (DenseNet). This includes capturing key features of marine oil spills by performing convolution operations on multi-band fluorescence images. The band features of fluorescence spectra extracted from differentiable pooling deep neural networks, oil film thickness features, and image features extracted from fluorescence image data are fused and then input into a Lasso regression model for analysis and prediction. The Lasso regression model integrates a Bayesian deep learning framework to further quantify the uncertainty of the prediction, thereby achieving accurate prediction and early warning of marine oil spills. Genetic algorithm (GA) and particle swarm optimization (PSO) methods are used to optimize the structure and hyperparameters of the neural network, thereby further improving the performance and accuracy of the model.

[0026] Specifically, it includes the following steps: Obtain fluorescence spectra and fluorescence image data of oil spills of different oil types and with different oil film thicknesses. Step 1: Acquire fluorescence spectral data and images of marine oil spills; Add 20 ml of seawater to a petri dish and let it stand. Then, collect the original spectrum of the seawater within the preset excitation wavelength range (250-550 nm) and emission wavelength range (260-1700 nm). Add oil samples to the petri dish according to the set concentration gradients (1-10 µL with a concentration gradient of 1 µL, 10-50 µL with a concentration gradient of 5 µL, 50-100 µL with a concentration gradient of 10 µL, 100-500 µL with a concentration gradient of 50 µL, 500 µL-1 ml with a concentration gradient of 100 µL, and 1-3 ml with a concentration gradient of 1 ml). After each addition, save 5 spectral data for the corresponding concentration under the set excitation-emission wavelength matrix, which are the center point and the four surrounding points. The average spectrum of 3 acquisitions is taken for each acquisition. Follow the same steps to collect fluorescence spectral data for the other three oil samples at different concentrations. Similarly, with the selected excitation wavelength and corresponding emission band filter combination, the original seawater image was acquired, and different filters were replaced to obtain fluorescence images in different band ranges; oil samples were dropped at the same concentration gradient, and the image at that concentration was saved after each drop, and the corresponding band image was acquired by replacing the filter; the acquisition of multi-band fluorescence images of the four oil samples at each concentration was completed in sequence. Step 2: Preprocess the fluorescence spectral data and images of marine oil spills; In data processing, moving average smoothing and logistic normalization are applied to marine oil spill fluorescence spectral data of different oil types and oil film thicknesses to ensure data consistency and integrity. The moving average smoothing formula is as follows: , in, Indicates in x i The result after smoothing the fluorescence spectrum intensity at the wavelength, f(x) i-1 ), f(x i ) and f(x i+1 ) respectively represent x i-1 x i and x i+1 The fluorescence spectral intensity measured at a specific wavelength, i.e., the smoothed value of the fluorescence spectral intensity at each point, is the average value of that point and one point before and after it.

[0027] The Logistic normalization process is represented as follows: , , Where e represents the base of the natural logarithm, g(x) i ) indicates that in x i The fluorescence spectral intensity after Logistic function calculation at the specified wavelength, f(x) i ) indicates that in x i Fluorescence spectral intensity measured at the specified wavelength. Indicates in x i The result of normalizing the fluorescence spectrum intensity at the specified wavelength, g(x). max and g(x) min These represent the maximum and minimum fluorescence spectral intensities after the Logistic function operation, respectively.

[0028] For the preprocessing of fluorescence images, Sobel operators in the horizontal and vertical directions were used for edge detection to detect the edges of glass containers containing marine oil spill samples. Based on the edge detection results, the fluorescence images were cropped to retain only the images of the simulated marine oil spill areas.

