A sea surface oil spill detection method based on a hyperspectral remote sensing detection 1D-CNN model

Abstract

A kind of sea surface oil spill hyperspectral remote sensing detection 1D-CNN model relates to safety risk assessment technical field, including 3 one-dimensional convolution layers, 3 one-dimensional maximum pooling layers and 2 fully connected layers, the model is equipped with adaptive long-term moment estimation (ALTME) optimizer, selects linear rectifier function (ReLU) as activation function, selects multi-classification cross-entropy as loss function.The present application is based on the gradient updating mechanism of ALTME optimizer second momentum term, gives adaptive cumulative learning weight to historical oil spill spectral gradient information, realizes the adaptive selection function of cumulative learning weight by iteration process.Thereby increase the learning proportion of historical spectral information, fully learn oil spill spectral characteristics, realize the cumulative learning function of different batches, long-term oil spill spectral characteristics.Further overcome the problem of poor spectral separability between different kinds of light oil and different thicknesses of heavy oil, accurately identify the type of sea surface oil spill and detect different thicknesses of oil film.

Description

Technical Field

[0001] This invention relates to the field of safety risk assessment technology, and in particular to a 1D-CNN model for hyperspectral remote sensing detection of oil spills on the sea surface. Background Technology

[0002] With the rapid development of the global maritime transport and offshore oil extraction industries, marine oil spills are frequent, seriously affecting the sustainable development of the marine ecological environment and marine resources. Accurate identification and analysis of oil spill types are crucial for determining liability in oil spill accidents and are of great significance for on-site emergency response and rapid and effective treatment of marine pollution. Obtaining accurate oil slick thickness and estimating the spill volume is an important basis for pollution compensation and accountability, and plays a vital role in scientific decision-making and determining the accident level at the scene of an oil spill.

[0003] Oil spills are considered weak targets in optical remote sensing research. The spectral curves of different types of light oil are similar and difficult to distinguish. Traditional remote sensing modeling methods based on characteristic bands are ineffective in identifying light oil types and are mostly conducted in laboratory settings. Heavy oil films strongly absorb sunlight, have low remote sensing reflectivity, and exhibit poor spectral separation between films of varying thicknesses, making it challenging to conduct research on the thickness of thick oil films using optical remote sensing. Currently, oil film thickness research primarily focuses on inverting the thickness of thin or relative oil films using traditional remote sensing modeling methods, with insufficient attention paid to the detection of the absolute thickness of thick oil films.

[0004] Hyperspectral remote sensing features high spectral resolution and a wide spectral response range, enabling it to carry richer spectral characteristics of oil spills compared to traditional multispectral remote sensing techniques. Deep learning, with its powerful ability to extract high-dimensional data features, has developed rapidly in recent years. Combining hyperspectral remote sensing with deep learning is beneficial for mining spectral characteristics of oil spills, accurately identifying oil spill types, and detecting oil films of varying thicknesses. Summary of the Invention

[0005] This invention provides a 1D-CNN model for hyperspectral remote sensing detection of marine oil spills. This model is constructed by establishing an Adaptive Long Time Moment Estimation (ALTME) optimizer and building upon a one-dimensional convolutional neural network (1D-CNN). Based on the gradient update mechanism of the second momentum term of the ALTME optimizer, the model assigns adaptive cumulative learning weights to historical oil spill spectral gradient information, achieving adaptive selection of these weights through an iterative process. This increases the learning weight of historical spectral information, fully learns the spectral characteristics of oil spills, and achieves cumulative learning of spectral characteristics from different batches and over long periods. This overcomes the problem of poor spectral separability between different types of light oil and different thicknesses of heavy oil, accurately identifying the types of marine oil spills and detecting oil films of different thicknesses.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A 1D-CNN model for hyperspectral remote sensing of oil spills on the sea surface includes three 1D convolutional layers, three 1D max pooling layers, and two fully connected layers. The model is equipped with an adaptive long-term moment estimation (ALTME) optimizer, selects the linear rectified function (ReLU) as the activation function, and selects multi-class cross-entropy as the loss function.

