A SAR ocean oil spill intelligent identification method and system based on a diffusion model

Abstract

The application discloses a SAR marine oil spill intelligent identification method and system based on a diffusion model, which converts oil spill detection into a conditional generation task, comprising: performing geometric correction, radiation calibration, noise suppression and normalization on a SAR image; adopting a feature confidence modeling module CEM to enhance the reliability of features through spatial and channel double confidence calculation; constructing a conditional diffusion model to guide by taking the SAR image as a condition, and iteratively denoising to generate a segmentation mask by using a noise estimation network of a U-Net architecture; performing five independent generations, and improving stability by pixel-level average fusion and uncertainty quantification; extracting oil spill area parameters through morphological processing, and outputting in a standard geographic information format. The system first applies a conditional denoising diffusion probability model to the marine remote sensing field, realizes pixel-level uncertainty estimation, is significantly superior to existing methods, and provides an intelligent identification scheme with high precision and high reliability for oil spill emergency response.

Description

Technical Field

[0001] This invention belongs to the field of marine environmental monitoring technology, specifically relating to a SAR-based intelligent identification method and system for marine oil spills based on a diffusion model. Background Technology

[0002] The ocean is the largest ecosystem on Earth, crucial for maintaining global biodiversity, regulating climate, and preserving ecological balance. However, with increased maritime transport and marine development activities, oil spills are frequent, posing a serious threat to the marine ecosystem. The Deepwater Horizon drilling platform explosion in the Gulf of Mexico in April 2010 caused a massive oil spill covering an area of ​​approximately 160 kilometers by 72 kilometers, resulting in long-term damage to the marine ecosystem. Timely and accurate detection of oil spill areas is of great importance for environmental protection and emergency response.

[0003] Synthetic Aperture Radar (SAR) remote sensing technology, due to its all-weather, all-time imaging capabilities, is not limited by meteorological conditions or lighting, and can penetrate cloud layers, has become a primary technical means for monitoring marine oil spills. SAR acquires surface information by actively transmitting microwave signals and receiving backscattered echoes from ground features. In SAR images, oil spill areas appear as dark areas because the oil slick suppresses shortwaves from the sea surface, resulting in reduced backscattering.

[0004] Early research primarily employed traditional image processing methods, such as thresholding, texture analysis, edge detection, and region growing, as well as machine learning methods like support vector machines and artificial neural networks. These methods heavily rely on manually designed features and manual parameter tuning, resulting in poor adaptability and unstable performance under varying sea conditions.

[0005] In recent years, deep learning methods have made significant progress in SAR oil spill detection. Semantic segmentation models based on fully convolutional neural networks, U-Net, DeepLabV3+, and Transformer-based methods can automatically learn feature representations. However, these discriminative models essentially perform deterministic mappings and cannot generatively reconstruct weak textures in SAR images. When faced with complex scenes with low signal-to-noise ratios and high uncertainties, they are prone to overfitting and insufficient generalization.

[0006] Diffusion probabilistic models, as a new generation of generative artificial intelligence technology, have demonstrated powerful capabilities in computer vision tasks such as image generation, restoration, and segmentation. Unlike traditional discriminative models, diffusion models, by modeling the implicit probability distribution of data, can generate high-quality, diverse results and naturally provide confidence information, i.e., uncertainty estimation. Conditional diffusion models introduce task-specific conditional information at each diffusion step, achieving controllable generation, and their effectiveness has been proven in fields such as medical image segmentation. However, no research has yet applied conditional diffusion models to SAR oil spill detection tasks. Summary of the Invention

[0007] This invention proposes a SAR-based intelligent identification method and system for marine oil spills, transforming the oil spill detection task into a condition generation task. It achieves high-precision and reliable oil spill area detection through generative artificial intelligence technology. The specific technical solution is as follows:

[0008] A SAR-based intelligent identification method for marine oil spills, based on a diffusion model, performs the following steps sequentially:

[0009] S1: Perform geometric correction, radiometric calibration, noise suppression, and data normalization on the acquired marine environment SAR remote sensing images to obtain the preprocessed SAR image S;

[0010] S2: Input the SAR image S into the feature confidence modeling module, calculate the spatial confidence level through the spatial confidence modeling unit, calculate the channel confidence level through the channel confidence modeling unit, and perform feature fusion to obtain high-confidence features. ,in Represents the real number field, where C is the number of channels, H is the height, and W is the width;

[0011] S3: The high-confidence features The SAR image S is used as a conditional input to the noise estimation network for diffusion. The diffusion process initializes and generates a noisy image. For time step t, iterate from T to 1: Construct conditional input Call NEU to predict noise Calculated according to the diffusion backsampling formula After T iterations, the generated segmentation mask is obtained. ;

[0012] S4: Repeat step S3 n times, using different initial noises, to obtain n independently generated segmentation masks. (1), (2), ..., (n)}, where (i) represents the final mask obtained by the i-th generation inference. The pixel-level average fusion probability P is calculated, and the final binarized mask M is obtained by using the threshold τ.