[0029] The Sobel operator sets 3x3 convolution kernels G(x) and G(y) in the horizontal and vertical directions, respectively: , , For each fluorescence image I(i, j), convolution operations are performed using G(x) and G(y) respectively to obtain the gradient components of each pixel (i, j) in the x and y directions. and ; For each pixel (i, j) in the image, the total gradient magnitude is calculated using Euclidean distance combined with the gradients in the x and y directions. The calculation process is expressed as follows: , set up The threshold is 150; after calculating any pixel in the fluorescence image If the value is greater than the threshold, the point is identified as an edge point; if the value is less than the threshold, it is identified as a non-edge point. Based on the edge detection results, the image is cropped, retaining only the central region containing the oil spill sample. Step 3: Construct a mathematical model; An exponential decay model between fluorescence intensity and oil film thickness was constructed using the pretreated oil spill fluorescence spectrum. Its mathematical expression is as follows: , Where I(d) is the fluorescence intensity when the oil film thickness is d, and I0 is the initial fluorescence when the oil film thickness approaches zero. The intensity attenuation coefficient, The oil film thickness (mm) represents the rate at which the fluorescence signal decays with increasing thickness.

[0030] To further construct three-dimensional fluorescence spectral data, specifically, fluorescence intensity was measured for each oil type at M excitation wavelengths (λ_ex1, λ_ex2, ..., λ_exM) and N emission wavelengths (λ_em1, λ_em2, ..., λ_emN), and D oil film thickness gradients were set to form a three-dimensional data tensor. This tensor comprehensively characterizes the three-dimensional response relationship of fluorescence intensity with excitation wavelength, emission wavelength, and oil film thickness. Further, a thickness difference matrix A is constructed. oil Each element in Represented as: , Among them, f t (x i ) and f t+1 (x i ) represent a certain type of oil film at x i The fluorescence spectral intensity between adjacent thicknesses t and t+1 at the wavelength, f j (x i ) indicates that a certain type of oil film exists at x i Fluorescence intensity at wavelength j for an oil film thickness of j.

[0031] Step 4: Extract band features and thickness features: In the stage of extracting band and thickness features from three-dimensional fluorescence spectral data and thickness difference matrix using a differentiable pooling deep neural network DP-DNN, Fourier transform is used to perform pooling operations on the three-dimensional fluorescence spectral data in the frequency domain to extract the band features of the fluorescence spectrum; a globally learnable pooling method is used to perform pooling operations on the thickness difference matrix to extract the oil film thickness features of the fluorescence spectrum. Specifically, the input layer receives three-dimensional fluorescence data F and thickness difference matrix A. oil The Fourier pooling layer performs a DFT on F and extracts the low-frequency components; the learnable pooling layer generates spatial weights through an attention mechanism, and outputs a thickness feature vector after weighted summation. Simultaneously, the attention weight generation network... This is specifically implemented as a combination of 1×1 convolutional layers, ReLU activation function, and Softmax layer to enhance operability.

[0032] Fourier transform was used to perform pooling operations on the three-dimensional fluorescence spectral data in the frequency domain to extract the band features of the fluorescence spectrum. This included using discrete two-dimensional Fourier transform to transform the fluorescence spectral curves of a certain oil with different oil film thicknesses. The transformed frequency domain curves are as follows: , Where M and N represent the data lengths of the wavelength range and oil film thickness range of the fluorescence spectrum, respectively; k and l represent the frequency indices of the fluorescence spectrum in the wavelength and oil film thickness directions, respectively. A globally learnable pooling method is used to perform pooling operations on the thickness difference matrix to extract oil film thickness features from the fluorescence spectrum. This includes generating spatial weights through an attention mechanism using globally learnable pooling and then performing weighted summation. , Where, x cij It is a thickness difference matrix The elements, output The attention weight a ij Through a small learning network of 1x1 convolution and ReLU Generation can be represented as: .

[0033] Step 5: Extract texture features from the fluorescence image: For fluorescence images, texture features of marine oil spill fluorescence images are extracted using convolutional neural networks (CNN), deep residual networks (ResNet), and densely connected networks (DenseNet); the specific process is as follows: The input is a multi-band fluorescence image after Sobel edge detection and cropping, with each image having dimensions H×W×C, where C is the number of bands. Each band image is then standardized. , in, and denoted as the mean and standard deviation of the c-th band, respectively.

