Moving ship detection method based on spatial-temporal characteristics

Abstract

The invention provides a moving ship detection method based on spatial-temporal characteristics, which comprises the following steps: firstly, eliminating land and thick cloud regions through dual-threshold segmentation preprocessing, and adaptively extracting weak ship candidate targets in a water area by combining top-hat transform with sequential statistical filtering; combining detection results of each wave band, expanding the area of a ship candidate target, realizing signal enhancement and eliminating static false alarms, combining multiple frames of candidate targets, and carrying out filtering processing by utilizing a multi-direction rotation linear kernel by utilizing position offset characteristics of a moving target in five wave bands of a Gaofen-4 satellite due to imaging time difference; according to the method, real ships in smooth track distribution are reserved, isolated and randomly distributed false alarms or non-coherent track false alarms are eliminated, and stable and high-precision ship detection is effectively guaranteed while the low false alarm rate is kept.

Description

Technical Field

[0001] This invention relates to the field of ship inspection technology, and in particular to a method for detecting moving ships based on spatiotemporal characteristics. Background Technology

[0002] Maritime ship detection plays a vital role in promoting marine economic development, ensuring navigational safety, and strengthening integrated maritime management. Compared to the Automatic Identification System (AIS), which relies on active ship cooperation, satellite remote sensing technology offers advantages such as strong autonomy and wide coverage, providing robust support for maritime ship detection. Compared to stationary ships, moving ships better reflect the current active maritime traffic situation, directly reflecting navigation density and dynamic distribution, providing more timely and effective data support for waterway management and port scheduling. Furthermore, moving targets typically pose higher safety risks, such as collisions and unauthorized intrusions; detecting moving ships helps identify potential hazards early, enhancing maritime safety early warning capabilities.

[0003] Geostationary orbit satellites, leveraging their high orbital advantage, can acquire continuous image sequences with temporal resolution ranging from seconds to minutes from a fixed observation perspective. This enables highly timely and continuous monitoring of dynamic targets over a wide area, providing crucial support for maritime situational awareness. Taking China's Gaofen-4 satellite as an example, its large-scale CMOS staring camera boasts a spatial resolution of 50 meters, revisit capability ranging from seconds to minutes, and a wide-area imaging range of up to 250,000 square kilometers. It can achieve near real-time, continuous, long-term temporal observation of moving targets within a fixed sea area, demonstrating significant potential for continuous detection of ships moving over large areas of the sea. Despite these advantages, using geostationary orbit image sequences for ship detection still faces numerous challenges. On one hand, limited by orbital altitude, ships appear small and have weak signals in images, making them susceptible to interference from complex sea surface backgrounds and leading to missed detections. On the other hand, false alarms such as fragmented clouds and islands exhibit similar characteristics to ships in images, easily triggering a series of false detections.

[0004] Moving target detection technology based on image sequences can be divided into methods that detect each frame of the image independently and methods that use multiple frames of images for joint detection.

[0005] Current methods for independent detection of each image frame mainly utilize detection methods based on target grayscale features and those based on deep learning networks. Target grayscale feature-based methods leverage statistical or artificially designed features that distinguish the target grayscale value from background noise to enhance the contrast between the ship and background signals, thus achieving target detection. Methods using a two-parameter constant alarm rate (CPNR) approach perform threshold segmentation based on the statistical distribution difference between the target and the Gaussian-distributed clutter in the surrounding background, thereby achieving target detection. Methods based on local peak signal-to-noise ratio (PSNR) enhance the target by quantifying the peak intensity of the target signal relative to the intensity of background noise. Li Xiaobo et al. used multi-structure, multi-scale morphological filtering for background suppression, followed by adaptive threshold segmentation and self-organizing clustering to obtain candidate targets. Additionally, some studies have fused ship automatic identification systems with Gaofen-4 imagery to improve target detection accuracy. Deep learning-based target detection methods extract features from different levels of the ship during training, thereby improving the robustness of target detection. Wang et al. utilized the sequential imaging characteristics of Gaofen-4 across different bands, employing convolutional neural networks to extract spatiotemporal joint features of the ship from multi-band images to achieve target detection. Liu et al. achieved accurate ship detection at the pixel level based on an improved Mask R-CNN model.

