A method for quantitatively inverting the thickness and emulsification degree of sea surface oil spill based on multi-scale airborne hyperspectral image

Abstract

The application discloses a kind of based on multiscale airborne hyperspectral image sea surface oil spill thickness and emulsification degree quantitative inversion method, basic steps are: using the robust normalization method based on quantile pre-processing input hyperspectral data, weaken the interference of abnormal spectral value to model training;Subsequently, construct spectral grouping embedding module, carry out local feature extraction and sequence reorganization to adjacent band, and utilize multiscale state space network to carry out recursive modeling to spectral feature under different scales, to further capture long-distance dependence in spectral sequence;Finally, realize the quantitative inversion of different thickness non-emulsified crude oil, different thickness and concentration water-in-oil emulsion through regression inversion module.The method proposed in the application can effectively depict the thickness concentration range and volume distribution characteristics of different scale oil spill scene, improve the precision of oil spill parameter quantitative inversion under complex sea surface environment, and is suitable for airborne hyperspectral image sea surface oil spill monitoring.

Description

Technical Field

[0001] This invention relates to hyperspectral image inversion, oil spill thickness inversion, and water-in-oil emulsion concentration inversion, specifically to a quantitative inversion method for oil spill thickness and emulsification degree based on multi-scale airborne hyperspectral images. Background Technology

[0002] Marine oil spills severely damage the marine environment, economic development, and human health. After oil spills enter the ocean, the surface oil slick, influenced by marine environmental dynamics such as wind, waves, and currents, forms an oil film of uneven thickness. Furthermore, the oil film continuously mixes with seawater, forming water-in-oil emulsions and oil-in-water emulsions of varying thicknesses and volume concentrations, with water-in-oil emulsions being more common. After an oil spill, the spill volume is a key indicator for assessing the severity of the incident; oil film thickness and extent can estimate the spill volume. The degree of emulsification of the oil spill emulsion, including its thickness and volume concentration, directly impacts emergency response plans and the implementation of pollution compensation and accountability. Therefore, conducting quantitative inversion of oil film thickness and emulsification degree can provide technical support for marine oil spill emergency response and disaster loss assessment.

[0003] Remote sensing technology boasts a wide coverage and rapid response capability, providing technical support to relevant departments during oil spill monitoring. Commonly used remote sensing methods include microwave remote sensing, thermal infrared remote sensing, and optical remote sensing. Microwave remote sensing can monitor around the clock and in all weather conditions, but it is susceptible to interference from marine environmental factors and oil spill-like substances, resulting in a high false alarm rate. Thermal infrared remote sensing has day and night monitoring capabilities, but its resolution is limited. In contrast, hyperspectral remote sensing technology offers advantages such as high resolution and rich information content, providing diverse spectral curve features. Some studies have made progress in quantizing and inverting hyperspectral data based on deep learning models. With the development of state-space models, network models based on selective state-space mechanisms are increasingly being applied to hyperspectral data processing tasks, demonstrating good performance in modeling long-sequence spectral features.

[0004] Currently, the monitoring of non-emulsified crude oil thickness and oil spill emulsion thickness and concentration is mostly based on inversion using different methods. There is a lack of a joint inversion method for thickness and concentration under different scale oil spill scenarios, and the inversion accuracy is insufficient to meet the requirements for precise inversion. Therefore, this invention proposes a quantitative inversion method for sea surface oil spill thickness and emulsification degree based on multi-scale airborne hyperspectral imagery. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] This invention proposes a quantitative inversion method for the thickness and emulsification degree of oil spills on the sea surface based on multi-scale airborne hyperspectral images. It mainly uses spectral grouping embedding and multi-scale state space network to fully capture the long-distance dependence in the spectral sequence, thereby realizing the inversion of the thickness and concentration of non-emulsified crude oil and water-in-oil emulsions on the sea surface.

[0007] (II) Technical Solution

[0008] This invention includes the following steps:

[0009] (1) Hyperspectral data preprocessing: Based on the acquired airborne hyperspectral oil spill data, a robust normalization method based on quantiles is used to standardize the original spectral data to obtain normalized spectral sample features, which effectively suppresses the influence of abnormal spectral values ​​on model training and improves the stability of feature distribution.

[0010] (2) Spectral sequence reconstruction: In order to extract local spectral structure information, the spectral features are first initially encoded using one-dimensional convolution, and local features are extracted from adjacent bands using a sliding window method. Then, the LeakyReLU activation function is used to enhance the nonlinear expression of spectral features and construct spectral sequence features with continuous spectral band correlation to provide input data for subsequent long-distance dependency modeling.