[0008] Preferably, the Adaptive Long Time Moments Estimation (ALTME) optimizer is implemented based on the Keras framework, and its algorithm steps include:

[0009] The ALTME optimizer, based on the second momentum term gradient update mechanism, assigns adaptive cumulative learning weights to historical oil spill spectral gradient information. Through an iterative process, it adaptively selects the cumulative learning weights of the second momentum term. Based on the selected cumulative learning weights, while ensuring model convergence, it increases the proportion of historical cumulative spectral information learning, enabling the model to fully learn oil spill spectral features. The ALTME optimizer also includes a momentum correction term, which effectively avoids the instability of the first and second momentum terms due to small gradients in the early stages of training, thus affecting the model's training performance. Theoretically, this allows the oil spill detection model to have the ability to memorize long-term oil spill feature information while maintaining stable oil spill detection capabilities.

[0010] Preferably, the 1D-CNN model uses a one-dimensional planar convolution kernel to convolve the receptive field of the oil spill spectral information, and the calculation process is shown in formula (1):

[0011]

[0012] In the formula, h i,j Let w represent the j-th output feature map of the i-th convolutional layer, M represent the number of feature maps in the i-th convolutional layer, and w represent the output feature map of the j-th convolutional layer. i ,m j Represents weight, b i,j represents the paranoia factor, and g(·) represents the activation function.

[0013] Preferably, the linear rectified function (ReLU) is calculated as shown in equation (2):

[0014] g(x) = max(0,x) (2)

[0015] In the formula, max(·) represents the max function, and x represents the input vector of the previous layer neuron.

[0016] Preferably, the pooling method of the 1D-CNN model is set to one-dimensional max pooling, and each pooling layer corresponds to the receptive field of the convolutional layer within the range of N×1. The max pooling formula is shown in equation (3):

[0017]

[0018] In the formula, max(·) represents the one-dimensional max pooling function, u(n,1) represents the window function of the convolutional layer, and a j It is the maximum value within the neighborhood.

[0019] Preferably, the 1D-CNN model selects multi-class cross-entropy as the loss function of the oil spill detection model, and conducts oil spill detection experiments based on the Softmax classifier. The calculation method of multi-class cross-entropy is shown in Equation (4):

[0020]

[0021] In the formula, Z i The logits function represents the softmax classifier, where n represents the number of targets detected, and P... i This represents the true sample distribution.

[0022] Preferably, the backpropagation (BP) algorithm is used in the back-propagation process of the 1D-CNN model. Combined with the calibrated sample label dataset, the weights and biases of the model are adjusted layer by layer to achieve the optimal mapping effect of the model to the sample label dataset.

[0023] The beneficial effects of the 1D-CNN model for hyperspectral remote sensing detection of oil spills on the sea surface in this invention are as follows:

[0024] 1. This invention utilizes a 1D-CNN model for hyperspectral remote sensing of oil spills on the sea surface equipped with an ALTME optimizer. This model has the ability to memorize spectral feature information from different batches over long periods. Compared with traditional optimizers, it can effectively avoid the loss of spectral feature information of oil spills, making the model perform better and more stable throughout the entire iterative training process. This results in a more accurate and stable oil spill detection capability.

[0025] 2. The ALTME optimizer of this invention has a momentum correction term, which can effectively avoid the situation where unstable first and second momentum terms are generated in the early stage of training due to small gradients, thus affecting the model training effect. Attached Figure Description

[0026] Figure 1 A schematic diagram of the 1D-CNN model for hyperspectral remote sensing detection of oil spills on the sea surface in this invention;

[0027] Figure 2Figure 1 shows the experimental results of oil spill detection at sea surface according to this invention; Detailed Implementation

[0028] The following description provides a detailed explanation of the embodiments of the present invention in a step-by-step manner. This description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0029] In the description of this invention, it should be noted that the terms "upper," "lower," "left," "right," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or a specific orientational structure and operation. Therefore, they should not be construed as limiting this invention.

[0030] Example 1

[0031] This invention is based on an Adaptive Long-Term Moments Estimation (ALTME) optimizer, constructed using a one-dimensional convolutional neural network (1D-CNN). The implementation of the ALTME optimizer algorithm is as follows:

[0032] The ALTME algorithm is implemented based on the Keras framework. By assigning adaptive weights to the historical spectral gradient information of the oil spill, the proportion of historical spectral cumulative learning is increased, avoiding the loss of effective spectral feature information. This overcomes the problem of poor spectral separability among light oils and deeply explores the spectral differences between oil films of different thicknesses. The ALTME algorithm is shown in Table 1.