[0006] Compared to single-scene images, sequential images contain multi-temporal features that are beneficial for expanding the dimension of target features, enabling better detection of moving targets. Currently, deep learning methods have achieved good results in this field. Xu et al. synthesized three consecutive frames of images, extracted target features based on an attention mechanism in COCO-Net for target enhancement, and used multi-scale feature fusion to achieve target detection, further addressing the problem of small ship targets making it difficult to utilize texture and shape information, and effectively eliminating noise interference. Feng et al. used CNN to extract temporal fusion features from three consecutive frames of images and combined them with an hourglass network and mismatch suppression for moving target detection. Pi et al. fused the motion features of targets in adjacent multi-frame images to promote semantic feature extraction and introduced a transformer to optimize target detection features. Commonly used non-deep methods for target detection in sequential images mainly include frame difference and background modeling. Frame difference effectively separates moving objects from the background by thresholding the differences between consecutive frames. For example, Wei et al. took the frame difference of three consecutive frames, modeled local noise in the frame difference results, and then extracted the target through local thresholding. Background modeling methods reconstruct the background using Gaussian mixture models, background extraction, low-rank and structured sparse decomposition, and extended low-rank and structured sparse decomposition, and then extract the target by subtracting the background. Some studies combine both methods for target detection; for example, Yin et al. extracted moving targets by pairwise difference analysis over three consecutive frames, and then combined this with low-rank matrix completion for background modeling to achieve target-background separation.

[0007] However, the above-mentioned technical solutions still have many problems when applied in real-world maritime scenarios, as follows:

[0008] For frame-by-frame independent target detection methods, in low-to-medium resolution remote sensing imagery, detection based on grayscale or manually designed features struggles to effectively distinguish ships from interfering targets such as scattered clouds and islands, easily leading to a high false alarm rate. Although depth features have stronger expressive power, ships typically lack obvious texture, structure, and shape features, and their appearance in images is easily affected by factors such as speed, heading, and background brightness, further reducing the generalization ability of the detection network. Furthermore, single-frame detection fails to fully utilize the spatiotemporal features of ships contained within the image sequence, limiting the improvement of overall detection accuracy.

[0009] While multi-frame joint detection can fully exploit the temporal motion characteristics of ships, the sheer volume of remote sensing sequence data means that deep learning-based detection networks often face high computational and storage costs, making it difficult to meet the demands of real-time processing and large-scale maritime applications. Traditional non-deep methods, such as frame differencing or background modeling, are easily affected by brightness variations and dynamic backgrounds, especially in wave-filled or complex sea environments, leading to numerous false alarms and impacting detection reliability.

[0010] Most current experiments are conducted only under ideal conditions of no clouds or large-area cloud cover, without fully considering the interference of widespread fragmented clouds on detection algorithms. In real-world marine environments where fragmented clouds are frequent, fragmented clouds and moving ships are often highly similar in size, brightness, and texture, which can easily generate a large number of false alarms, thus significantly reducing the stability and robustness of detection algorithms. Summary of the Invention

[0011] The purpose of this invention is to solve the problems of high false alarm rate and low detection rate in the prior art, and to propose a moving ship detection method based on spatiotemporal features.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] A method for detecting moving ships based on spatiotemporal features includes:

[0014] Step 1: For the five band images of Gaofen-4 satellite, preprocessing is performed on each band using double threshold segmentation to remove land and thick cloud areas from the images. Then, top-hat transform combined with sequential statistical filtering is used to adaptively extract candidate ship targets from each band image.

[0015] Step 2: Utilize the positional shift characteristics of moving targets in the five bands of the Gaofen-4 satellite due to imaging time differences, merge the ship candidate targets extracted from each band, and remove static false alarms to obtain the ship candidate targets extracted from a single frame image.