[0011] (3) Multi-scale spectral feature extraction: The recombined spectral sequence is input into a multi-scale state space network, and a four-layer network is used for feature modeling. Root mean square normalization is applied to the input spectral features, and the feature dimension is expanded by linear mapping, which is used for feature modeling and fusion after modeling. In the network encoding process, spectral features of different resolutions are obtained step by step by three one-dimensional convolution downsampling, and the SiLU activation function is used to enhance the nonlinear expression capability. The activated features are modeled for long distance by the S6 state space module. In the decoding process, in order to fuse information at different scales, the initial resolution of the spectral sequence is restored by layer-by-layer upsampling and skip connections, and the effective features are screened and redundant information is suppressed by using a gating mechanism. Finally, the original dimension of the spectral features is restored, so as to realize the effective extraction of long-distance dependencies in the spectral sequence.

[0012] (4) Quantitative inversion of oil spill parameters: The extracted multi-scale spectral features are input into the regression inversion module, the sequence features are flattened, and nonlinear feature mapping is performed by combining two fully connected networks to realize the quantitative inversion of non-emulsified crude oil of different thicknesses and water-in-oil emulsions of different thicknesses and concentrations. Finally, the corresponding thickness distribution result map and concentration distribution result map are output.

[0013] (III) Beneficial Effects

[0014] The advantages of this invention are as follows:

[0015] Marine oil spill environments are complex, containing non-emulsified crude oil of varying thicknesses and water-in-oil emulsions of varying thicknesses and concentrations, with scale differences existing between different spill scenarios. This invention utilizes a combination of spectral grouping embedding and multi-scale state-space networks to fully extract long-range dependency features from spectral sequences and learn from local spectral features, thereby improving the spectral feature representation capability of monitored objects in complex marine environments and achieving quantitative inversion of the thickness and concentration of non-emulsified crude oil and water-in-oil emulsions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the network architecture for a method to quantify and invert the thickness and emulsification degree of oil spills on the sea surface based on multi-scale airborne hyperspectral imagery.

[0017] Figure 2 A graph showing the inversion results of non-emulsified crude oil thickness compared with other methods using the developed model;

[0018] Figure 3 A graph showing the inversion results of the thickness of water-in-oil emulsions compared with other methods using the developed model;

[0019] Figure 4 The inversion results are shown in a large-scale oil spill scenario.

[0020] Figure 5 This is a diagram showing the inversion results of AVIRIS remote sensing imagery. Detailed Implementation

[0021] To make the objectives, content, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples:

[0022] Referring to Figure 1, the specific implementation steps of the present invention are as follows:

[0023] (1) Hyperspectral data preprocessing

[0024] The data used in this embodiment are remote sensing images of non-emulsified crude oil and water-in-oil emulsions acquired by the UAV-borne Curt S185 hyperspectral imaging system, the UAV-borne GaiaSky-Mini3-VN hyperspectral imaging system, and the AVIRIS hyperspectral imaging system. These include remote sensing data from a 7 cm diameter annular oil spill experiment, a 5*5 m oil spill pool experiment, a 10*10 m oil spill pool experiment, and AVIRIS remote sensing data. The original hyperspectral data contains 126 bands, ranging from 450 nm to 950 nm, with a band interval of 4 nm. First, the input data is preprocessed using a quantile-based robust normalization method. For each spectral band, the upper and lower quantiles are set to 1% and 99%, respectively. The scaling factor is calculated using the difference between the high and low quantiles, and the normalized spectral features are obtained to reduce the interference of abnormal spectral values ​​on model training.

[0025] (2) Spectral sequence recombination

[0026] 1) Input the normalized hyperspectral data into the spectral grouping embedding module. In order to extract the local structural information of the spectrum, first use one-dimensional convolution to encode the spectral features in groups. The input spectral dimension is 126, and the number of output channels is set to 16 after convolution.

[0027] 2) A sliding window approach is used to extract local features from adjacent bands. The window length for spectral grouping is set to 15 and the sliding step size is set to 3, thereby enhancing the correlation expression ability between adjacent spectral bands.

[0028] 3) The extracted local spectral features are nonlinearly mapped using the LeakyReLU activation function to enhance the ability to express the differences between different bands. The negative slope parameter is set to 0.1.

[0029] 4) Finally, the extracted local spectral features are rearranged according to the batch size, sequence length, and feature dimension to obtain a spectral feature sequence with a length of 38 and a feature dimension of 16.