[0033] Table 1. Adaptive Long-Term Moment Estimation Optimizer (ALTME) Algorithm

[0034]

[0035] Where β represents the first moment, λ represents the adaptive cumulative learning weights, iter represents the number of iterations, and g t m represents the gradient at time t. t Representing the first momentum term, v t Representing the second momentum term, m t correct Representing the first momentum correction term, v t correct Represents the second momentum correction term, Loss(·) represents the loss function, and θ t α represents the weights of the deep learning model, α represents the learning rate, and ε represents the weight adjustment value.

[0036] Example 2

[0037] The specific structure of the 1D-CNN model for hyperspectral remote sensing detection of oil spills on the sea surface in this invention is as follows:

[0038] The ALTME optimizer was integrated into a 1D-CNN to construct a 1D-CNN model for hyperspectral remote sensing of oil spills on the sea surface. This model consists of three 1D convolutional layers, three 1D max-pooling layers, and two fully connected layers, as shown in the model structure below. Figure 1 As shown.

[0039] Unlike traditional convolution methods, 1D-CNN is based on the operation of convolution of the receptive field of oil spill spectral information using a one-dimensional planar convolution kernel. It has strong feature extraction capabilities and nonlinear data fitting capabilities. The calculation process is shown in Equation (1).

[0040]

[0041] Among them, h i,j Let w represent the j-th output feature map of the i-th convolutional layer, M represent the number of feature maps in the i-th convolutional layer, and w represent the output feature map of the j-th convolutional layer. i,mj Represents weight, b i,j represents the paranoia factor, and g(·) represents the activation function.

[0042] This invention significantly reduces the number of parameters in the marine oil spill detection model by using sparse representation and weight sharing, thereby reducing the redundancy of hyperspectral remote sensing data and improving oil spill detection efficiency.

[0043] This invention selects the Rectified Linear Function (ReLU) as the activation function for the marine oil spill detection model. This function enables the model to perform gradient descent more efficiently while maintaining a relatively fast computation speed. Combined with the ALTME optimizer proposed in this paper, it can effectively suppress the phenomena of gradient vanishing and gradient exploding.

[0044] The formula for the ReLU function is shown in equation (2):

[0045] g(x) = max(0,x) (2)

[0046] In the formula, max(·) represents the max function, and x represents the input vector of the previous layer neuron.

[0047] The pooling method of the model is set to one-dimensional max pooling. This module can reduce the risk of overfitting in the oil spill detection model and enhance the robustness of the model through the pooling process. Each pooling layer corresponds to the receptive field of the convolutional layer within the range of N×1. The max pooling formula is shown in equation (3):

[0048]

[0049] Where max(·) represents the one-dimensional max pooling function, u(n,1) represents the window function of the convolutional layer, and a j It is the maximum value within the neighborhood.

[0050] This invention selects multi-class cross-entropy as the loss function of the oil spill detection model, and conducts oil spill detection experiments based on the Softmax classifier, as shown in equation (4):

[0051]

[0052] Among them, Z i The logits function represents the softmax classifier, where n represents the number of targets detected, and P... i This represents the true sample distribution.

[0053] The reverse fine-tuning process of this invention adopts the BP (Back Propagation) algorithm. The oil spill detection model is combined with the calibrated sample label dataset, and the weights and biases of the model are adjusted layer by layer to achieve the best mapping effect of the model to the sample label dataset.

[0054] Example 3: The experimental process for detecting oil spills on the sea surface based on a 1D-CNN model for hyperspectral remote sensing is as follows:

[0055] An oil spill detection experiment was conducted based on a 1D-CNN model for hyperspectral remote sensing of oil spills on the sea surface. Overall accuracy (OA) and the Kappa coefficient were selected to measure the overall detection accuracy and consistency of the oil spill detection experiment. The F1-Score, which balances recall and precision, was chosen to measure the single-target detection accuracy of the oil spill detection model for each experimental group. The calculation formulas for each accuracy evaluation index are shown in equations (5) to (9).

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] In the above formulas, TP represents the number of pixels of true positive examples, TN represents the number of pixels of true negative examples, FP represents the number of pixels of false positive examples, and FN represents the number of pixels of false negative examples. The oil spill detection accuracy is shown in Tables 2-4, and the detection results are as follows: Figure 2 As shown.