[0013] S5: Perform morphological operations and connected component analysis on the final binarized mask M to extract the geometric parameters, location information and confidence information of the oil spill area, and output it in a standard geographic information format.

[0014] A SAR marine oil spill intelligent identification system based on a diffusion model includes:

[0015] The preprocessing module performs geometric correction, radiometric calibration, noise suppression, and data normalization on the acquired marine environment SAR remote sensing images to obtain the preprocessed SAR image S.

[0016] The feature confidence modeling module inputs the SAR image S, calculates spatial confidence through the spatial confidence modeling unit, calculates channel confidence through the channel confidence modeling unit, and performs feature fusion to obtain high-confidence features. ,in Represents the real number field, where C is the number of channels, H is the height, and W is the width;

[0017] The conditional diffusion generation inference module generates the high-confidence features. The SAR image S is used as a conditional input to the noise estimation network for diffusion. The diffusion process initializes and generates a noisy image. For time step t, iterate from T to 1: Construct conditional input Call NEU to predict noise Calculated according to the diffusion backsampling formula After T iterations, the generated segmentation mask is obtained. ;

[0018] The fusion module repeats the diffusion process n times, using different initial noises, to obtain n independently generated segmentation masks. (1), (2), ..., (n)}, where (i) indicates the first The final mask obtained by secondary generation inference is used to calculate the pixel-level average fusion probability P, and the final binarized mask M is obtained by using the threshold τ.

[0019] The results output module performs morphological operations and connected component analysis on the final binary mask M to extract the geometric parameters, location information, and confidence information of the oil spill area, and outputs them in a standard geographic information format.

[0020] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0021] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0022] The present invention has the following beneficial effects:

[0023] 1. This invention innovatively applies a conditional diffusion model to SAR marine oil spill detection. Addressing the unique characteristics of SAR images (strong speckle noise, low signal-to-noise ratio, and blurred oil spill boundaries), this invention introduces a conditional denoising diffusion probability model into the oil spill detection task. By transforming the detection problem into a conditional generation problem, SAR images are used as conditions to guide each diffusion step, achieving controllable and accurate generation of the oil spill area. Experimental results show that on the PALSAR and SENTINEL datasets, the Dice coefficients of this invention reach 0.764 and 0.766, respectively, significantly outperforming eight existing methods, including the deterministic mapping methods of traditional discriminative models such as FCN (0.723 / 0.741), U-Net (0.716 / 0.733), and Swin-Unet (0.716 / 0.755).

[0024] 2. Dual-Reset Confidence Enhancement Mechanism Improves Feature Reliability. The Feature Confidence Modeling Module (CEM) designed in this invention includes a spatial confidence modeling unit and a channel confidence modeling unit, which solves the problem of low feature confidence in SAR images through an adaptive weight allocation mechanism. In the spatial dimension, attention weights are calculated using average pooling and max pooling to highlight the target region and suppress background noise; in the channel dimension, channel weights are calculated using global average pooling and multilayer perceptron to strengthen key feature channels and weaken noisy channels, significantly improving feature reliability. Ablation experiments show that, compared with using the RU module, using the CEM module improves the Dice coefficient from 0.725 / 0.726 to 0.764 / 0.766, and the accuracy from 0.911 / 0.783 to 0.917 / 0.848, verifying the effectiveness of the dual-reset confidence enhancement mechanism.

[0025] 3. Multiple generation fusion strategy improves detection stability. This invention generates n segmentation masks by performing n independent diffusion inferences and using different initial noises, and calculates the pixel-level average fusion probability to eliminate the uncertainty of a single random sampling. Experimental verification: (1) In five batches of PALSAR and SENTINEL data, the five fusion results (e.g., 0.832, 0.864) are all higher than the average of the five generation (0.809, 0.811), which is about 2-3 percentage points higher; (2) Compared with the experiments from n=1 to n=10, n=5 achieves the best balance in multiple indicators of the two datasets, with the Dice coefficient reaching 0.764 / 0.766. Further increasing the n value yields diminishing returns; (3) Using a parallel acceleration strategy, the processing time for five generations is about 8 seconds, which is reasonable compared to 1.5 seconds for a single generation.

