Improved resnet-50 intelligent identification method for ocean internal wave based on fusion of multi-scale features

Abstract

The application discloses a modified ResNet-50 ocean internal wave intelligent recognition method fusing multi-scale features, and belongs to the technical field of cross of marine satellite remote sensing and computer vision, and comprises the following steps: acquiring an original high-resolution remote sensing measured image, and obtaining an enhanced satellite remote sensing image through pretreatment; manually labeling and correcting the labeling discontinuity and adhesion problems; through morphological dilation, adaptive sliding window cropping, data enhancement and data set division, a balanced data set of positive and negative samples is constructed; a single-channel backbone network is reconstructed, an ASPP module is introduced to build a multi-scale feature network; an optimal detection model is obtained through iterative optimization of a coupled loss function; and a remote sensing image to be detected is input, multi-dimensional evaluation is carried out, and an internal wave pixel-level prediction mask is obtained. The application is suitable for single-polarization remote sensing image features, can suppress sea clutter interference, reduce internal wave false alarm and false alarm, and improve the internal wave recognition accuracy and topological restoration robustness under complex sea conditions.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of marine satellite remote sensing and computer vision, and in particular to an improved ResNet-50 intelligent identification method for marine internal waves that integrates multi-scale features. Background Technology

[0002] Existing technologies rely on the advantages of satellite remote sensing in all-weather, all-time, and high spatial resolution observation, using it as the main means for large-scale monitoring and spatial distribution analysis of ocean internal waves. At the same time, traditional image processing methods and basic convolutional neural network models are used to carry out automatic detection of ocean internal waves and extraction of wave crest features on high-resolution satellite remote sensing images.

[0003] However, existing technologies for satellite remote sensing images employ a holistic approach to reading and processing the entire scene. This approach is susceptible to the large data volume of individual scenes, easily leading to memory and video memory overflows in computing devices and hindering efficient batch data processing. Traditional image processing methods, in complex marine environments, suffer from high false alarm and false alarm rates due to interference from weak backscattered signals from internal waves, sea surface clutter such as wind shadows and ship wakes. Furthermore, the inherent operating mechanism of basic convolutional neural networks, which adapts to RGB three-channel input, can disrupt the grayscale physical properties of single-polarization images and generate redundant calculations. At the same time, conventional deep learning models are affected by the extremely unbalanced topological structure and sample distribution of the elongated strips of ocean internal waves, making it difficult to focus on weak internal wave signals, resulting in broken wave crest lines and insufficient model generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide an improved ResNet-50 intelligent identification method for ocean internal waves that integrates multi-scale features, thereby solving the aforementioned technical problems.

[0005] To achieve the above objectives, this invention provides an improved ResNet-50 intelligent identification method for ocean internal waves that integrates multi-scale features, comprising the following steps: S1. Obtain marine remote sensing image data, and obtain enhanced satellite remote sensing images through invalid value filtering, large-scale local slice extraction, logarithmic transformation backscatter signal enhancement and percentile truncation normalization preprocessing operations. S2. Based on the enhanced satellite remote sensing image of S1, correct the discontinuity and adhesion problems in the internal wave crest line data to obtain standardized internal wave crest line data; S3. Based on the normalized internal wave crest line annotation data of S2, a balanced positive and negative sample dataset is obtained through morphological dilation, adaptive sliding window cropping, random data augmentation and dataset partitioning operations. S4. By reconstructing the single-channel backbone network through cross-channel weight fusion and introducing the hollow spatial pyramid pooling ASPP module to construct a multi-scale feature extraction structure operation, and combining it with ResNet-50, an improved deep learning algorithm is obtained, which is adapted to a deep learning multi-scale network for single-polarity features. S5. Based on the balanced positive and negative sample dataset of S3 and the deep learning multi-scale network of S4, the model weights are iteratively optimized through a joint loss function coupled with the focal loss function FocalLoss and the dice loss function DiceLoss to obtain the optimal detection model. S6. Based on the optimal detection model of S5, the original satellite remote sensing image to be detected is input for inference after image preprocessing in S1. Sigmoid probability generation, threshold binarization processing, and output of pixel-level prediction mask of internal wave are performed. Then, the recognition accuracy is quantified through multi-dimensional index evaluation operation.

[0006] Preferably, the specific steps of S1 include: S11. Based on the acquired original high-resolution satellite remote sensing image, the satellite remote sensing local slice image is obtained by filtering invalid pixel values ​​and extracting local high-resolution physical slices at a preset scale. S12. Based on local slice images from satellite remote sensing, logarithmic transformation is used to map backscattered signals, compress bright clutter, and stretch the contrast of weak signals. The logarithmic transformation formula is as follows: ; in, This represents the original radar backscatter intensity. This is the converted decibel value. It is a smoothing factor; To obtain logarithmic domain-enhanced satellite remote sensing images; S13. Based on the logarithmic domain-enhanced satellite remote sensing image, the enhanced satellite remote sensing image is obtained by removing extreme pixel interference through percentile truncation and normalizing to the standard grayscale range through linear mapping.