[0034] Furthermore, shallow feature extraction (CNN) is used to extract texture features from marine oil spill fluorescence images. This includes using a k×k convolution kernel to gradually slide across the image from the top left corner to cover the entire image when performing convolution operations on filtered multi-band fluorescence images. Furthermore, deep feature extraction (ResNet + DenseNet) is used to extract mid-to-high-level texture features. Specifically, skip connections in the ResNet module alleviate gradient vanishing, where the output of the Lth layer is: , The DenseNet module receives features from all preceding layers as input at each layer, enhancing feature reuse. , After the convolution operation, each pixel in the image is weighted according to the band characteristics of the fluorescence spectrum. The weighted sum of each pixel's value can be represented as the pixel value of that point: , Among them, w c This represents the weight value extracted based on the fluorescence spectral band features; I(x,y) is the pixel value of the input fluorescence image at (x,y); K(i,j) is the weight of the convolution kernel at position (i,j). P(x,y) represents the pixel value of the convolved feature map at position (x,y); P(x,y) represents the pixel value of the fluorescence image at position (x,y) of a certain band after fusing fluorescence spectral features.

[0035] Finally, the feature map is transformed into a feature vector F through global average pooling (GAP). i ∈ to R Di This is a vector representation of image texture features, used for subsequent feature fusion and predictive analysis.

[0036] Step 6: Combine and merge the three features: The band features of fluorescence spectra extracted from differentiable pooling deep neural networks, oil film thickness features, and image features extracted from fluorescence image data are spliced ​​together according to latitude features to form fused features. The fused features are then input into a Lasso regression model to output the confidence interval for marine oil spill monitoring and prediction.

[0037] 1. First, encode the three types of features separately to obtain vector representations: F s : Fluorescence spectral band feature vector (from DP-DNN); Ft Oil film thickness feature vector (derived from the thickness difference matrix); F i Fluorescent image texture feature vector (from joint CNN); Use a fully connected layer to map each feature vector to the same dimension D: , Subsequently, these three types of features are concatenated along the feature dimension to form a unified high-dimensional feature representation:

[0038] in This represents a vector concatenation operation; 2. Finally, the concatenated and fused features are input into the Lasso regression (Least Absolute Shrinkage and Selection Operator) model for analysis and prediction. Lasso regression effectively achieves feature selection and sparse modeling by introducing L1 regularization into the loss function. Its optimization objective function is: , in, This represents the actual oil spill monitoring value for the j-th sample. For bias terms, The regression coefficients corresponding to the i-th feature are... This is a regularization parameter used to control the sparsity of the model.

[0039] 3. To quantify prediction uncertainty, a Bayesian deep learning framework is integrated into the Lasso regression model, assuming the regression coefficients... Follows prior distribution: The posterior distribution p is estimated through variational inference or MCMC sampling. The predicted distribution is as follows: , By analyzing the regression coefficients By establishing probability distribution assumptions, the uncertainty of prediction results is quantified, and the confidence interval of marine oil spill monitoring and prediction is output.

[0040] Step 7: Optimize the parameters: To further improve model performance and robustness, this method employs a genetic algorithm (GA) and particle swarm optimization (PSO) to jointly optimize the structural parameters of the deep neural network (including the number of convolutional layers, kernel size, and pooling method) and the regularization hyperparameter λ of Lasso regression. The optimization process aims to find the optimal combination of parameters to minimize the model's prediction error on the validation set.

[0041] First, the parameters to be optimized are encoded into a vector form that the optimization algorithm can process. The parameter vector is defined as follows: Number of convolutional layers: L∈{1, 2, ... L} max} Kernel size: K∈{3, 5, 7} Pooling method: P∈{MaxPooling,AveragePooling,GlobalAveragePooling} Lasso regularization hyperparameters: λ∈[0.001,1.0] For the above mixed-type parameters (discrete and continuous coexisting), such as Figure 2 As shown, this embodiment designs encoding strategies: a mixed encoding of real numbers and integers is used for GA; continuous real number encoding is used for PSO, and discrete parameters are rounded or mapped to the nearest value in a predefined set. A two-stage hybrid optimization strategy is adopted: on the one hand, in the GA (Global Exploration) stage, a genetic algorithm is used to perform a coarse search in the global scope, and potential parameter regions are selected through selection, crossover, and mutation operations to avoid the optimization process from getting trapped in local optima too early; on the other hand, in the PSO (Local Fine Search) stage, the best individuals obtained in the GA stage are used as the initial particle swarm for PSO. Taking advantage of the fast convergence speed and strong local search ability of the particle swarm algorithm, a fine search is performed in the neighborhood of the high-quality solution, thereby efficiently approximating the global optimum.