[0026] 4. Deep integration of the U-Net architecture and CEM module. CEM is used for feature extraction in the encoder stage of the Noise Estimation Network (NEU), and residual units (RUs) are used for feature recovery in the decoder stage. Combined with skip connections and temporal step embedding, this achieves an organic unity of multi-scale feature extraction and noise prediction. The 4-layer encoder-decoder structure effectively balances feature representation capability and computational efficiency. Ablation experiments show that removing the Conditional Diffusion Model (CDM) for direct segmentation reduces the Dice coefficient to 0.713 / 0.712, validating the necessity of the diffusion generation strategy.

[0027] 5. Provides a complete mathematical framework and implementation method. This invention elaborates on the complete mathematical derivations of the forward diffusion process, the reverse denoising process, conditional diffusion expansion, noise variance scheduling (linear scheduling), and diffusion backsampling formula, ensuring that those skilled in the art can implement this invention and meeting the requirements for full patent disclosure.

[0028] 6. Improved boundary positioning accuracy. This invention performs well in boundary recognition: (1) On the SENTINEL dataset, the HD95 index is 10.73 pixels, which is better than the benchmark methods (FCN 11.17, U-Net 11.05, Transunet 11.94) and second only to SegNet's 10.76; (2) U-Net skip connections preserve detailed information, generative reconstruction enhances weak boundary features, and multiple fusions optimize boundary positioning; (3) On the precision index, the PALSAR dataset reaches 0.816, which is significantly better than all the comparison methods (highest 0.790), indicating a low false alarm rate and the ability to effectively distinguish between oil spills and oil spill-like phenomena (low wind areas, biofilms, etc.).

[0029] 7. Provide an end-to-end system-level solution. This invention provides a complete process from input to output: (1) The preprocessing module performs geometric correction (terrain correction, slant range to ground distance conversion), radiometric calibration, noise suppression (Lee / Frost / Sigma filtering) and normalization; (2) The feature confidence modeling module extracts high-confidence features. (3) The conditional diffusion generation inference module generates a segmentation mask; (4) The fusion module calculates the pixel-level average and applies a threshold τ; (5) The result output module extracts geometric parameters and location information, and outputs them in GeoJSON, Shapefile or GeoTIFF format. The modular design of the system facilitates expansion and deployment.

[0030] 8. Computational Efficiency and Practicality. The computational performance of this invention meets the requirements of practical applications: (1) It takes about 1.5 seconds to generate and process a 512×512 pixel SAR image in a single batch, and about 8 seconds for parallel processing of 5 batches; (2) The peak memory for inference is about 3.8GB, supporting conventional GPU hardware; (3) The number of parameters in the noise estimation network NEU is about 45M, and the model size is moderate; (4) It supports large image block processing, batch processing acceleration, and multi-GPU distributed computing. A good balance is achieved between accuracy and efficiency.

[0031] 9. Strong generalization ability. The present invention performs stably on different datasets: (1) The PALSAR and SENTINEL datasets are from different sensors (ALOS-2 and Sentinel-1) and have different imaging parameters and noise characteristics; (2) The Dice coefficients on both datasets exceed 0.76, and the accuracies are 0.917 and 0.848, respectively, demonstrating good generalization ability; (3) Generative modeling learns the deep structure distribution rather than the surface texture, the CEM module adaptively adjusts the feature weights, and the probabilistic modeling method naturally adapts to changes in data distribution, and can cope with different sea conditions and imaging conditions. Attached Figure Description

[0032] Figure 1 This is a flowchart of the present invention;

[0033] Figure 2 This is a basic process diagram of the conditional diffusion model;

[0034] Figure 3 The framework diagram for the fusion strategy is generated multiple times; Noise 1-n represents the noise to be added, Pre-trained Conditional Diffusion Model represents the pre-trained conditional diffusion model, Segmentation 1-n represents the N segmentation results obtained, mean represents taking the average value, and Thresholding represents thresholding. Final Segmentation represents the result after thresholding.

[0035] Figure 4 This is the CEM framework diagram. Conv2d represents a convolutional neural network, linear represents a linear layer, spa-confi represents a spatial confidence enhancement module, Cha-confi represents a channel confidence enhancement module, BatchNorm represents batch normalization, SiLU represents the SiLU activation function, Position Embedding represents positional encoding embedding, and Diffusion Step represents the diffusion embedding part, used to indicate the current noise intensity and assist the network in predicting the current noise. Input represents the input, and Output represents the output.

[0036] Figure 5 This is a framework diagram of the residual unit (RU); Conv2d represents a convolutional neural network, linear represents a linear layer, BatchNorm represents batch normalization, SiLU represents the SiLU activation function, Position Embedding represents position encoding embedding, Diffusion Step represents the diffusion step embedding part, which is used to indicate the current noise intensity and assist the network in predicting the current noise; Input represents the input, and Output represents the output.