[0007] Preferably, the specific steps of S2 include: S21. Based on enhanced satellite remote sensing images, initial internal wave crest line annotation data is obtained through manual delineation of internal wave crest line trajectories and structured storage of annotation information. S22. Based on the initial internal wave crest line annotation data of S21, the internal wave crest line discontinuity breakpoint logical connection operation is used to obtain the breakpoint corrected internal wave crest line annotation data. S23. Based on the breakpoint correction of the inner wave crest line annotation data in S22, the standard inner wave crest line annotation data is obtained by pixel-level separation of the adhesion area of ​​the inner wave crest line.

[0008] Preferably, the specific steps of S3 include: S31. Based on the standard inner wave peak line annotation data, the inner wave target pixel coverage area is expanded by morphological dilation operation to obtain the dilated inner wave annotation mask data. S32. Based on the dilated internal wave annotation mask data of S31, the initial image slice data is obtained through adaptive sliding window cropping and dynamic removal of pure background slices. S33. Based on the initial image slice data in S32, data augmentation operations such as random geometric transformation and pixel feature perturbation are used to obtain enhanced image slice data; S34. Based on the enhanced image slice data of S33, the training subset and validation subset are divided by a preset ratio to obtain a balanced dataset of positive and negative samples.

[0009] Preferably, the morphological dilation operation in S31 is specifically performed as follows: based on the single-pixel peak line mask corresponding to the standard internal wave peak line annotation data, the operation expands the coverage area of ​​the internal wave target pixel to avoid the loss of internal wave features during the multi-layer downsampling process of the network, and strengthens the response intensity of internal wave features in the deep network, so as to enhance the anti-interference and suppression capability of radar sea clutter.

[0010] Preferably, the specific steps of S4 include: S41. By extracting the weights of the first-layer three-channel convolution kernel, the first-layer convolution weight data of ResNet-50 is obtained. S42. Based on the ResNet-50 first-layer convolutional weight data from S41, perform cross-channel weight fusion to reconstruct single-channel convolutional kernels. The weight fusion formula is as follows: ; in, These are the original weights of the three-channel convolution kernel; These are the weights of the single-channel convolutional kernel obtained after cross-channel fusion; S43. Based on the deep feature map output by the single-channel ResNet-50 backbone network, a deep learning multi-scale network adapted to single-polarization features is obtained by introducing the hollow spatial pyramid pooling (ASPP) module, parallel multi-branch feature extraction and feature concatenation integration operations.

[0011] Preferably, the specific process of introducing the Aperture Spatial Pyramid Pooling (ASPP) module in S43 is as follows: Based on the deep feature map output by the single-channel ResNet-50 backbone network, the inner wave detail features, multi-scale context features, and global image prior information are extracted by setting 1×1 standard convolution branches, 3×3 dilated convolution branches with different dilation rates, and global average pooling branches in parallel. Then, the features of each branch are spliced ​​and integrated to obtain a topological structure that accurately adapts to the slender inner waves and variable scales.

[0012] Preferably, the specific steps of S5 include: S51. Based on a deep learning multi-scale network, the coupling relationship between FocalLoss and DiceLoss is constructed, and a joint loss function expression is defined. The formula for FocalLoss is: ; in, For predicting probability values; To balance the weighting coefficients of positive and negative samples; For focusing parameters; The formula for DiceLoss is: ; in, For pixels The predicted probability; This is a real label; A smoothing coefficient to prevent division by zero errors; The formula for the joint loss function is: ; S52. Based on the balanced positive and negative sample dataset, the deep learning multi-scale network, and the joint loss function of S51, the network weights are iteratively updated through the backpropagation algorithm to obtain the network weight parameters after multiple iterations. S53. Based on the network weight parameters after multiple iterations in S52, the optimal detection model is obtained by evaluating and selecting the optimal weight parameters through validation set performance indicators.

[0013] Preferably, the specific steps of S6 include: S61. Based on the optimal detection model, the original feature data output by the model is obtained by inputting the satellite remote sensing image to be detected after image preprocessing in S1 and performing model inference operation. S62. The original feature data output by the model based on S61 is used to generate a pixel-level prediction probability map through a Sigmoid probability generation operation. S63. The pixel-level prediction probability map based on S62 is used to obtain the inner wave pixel-level prediction mask through a threshold binarization process. S64, based on S63, uses an internal wave pixel-level prediction mask. Through multi-dimensional index quantification evaluation, the accuracy of internal wave recognition is quantified.

[0014] Preferably, the specific process of the multi-dimensional index quantification evaluation operation in S64 is as follows: Based on the pixel-level prediction mask of the inner wave and the corresponding real labeled data, the model’s recognition performance of the inner wave is comprehensively quantified by introducing the cross-union ratio, precision, recall and F1 score. Among them, the cross-union ratio is used to measure the spatial overlap between the prediction mask and the real label, the precision is used to evaluate the model’s ability to suppress false positives, the recall is used to evaluate the model’s ability to reduce false negatives, and the F1 score is used to comprehensively balance the model’s recognition accuracy.

[0015] Therefore, the improved ResNet-50 ocean internal wave intelligent identification method that integrates multi-scale features, as described above, has the following beneficial effects: 1. By combining invalid value filtering, large-scale local slicing, logarithmic transform signal enhancement, and percentile truncation normalization preprocessing techniques, logarithmic transform compresses bright sea clutter and stretches the contrast of weak internal wave backscattered signals, while percentile truncation eliminates extreme pixel interference. The entire process can improve the identification of weak ocean internal wave features in single-polarization remote sensing images, reduce sea surface clutter noise interference, and provide a high-quality image base with higher signal-to-noise ratio and uniform grayscale distribution for subsequent annotation correction and model training, thereby reducing recognition errors from the data source.