[0016] Step 3: Merge the candidate targets extracted from multiple frames of images. Based on the merged results, use a multi-directional rotating linear kernel for filtering to retain real ships with smooth trajectory distribution and remove false alarms that are isolated, randomly distributed, or have inconsistent trajectories, thus obtaining the multi-frame detection results.

[0017] Furthermore, step 1 specifically includes the following steps:

[0018] Step 11: Calculate the global land-sea segmentation threshold T in the image using the Otsu algorithm. Otsu ;

[0019] Step 12: Divide the image into several 20*20 pixel local windows, and use the sequential statistical filtering algorithm to solve the local threshold T in each local window. local ;

[0020] Step 13: Compare T in each local window Otsu and T local If T Otsu <T local Then T Otsu To divide the area into bright and dark regions using a segmentation threshold, if T local <T Otsu Then T local Bright and dark areas are divided according to the segmentation threshold. Bright areas with an area greater than 1000 pixels are identified as land and bright areas with an outline length greater than 20 pixels are identified as thick clouds. Land and thick clouds are removed to obtain the preprocessed image.

[0021] Step 14: Use morphological top-hat transformation to process the imaging information of each band in the preprocessed image to obtain the enhanced image;

[0022] Step 15: In the enhanced image, take the 90th percentile pixel value in brightness ranking within each 20*20 pixel local window as the background estimation threshold, and identify pixels with brightness above the background estimation threshold as candidate targets. After the identification is completed, identify the candidate ship targets in the detected image.

[0023] Furthermore, the solution local threshold T local The formula is:

[0024] T local =DN 0～90％ +β

[0025] Among them, DN 0～90％ This represents the pixel value ranking 90th in brightness within the window, where β is an adaptive threshold. Further, the formula for the adaptive threshold β is:

[0026] β=DN 50％+3*std(DN)

[0027] Among them, DN 50％ std(DN) represents the median brightness of the pixels within the window, and std(DN) represents the standard deviation of the brightness of the pixels within the window.

[0028] Furthermore, step 2 specifically includes the following steps:

[0029] Step 21: Merge the candidate ship targets extracted from each waveband;

[0030] Step 22: Perform conditional judgment on the candidate ship targets in the multi-band merging results. If the area of ​​the candidate ship target does not expand or is less than the set threshold, it is judged as a static false alarm and is removed. After the removal is completed, the candidate ship targets extracted from the single frame image are obtained.

[0031] Furthermore, the thick cloud extracted in step 1 is expanded using a mathematical morphological dilation operation, and the portion of the ship candidate target extracted in step 2 that has spatial intersection with the expanded cloud region is removed.

[0032] Furthermore, step 3 specifically includes the following steps:

[0033] Step 31: For each frame of extracted candidate ship targets, calculate its center coordinates as the center point of the corresponding candidate ship target;

[0034] Step 32: Merge the center points extracted from all frames to construct a binary spatiotemporal point set. Then, perform filtering based on a multi-directional rotating linear kernel and conduct directional statistical analysis on the spatiotemporal point set. Ship candidate targets that do not have linear and consistent motion characteristics are regarded as false alarms and are eliminated. The obtained series of ship candidate targets with smooth trajectories are restored to the corresponding frame images to obtain multi-frame detection results.

[0035] Furthermore, the specific steps of performing filtering based on a multi-directional rotating linear kernel, conducting directional statistical analysis on the spatiotemporal point set, and filtering out candidate ship targets that do not have linearly consistent motion characteristics as false alarms include:

[0036] Step 321: The multi-directional rotating linear kernel rotates and scans between 0° and 180° with a step size of 10° to cover different possible ship movement directions, and the kernel size corresponding to each direction is adaptively set according to the average time interval between adjacent frames in the image sequence.

[0037] Step 322: For each center point, count the number of other center points falling within the core range in each direction, and select the direction with the most points from the 18 directions as the main motion direction of the candidate ship target corresponding to that center point. If the number of valid points in the main motion direction is less than three, the candidate target does not have linear and consistent motion characteristics, is judged as a false alarm and is eliminated.