[0037] Figure 6 This is the NEU framework diagram; CEM represents the feature confidence modeling module; DS represents the downsampling module, which consists of a convolutional neural network, batch normalization, and the SiLU activation function. US represents the upsampling module, which consists of a convolutional neural network, batch normalization, and the SiLU activation function. Predictive noise represents the predicted noise, mask represents the mask, SAR image represents the input SAR image, and Noise represents noise.

[0038] Figure 7 To generate the relationship between the number of fusions n and the mask Dice score;

[0039] Figure 8 Visualize and compare segmentation masks. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0041] like Figure 1 As shown, this invention provides a SAR-based intelligent identification method for marine oil spills using a diffusion model. The complete workflow begins with the input and preprocessing of SAR images and ends with the output of an oil spill area mask and related geographic information. The specific principle of its core generation step S3 (conditional diffusion generation inference) is explained by... Figure 2 Detailed demonstration. Includes the following steps:

[0042] Step S1: SAR image preprocessing. Acquire SAR remote sensing images of the marine environment, perform geometric correction, radiometric calibration, noise suppression, and data normalization to obtain the preprocessed SAR image S.

[0043] Step S2: Feature confidence modeling. The preprocessed SAR image S is input into the feature confidence modeling module, and the spatial confidence level is calculated by the spatial confidence modeling unit. Channel confidence is calculated using the channel confidence modeling unit. High-confidence features are obtained by performing feature fusion. ,in Representing the real number field, C is the number of channels, H is the height, and W is the width; the above process is as follows: Figure 4 As shown, this module receives the input feature map and feeds it into both the Spatial Confidence Enhancement Module (spa-confi) and the Channel Confidence Enhancement Module (cha-confi). The spatial module generates spatial attention weights by aggregating contextual information, while the channel module generates channel attention weights by analyzing the relationships between channels. These two weights are adaptively fused with the original features and combined with diffusion step embedding information to finally output enhanced high-confidence features: the spatial confidence enhancement module calculates spatial attention weights through average pooling and max pooling of the feature map, followed by convolutional layers, highlighting the target region and suppressing background noise; the channel confidence enhancement module calculates channel attention weights through global average pooling and fully connected layers, strengthening key feature channels and weakening noisy channels. The outputs of the two modules are weighted and fused, then combined with the original features, diffusion step embedding, and other information to finally output high-confidence enhanced features. Figure 4 In this context, Conv2d represents a convolutional layer, Linear represents a linear layer, BatchNorm represents a batch normalization layer, and SiLU represents an activation function.

[0044] Step S3: Conditional diffusion generation inference, using high-confidence features F and SAR image S as conditional input to the noise estimation network NEU. The diffusion process initializes and generates a noisy image. For time step t, iterate from T to 1: Construct conditional input Call NEU to predict noise Calculated according to the diffusion backsampling formula After T iterations, the generated segmentation mask is obtained. This step is based on a conditional denoising diffusion probability model, and its core principles of forward denoising and inverse denoising are as follows: Figure 2As shown, the forward diffusion process on the left illustrates how noise is gradually added to the clear ground truth mask (X0) as the time step (t) increases, eventually transforming it into pure Gaussian noise (X_T). The reverse denoising process on the right shows how, starting with noise X_T, guided by a SAR image, noise is predicted and gradually removed through a noise estimation network (NEU), ultimately reconstructing the predicted segmentation mask (X0). The conditional arrows emphasize the crucial role of the SAR image in the entire denoising process.

[0045] Step S4: Generate consistent fusion multiple times, repeating step S3 for a total of n times, using different initial noises to obtain n independently generated segmentation masks. (1), (2), ..., (n)}, where (i) represents the final mask obtained from the i-th generation inference, calculates the pixel-level average fusion probability P, and applies a threshold τ to obtain the final binarized mask M; the above process is as follows Figure 3 As shown, this framework demonstrates how, to improve the stability of the results, it employs a parallel or serial approach, using n different initial noise sets (Noise 1-n) as inputs to a pre-trained Conditional Diffusion Model, independently performing n inverse denoising inferences to obtain n potentially different segmentation results (Segmentation 1-n). Subsequently, these results are averaged at the pixel level to obtain a probability map. Finally, thresholding transforms the probability map into a final binary segmentation mask (FinalSegmentation). This strategy effectively eliminates the uncertainty introduced by a single random sampling.

[0046] Step S5: Output the oil spill detection results. Perform post-processing on the final mask M, including morphological operations and connected component analysis, to extract the geometric parameters, location information, and confidence information of the oil spill area. Output the results in GeoJSON, Shapefile, or GeoTIFF format.