[0016] 2. To address the defects of discontinuous breaks and pixel adhesion in the initial manually labeled wave crest lines, we sequentially implemented logical patching of breaks and pixel separation in adhesion areas, followed by morphological dilation to widen the internal wave target pixel region. This not only restores the true continuous band-like topology of internal waves and corrects the human error in the labeled data, but the dilation operation can also resist the problem of loss of narrow internal wave features caused by multi-layer downsampling of the network, strengthen the response of deep networks to internal wave features, improve the model's anti-interference ability for weak internal wave targets in chaotic sea conditions, and ensure the accuracy and reliability of the labeled ground truth values ​​of the dataset.

[0017] 3. Adaptive sliding window cropping and dynamic screening of pure background slices are carried out based on the corrected peak mask. Combined with random geometry and pixel perturbation enhancement, and proportional division of training and validation sets, a balance between the number of positive and negative samples is achieved. The adaptive sliding window can adapt to internal wave bodies of varying lengths and scales. Random enhancement enriches the diversity of sample morphology. The balanced sample distribution can alleviate the sample skew problem caused by the fact that the background in the marine scene is much larger than the internal wave targets. It can avoid the tendency of model training to be biased towards the background and to miss internal waves, and improve the model's ability to generalize and adapt to remote sensing images of different sea areas.

[0018] 4. By employing cross-channel weight fusion, the original three-channel ResNet-50 is reconstructed into a single-channel backbone adapted to single-polarization remote sensing data, solving the problems of traditional three-channel networks being unable to match single-band radar imagery and weight redundancy mismatch. The ASPP multi-branch hollow pooling structure is superimposed to extract fine details, multi-scale context, and global prior features in parallel, accurately matching the topological characteristics of long, thin, and large-scale internal waves in the ocean. The overall network reduces the computational power loss of invalid channels, and the multi-scale feature fusion takes into account both the ability to capture small and fragmented internal waves and large-scale continuous internal waves, significantly improving the completeness of multi-morphological internal wave feature extraction compared to the original ResNet-50.

[0019] 5. The joint loss method combines the advantages of FocalLoss in balancing positive and negative samples and focusing on easy and difficult samples with the advantages of DiceLoss in optimizing segmentation overlap. FocalLoss reduces the loss weight of massive simple background samples and focuses on difficult-to-identify weak internal wave samples, while DiceLoss directly constrains the spatial overlap between the predicted mask and the real wave peak region. The two are coupled and work together to optimize the model weight iteration direction, which not only improves the training bias of sample imbalance, but also accurately constrains the pixel-level segmentation matching degree, accelerates the model convergence speed, reduces the two core identification defects of internal wave missed detection and background false detection, and quickly selects the optimal detection weight model with high fit.

[0020] 6. A unified preprocessing workflow for reusing test images ensures input consistency. After model inference, Sigmoid probability mapping, and threshold binarization, pixel-level inner wave masks are output, along with a four-dimensional evaluation system of IoU, precision, recall, and F1 score. The standardized inference link unifies the training and testing data processing logic, eliminating precision fluctuations caused by differences in pre- and post-processing. Multiple indicators perform their respective functions, measuring spatial overlap, false alarm suppression, false negative prevention, and overall recognition level. This allows for comprehensive and objective quantification of model recognition performance, accurate identification of model weaknesses, and support for horizontal comparison and iterative optimization of method effectiveness.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 A flowchart of the improved ResNet-50 intelligent identification method for ocean internal waves that integrates multi-scale features provided by the present invention; Figure 2 The image processing and inner wave fringe annotation flowchart provided by this invention; Figure 3 This is a schematic diagram of the morphological dilation of the labeled data provided by the present invention; Figure 4 This invention provides a schematic diagram of the single-channel reconstruction and network structure. Figure 5A global overview of the original satellite remote sensing image provided for the second embodiment of the present invention; Figure 6 To Figure 5 The diagram shows the processing steps of the original satellite remote sensing image global overview map. Figure (a) is the preprocessed original satellite image global overview map, and Figure (b) is a local high-resolution slice map of the enhanced satellite image. Figure 7 The second embodiment of the present invention is shown in the schematic diagram of model reasoning and recognition, wherein Figure (a) is a slice of satellite remote sensing image used as input to the model, Figure (b) is a probability heat map of the model after internal wave recognition, and Figure (c) is the output result of automatic internal wave recognition. Figure 8 The truth mask diagram provided in the second embodiment of the present invention is shown in Figure (a), which is a single pixel mask diagram and Figure (b) is a dilation mask diagram. Figure 9 This is a verification diagram of the model recognition result in the second embodiment of the present invention, wherein Figure (a) is a local high-resolution slice of the enhanced satellite remote sensing image, and Figure (b) is the internal wave recognition result and... Figure 8 The results of the comparison between the ground truth mask and the actual mask are shown in the figure. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0024] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0025] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] Based on the above analysis, this invention is designed. (See appendix.) Figures 1-2 An improved ResNet-50 intelligent identification method for ocean internal waves, incorporating multi-scale features, includes the following steps: S1. Obtain marine remote sensing image data, and obtain enhanced satellite remote sensing images through invalid value filtering, large-scale local slice extraction, logarithmic transformation backscatter signal enhancement, and percentile truncation normalization preprocessing operations. The specific steps of S1 include: S11. Based on the acquired original high-resolution satellite remote sensing image, a local slice image is obtained by filtering invalid pixel values ​​and extracting local high-resolution physical slices at a preset scale. In one embodiment of the present invention, target data is extracted from raw high-resolution satellite remote sensing images covering the South China Sea and the western Pacific Ocean. The raw data is usually VV-polarized floating-point data, containing geographic coordinate information and corresponding radar backscattering intensity values. For the ultra-large-scale original images at the GB level, this embodiment adopts a dynamic sliding window algorithm to perform boundary security checks based on preset starting coordinates and accurately extract local high-resolution slices with a physical size of 6000×6000 pixels from the original image, thereby obtaining satellite remote sensing local slice images to avoid memory overflow and preserve the fine structure of the internal wave signal. At the same time, invalid pixel values ​​are filtered out, that is, pixel values ​​representing land or echo-free areas are marked as NaN states other than valid values.