[0038] Furthermore, the formula for filtering using the multi-directional rotating linear kernel is as follows:

[0039] Θ = {θ1, θ2, ..., θ N}, N=18

[0040]

[0041] Where Θ represents the set of directions, Indicates direction θ k The corresponding filter kernel The function, if the center point (x,y) falls on the filter kernel In the template, the value is 1 if it is true and 0 otherwise, and I represents a binary spatiotemporal point set. Indicates direction θ k The corresponding convolution response, θ max (x c ,y c ) represents each center point (x) c ,y c The corresponding main motion direction, R max (x c ,y c ) indicates the number of valid points in the main direction.

[0042] Compared with existing technologies, the advantages of this invention are:

[0043] 1. This invention utilizes the temporal imaging characteristics of the five spectral bands of the Gaofen-4 satellite across panchromatic, visible, and near-infrared wavelengths through multi-band signal fusion. This merges signals from the same ship across different bands, significantly improving the overall strength of the ship's signal. This strategy naturally distinguishes between moving targets and static false alarms. Since fixed objects such as islands or platforms maintain a consistent position across all bands, their area does not significantly expand after merging, making it easier to eliminate static false alarms. Simultaneously, this multi-band fusion strategy also integrates the different reflection characteristics of different bands towards clouds, water, and metallic targets, improving the detection stability of ship signals against complex backgrounds.

[0044] 2. This invention relies on the continuous motion trajectory of ships in time-series images. By designing linear filter kernels in different directions, it performs linear distribution statistics on candidate targets in the merged multi-frame images, identifies real ships with smooth trajectory distribution, and eliminates false alarms that are isolated, randomly distributed, or have inconsistent trajectories. This strategy effectively ensures stable and high-precision ship detection while maintaining a low false alarm rate. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the overall technical process of a motion ship detection method based on spatiotemporal features proposed in this invention.

[0046] Figure 2 This is a schematic diagram of the multi-directional rotating kernel linear filter proposed in this invention. Detailed Implementation

[0047] The invention will now be further explained with reference to the accompanying drawings.

[0048] like Figure 1 As shown, the present invention provides a method for detecting moving ships based on spatiotemporal features, including:

[0049] Step 1: Ship signal morphological enhancement and adaptive segmentation: For the five band images of Gaofen-4 satellite, preprocessing is performed on each band using double threshold segmentation to remove land and thick cloud regions from the images. Then, top-hat transform combined with sequential statistical filtering is used to adaptively extract candidate ship targets from each band image.

[0050] Step 2: Enhancement of candidate target signals based on multi-band imaging time migration: Utilizing the position migration characteristics of moving targets in the five bands of the Gaofen-4 satellite due to imaging time differences, the ship candidate targets extracted from each band are merged and static false alarms are removed to obtain the ship candidate targets extracted from a single frame image.

[0051] Step 3: False alarm removal based on multi-frame motion coherence multi-directional linear filtering: The candidate targets extracted from multiple frames of images are merged. Based on the multi-frame merging results, multi-directional rotating linear kernels are used for filtering to retain real ships with smooth trajectory distribution and remove isolated, randomly distributed, or non-coherent false alarms to obtain multi-frame detection results.

[0052] In step 1, before conducting ship target detection, a dual-threshold segmentation algorithm combining global and local thresholding is first used to effectively extract land and cloud regions from the image. Specific steps include:

[0053] First, the global land-sea segmentation threshold T of the image is calculated using the Otsu algorithm. Otsu The image was initially divided into bright areas (land and clouds) and dark areas (water).

[0054] The image is then divided into several 20*20 pixel local windows, and the local threshold is solved in each window using a sequential statistical filtering algorithm.