[0047] The feature confidence modeling module in step S2, i.e., CEM, is as follows: Figure 4As shown. It includes spatial confidence modeling units and channel confidence modeling units. The implementation method of the spatial confidence modeling unit is as follows:

[0048] Average pooling and max pooling are performed on the input feature x to obtain and ;

[0049] Will and After concatenation along the channel dimension, spatial confidence is calculated using a two-layer convolutional network:

[0050] ;

[0051] in, , This is the weight matrix. , As the bias vector, a convolution operation is performed on the original feature x to obtain supplementary features. ;

[0052] Weighted fusion based on spatial confidence:

[0053] ;

[0054] in, This indicates multiplication of the primary element. This enhances spatial features.

[0055] The implementation method of the channel confidence modeling unit is as follows:

[0056] Spatial enhancement features Global average pooling is performed to obtain ;

[0057] Predicting channel confidence using a multilayer perceptron:

[0058] ;

[0059] , This is the weight matrix. , For the bias vector, Perform convolution operations to obtain supplementary features ;

[0060] We obtain the following by weighted fusion based on channel confidence:

[0061] ;

[0062] in, To output high-confidence features F= using weighted fusion, .

[0063] The conditional diffusion generation inference in step S3 uses the following diffusion backsampling formula:

[0064] ;

[0065] in, To control the noise intensity added at diffusion step t, for arrive The product of consecutive products, Noise predicted for the NEU network. Sampling is performed from a standard normal distribution N(0,I) (where I is the identity matrix, indicating that each pixel is sampled independently). This is the variance parameter.

[0066] The noise variance scheduling employs linear scheduling, resulting in a gradual increase in noise intensity, which in turn correlates with... Related to (noise variance scheduling parameters);

[0067] The multiple generation consistency fusion in step S4 includes:

[0068] Pixel-level averaging fusion: uncertainty quantification: confidence leveling: threshold decision.

[0069] The SAR image preprocessing in step S1 includes:

[0070] Supports multi-source SAR data access: Sentinel-1, GF-3, ALOS-2 PALSAR satellite data, and supports HH, HV, VV, VH multi-polarization modes;

[0071] Geometric correction includes terrain correction and slope range to ground distance conversion;

[0072] Radiation calibration uses the following formula:

[0073] ;

[0074] in, It is the backscattering coefficient, and DN is the digital quantization value, which is the raw pixel value output by the sensor; This represents the noise value; where K is a scaling constant.

[0075] Noise suppression employs Lee filtering, Frost filtering, or a modified Sigma filter;

[0076] Data normalization uses the following formula:

[0077] ;

[0078] Where S is the original pixel value of the SAR image; The minimum pixel value; Maximum pixel value; The normalized pixel value, ranging from [0,1].

[0079] The noise estimation network NEU in step S3 adopts a U-Net architecture, and its overall structure diagram is shown below. Figure 6 As shown, the network takes a noisy mask and a SAR image as input. The encoder path contains multiple down-sampling modules (DS), each integrating a CEM module for feature enhancement. The decoder path contains multiple up-sampling modules (US), each using an RU module for feature recovery. Skip connections pass multi-scale features from the encoder to the decoder. The bottleneck layer fuses temporal information. The network's final output is predicted noise, including:

[0080] Encoder: 4-layer downsampling path, each layer uses the feature confidence modeling module CEM for feature extraction, and the feature map size is downsampled from H×W to H / 16×W / 16; the structure of the feature confidence modeling module CEM is the same as that of the feature confidence modeling module CEM in S2.

[0081] Bottleneck layer: Located at the deepest layer of the network, it incorporates time step embeddings;

[0082] Decoder: 4-layer upsampling path, each layer uses Residual Units (RUs) for feature extraction; the structure of the RU module is as follows... Figure 5 As shown, the input features are added and fused with the corresponding layer features from the encoder (via skip connections), position embeddings, and diffusion step embeddings. The fused features are then processed through a residual block consisting of a convolutional layer (Conv2d), batch normalization (BatchNorm), and the SiLU activation function (SiLU), ultimately outputting refined features for upsampling. This design facilitates feature reconstruction and gradient flow. After receiving the feature map from the previous layer, the model adds it to the position encoding information and diffusion step information from the linear layer, and then outputs the upsampled result after passing through a convolutional neural network with residuals.

[0083] Skip connections: connect the corresponding layers of the encoder and decoder, preserving detailed information;

[0084] Time step embedding uses sinusoidal position coding:

[0085] ;

[0086] ;

[0087] In both formulas, d represents the total number of image token embedding dimensions; i represents the index of the current image token embedding dimension; and t represents the position of the image token on the image.

[0088] In step S4, the number of generation times n is set to 5, which, according to experimental verification, achieves the best balance between accuracy and efficiency.