[0027] S12. Based on local slice images from satellite remote sensing, logarithmic transformation is used to map backscattered signals, compress bright clutter, and stretch the contrast of weak signals. The logarithmic transformation formula is as follows: ; in, This represents the original radar backscatter intensity. This is the converted decibel value. It is a smoothing factor; To obtain logarithmic domain-enhanced satellite remote sensing images; By converting linear intensity radar echoes to decibel scales, it is possible to effectively compress bright clutter and stretch the contrast of weak ripples on the water surface; and this logarithmic mapping process can highlight the subtle Bragg scattering changes on the sea surface caused by internal waves.

[0028] S13. Based on the logarithmic domain-enhanced satellite remote sensing image, the enhanced satellite remote sensing image is obtained by removing extreme pixel interference through percentile truncation and normalizing to the standard grayscale range through linear mapping.

[0029] Specifically, this embodiment combines a percentile truncation strategy with... The image undergoes adaptive dynamic range adjustment by calculating the cumulative distribution of pixel values ​​within the slice and extracting the 2% quantile. and 98th percentile As the truncation boundary for brightness, its specific truncation logic is as follows: ; in, These are the image pixel values ​​after percentile truncation. Subsequently, the truncated decibel values ​​were normalized to an 8-bit grayscale range using a linear mapping formula to generate enhanced SAR satellite images. The normalization formula is as follows: ; in, The image pixel value after linear normalization to the 8-bit grayscale range, with a value range of [0, 255]; The enhanced satellite remote sensing image obtained above is a single-channel grayscale image, which can represent the enhanced physical characteristics of radar echoes over the sea surface. This process can effectively capture the local texture structure in the image, which not only has good noise resistance, but also maintains the physical consistency of the image, so that the inner wave peaks that were originally hidden in the background can obtain the maximum detail in the [0,255] grayscale space.

[0030] S2. Based on the enhanced satellite remote sensing image of S1, correct the discontinuity and adhesion problems in the internal wave crest line data to obtain standardized internal wave crest line data; The specific steps of S2 include: S21. Based on enhanced satellite remote sensing images, initial internal wave crest line annotation data is obtained through manual delineation of internal wave crest line trajectories and structured storage of annotation information. In this implementation, manual annotation is performed using the image annotation tool Labelme. Combined with the unique physical characteristics of alternating light and dark in the internal wave stripes in satellite remote sensing images, the trajectory of the wave crest line of the internal wave beam is accurately delineated. Subsequently, the annotation results are extracted and made into a standard JSON format file to serve as the ground truth labels required for training the machine learning model. S22. Based on the initial internal wave crest line annotation data of S21, the internal wave crest line discontinuity breakpoint logical connection operation is used to obtain the breakpoint corrected internal wave crest line annotation data. Specifically, the logic is used to fill in the gaps caused by speckle noise masking through manual inspection; S23. Based on the breakpoint correction of the internal wave crest line annotation data in S22, pixel-level separation of the internal wave crest line adhesion region is performed to obtain standardized internal wave crest line annotation data; this ensures that the annotation samples can accurately reflect the real spatial topology of ocean internal waves and provide high-quality data support for subsequent model learning of the geometric continuity of internal wave fringes.

[0031] S3. Based on the normalized internal wave crest line data of S2, a balanced positive and negative sample dataset is obtained through morphological dilation, adaptive sliding window clipping, random data augmentation, and dataset partitioning. See Figure 3 As shown, the specific steps of S3 include: S31. Based on the standard inner wave peak line annotation data, the inner wave target pixel coverage area is expanded by morphological dilation operation to obtain the dilated inner wave annotation mask data. The morphological dilation operation in S31 is as follows: Based on the single-pixel peak line mask corresponding to the standard inner wave peak line annotation data, the operation expands the coverage area of ​​the inner wave target pixel to avoid the loss of inner wave features during the multi-layer downsampling process of the network and strengthens the response intensity of inner wave features in the deep network. Specifically, a morphological dilation operation is performed on the original mask using a 5×5 convolution kernel. This operation enhances the response intensity of the slender strip targets in the high-level feature map of the network by artificially increasing the target area of ​​the target pixel cluster, thereby improving the model's feature extraction efficiency for fine stripes.

