An Iceberg Detection Method for Synthetic Aperture Radar Images in the Antarctic Sea Area

Abstract

The present invention discloses an iceberg detection method for synthetic aperture radar images in the Antarctic Sea area, comprising the following steps: obtaining SAR image data of the Antarctic Sea area, screening Level-1 SLC products with latitudes of 40°S - 90°S, the data mode being Wave Mode, and the polarization modes including HH and VV; manually annotating the SAR images, visually interpreting them in combination with geographical location information, shape and texture features, and generating an Antarctic iceberg labeled dataset in COCO format; constructing a cascaded convolutional neural network model IB-Detector; evaluating the performance of the model using the test set and outputting iceberg position, contour and area information; the present invention improves the detection accuracy of icebergs in complex scenarios and meets the task requirements of iceberg target detection.

Description

Technical Field

[0001] The present invention relates to the field of marine science and technology, and particularly to an iceberg detection method for synthetic aperture radar images in the Antarctic Sea area. Background Art

[0002] Observation means such as observatories, shipborne, and aircraft can extract iceberg size and shape information, but are limited to typical large icebergs and cannot be applied to extensive iceberg detection. With the continuous development of spaceborne remote sensing technology, remote sensing data with a wide spatial range and high spatio-temporal resolution have become a powerful means for observing polar icebergs. Spaceborne remote sensing technology is mainly divided into optical remote sensing and microwave remote sensing. Among them, optical sensors are difficult to play a long-term role in iceberg monitoring due to thick clouds and limited sunshine time in polar regions. Microwave imaging sensors are not affected by weather and have the ability to observe all day and all weather, playing an important role in polar research. The swath width of a scatterometer reaches 1,000 kilometers, but its spatial resolution is relatively low (generally 25 kilometers), and it is only suitable for detecting huge icebergs with a length of more than 18.5 kilometers. Radar altimeters are limited to sub-satellite point observations and are often used to detect large icebergs in the open ocean. Therefore, both are not suitable for small and medium-scale icebergs, especially for iceberg monitoring tasks in areas close to the ice sheet.

[0003] The swath width of spaceborne synthetic aperture radar (SAR) is 5 - 500 kilometers, and the spatial resolution is 5 - 100 meters. Its fine resolution provides valuable data support for iceberg monitoring at different scales. Scholars at home and abroad have carried out a series of studies on iceberg detection in synthetic aperture radar images. However, iceberg detection based on SAR data is carried out based on wide-swath images, and the geographical location is mostly limited to the periphery of the Antarctic continent, and the target focuses on relatively large icebergs. However, small and medium-sized icebergs floating in the open ocean north of Antarctica are the main types of Antarctic icebergs. How to overcome the high false alarm rate of existing models for iceberg image classification, mark the positions of icebergs, extract iceberg contours, estimate iceberg areas, etc. from the massive classified iceberg data are urgent problems to be solved at present. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide an iceberg detection method for synthetic aperture radar images in the Antarctic Sea area, which overcomes the high false alarm rate of existing models for iceberg image classification, and solves problems such as marking the positions of icebergs, extracting iceberg contours, and estimating iceberg areas from the massive classified iceberg data.

[0005] Technical Solution: An iceberg detection method for synthetic aperture radar images in the Antarctic Sea area according to the present invention includes the following steps:

[0006] (1)Obtain SAR image data of the Antarctic Sea Area, and screen Level-1 SLC products with latitudes from 40°S to 90°S. The data mode is Wave Mode, and the polarization modes include HH and VV;

[0007] (2)Manually annotate the SAR images, and perform visual interpretation by combining geographical location information, shape, and texture features to generate an Antarctic iceberg marking dataset in COCO format;

[0008] (3)Construct a cascaded convolutional neural network model IB-Detector, including: Classification model: It consists of a Stem layer, Inception layers, and a fully connected layer. Among them, the Inception layer uses Inception modules with multiple groups of parallel convolutional kernels of different sizes to extract multi-scale features; Detection model: It consists of a backbone layer, a neck layer, and a head layer. The backbone layer contains a C3k2 module, the neck layer fuses multi-scale features, and the head layer integrates the CBAM attention mechanism to improve the detection accuracy by combining channel and spatial attention; Use the marked dataset to train the cascaded model, and use the output of the classification model as the input of the detection model;

[0009] (4)Use the test set to evaluate the model performance and output iceberg location, contour, and area information.