[0042] To verify the effectiveness of the GA-PSO hybrid optimization strategy, such as... Figure 3 As shown, this embodiment conducted a systematic experiment on a marine oil spill SAR image dataset. The experimental setup included dividing the dataset into training, validation, and test sets in a 7:2:1 ratio; comparing the accuracy, F1 score, and AUC of the models before and after optimization. The results show that the model optimized with GA-PSO achieved an accuracy of 98.2% and an F1 score of 0.976 on the test set, significantly outperforming the unoptimized model (accuracy 94.5%, F1 score 0.932) and the results of a single optimization algorithm (GA or PSO). Furthermore, ablation experiments further demonstrate that the hybrid strategy outperforms the single optimization method in both convergence speed and stability.

[0043] After multiple rounds of iterative optimization, the trained model utilizes the fused feature vectors. The system will monitor and issue early warnings, outputting predictions of marine oil spills and corresponding uncertainty estimates, providing a reliable basis for accurate identification and early warning.

[0044] Example 2 This embodiment provides a marine oil spill monitoring system based on a multi-band fluorescence spectral neural network, including: The data acquisition module is configured to acquire fluorescence spectrum and fluorescence image data of oil spills of different oil types and different oil film thicknesses; The preprocessing module is configured to preprocess the fluorescence spectral and fluorescence image data of marine oil spills, respectively. The difference module is configured to construct three-dimensional fluorescence spectral data and a thickness difference matrix using preprocessed fluorescence spectra; The spectral feature module is configured to extract band features and thickness features from three-dimensional fluorescence spectral data and thickness difference matrix using a deep neural network. The image feature module is configured to extract features from fluorescence image data based on a joint convolutional neural network. The prediction module is configured to analyze and predict the extracted features after feature fusion based on the Lasso regression model. The early warning module is configured to output early warning results.

[0045] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for monitoring marine oil spills based on a multi-band fluorescence spectral neural network.

[0046] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor as described in the marine oil spill monitoring method based on a multi-band fluorescence spectral neural network.

[0047] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring marine oil spills based on a multi-band fluorescence spectral neural network, characterized in that, include: Obtain fluorescence spectra and fluorescence image data of oil spills of different oil types and with different oil film thicknesses; The fluorescence spectra and fluorescence image data of marine oil spills were preprocessed respectively; Three-dimensional fluorescence spectral data and thickness difference matrix were constructed using the preprocessed fluorescence spectra. Deep neural networks were used to extract band features and thickness features from three-dimensional fluorescence spectral data and thickness difference matrices. Feature extraction from fluorescence image data based on a joint convolutional neural network; The extracted features are analyzed and predicted based on the Lasso regression model after feature fusion; Output the warning results.

2. The marine oil spill monitoring method based on a multi-band fluorescence spectral neural network according to claim 1, characterized in that, The preprocessing of the fluorescence spectral and fluorescence image data of marine oil spills includes moving average smoothing and logistic normalization of the fluorescence spectral data of marine oil spills with different oil types and oil film thicknesses to ensure data consistency and integrity. The average smoothing is expressed as follows: ， in, Indicates in x i The result after smoothing the fluorescence spectrum intensity at the wavelength, f(x) i-1 ), f(x i ) and f(x i+1 ) respectively represent x i-1 x i and x i+1 The fluorescence spectral intensity measured at a given wavelength is expressed by the Logisti function as follows: The Logistic normalization process is represented as: ， Where e represents the base of the natural logarithm, g(x) i ) indicates that in x i The fluorescence spectral intensity after Logistic function calculation at the specified wavelength, f(x) i ) indicates that in x i Fluorescence spectral intensity measured at a specific wavelength Indicates in x i The result of normalizing the fluorescence spectrum intensity at the specified wavelength, g(x). max and g(x) min These represent the maximum and minimum fluorescence spectral intensities after the Logistic function operation, respectively.