[0089] The method employs a parallel acceleration strategy, allowing multiple generation tasks to be executed in parallel. This enables the simultaneous processing of multiple generation tasks on the GPU, making the total time for n=5 close to the time for n=1.

[0090] This invention provides an intelligent marine oil spill detection system based on a conditional diffusion model using SAR images, comprising the following modules:

[0091] SAR image preprocessing module: responsible for receiving and preprocessing multi-source SAR data;

[0092] Feature confidence modeling module: Improves feature reliability through dual-reset confidence modeling of spatial and channel dimensions;

[0093] Noise estimation network: A U-Net architecture is used to predict noise during the diffusion process;

[0094] Conditional diffusion generation inference module: Guided by SAR images, it generates a segmentation mask through a diffusion model;

[0095] Multiple Generation Consistency Fusion Module: Improves result stability through multiple independent generation and probabilistic fusion;

[0096] Oil spill detection result output module: provides structured detection results and uncertainty quantification.

[0097] The modules work together to form a complete processing chain from SAR image input to oil spill mask output.

[0098] A smart marine oil spill detection system based on a conditional diffusion model using SAR images, comprising:

[0099] The SAR image preprocessing module is used to receive multi-source SAR data and perform geometric correction, radiometric calibration, noise suppression, and data normalization.

[0100] The feature confidence modeling module includes a diffusion step embedding unit, a basic feature extraction unit, a spatial confidence modeling unit, and a channel confidence modeling unit, which are used to improve the reliability of SAR features.

[0101] The noise estimation network adopts a U-Net encoder-decoder architecture, using a feature confidence modeling module in the encoder stage and a residual unit in the decoder stage, with skip connections and time step embeddings, to predict noise during the diffusion process;

[0102] The conditional diffusion generation inference module uses SAR images as conditions to generate a segmentation mask from pure noise through an iterative denoising process.

[0103] The consistency fusion module is generated multiple times, and n independent diffusion inferences are performed. The final mask is obtained through pixel-level average fusion, uncertainty quantization and threshold decision.

[0104] The oil spill detection result output module performs post-processing on the final mask, extracts geometric parameters, location information, and confidence information, and outputs the results in a standard geographic information format.

[0105] In the feature confidence modeling module:

[0106] The spatial confidence modeling unit calculates spatial confidence through a convolutional network. The formula is:

[0107] ;

[0108] The channel confidence modeling unit calculates channel confidence using a multilayer perceptron. The formula is:

[0109] ;

[0110] Feature fusion is achieved through weighted fusion:

[0111] ;

[0112] ;

[0113] A smart marine oil spill detection device based on a conditional diffusion model using SAR images, comprising:

[0114] processor;

[0115] Memory, which stores computer programs;

[0116] When the computer program is executed by the processor, it implements the intelligent detection method for marine oil spills in SAR images.

[0117] The device further includes:

[0118] A communication interface for receiving SAR data from satellite, airborne, or shipboard platforms;

[0119] A display screen is used to visually show the test results;

[0120] The model parameters stored in the memory include:

[0121] The weight parameters of the noise estimation network NEU have a total of approximately 45M parameters.

[0122] Noise variance scheduling parameters ;

[0123] Hyperparameters such as the number of generation times n and the threshold τ.

[0124] The device is equipped with a GPU acceleration unit and supports:

[0125] The processing time for a single generation is approximately 1.5 seconds;

[0126] The parallel processing time for 5 generation iterations is approximately 8 seconds.

[0127] Peak inference memory usage is approximately 3.8GB;

[0128] Supports real-time processing of 512×512 pixel SAR images.