[0010] Furthermore, the SAR data preprocessing includes: Normalization processing, and the radar cross-section σ0 calculation formula is:

[0011] ;

[0012] where DN represents the original signal intensity received by the radar, and K represents the calibration constant; Randomly divide the labeled dataset into a training set, a validation set, and a test set according to a certain proportion.

[0013] Furthermore, the Inception layer of the classification model includes: 3 groups of Inception-A modules, which are parallelly stacked by multi-branch 3×3 and 5×5 convolutions to expand the channels to 256; 5 groups of Inception-B modules, which use asymmetric decomposition convolutions to increase the channels to 512; 2 groups of Inception-C modules, which fuse multi-scale features into 1280 channels; The auxiliary classifiers are located after the 11th layer and the 17th layer respectively, and are used for backpropagation gradient optimization.

[0014] Furthermore, the C3k2 module of the detection model extracts deep features through variable convolutional kernels and channel separation operations; The CBAM attention mechanism applies channel attention and spatial attention in sequence; The detection head adopts a bidirectional feature pyramid structure to fuse P3-P5 multi-scale features.

[0015] Furthermore, manual annotation includes: bounding box and pixel-level contour marking for iceberg targets; combining radar scattering, texture, and shape to eliminate false alarm targets.

[0016] An iceberg detection system for synthetic aperture radar images in the Antarctic Sea area according to the present invention includes:

[0017] A data acquisition module for screening and preprocessing Antarctic Sea area Wave Mode SAR data;

[0018] A labeling module for generating an iceberg labeling dataset in COCO format;

[0019] A cascaded model training module for jointly optimizing a classification model and a detection model;

[0020] An evaluation module for outputting the precision, recall rate, and F1 score of the detection result.

[0021] An electronic device according to the present invention includes a memory, a processor, and a computer program stored on the memory. When the processor executes the program, the steps of any one of the methods are implemented.

[0022] A computer-readable storage medium according to the present invention stores a computer program. When the program is executed by a processor, the steps of any one of the methods are implemented.

[0023] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: Using synthetic aperture radar WaveMode mode data has a higher resolution compared to other data; adopting a cascaded neural network solves the problem of extracting icebergs in specific scenarios and improves the detection accuracy of icebergs in complex scenarios; adopting a variety of Inception modules, using parallel convolutional kernels and pooling of different sizes, reduces the number of network parameters, reduces computational complexity, and maintains good performance at the same time; introducing the C3k2 module and the CBAM attention mechanism. C3k2 effectively extracts deep features by combining variable convolutional kernels and channel separation. The CBAM attention mechanism combines channel attention and spatial attention to improve recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flowchart of the present invention;

[0025] Figure 2 is a classification network structure diagram of the Antarctic Sea area iceberg detection model IB-Detector of the present invention;

[0026] Figure 3 is a detection network structure diagram of the Antarctic Sea area iceberg detection model IB-Detector of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0027] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings.

[0028] As Figure 1 shown, an iceberg detection method for synthetic aperture radar images in the Antarctic Sea area provided by an embodiment of the present invention includes the following steps:

[0029] Step 1: Obtain SAR Wave Mode image data, perform latitude screening on the original data, and extract SAR satellite data with a latitude of 40° - 90°S.

[0030] Step 2: Manually annotate the SAR images using Labelme and convert them into the COCO format to establish a labeled dataset of Antarctic iceberg SAR images.