[0032] S32. Based on the dilated internal wave annotation mask data of S31, the initial image slice data is obtained through adaptive sliding window cropping and dynamic removal of pure background slices. Specifically, the adaptive sliding window cropping process includes: dynamically calculating the proportion of positive sample pixels in each slice, and removing pure background slices that do not contain internal wave information in real time, thereby ensuring that the proportion of positive sample slices in the training set is maintained at a preset high level, such as 70%.

[0033] S33. Based on the initial image slice data in S32, data augmentation operations such as random geometric transformation and pixel feature perturbation are used to obtain enhanced image slice data; Specifically, in order to improve the generalization ability and robustness of the model, random horizontal / vertical flips, multi-angle rotations, and linear brightness and contrast perturbations are applied to the cropped image slices. This perturbation strategy forces the model to no longer rely on absolute pixel intensity, but to learn the relative geometry and spatial gradient features of the inner wave ripples.

[0034] S34. Based on the enhanced image slice data of S33, the training subset and validation subset are divided by a preset ratio to obtain a balanced dataset of positive and negative samples. Specifically, the processed and enhanced image slice data is divided into training and validation datasets according to a predetermined ratio. In this embodiment, the training set accounts for 80% and is used for updating model weights; the validation set accounts for 20% and is used to evaluate the model's recognition accuracy in real time and to optimize and fine-tune hyperparameters during training.

[0035] S4. By reconstructing the single-channel backbone network through cross-channel weight fusion and introducing the hollow spatial pyramid pooling ASPP module to construct a multi-scale feature extraction structure operation, and combining it with ResNet-50, an improved deep learning algorithm is obtained, which is adapted to a deep learning multi-scale network for single-polarity features. See Figure 4 As shown, the specific steps of S4 include: S41. By extracting the weights of the first-layer three-channel convolution kernel, the first-layer convolution weight data of ResNet-50 is obtained. S42. Based on the ResNet-50 first-layer convolutional weight data from S41, perform cross-channel weight fusion to reconstruct single-channel convolutional kernels. The weight fusion formula is as follows: ; in, These are the original weights of the three-channel convolution kernel; These are the weights of the single-channel convolutional kernel obtained after cross-channel fusion; Specifically, extract the first layer of the pre-trained ResNet-50 backbone network with a size of... The original weight tensor To fully preserve the model's low-level edge and texture perception capabilities learned on large-scale datasets while changing the input dimension, a weight fusion strategy is used. The weights of the three channels are accumulated element-wise along the channel dimension to reconstruct the generated image of size [size missing]. Single-polarization dedicated convolution kernel This fusion operation activates the initial perception accuracy of single-polarization enhanced features, improving the convergence speed of the model in complex radar contexts.

[0036] S43. Based on the deep feature map output by the single-channel ResNet-50 backbone network, a deep learning multi-scale network adapted to single-polarization features is obtained by introducing the hollow spatial pyramid pooling (ASPP) module, parallel multi-branch feature extraction and feature concatenation and integration operations. The specific process of introducing the Aperture Spatial Pyramid Pooling (ASPP) module in S43 is as follows: Based on the deep feature map output by the single-channel ResNet-50 backbone network, the inner wave detail features, multi-scale context features, and global image prior information are extracted by setting 1×1 standard convolution branches, 3×3 dilated convolution branches with different dilation rates, and global average pooling branches in parallel. Then, the features of each branch are spliced ​​and integrated to obtain a topological structure that accurately adapts to the slender inner waves and variable scales.

[0037] Specifically, in this embodiment, a void ratio is set. First, an adaptive method is used to capture internal wave beams of different scales and widths. Second, a global average pooling branch is introduced in parallel to obtain global image-level contextual prior information, enhancing the model's ability to suppress overall ocean background noise. Finally, the feature maps output by all the above parallel branches are concatenated along the channel dimension and applied... The standard convolutional branch performs in-depth feature dimensionality reduction and information integration; this process enables the effective fusion of multi-scale network features, thereby accurately capturing the slender and scale-variable spatial topology of ocean internal waves under complex sea conditions.

[0038] S5. Based on the balanced positive and negative sample dataset of S3 and the deep learning multi-scale network of S4, the model weights are iteratively optimized through a joint loss function coupled with the focal loss function FocalLoss and the dice loss function DiceLoss to obtain the optimal detection model. The specific steps of S5 include: S51. Based on a deep learning multi-scale network, a coupling relationship between FocalLoss and DiceLoss is constructed, and a joint loss function expression is defined. FocalLoss uses a modulating factor to mine hard negative samples, and its formula is: ; in, For predicting probability values; To balance the weighting coefficients of positive and negative samples; For focusing parameters; This formula allows the model to dynamically reduce the gradient weights of large, easily classifiable background pixels, forcing computational resources to be concentrated on identifying extremely weak peak line pixels, thus alleviating the classification shift problem caused by extreme class imbalance. DiceLoss is used to enforce topological connectivity constraints. It directly measures the regional similarity between the predicted results and the ground truth labels from the perspective of the global spatial intersection-union ratio. Its formula is as follows: ; in, For pixels The predicted probability; This is a real label; A smoothing coefficient to prevent division by zero errors; The DiceLoss algorithm mathematically encourages the network to output a linear mask with a complete structure and strong connectivity, which suppresses isolated false alarms caused by complex radar background noise, thereby ensuring the physical continuity of the extracted ocean internal wave stripes in terms of spatial morphology.