[0055] Finally, the segmentation threshold for each window is set to T. loacl With T Otsu The smaller one, to achieve the distinction extraction of different goals: when T Otsu <T local Then T Otsu Dividing the area into bright and dark regions using a segmentation threshold enables complete extraction of bright targets. When T local <T Otsu Then T local Dividing bright and dark areas by setting a segmentation threshold helps to preserve weak targets. Then, based on the generated binary image, all connected regions are further extracted: bright areas with an area of ​​more than 1,000 pixels are identified as land, bright areas with a contour length of more than 20 pixels are identified as thick clouds, and other small bright areas (ships or fragmented clouds) are retained for subsequent detection. This preprocessing process can effectively remove the interference of land and thick clouds on subsequent ship detection.

[0056] In the above embodiments, the local threshold T is solved. local The formula is:

[0057] T local =DN 0～90％ +β

[0058] Among them, DN 0～90％ This represents the pixel value ranking 90th in brightness within the window, where β is an adaptive threshold.

[0059] The formula for the adaptive threshold β is:

[0060] β=DN 50％ +3*std(DN)

[0061] Among them, DN 50％ std(DN) represents the median brightness of the pixels within the window, and std(DN) represents the standard deviation of the brightness of the pixels within the window.

[0062] The introduction of an adaptive threshold β effectively prevents water bodies from being mistakenly extracted. The adaptive threshold β is adaptively set to the sum of the median brightness of pixels within the window and three times the standard deviation of the brightness of that window, thereby achieving dynamic adjustment of the water body background threshold.

[0063] Next, morphological top-hat transform was applied to the five spectral band images of the Gaofen-4 satellite to enhance the signals of darker, smaller-scale ships against the sea background. This transform helps to highlight the brightness difference between the ship target and its surrounding background and suppress background interference. On the enhanced image, an adaptive thresholding method based on sequential statistical filtering was used to extract candidate targets. Specifically, the 90th percentile pixel value in brightness within a 20*20 pixel local window was used as the background estimation threshold, and pixels above this threshold were identified as candidate targets. To prevent the water background from being mistakenly extracted, an adaptive threshold β was also introduced, which was adaptively set to the sum of the median pixel brightness within the window and three times the standard deviation of the window's brightness.

[0064] Ultimately, all false alarms that were identified as being in land or thick cloud areas during the preprocessing stage were eliminated, leaving only valid candidate ship targets located in water.

[0065] During the five-band imaging process, the five rotating filter wheels on the Gaofen-4 satellite separate the light into five bands in the order of panchromatic, blue, green, red, and near-infrared. This means that after imaging through a filter in one band, the filter wheel switches filters to image the next band. Because the rotation of the filter wheels takes time, Gaofen-4 exhibits both sequential and time-division imaging characteristics across the five bands. Therefore, the imaging position of moving targets varies across different bands.

[0066] Existing research, combining AIS recorded data with ship information extracted from the Gaofen-4 satellite, has calculated that the imaging time interval for each band of the Gaofen-4 satellite is approximately 10 seconds. Therefore, the interval between the panchromatic and near-infrared bands can reach 40 seconds. Based on the estimated speed of a typical ship, this time difference can cause a spatial shift of approximately 8 pixels in the panchromatic and near-infrared images.

[0067] Therefore, in step 2, based on the positional shift characteristics of the five bands of the Gaofen-4 satellite caused by imaging time differences, a candidate target signal enhancement method based on multi-band imaging time-series shift is provided. Specifically, for enhancing candidate ship targets, this method includes:

[0068] First, merge the candidate ship targets extracted from each waveband.

[0069] Next, conditional judgments are made on the candidate ship targets in the multi-band merging results. If the area of ​​the candidate ship target does not expand or is less than the set threshold, it is judged as a static false alarm and is removed. After the removal is completed, the candidate ship targets extracted from the single frame image are obtained.

[0070] The extracted candidate ship targets may still contain some fragmented clouds with similar size and brightness characteristics to ships. Some of these fragmented clouds are distributed at the edges of large-scale thick clouds. To further suppress such interference:

[0071] This invention uses mathematical morphological dilation to expand the thick cloud extracted in step 1, and removes the portion of the ship candidate target extracted in step 2 that has spatial intersection with the expanded cloud region.