[0031] Step 3: Construct an iceberg detection model IB-Detector for the Antarctic Sea area; the steps are as follows:

[0032] Step 3.1: As Figure 2 shown, Model 1 is a classification model, and the complete process from its input to output is as follows: The input layer (with a size of 299×299×3) first passes through the Stem backbone structure (composed of multiple layers of 3×3 convolutional and pooling operations, gradually reducing the dimension to a 35×35×192 feature map), and then enters 3 groups of Inception-A modules (repeatedly stacked 3 times through multi-branch parallel convolutions to expand the number of channels to 256). After that, the first auxiliary classifier is connected (after the 11th layer); then, it passes through 5 groups of Inception-B modules (repeatedly processed 5 times using the asymmetric decomposition convolution technique to increase the number of channels to 512), and the second auxiliary classifier is connected (after the 17th layer); then, it enters 2 groups of Inception-C modules (further fusing multi-scale features to expand the number of channels to 1280). Finally, it passes through the global average pooling layer (compressing the features into a 1×1×2048 dimension), the Dropout layer (to prevent overfitting), and the fully connected classification layer (outputting Softmax probabilities). The entire process realizes an efficient image classification task through modular stacking, multi-scale feature fusion, and auxiliary classification supervision.

[0033] Step 3.2: As Figure 3As shown in the figure, Model 2 is a detection model, and the input is the category image output by the classification model. Its overall structure is as follows: The Backbone main network takes multi-stage downsampling and feature enhancement as the core. First, it quickly realizes 4-fold downsampling through two 3×3 convolutions with a stride of 2 (channels 64→128), and then stacks the C3k2 module twice (channels 256) to extract intermediate features and further downsample to P3 / 8; then continue to generate P4 / 16 and P5 / 32 deep features through the C3k2 module (channels 512) and convolutional downsampling, and introduce the CBAM attention module after P5 (repeated 2 times) to enhance global context awareness; The Head detection head adopts a two-way feature fusion strategy. First, it upsamples P5 and concatenates it with P4, and refines it through the C3k2 module (channels 512→256) to generate P3 / 8 small target features. Then, after downsampling P3, it is concatenated with P4 and P5 features level by level, and the C3k2 module (channels 512→1024) is used to optimize the medium / large target detection layer. Finally, the Detect layer integrates the multi-scale features of P3-P5 and outputs the detection results.

[0034] Step 4: Use the Antarctic iceberg SAR image labeled dataset as the training input to train the model.

[0035] Step 5: Input the test set data into the trained model to evaluate the model effect. The performance evaluation of the model needs to combine multiple key indicators: Accuracy measures the overall prediction accuracy, and the calculation formula is:

[0036] ;

[0037] Among them, True Positive (TP) represents the number of targets correctly identified by the model (for example, correctly detecting icebergs), True Negative (TN) represents the number of non-targets correctly excluded by the model (for example, accurately determining non-iceberg areas), False Positive (FP) refers to the situation where the model misjudges non-targets as targets (such as mistaking waves for icebergs), and False Negative (FN) is the situation where the target is missed by the model (such as real icebergs not being detected).

[0038] However, it is prone to distortion in the case of class imbalance (for example, when the negative samples account for 90%, predicting all negative samples can obtain 90% accuracy). Precision focuses on the reliability of predicting positive classes and emphasizes reducing false positives (FP). Its formula is:

[0039] ;

[0040] It is applicable to scenarios that require high confidence such as spam detection; while Recall focuses on the coverage rate of real positive classes to reduce missed detections (FN), and is defined as:

[0041] ;

[0042] In the object detection task, AP50 calculates the average precision with an intersection over union (IoU) threshold of 0.5. The IoU formula is:

[0043] ;

[0044] It reflects the detection ability under loose localization; while AP50-95 (AP@0.5:0.95) comprehensively evaluates the robustness of the model under strict localization requirements through the mean value of IoU thresholds from 0.5 to 0.95 (step size 0.05).