[0039] The formula for the joint loss function is: ; A joint loss function is used to achieve dual constraints on pixel-level classification accuracy and global geometric connectivity.

[0040] S52. Based on the balanced positive and negative sample dataset, the deep learning multi-scale network, and the joint loss function of S51, the network weights are iteratively updated through the backpropagation algorithm to obtain the network weight parameters after multiple iterations. S53. Based on the network weight parameters after multiple iterations in S52, the optimal detection model is obtained by evaluating and selecting the optimal weight parameters through validation set performance indicators.

[0041] S6. Based on the optimal detection model of S5, the original satellite remote sensing image to be detected, which has undergone image preprocessing in S1, is input and subjected to inference, Sigmoid probability generation, threshold binarization processing, and multi-dimensional index quantification evaluation to obtain the inner wave pixel-level prediction mask and complete the quantification of recognition accuracy. S61. Based on the optimal detection model, the original feature data output by the model is obtained by inputting the satellite remote sensing image to be detected after image preprocessing in S1 and performing model inference operation. S62. The original feature data output by the model based on S61 is used to generate a pixel-level prediction probability map through a Sigmoid probability generation operation. S63. The pixel-level prediction probability map based on S62 is used to obtain the inner wave pixel-level prediction mask through a threshold binarization process. S64, based on S63, uses an inner wave pixel-level prediction mask and performs multi-dimensional index quantification evaluation to complete the quantification of inner wave recognition accuracy. The specific process of multi-dimensional index quantification evaluation in S64 is as follows: Based on the pixel-level prediction mask of the inner wave and the corresponding real labeled data, the model’s recognition performance of the inner wave is comprehensively quantified by introducing the crossover ratio, precision, recall and F1 score. Among them, the crossover ratio is used to measure the spatial overlap between the prediction mask and the real label, the precision is used to evaluate the model’s ability to suppress false positives, the recall is used to evaluate the model’s ability to reduce false negatives, and the F1 score is used to comprehensively balance the model’s recognition accuracy.

[0042] Specifically, precision refers to the proportion of pixels that the model identifies as positive, but which are actually positive. It measures the accuracy of the model in identifying positive classes, and the calculation formula is: ; in, Accuracy refers to the proportion of pixels that the model predicts as internal waves (positive class) but are actually internal waves. It is used to evaluate the model's ability to suppress false positives when identifying internal waves and to avoid misjudging background noise such as sea surface shadows and ship wakes as internal waves. For true examples, it represents the number of pixels that the model predicts to be inner wave pixels and that are actually inner wave pixels, i.e., the number of correctly identified inner wave pixels. False positives are the number of pixels that the model predicts to be inner wave pixels but are actually background pixels, i.e., the number of false positive pixels. Recall rate refers to the proportion of all actual positive pixels that the model correctly identifies as positive. It measures the model's ability to extract positive pixels, and the calculation formula is: ; in, Recall rate refers to the proportion of pixels that are actually internal waves that are correctly identified as internal waves by the model. It is used to evaluate the model's ability to capture weak internal wave signals and reduce false negatives caused by weak internal wave signals and low contrast. This is the number of pixels that the model predicts to be background pixels but are actually inner wave pixels, i.e., the number of missed pixels. The F1 score is the harmonic mean of precision and recall, which comprehensively considers the balance between suppressing false positives and reducing false negatives. The calculation formula is as follows: ; in, F1 score is the harmonic mean of precision and recall, used to comprehensively measure the model's performance in balancing the suppression of false positives and the reduction of false negatives. Intersection over Union (IoU) measures the degree of geometric overlap between the predicted mask and the ground truth annotations. The formula is: ; in, The crossover ratio (CR) is used to reflect the morphological integrity and positional accuracy of the internal waves identified by the model. A higher value indicates a better spatial match between the predicted results and the actual internal waves.

[0043] Based on the above multi-dimensional evaluation system, the morphological integrity, positional accuracy, and robustness of the model in identifying internal waves under complex sea conditions are comprehensively quantified from multiple dimensions, including spatial overlap, false alarm suppression capability, false alarm reduction capability, and overall performance balance. This provides an objective, comprehensive, and reliable performance evaluation basis for model iterative optimization and engineering applications related to ocean internal waves, ensuring that the evaluation results can truly reflect the model's ability to capture weak internal wave signals and avoiding the one-sidedness of single-dimensional evaluation.

[0044] The second embodiment of this invention: A high-resolution satellite remote sensing image in VV polarimetric interferometry wide swath (IW) mode of a certain satellite is selected, located at longitude 116.3°E-119°E and latitude 7.2°N-9.2°N. See [link to relevant documentation]. Figure 5 As shown; For ultra-large-scale raw images, a single-polarization physical feature enhancement preprocessing is first performed: invalid values ​​≤0 are filtered for NaN and decibel conversion is performed, and then truncation and linear mapping are performed based on the 2% and 98th percentiles of the pixel cumulative distribution. The global overview of the preprocessed image is as follows: Figure 6 As shown in Figure a; subsequently, based on the set starting pixel coordinates, column coordinate 19505, row coordinate 10000, a 6000×6000 pixel local high-resolution slice is extracted from the image, see [reference]. Figure 6 As shown in Figure b, preprocessing effectively highlights the details of the inner wavy stripes hidden in the background, providing a high-quality data foundation for subsequent deep learning feature extraction.