[0030] (3) Multiscale spectral feature extraction

[0031] 1) The recombined spectral sequence is input into a multi-scale state-space network, and a four-layer state-space structure is used to recursively model the spectral features at different scales. First, the spectral sequence with a length of 38 and a feature dimension of 16 is normalized by root mean square, and the feature dimension is expanded by linear mapping to 32. The expanded feature dimension is then divided into two branches, X and Z, along the channel dimension, with each branch having a feature dimension of 16. The X branch is used for subsequent multi-scale state-space feature modeling, and the Z branch will participate in the fusion process after modeling as a gated feature.

[0032] 2) In the four-layer state space structure, X is the first-layer spectral feature X1, with a sequence length of 38 and a feature dimension of 16. Subsequently, features X2, X3, and X4 are generated layer by layer through one-dimensional convolutional downsampling. The downsampling process is implemented using one-dimensional convolution with a kernel size of 3 and a stride of 2. In the four-layer features, the sequence length decreases layer by layer, while the channel dimension increases layer by layer. That is, downsampling X1 yields the second-layer feature X2, with a sequence length of 19 and a feature dimension of 32. Then, downsampling X2 yields the third-layer feature X3, with a sequence length of 10 and a feature dimension of 64. Finally, downsampling X3 yields the fourth-layer feature X4, with a sequence length of 5 and a feature dimension of 128.

[0033] 3) For the four-layer features X1, X2, X3, and X4, local spectral modeling and long-range dependency modeling are performed on the features at each scale in each layer. First, the spectral sequence is captured by one-dimensional convolution to capture local features, and the size of the convolution kernel is set to 3. Then, the SiLU activation function is used to enhance the nonlinear expressive power. The activated features are input into the S6 state space module, and a dynamic step size parameter is introduced to dynamically adjust the state update process according to the input features, thereby effectively capturing long-range spectral dependencies. Finally, the features X1', X2', X3', and X4' at each scale are obtained.

[0034] 4) In the decoding stage, a layer-by-layer upsampling method is used to gradually restore the resolution of the spectral sequence. Skip connections are then used to fuse shallow detail features with deep semantic features to achieve multi-scale spectral information reconstruction. First, the fourth-layer feature X4' is upsampled to restore the spectral resolution, restoring the sequence length to 10. This is then fused element-by-element with the third-layer feature X3'. The fused result is further upsampled to restore the sequence length to 19 and fused with the second-layer feature X2'. After fusion, a third layer of upsampling restores the sequence length to 38, which is then fused with the first-layer feature X1'. Finally, the initial resolution of the spectral sequence is restored layer by layer, resulting in the fused high-resolution features.

[0035] 5) Using the Z-branch divided along the channel dimension, the Z-branch is first used to generate gated features through the SiLU activation function; then the fused high-resolution features are multiplied element-wise with the gated features to achieve effective spectral information enhancement and redundancy information suppression; finally, the original dimension of the spectral features is restored through linear mapping, and the output sequence has a length of 38 and a feature dimension of 16 multi-scale spectral features.

[0036] (4) Quantitative inversion of oil spill parameters

[0037] 1) Input the spectral features extracted by the multi-scale state space network into the regression inversion module to realize the quantitative inversion of the thickness and emulsification degree of the oil spill on the sea surface; First, flatten the output spectral sequence features to obtain a one-dimensional feature vector with a length of 608.

[0038] 2) The flattened one-dimensional feature vector is input into a two-layer fully connected network. The output dimension of the first fully connected layer is set to 60, and the LeakyReLU activation function is used for non-linear mapping. Then, through layer normalization and the second fully connected network, the feature dimension is mapped from 60 to 32, and further mapped to the 1-dimensional inversion result.

[0039] 3) Finally, the quantitative inversion results of non-emulsified crude oil of different thicknesses and water-in-oil emulsions of different thicknesses and concentrations are output, and the corresponding oil spill thickness distribution map and concentration distribution map are generated.

[0040] (5) Evaluation of the accuracy of oil spill thickness inversion at sea surface

[0041] To verify the effectiveness of the proposed method (HSQR-Mamba), this invention employs a Deep Belief Network (DBN), a Convolutional Neural Network combined with a Gated Recurrent Unit (CNN_GRU), and a Transformer model for comparative experiments. Accuracy is evaluated using preset thickness values ​​for non-emulsified crude oil and water-in-oil emulsion annexes of varying thicknesses as benchmarks. The thickness inversion results for non-emulsified crude oil and water-in-oil emulsions obtained by different models are shown below. Figure 2 and Figure 3 As shown, the inversion accuracy is detailed in Tables 1 and 2.