[0062] Table 2. Accuracy of oil spill detection at time T1

[0063]

[0064] Table 3. Accuracy of oil spill detection at time T2

[0065]

[0066] Table 4. Accuracy of oil spill detection at time T3

[0067]

[0068]

[0069] As shown in Tables 2 to 4, the oil spill detection model equipped with the ALTME optimizer has higher OA and Kappa coefficients at times T1 to T3 than the SVM and traditional 1D-CNN models. The single-target detection accuracy F1-Score for different oil spill experimental groups is also higher than the other two types of models.

[0070] like Figure 2 As shown, the 1D-CNN model for oil spill detection on the sea surface equipped with the ALTME optimizer can overcome the problem of poor spectral separability between different types of oil spills and between oil films of different thicknesses. Through cumulative learning, it can fully explore and utilize the spectral feature data of oil spills from different batches and over long periods, thereby effectively improving the detection effect of oil spills on the sea surface.

Claims

1. A method for detecting oil spills on the sea surface based on a 1D-CNN model of hyperspectral remote sensing, characterized by: The model includes three one-dimensional convolutional layers, three one-dimensional max pooling layers, and two fully connected layers. It is equipped with an adaptive long-term moment estimation optimizer, which selects the linear rectified function as the activation function and the multi-class cross-entropy as the loss function. The adaptive long-term moment estimation optimizer described above is implemented based on the Keras framework, and its algorithm process includes: Based on the gradient update mechanism of the second momentum term, adaptive cumulative learning weights are assigned to historical oil spill spectral gradient information. The adaptive selection function of the cumulative learning weights of the second momentum term of the optimizer is realized through an iterative process. Based on the selected cumulative learning weights, the proportion of historical cumulative spectral information learning is increased while ensuring model convergence, so that the model can fully learn the characteristics of oil spill spectra. By setting a momentum correction term, the unstable first and second momentum terms caused by small gradients in the early stage of training are avoided, which may affect the model training effect. The gradient update mechanism for the second momentum term is expressed as follows: v t = (1 + λ ) v t-1 +λg t 2 ; in, λ Represents adaptive cumulative learning weights. g t represent t Time gradient, v t This represents the second momentum term.

2. The method for detecting oil spills on the sea surface based on a 1D-CNN model of hyperspectral remote sensing as described in claim 1, characterized in that: The 1D-CNN model described above uses a one-dimensional planar convolution kernel to convolve the receptive field of the oil spill spectral information. The calculation process is shown in formula (1): （1） In the formula, h i,j Representing the i The first convolutional layer j Each output feature map M Representing the i The number of feature maps in each convolutional layer. w i , m j Represents weight, b i,j Represents the degree of paranoia. g (·) represents the activation function.

3. The method for detecting oil spills on the sea surface based on a hyperspectral remote sensing 1D-CNN model as described in claim 1, characterized in that: The calculation method for the linear rectifier function is shown in equation (2): （2） In the formula, max(·) represents the max function. x This represents the input vector of the neuron in the previous layer.

4. The method for detecting oil spills on the sea surface based on a hyperspectral remote sensing 1D-CNN model as described in claim 1, characterized in that: The pooling method of the 1D-CNN model is set to one-dimensional max pooling, and each pooling layer corresponds to... N The receptive field of the convolutional layer within the range of ×1 is given by the max pooling formula as shown in equation (3): （3） In the formula, max(·) represents a one-dimensional max pooling function. u ( n , 1) represents the window function of the convolutional layer, a j It is the maximum value within the neighborhood.

5. The method for detecting oil spills on the sea surface based on a hyperspectral remote sensing 1D-CNN model as described in claim 1, characterized in that: The 1D-CNN model selected multi-class cross-entropy as the loss function of the oil spill detection model, and carried out oil spill detection experiments based on the Softmax classifier. The calculation method of multi-class cross-entropy is shown in Equation (4): （4） In the formula, Z i The logits function represents the Softmax classifier. n P represents the number of targets detected. i This represents the true sample distribution.

6. The method for detecting oil spills on the sea surface based on a hyperspectral remote sensing 1D-CNN model as described in claim 1, characterized in that: The backpropagation fine-tuning process of the 1D-CNN model adopts the BP algorithm, which, combined with the calibrated sample label dataset, adjusts the weights and biases of the model layer by layer, so that the model's mapping to the sample label dataset achieves the optimal effect.

Images