[0129] To verify the effectiveness of the present invention, the following four evaluation indicators were used for performance evaluation: (1) Dice coefficient (Dice Score): measures the degree of overlap between the predicted mask and the real mask, and the calculation formula is as follows: , where A is the predicted region and B is the real region, with a value range of [0,1]. The larger the value, the higher the segmentation accuracy; (2) Hausdorff Distance 95th percentile (HD95): measures the distance error between the predicted boundary and the real boundary, in pixels. The 95th percentile is used to exclude extreme outliers. The smaller the value, the more accurate the boundary localization; (3) Precision: the proportion of pixels predicted as oil spill that are actually oil spill. The calculation formula is TP represents true positives and FP represents false positives, with a value range of [0,1]. The larger the value, the lower the false alarm rate. (4) Accuracy: The proportion of correctly predicted pixels among all pixels, calculated using the formula: Where TN represents a true negative and FN represents a false negative, with values ​​ranging from [0,1]. Higher values ​​indicate higher overall accuracy. The OSDDiff method of this invention is compared with eight other similar methods on four metrics: Dice Score, HD95, Precision, and Accuracy. The results are shown in the table below. Figure 8 As shown in the figure, bold indicates the best value, underline indicates the second best value, and in the metrics, ↑ indicates that the higher the value of the metric, the better the model performance, and ↓ indicates that the lower the value of the metric, the better the model performance. PALS represents the PALSAR dataset, and SENT represents the SENTINEL dataset. Although it did not achieve the best scores on the HD95 evaluation metric of the PALSAR dataset and the Precision evaluation metric of SENTINEL, it achieved the best results compared to other methods on the other metrics. OSDDiff achieved the best overall performance on both the PALSAR and SENTINEL datasets, especially significantly outperforming existing methods in Dice coefficient and Accuracy. Although some methods have slight advantages on individual metrics (such as HD95 on the PALSAR dataset of SegNet and Precision on the SENTINEL dataset of Att-Unet), none of them can comprehensively outperform OSDDiff on multiple core metrics. The results verify the effectiveness of the architecture adopted by OSDDiff in improving segmentation robustness and generalization ability.

[0130]

[0131] The following experiments further validated the effectiveness of the multiple iteration strategy. Five batches (batch 1-5) of SAR images were selected from the PALSAR and SENTINEL datasets to verify whether the mean fusion accuracy improved compared to the Dice score of each mask generation. The results are shown in the table below, with bold data representing the best accuracy. Compared to the mean Dice score of the five mask generation iterations, the strategy of fusing five masks consistently achieved higher scores, demonstrating the effectiveness of the multiple generation fusion strategy in improving the accuracy of the segmentation mask.

[0132]

[0133] The following experiments will conduct 1 to 10 generation fusion strategy experiments on two datasets respectively to find the optimal number of generation attempts. The experimental results are shown in the table below. Bold text indicates the optimal value, underlined text indicates the second-best value, and in the metrics, ↑ indicates a higher value and better model performance, while ↓ indicates a lower value and better model performance. From the experimental results, it can be seen that in... The OSDDiff model achieved higher scores on more metrics across both datasets.

[0134]

[0135] The following ablation experiments were conducted on the confidence enhancement module CEM, and the results are shown in the table below. Specifically, the CEM module used in the encoder stage was replaced with the RU module used in the decoder stage, also employing a 5-pass generation fusion strategy. This indicates that the module is not used. This indicates the use of the module. Experimental results show that: (1) Among the 8 indicators, the CEM module achieves better results in 6 indicators, including the Dice coefficient (improved by 5.4% / 5.5%), Accuracy (improved by 0.7% / 8.3%), and Precision (improved by 29.1%) of the PALSAR dataset; (2) The configuration without CEM has a slight advantage in HD95 (9.01 vs 9.19) of the PALSAR dataset and Precision (0.746 vs 0.671) of the SENTINEL dataset; (3) Considering all factors, the CEM module brings significant improvements to the core evaluation indicators Dice and Accuracy, and its contribution to the overall performance is positive, verifying the effectiveness of the double-reset reliability enhancement mechanism.

[0136]

[0137] Ablation experiments were conducted on the Conditional Diffusion Model (CDM), and the results are shown in the table below. SARimages represents the original SAR image, Ground truth represents the true result, and the rest are the prediction results of the comparison model. Specifically, the same network model as OSDDiff was used directly to segment the SAR image, without using the generation strategy of the Conditional Diffusion Model. The experimental results show that directly using this model to segment SAR images cannot cope with their high noise, low resolution, and blurred boundaries, resulting in low segmentation performance. It also demonstrates the effectiveness of the CDM generation strategy; in addition to the HD95 score on the PALSAR dataset, the iterative denoising and generation strategy through the reverse process yields a segmentation mask with higher accuracy across all evaluation metrics.

[0138]

[0139] Furthermore, the impact of the optimal number of generation fusions, n, on the model was explored. The experiment consisted of two parts. The first part used three batches of data to test the Dice scores for fusions from 1 to 99, with the Dice scores of the segmentation mask varying as follows: Figure 7As shown in the figure, it is clear that the three batches of data follow the same trend, that is, the Dice score tends to stabilize afterward, so the optimal value should be within 10.