3. The marine oil spill monitoring method based on a multi-band fluorescence spectral neural network according to claim 2, characterized in that, The preprocessing of the fluorescence spectrum and fluorescence image data of marine oil spills includes preprocessing of the fluorescence image by using Sobel operators in the horizontal and vertical directions for edge detection to detect the edges of the glass containers containing marine oil spill samples, and cropping the fluorescence image based on the edge detection results to retain only the image of the simulated marine oil spill area.

4. The marine oil spill monitoring method based on a multi-band fluorescence spectral neural network according to claim 3, characterized in that, The process of constructing three-dimensional fluorescence spectral data and a thickness difference matrix using preprocessed fluorescence spectra includes: constructing a mathematical model between fluorescence spectral intensity and oil film thickness using the preprocessed fluorescence spectra; the relationship between fluorescence spectral intensity and oil film thickness follows an exponential decay model; constructing three-dimensional fluorescence spectral data based on the model and combined with fluorescence spectral data of different oil types and oil film thicknesses; constructing a thickness difference matrix using the band characteristics in the fluorescence spectral data and the mathematical relationship between oil film thickness; and calculating the fluorescence spectral intensity difference of adjacent oil film thicknesses at different wavelengths using Euclidean distance. The expression for the exponential decay model is as follows: ， Where I(d) is the fluorescence intensity when the oil film thickness is d, and I0 is the initial fluorescence intensity when the oil film thickness approaches zero. The intensity attenuation coefficient, The oil film thickness (mm) represents the rate at which the fluorescence signal decays with increasing thickness.

5. The marine oil spill monitoring method based on a multi-band fluorescence spectral neural network according to claim 4, characterized in that, The process of constructing three-dimensional fluorescence spectral data and a thickness difference matrix using preprocessed fluorescence spectra also includes constructing three-dimensional fluorescence spectral data for each oil type with different oil film thicknesses as the distinction; and constructing a thickness difference matrix based on Euclidean distance, where the thickness difference matrix A... oil Each element in Represented as: ， Among them, f t (x i ) and f t+1 (x i ) represent a certain type of oil film at x i The fluorescence spectral intensity between adjacent thicknesses t and t+1 at the wavelength, f j (x i ) indicates that a certain type of oil film exists at x i The fluorescence spectral intensity at wavelength j for an oil film thickness of j; Fourier transform is used to perform pooling operations on the three-dimensional fluorescence spectral data in the frequency domain to extract the band features of the fluorescence spectrum, including using discrete two-dimensional Fourier transform to transform the fluorescence spectral curves of a certain oil with different oil film thicknesses. The transformed frequency domain curves are as follows: ， Where M and N represent the data lengths of the wavelength range and oil film thickness range of the fluorescence spectrum, respectively; k and l represent the frequency indices of the fluorescence spectrum in the wavelength and oil film thickness directions, respectively.

6. The marine oil spill monitoring method based on a multi-band fluorescence spectral neural network according to claim 5, characterized in that, The method of extracting band and thickness features from three-dimensional fluorescence spectral data and thickness difference matrix using a deep neural network includes using a differentiable pooling deep neural network (DP-DNN) to extract band and thickness features from the three-dimensional fluorescence spectral data and thickness difference matrix. Specifically, Fourier transform is used to perform pooling operations on the three-dimensional fluorescence spectral data in the frequency domain to extract band features of the fluorescence spectrum; a globally learnable pooling method is used to perform pooling operations on the thickness difference matrix to extract oil film thickness features from the fluorescence spectrum; the globally learnable pooling method generates spatial weights through an attention mechanism and performs weighted summation, as shown below: ， Where, x cij It is a thickness difference matrix The elements, output The attention weight a ij Through a small learning network of 1x1 convolution and ReLU Generate, represented as: .