[0042] Table 1. Accuracy of thickness inversion for non-emulsified crude oil using different models

[0043]

[0044] Table 2. Accuracy of different models in thickness inversion for water-in-oil emulsions

[0045]

Claims

1. A method for quantitative inversion of sea surface oil spill thickness and emulsification degree based on multi-scale airborne hyperspectral imagery, characterized in that, Includes the following steps: (1) Hyperspectral data preprocessing: Based on the acquired airborne hyperspectral oil spill data, a robust normalization method based on quantiles is used to standardize the original spectral data to obtain normalized spectral sample features, which effectively suppresses the influence of abnormal spectral values ​​on model training and improves the stability of feature distribution. (2) Spectral sequence reconstruction: In order to extract local spectral structure information, the spectral features are first initially encoded using one-dimensional convolution, and local features are extracted from adjacent bands using a sliding window method. Then, the LeakyReLU activation function is used to enhance the nonlinear expression of the spectral features and construct spectral sequence features with continuous spectral band correlation to provide input data for subsequent long-distance dependency modeling. (3) Multi-scale spectral feature extraction: The recombined spectral sequence is input into a multi-scale state space network, and a four-layer network is used to carry out feature modeling. Root mean square normalization is applied to the input spectral features, and the feature dimension is expanded by linear mapping, which is used for feature modeling and the fusion process after modeling, respectively. In the network encoding process, spectral features of different resolutions are obtained step by step by three one-dimensional convolution downsampling, and the SiLU activation function is used to enhance the nonlinear expression capability. The activated features are then modeled over a long distance by the S6 state space module. In the decoding process, in order to fuse information at different scales, the initial resolution of the spectral sequence is restored by layer-by-layer upsampling and skip connections. Then, a gating mechanism is used to filter effective features and suppress redundant information. Finally, the original dimension of the spectral features is restored, so as to effectively extract long-distance dependencies in the spectral sequence. (4) Quantitative inversion of oil spill parameters: The extracted multi-scale spectral features are input into the regression inversion module, the sequence features are flattened, and nonlinear feature mapping is performed by combining two fully connected networks to realize the quantitative inversion of non-emulsified crude oil of different thicknesses and water-in-oil emulsions of different thicknesses and concentrations. Finally, the corresponding thickness distribution result map and concentration distribution result map are output.

2. The method according to claim 1, characterized in that, In step (1), the data standardization process adopts a robust normalization method based on quantiles. During normalization, the upper and lower quantiles are set to 1% and 99% respectively, and the number of bands in the original hyperspectral data is 126.

3. The method according to claim 1, characterized in that, In step (2), one-dimensional convolution is used to encode local features of the spectral sequence. The number of convolution output channels is set to 16, the length of the spectral grouping window is set to 15, and the sliding step size is set to 3. The LeakyReLU activation function is used for nonlinear mapping, and the negative slope parameter is set to 0.1 to obtain a spectral feature sequence with a length of 38 and a feature dimension of 16.

4. The method according to claim 1, characterized in that, The multi-scale state space network in step (3) uses a four-layer coding structure to model the spectral features at multiple scales, where the sequence length of the first layer features is 38 and the feature dimension is 16. The second layer of features has a sequence length of 19 and a feature dimension of 32; the third layer of features has a sequence length of 10 and a feature dimension of 64; and the fourth layer of features has a sequence length of 5 and a feature dimension of 128.

5. The method according to claim 1, characterized in that, The regression inversion module in step (4) includes a multi-layer fully connected network. First, the spectral sequence features are flattened to obtain a one-dimensional feature vector with a length of 608. The output dimension of the first fully connected layer is set to 60, and the LeakyReLU activation function is used for nonlinear mapping. The output dimension of the second fully connected layer is set to 32, and finally a one-dimensional thickness or concentration inversion result is obtained through the linear output layer.

6. The method according to claim 1, characterized in that, One-dimensional convolution is used to downsample multi-scale features, with the kernel size set to 3 and the stride set to 2.

7. The method according to claim 1, characterized in that, The AdamW optimizer is used for parameter optimization during network training.

8. The method according to claim 1, characterized in that, SmoothL1Loss is used as the loss function.

9. The method according to claim 1, characterized in that, The learning rate is 0.

001.

10. The method according to claim 1, characterized in that, The batch training size is 128.

Images