[0045] The enhanced single-polarization satellite remote sensing image is input into the optimal recognition model. During the inference stage, a 512×512 sliding window and a step size of 384 pixels are used for local feature extraction. After region stitching, a probability distribution heatmap of the same size as the original image is generated. See [link to relevant documentation]. Figure 7 As shown in Figure b, the heatmap is then binarized and subjected to morphological closing operations to generate the final prediction mask. See Figure b for details. Figure 7 As shown in Figure c, the results indicate that the model can effectively locate the internal wave region and clearly delineate its morphological outline from a complex sea clutter background.

[0046] To verify the accuracy of the recognition results, this embodiment constructs a corresponding ground truth mask image, see [link / reference]. Figure 8 As shown, the center line is manually delineated along the inner wave fringe direction in the image, and then morphological dilation is performed on the original annotation using a 5×5 all-one convolution kernel to enhance the spatial response of the slender target and generate a binary ground truth mask. Accuracy evaluation is then performed, considering the slender topological features of the inner wave crest line, and a spatial tolerance buffer mechanism with a 3-pixel tolerance radius is introduced. Specifically, a 7×7 convolution kernel is used to generate a buffer validation area, and the predicted results are compared with the ground truth mask at the pixel level. The specific model performance metrics are as follows: , , , .

[0047] Compare the predicted results with the ground truth mask, see [link / reference] Figure 9 As shown, the red predicted area accurately covers most of the green true labeled area, and the high recall rate verifies the ability of single-channel feature enhancement and FocalLoss to capture weak internal wave signals; although the precision is affected by the model edge inflation effect and has a tolerance penalty, the high precision of 60.02% is still achieved. The results show that the predicted results are highly consistent with the actual internal waves in terms of spatial morphology, geometric continuity and topological location. This fully demonstrates the advantages of the ASPP module and DiceLoss in suppressing sea surface wind shadows, wake false alarms and ensuring the robustness of internal wave extraction, and provides reliable automated technical support for marine disaster early warning and dynamics research.

[0048] In summary, this invention addresses the challenges of weak signals, low signal-to-noise ratios, elongated topologies, and easy underreporting of internal waves in single-polarization satellite remote sensing images. Furthermore, it addresses the strong background clutter interference and extreme imbalance between positive and negative samples, which traditional identification methods often rely heavily on background noise and struggle to accurately capture internal wave features. This invention effectively highlights internal wave details hidden within complex sea clutter through single-polarization physical feature enhancement preprocessing. A balanced dataset is constructed using morphological dilation and adaptive sliding windows. A single-channel ResNet backbone network with cross-channel weight fusion, coupled with an ASPP module, accurately extracts multi-scale internal wave topological features. The model is further optimized using a joint loss function coupled with FocalLoss and DiceLoss. Validated by examples, the model achieves a recall of 85.95% and an IoU of 60.02% under complex sea conditions. It suppresses false alarms caused by wind shadows and ship wakes, reduces underreporting of weak internal wave signals, and improves the location accuracy and morphological integrity of internal wave identification. This provides efficient and reliable automated identification technology support for research on ocean internal wave dynamics and disaster early warning.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An improved ResNet-50 intelligent identification method for ocean internal waves that integrates multi-scale features, characterized in that, Includes the following steps: S1. Obtain marine remote sensing image data, and obtain enhanced satellite remote sensing images through invalid value filtering, large-scale local slice extraction, logarithmic transformation backscatter signal enhancement and percentile truncation normalization preprocessing operations. S2. Based on the enhanced satellite remote sensing image of S1, correct the discontinuity and adhesion problems in the internal wave crest line data to obtain standardized internal wave crest line data; S3. Based on the normalized internal wave crest line data of S2, a balanced positive and negative sample dataset is obtained through morphological dilation, adaptive sliding window clipping, random data augmentation, and dataset partitioning. S4. By reconstructing the single-channel backbone network through cross-channel weight fusion and introducing the hollow spatial pyramid pooling ASPP module to construct a multi-scale feature extraction structure operation, and combining it with ResNet-50, an improved deep learning algorithm is obtained, which is adapted to a deep learning multi-scale network for single-polarity features. S5. Based on the balanced positive and negative sample dataset of S3 and the deep learning multi-scale network of S4, the model weights are iteratively optimized through a joint loss function coupled with the focal loss function FocalLoss and the dice loss function DiceLoss to obtain the optimal detection model. S6. Based on the optimal detection model of S5, the original satellite remote sensing image to be detected is input for inference after image preprocessing in S1. Sigmoid probability generation, threshold binarization processing, and output of pixel-level prediction mask of internal wave are performed. Then, the recognition accuracy is quantified through multi-dimensional index evaluation operation.

2. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 1, characterized in that: The specific steps of S1 include: S11. Based on the acquired original high-resolution satellite remote sensing image, the satellite remote sensing local slice image is obtained by filtering invalid pixel values ​​and extracting local high-resolution physical slices at a preset scale. S12. Based on local slice images from satellite remote sensing, logarithmic transformation is used to map backscattered signals, compress bright clutter, and stretch the contrast of weak signals. The logarithmic transformation formula is as follows: ； in, This represents the original radar backscatter intensity. This is the converted decibel value. It is a smoothing factor; To obtain logarithmic domain-enhanced satellite remote sensing images; S13. Based on the logarithmic domain-enhanced satellite remote sensing image, the enhanced satellite remote sensing image is obtained by removing extreme pixel interference through percentile truncation and normalizing to the standard grayscale range through linear mapping.

3. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 2, characterized in that: The specific steps of S2 include: S21. Based on enhanced satellite remote sensing images, initial internal wave crest line annotation data is obtained through manual delineation of internal wave crest line trajectories and structured storage of annotation information. S22. Based on the initial internal wave crest line annotation data of S21, the internal wave crest line discontinuity breakpoint logical connection operation is used to obtain the breakpoint corrected internal wave crest line annotation data. S23. Based on the breakpoint correction of the inner wave crest line annotation data in S22, the standard inner wave crest line annotation data is obtained by pixel-level separation of the adhesion area of ​​the inner wave crest line.

4. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 3, characterized in that: The specific steps of S3 include: S31. Based on the standard inner wave peak line annotation data, the inner wave target pixel coverage area is expanded by morphological dilation operation to obtain the dilated inner wave annotation mask data. S32. Based on the dilated internal wave annotation mask data of S31, the initial image slice data is obtained through adaptive sliding window cropping and dynamic removal of pure background slices. S33. Based on the initial image slice data in S32, data augmentation operations such as random geometric transformation and pixel feature perturbation are used to obtain enhanced image slice data; S34. Based on the enhanced image slice data of S33, the training subset and validation subset are divided by a preset ratio to obtain a balanced dataset of positive and negative samples.

5. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 4, characterized in that: The morphological dilation operation in S31 is as follows: Based on the single-pixel peak line mask corresponding to the standard internal wave peak line annotation data, the operation expands the coverage area of ​​the internal wave target pixel to avoid the loss of internal wave features during the multi-layer downsampling process of the network, and strengthens the response intensity of internal wave features in the deep network, so as to enhance the anti-interference and suppression capability of radar sea clutter.

6. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 4, characterized in that: The specific steps of S4 include: S41. By extracting the weights of the first-layer three-channel convolution kernel, the first-layer convolution weight data of ResNet-50 is obtained. S42. Based on the ResNet-50 first-layer convolutional weight data from S41, perform cross-channel weight fusion to reconstruct single-channel convolutional kernels. The weight fusion formula is as follows: ； in, These are the original weights of the three-channel convolution kernel; These are the weights of the single-channel convolutional kernel obtained after cross-channel fusion; S43. Based on the deep feature map output by the single-channel ResNet-50 backbone network, a deep learning multi-scale network adapted to single-polarization features is obtained by introducing the hollow spatial pyramid pooling (ASPP) module, parallel multi-branch feature extraction and feature concatenation integration operations.

7. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 6, characterized in that: The specific process of introducing the Aperture Spatial Pyramid Pooling (ASPP) module in S43 is as follows: Based on the deep feature map output by the single-channel ResNet-50 backbone network, the inner wave detail features, multi-scale context features, and global image prior information are extracted by setting 1×1 standard convolution branches, 3×3 dilated convolution branches with different dilation rates, and global average pooling branches in parallel. Then, the features of each branch are spliced ​​and integrated to obtain a topological structure that accurately adapts to the slender inner waves and variable scales.

8. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 6, characterized in that: The specific steps of S5 include: S51. Based on a deep learning multi-scale network, the coupling relationship between FocalLoss and DiceLoss is constructed, and a joint loss function expression is defined. The formula for FocalLoss is: ； in, For predicting probability values; To balance the weighting coefficients of positive and negative samples; For focusing parameters; The formula for DiceLoss is: ； in, For pixels The predicted probability; This is a real label; A smoothing coefficient to prevent division by zero errors; The formula for the joint loss function is: ； S52. Based on the balanced positive and negative sample dataset, the deep learning multi-scale network, and the joint loss function of S51, the network weights are iteratively updated through the backpropagation algorithm to obtain the network weight parameters after multiple iterations. S53. Based on the network weight parameters after multiple iterations in S52, the optimal detection model is obtained by evaluating and selecting the optimal weight parameters through validation set performance indicators.

9. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 8, characterized in that: The specific steps of S6 include: S61. Based on the optimal detection model, the original feature data output by the model is obtained by inputting the satellite remote sensing image to be detected after image preprocessing in S1 and performing model inference operation. S62. The original feature data output by the model based on S61 is used to generate a pixel-level prediction probability map through a Sigmoid probability generation operation. S63. The pixel-level prediction probability map based on S62 is used to obtain the inner wave pixel-level prediction mask through a threshold binarization process. S64, based on S63, uses an internal wave pixel-level prediction mask. Through multi-dimensional index quantification evaluation, the accuracy of internal wave recognition is quantified.

10. The improved ResNet-50 ocean internal wave intelligent identification method integrating multi-scale features according to claim 9, characterized in that: The specific process of multi-dimensional index quantification evaluation in S64 is as follows: Based on the pixel-level prediction mask of the inner wave and the corresponding real labeled data, the model’s recognition performance of the inner wave is comprehensively quantified by introducing the crossover ratio, precision, recall and F1 score. Among them, the crossover ratio is used to measure the spatial overlap between the prediction mask and the real label, the precision is used to evaluate the model’s ability to suppress false positives, the recall is used to evaluate the model’s ability to reduce false negatives, and the F1 score is used to comprehensively balance the model’s recognition accuracy.

Images