[0072] like Figure 2 As shown, real ship targets typically move along continuous, smooth trajectories, exhibiting significant coherence and directional consistency in their motion across time and space. Therefore, this design proposes a false alarm rejection method based on multi-frame linear motion consistency, aiming to effectively identify and eliminate isolated false alarms lacking coherent trajectories.

[0073] Therefore, in step 3, the center coordinates of the candidate ship targets detected in each frame of the continuous image sequence are first extracted. These center coordinates are then used as the center points, and the center points extracted from all frames are merged to construct a binary spatiotemporal point set. Based on this, a multi-directional rotating linear kernel is designed to perform filtering processing, and directional statistical analysis is conducted on the spatiotemporal point set to determine whether the candidate targets have linearly consistent motion characteristics within a certain time window.

[0074] The rotation kernel scans in 10° increments, ranging from 0° to 180°, covering various possible ship movement directions. The kernel size for each direction is adaptively set based on the average time interval between adjacent frames in the image sequence to ensure that the filter kernel size adapts to the actual velocity range of the target on a time scale.

[0075] For each center point, the number of other center points falling within the kernel range in each direction is counted, and the direction with the most points from the 18 directions is selected as the main motion direction of the candidate ship target corresponding to that center point. If the number of valid points in the main direction is less than three, it indicates that the target lacks the necessary motion continuity, is judged as a false alarm, and is eliminated, i.e., multi-directional filtering.

[0076] After the removal is completed, the series of ship candidate targets with smooth trajectories are restored to the corresponding frame images to obtain multi-frame detection results.

[0077] This method fully utilizes the spatiotemporal consistency of ship targets, enhances the ability to identify real motion trajectories, effectively suppresses random false alarms caused by cloud fragments, wave reflections, or imaging noise, and ensures the accuracy of ship detection results.

[0078] This invention first identifies weak candidate ship targets in the detection image through ship signal morphology enhancement and adaptive segmentation. Then, by using multi-band signal fusion, it leverages the temporal imaging characteristics of the five spectral bands of the Gaofen-4 satellite in panchromatic, visible, and near-infrared light to merge the signals of the same ship in different bands, significantly improving the overall strength of the ship signal and eliminating static false alarms. Finally, it eliminates false alarms by using linear filtering based on multi-frame motion coherence, removing isolated, randomly distributed, or non-coherent false alarms. While maintaining a low false alarm rate, it effectively ensures stable and high-precision ship detection.

[0079] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative and not exhaustive. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A method for detecting moving ships based on spatiotemporal features, characterized in that, include: Step 1: For the five band images of Gaofen-4 satellite, preprocessing is performed on each band using double threshold segmentation to remove land and thick cloud areas from the images. Then, top-hat transform combined with sequential statistical filtering is used to adaptively extract candidate ship targets from each band image. Step 2: Utilize the positional shift characteristics of moving targets in the five bands of the Gaofen-4 satellite due to imaging time differences, merge the ship candidate targets extracted from each band, and remove static false alarms to obtain the ship candidate targets extracted from a single frame image. Step 3: Merge the candidate targets extracted from multiple frames of images. Based on the merged results, use a multi-directional rotating linear kernel for filtering to retain real ships with smooth trajectory distribution and remove false alarms that are isolated, randomly distributed, or have inconsistent trajectories, thus obtaining the multi-frame detection results.