Claims

1. A SAR-based intelligent identification method for marine oil spills using a diffusion model, characterized in that, Perform the following steps in sequence: S1: Perform geometric correction, radiometric calibration, noise suppression, and data normalization on the acquired marine environment SAR remote sensing images to obtain the preprocessed SAR image S; S2: Input the SAR image S into the feature confidence modeling module, calculate the spatial confidence level through the spatial confidence modeling unit, calculate the channel confidence level through the channel confidence modeling unit, and perform feature fusion to obtain high-confidence features. ,in Represents the real number field, where C is the number of channels, H is the height, and W is the width; S3: The high-confidence features The SAR image S is used as a conditional input to the noise estimation network for diffusion. The diffusion process initializes and generates a noisy image. For time step t, iterate from T to 1: Construct conditional input The noise estimation network NEU is invoked to predict noise. Calculated according to the diffusion backsampling formula After T iterations, the generated segmentation mask is obtained. ; S4: Repeat step S3 n times, using different initial noises, to obtain n independently generated segmentation masks. (1), (2), ..., (n)}, where (i) represents the final mask obtained by the i-th generation inference. The pixel-level average fusion probability P is calculated, and the final binarized mask M is obtained by using the threshold τ. S5: Perform morphological operations and connected component analysis on the final binarized mask M to extract the geometric parameters, location information and confidence information of the oil spill area, and output it in a standard geographic information format.

2. The method according to claim 1, characterized in that, In step S1, geometric correction includes terrain correction and slant range-to-ground distance conversion; radiometric calibration uses the formula: ； in, It is the backscattering coefficient. It is a digital quantization value, which is the raw pixel value output by the sensor; This represents the noise value; where K is a scaling constant. Noise suppression employs Lee filtering, Frost filtering, or a modified Sigma filter; Data normalization uses the following formula: ； Where S represents the original pixel value of the SAR image S; The minimum pixel value; Maximum pixel value; The normalized pixel value, ranging from [0,1].

3. The method according to claim 1, characterized in that, The implementation method of the spatial confidence modeling unit is as follows: Average pooling and max pooling are performed on the input feature x to obtain and ; Will and After concatenation along the channel dimension, spatial confidence is calculated using a two-layer convolutional network: ； in, , This is the weight matrix. , As the bias vector, a convolution operation is performed on the original feature x to obtain supplementary features. ; Weighted fusion based on spatial confidence: ； in, This indicates multiplication of the primary element. This enhances spatial features.

4. The method according to claim 1, characterized in that... The implementation method of the channel confidence modeling unit is as follows: Spatial enhancement features Global average pooling is performed to obtain ; Predicting channel confidence using a multilayer perceptron: ； , This is the weight matrix. , For the bias vector, Perform convolution operations to obtain supplementary features ; Weighted fusion based on channel confidence: ； in, To output high-confidence features F= using weighted fusion, .

5. The method according to claim 1, characterized in that, The noise estimation network NEU in step S3 adopts the U-Net architecture, which includes a 4-layer encoder downsampling path, a bottleneck layer, and a 4-layer decoder upsampling path. Each layer of the encoder uses the feature confidence modeling module CEM to extract features, and each layer of the decoder uses the residual unit RU to recover features. The multi-scale detail information of the corresponding layer of the encoder is preserved through skip connections, and the bottleneck layer fuses the time step embedding information.

6. The method according to claim 1, characterized in that, Time step embedding uses sinusoidal position coding: ； ； In both formulas, d represents the total number of image token embedding dimensions; i represents the index of the current image token embedding dimension; and t represents the position of the image token on the image.

7. The method according to claim 1, characterized in that, Extract the geometric parameters, location information, and confidence information of the oil spill area, and output them in GeoJSON, Shapefile, or GeoTIFF format.

8. A SAR marine oil spill intelligent identification system based on a diffusion model, characterized in that, include: The preprocessing module performs geometric correction, radiometric calibration, noise suppression, and data normalization on the acquired marine environment SAR remote sensing images to obtain the preprocessed SAR image S. The feature confidence modeling module inputs the SAR image S, calculates spatial confidence through the spatial confidence modeling unit, calculates channel confidence through the channel confidence modeling unit, and performs feature fusion to obtain high-confidence features. ,in Represents the real number field, where C is the number of channels, H is the height, and W is the width; The conditional diffusion generation inference module generates the high-confidence features. The SAR image S is used as a conditional input to the noise estimation network for diffusion. The diffusion process initializes and generates a noisy image. For time step t, iterate from T to 1: Construct conditional input Call NEU to predict noise Calculated according to the diffusion backsampling formula After T iterations, the generated segmentation mask is obtained. ; The fusion module repeats the diffusion process n times, using different initial noises, to obtain n independently generated segmentation masks. (1), (2), ..., (n)}, where (i) indicates the first The final mask obtained by secondary generation inference is used to calculate the pixel-level average fusion probability P, and the final binarized mask M is obtained by using the threshold τ. The results output module performs morphological operations and connected component analysis on the final binary mask M to extract the geometric parameters, location information, and confidence information of the oil spill area, and outputs them in a standard geographic information format.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1 to 7.

Images