Claims

1. An iceberg detection method for synthetic aperture radar images in the Antarctic Sea area, characterized in that, Including the following steps: (1) Obtain SAR image data of the Antarctic Sea Area, and screen Level-1 SLC products with latitudes from 40°S to 90°S. The data mode is Wave Mode, and the polarization modes include HH and VV. The SAR data preprocessing includes: the formula for normalizing the radar cross-section σ0 is: ； where DN represents the original signal intensity received by the radar, and K represents the calibration constant; randomly divide the labeled dataset into a training set, a validation set, and a test set according to a certain proportion; (2) Manually annotate the SAR images, visually interpret them in combination with geographical location information, shape, and texture features, and generate an Antarctic iceberg labeled dataset in COCO format. Manual annotation includes: based on synthetic aperture radar (SAR) remote sensing images, draw pixel-level bounding boxes along the iceberg contour to ensure coverage of the main body of the iceberg and the marginal brash ice area; exclude false alarm targets by combining radar scattering, texture, and shape; (3) Build a cascaded convolutional neural network model IB-Detector, including: a classification model: composed of a Stem layer, Inception layers, and a fully connected layer. Among them, the Inception layer uses Inception modules with multiple groups of parallel convolutional kernels of different sizes to extract multi-scale features; a detection model: composed of a backbone layer, a neck layer, and a head layer. The backbone layer contains a C3k2 module, the neck layer fuses multi-scale features, and the head layer integrates the CBAM attention mechanism to improve the detection accuracy by combining channel and spatial attention. Train the cascaded model using the labeled dataset, and use the output of the classification model as the input of the detection model. It includes the following steps: Step 3.1: The complete process from input to output of the classification model is as follows: The input layer size is 299×299×3. First, it passes through the Stem backbone structure composed of multiple layers of 3×3 convolutional and pooling operations, gradually reducing the dimension to a 35×35×192 feature map. Then, it enters 3 groups of Inception-A modules and repeats the stacking of multi-branch parallel convolutions 3 times to expand the number of channels to 256. After that, the first auxiliary classifier is connected after the 11th layer; then, through 5 groups of Inception-B modules, the asymmetric decomposition convolution technology is used to process 5 times, and the number of channels is increased to 512, and the second auxiliary classifier is connected after the 17th layer; then, it enters 2 groups of Inception-C modules to further fuse multi-scale features, and the number of channels is expanded to 1280. Finally, through the global average pooling layer, the features are compressed into a 1×1×2048 dimension, the Dropout layer is used to prevent overfitting, and the fully connected classification layer outputs the Softmax probability; the entire process realizes an efficient image classification task through modular stacking, multi-scale feature fusion, and auxiliary classification supervision; Step 3.2: Detection model. The input is the category image output by the classification model. The overall structure is as follows: The Backbone main network focuses on multi-stage downsampling and feature enhancement. First, it quickly realizes 4-fold downsampling through two 3×3 convolutions with a stride of 2, increasing the number of channels from 64 to 128. Subsequently, the C3k2 module is stacked twice to extract intermediate features with 256 channels and further downsample to P3 / 8. Then, through the C3k2 module with 512 channels and convolutional downsampling, deep features P4 / 16 and P5 / 32 are generated. After P5, the CBAM attention module is introduced and repeated twice to enhance global context awareness. The Head detection head adopts a two-way feature fusion strategy. First, P5 is upsampled and concatenated with P4, and after being refined by the C3k2 module with the number of channels reduced from 512 to 256, small target features of P3 / 8 are generated. Then, after P3 is downsampled and concatenated with P4 and P5 features level by level, the C3k2 module is used to optimize the medium / large target detection layer with the number of channels increasing from 512 to 1024. Finally, the Detect layer integrates multi-scale features of P3 - P5 to output the detection results. (4) Use the test set to evaluate the model performance and output information about the iceberg position, contour, and area.

2. A system for implementing the method according to claim 1, characterized in that, Including: A data acquisition module for screening and preprocessing Antarctic Sea Area Wave Mode SAR data; A labeling module for generating an iceberg labeled dataset in COCO format; A cascaded model training module for jointly optimizing the classification model and the detection model; An evaluation module for outputting the precision, recall rate, and F1 score of the detection results.

3. An electronic device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, When the processor executes the program, it implements the steps of the method described in Claim 1.

4. A computer-readable storage medium storing a computer program, characterized in that, When the program is executed by the processor, it implements the steps of the method described in Claim 1.

Images