2. The method for detecting moving ships based on spatiotemporal features according to claim 1, characterized in that: The specific steps of step 1 include: Step 11: Calculate the global land-sea segmentation threshold T in the image using the Otsu algorithm. Otsu ; Step 12: Divide the image into several 20*20 pixel local windows, and use the sequential statistical filtering algorithm to solve the local threshold T in each local window. local ; Step 13: Compare T in each local window Otsu and T local If T Otsu <T local Then T Otsu To divide the area into bright and dark regions using a segmentation threshold, if T local <T Otsu Then T local Bright and dark areas are divided according to the segmentation threshold. Bright areas with an area greater than 1000 pixels are identified as land and bright areas with a length greater than 20 pixels are identified as thick clouds. Land and thick clouds are removed to obtain the preprocessed image. Step 14: Use morphological top-hat transformation to process the imaging information of each band in the preprocessed image to obtain the enhanced image; Step 15: In the enhanced image, take the 90th percentile pixel value in brightness ranking within each 20*20 pixel local window as the background estimation threshold, and identify pixels with brightness above the background estimation threshold as candidate targets. After the identification is completed, identify the candidate ship targets in the detected image.

3. The method for detecting moving ships based on spatiotemporal features according to claim 2, characterized in that: The local threshold T local The formula is: T local =DN 0～90％ +β Among them, DN 0～90％ This represents the pixel value ranking 90th in brightness within the window, where β is an adaptive threshold.

4. The method for detecting moving ships based on spatiotemporal features according to claim 3, characterized in that: The formula for the adaptive threshold β is: β=DN 50％ +3*std(DN) Among them, DN 50％ std(DN) represents the median brightness of the pixels within the window, and std(DN) represents the standard deviation of the brightness of the pixels within the window.

5. The method for detecting moving ships based on spatiotemporal features according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 21: Merge the candidate ship targets extracted from each waveband; Step 22: Perform conditional judgment on the candidate ship targets in the multi-band merging results. If the area of ​​the candidate ship target does not expand or is less than the set threshold, it is judged as a static false alarm and is removed. After the removal is completed, the candidate ship targets extracted from the single frame image are obtained.

6. The method for detecting moving ships based on spatiotemporal features according to claim 5, characterized in that: The thick cloud extracted in step 1 is expanded using a mathematical morphological dilation operation, and the portion of the ship candidate target extracted in step 2 that has spatial intersection with the expanded cloud region is removed.

7. The method for detecting moving ships based on spatiotemporal features according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 31: For each frame of extracted candidate ship targets, calculate its center coordinates as the center point of the corresponding candidate ship target; Step 32: Merge the center points extracted from all frames to construct a binary spatiotemporal point set. Then, perform filtering based on a multi-directional rotating linear kernel and conduct directional statistical analysis on the spatiotemporal point set. Ship candidate targets that do not have linear and consistent motion characteristics are regarded as false alarms and are eliminated. The obtained series of ship candidate targets with smooth trajectories are restored to the corresponding frame images to obtain multi-frame detection results.

8. The method for detecting moving ships based on spatiotemporal features according to claim 7, characterized in that: The specific steps of filtering based on a multi-directional rotating linear kernel, performing directional statistical analysis on the spatiotemporal point set, and filtering out candidate ship targets that do not have linearly consistent motion characteristics as false alarms include: Step 321: The multi-directional rotating linear kernel rotates and scans between 0° and 180° with a step size of 10° to cover different possible ship movement directions, and the kernel size corresponding to each direction is adaptively set according to the average time interval between adjacent frames in the image sequence. Step 322: For each center point, count the number of other center points falling within the core range in each direction, and select the direction with the most points from the 18 directions as the main motion direction of the candidate ship target corresponding to that center point. If the number of valid points in the main motion direction is less than three, the candidate target does not have linear and consistent motion characteristics, is judged as a false alarm and is eliminated.

9. The method for detecting moving ships based on spatiotemporal features according to claim 7, characterized in that: The formula for filtering using the multi-directional rotating linear kernel is: Θ={θ1,θ2,…,θ N },N=18 Where Θ represents the set of directions, Indicates direction θ k The corresponding filter kernel The function, if the center point (x,y) falls on the filter kernel In the template, the value is 1 if it is true and 0 otherwise, and I represents a binary spatiotemporal point set. Indicates direction θ k The corresponding convolution response, θ max (x c ,y c ) represents each center point (x) c ,y c The corresponding main motion direction, R max (x c ,y c ) indicates the number of valid points in the main direction.

Images