7. The marine oil spill monitoring method based on a multi-band fluorescence spectral neural network according to claim 6, characterized in that, The feature extraction of fluorescence image data based on a joint convolutional neural network includes using a convolutional neural network (CNN), a deep residual network (ResNet), and a densely connected network (DenseNet) to extract texture features of marine oil spill fluorescence images. Key features of marine oil spills are captured by performing convolution operations on multi-band fluorescence images. Specifically, when performing convolution operations on filtered multi-band fluorescence images, a k×k convolution kernel is gradually slid across the image from the top left corner, covering the entire image. After the convolution operation, each pixel in the image is weighted according to the band characteristics of the fluorescence spectrum, and the weighted sum of each pixel is represented as its pixel value. ， Among them, w c This represents the weight values ​​extracted based on the fluorescence spectral band features; I(x, y) is the pixel value at (x, y) of a certain band of the input fluorescence image; K(i, j) is the weight of the convolution kernel at position (i, j). P(x, y) represents the pixel value of the convolved feature map at position (x, y); P(x, y) represents the pixel value of the fluorescence image at position (x, y) of a certain band after fusing fluorescence spectral features.

8. The marine oil spill monitoring method based on a multi-band fluorescence spectral neural network according to claim 7, characterized in that, The Lasso regression model-based analysis and prediction of extracted features after feature fusion includes fusing the band features of fluorescence spectra, oil film thickness features, and image features extracted from differentiable pooling deep neural networks with the images extracted from fluorescence image data, and then inputting these features into the Lasso regression model for analysis and prediction. The Lasso regression model integrates a Bayesian deep learning framework to further quantify the uncertainty of prediction, thereby achieving accurate prediction and early warning of marine oil spills. First, the three types of features are encoded separately to obtain vector representations: Fs: fluorescence spectrum band feature vector; Ft: oil film thickness feature vector; Fi: fluorescence image texture feature vector. Then, the three types of features are concatenated along the feature dimension to form a unified high-dimensional feature representation. ,in This represents a vector concatenation operation; the concatenated and fused features are input into a Lasso regression model for analysis and prediction. Lasso regression achieves feature selection and sparse modeling by introducing L1 regularization into the loss function. Its optimization objective function is: ， in, This represents the actual oil spill monitoring value for the j-th sample. For bias terms, The regression coefficients corresponding to the i-th feature are... This is a regularization parameter used to control the sparsity of the model.

9. A method for monitoring marine oil spills based on a multi-band fluorescence spectral neural network according to claim 8, characterized in that, The Lasso regression model is used to analyze and predict the extracted features after feature fusion. It also includes using the genetic algorithm (GA) and particle swarm optimization (PSO) to jointly optimize the structural parameters of the deep neural network and the regularization hyperparameter λ of the Lasso regression. After multiple rounds of iterative optimization, the trained model uses the fused feature vector F_concat for monitoring and early warning, and outputs the prediction results of marine oil spills and the corresponding uncertainty estimates, providing a reliable basis for accurate identification and early warning.

10. A marine oil spill monitoring system based on a multi-band fluorescence spectral neural network, characterized in that, include: The data acquisition module is configured to acquire fluorescence spectrum and fluorescence image data of oil spills of different oil types and different oil film thicknesses; The preprocessing module is configured to preprocess the fluorescence spectral and fluorescence image data of marine oil spills, respectively. The difference module is configured to construct three-dimensional fluorescence spectral data and a thickness difference matrix using preprocessed fluorescence spectra; The spectral feature module is configured to extract band features and thickness features from three-dimensional fluorescence spectral data and thickness difference matrix using a deep neural network. The image feature module is configured to extract features from fluorescence image data based on a joint convolutional neural network. The prediction module is configured to analyze and predict the extracted features after feature fusion based on the Lasso regression model. The early warning module is configured to output early warning results.

Images