Dark ship identification method for satellite image and AIS ship association driven by position uncertainty

Abstract

The present application relates to a kind of dark ship identification method for satellite image and AIS ship association driven by position uncertainty, belong to satellite remote sensing image recognition technical field.The method includes: AIS data is cleaned and static information is backfilled, the instantaneous position of each ship at satellite imaging moment is calculated;Position error is decomposed into dynamic calculation error and antenna offset error and is orthogonal synthesis, the position uncertainty of each AIS ship is obtained;Inverse variance self-adaptive weighted association cost function driven by the position uncertainty is constructed, global optimal matching is carried out using Hungarian algorithm;For the image spot that has not been successfully matched to AIS ship, generate dark ship list.The present application realizes the dynamic concession of position and size discrimination right by position uncertainty, effectively improves the association accuracy of satellite image and AIS ship and dark ship identification precision.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method for identifying dark ships by associating satellite imagery with AIS vessels driven by position uncertainty. Background Technology

[0002] Currently, the Automatic Identification System (AIS) is the most commonly used method for monitoring the dynamics of ships at sea. AIS is a digital message actively broadcast by ships, providing clear information including the ship's identity, nationality, tonnage, and historical trajectory. However, it is also a non-intuitive digital stream, unable to visually display the spatial distribution of ships on the sea surface. Furthermore, as a cooperative observation system that relies on ships actively transmitting data, it often faces the challenge of ships suspected of illegal activities deliberately shutting down their systems or tampering with information to evade supervision, thus becoming undetected vessels.

[0003] Satellite optical remote sensing imagery can visually display a ship's position and accurately extract its centroid, length, beam, and heading. Therefore, satellite remote sensing imagery with non-cooperative observation capabilities has become a crucial tool for detecting ships in the shadows.

[0004] However, satellite remote sensing imagery is a static pixel map and cannot directly provide ship identification information. In practical applications, there is a significant spatiotemporal asynchrony between satellite imaging and AIS message reception. Coupled with the influence of ocean currents, wind direction, and ship movement, there is a significant positional uncertainty between the AIS extrapolated position and the actual observed position in the satellite imagery. This is especially true in densely populated sea areas such as ports and busy waterways, where traditional nearest neighbor matching algorithms are prone to mismatches and omissions, leading to false alarms of unknown ships.

[0005] Therefore, accurately decoupling the spatiotemporal asynchrony and achieving matching between satellite remote sensing patches and AIS vessels, as well as precise identification of smuggled vessels, is a current technical bottleneck in intelligent maritime traffic supervision and non-cooperative target identification.

[0006] This invention proposes a method for identifying dark ships by associating satellite imagery with AIS-independent vessels driven by position uncertainty. It aims to achieve reliable fusion of heterogeneous data under complex sea conditions through error modeling and optimized assignment, providing an effective technical approach for accurate association of maritime targets and screening of ships without AIS. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for identifying dark ships by associating satellite imagery with AIS ships driven by position uncertainty.

[0008] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0009] A method for identifying dark ships by associating satellite imagery with AIS vessels driven by position uncertainty, the key technology of which includes the following steps:

[0010] Step S1: Clean the AIS data to remove invalid messages, and backfill the ship size characteristics from the static messages into the corresponding dynamic messages based on MMSI; calculate the satellite imaging time for each AIS ship j. instantaneous position ;

[0011] Step S2: Extract the dimensional features from the AIS ship message: distance A from the antenna to the bow, distance B to the stern, distance C to the port side, and distance D to the starboard side. Combine this with the heading COG, calculate the eastward correction ΔE and northward correction ΔN of the antenna position relative to the ship's centroid. By superimposing the aforementioned corrections, the centroid of the hull is obtained at... Corrected instantaneous position coordinates , ;

[0012] Step S3: Determine the uncertainty of dynamic estimation error based on the dynamic message quality of each AIS vessel j. Based on the execution status of antenna position correction, determine the uncertainty of antenna offset error. ,Will and As two independent uncertainty components, the orthogonally combined positional uncertainty ;

[0013] Step S4: Construct the association cost function between satellite remote sensing image patch i and AIS ship j ,

[0014] ,in, The square of the positional residual.

[0015] ,

[0016] In the formula, , The coordinates of the geometric center of satellite remote sensing image patch i; and The measured length and width of satellite remote sensing image patch i. and The theoretical length and width of the ship carried in the AIS ship j message. and Measurement uncertainty when extracting length and width from remote sensing images;

[0017] Based on the aforementioned association cost function, a cost matrix is ​​constructed. The Hungarian algorithm is used to solve the global one-to-one optimal assignment between satellite remote sensing image patches and AIS vessels, and the matching association pairs and unmatched image patches are output. For image patches that fail to match AIS vessels, a list of unmatched vessels is generated.

[0018] In one embodiment of the present invention, in step S1, the AIS ship j at the satellite imaging time instantaneous position The calculation method is as follows:

[0019] for For vessels that send messages both forward and backward, linear interpolation is used to calculate their... Position coordinates at that moment;

[0020] for For vessels with messages sent from only one side, their location can be estimated using a trajectory extrapolation method. Position coordinates at that moment;

[0021] For stationary ships with a ground speed (SOG) below a preset speed threshold, ±T is the time window. All AIS reporting locations within the window are extracted, and the median is calculated as the median. The location coordinates at a given time, where T is 30~50 minutes.

[0022] As one embodiment of the present invention, the calculation method for the eastward correction ΔE and the northward correction ΔN in step S2 is as follows:

[0023] ;

[0024] ;

[0025] The ship's centroid is The corrected instantaneous position coordinates:

[0026] , ,in , for The coordinates.

[0027] In one embodiment of the present invention, in step S2, the centroid of the hull is... The time correction is only performed when the heading to the ground is valid and the ship length is not less than a preset ship length threshold, wherein the preset ship length threshold ranges from 55 to 65 meters.

[0028] As one embodiment of the present invention, for For vessels with messages on one side, the extrapolation time should not exceed 10 minutes; otherwise, the corresponding MMSI-related messages will be removed.

[0029] In step S3, for a ship at sea, the error uncertainty is dynamically calculated. ,in, The uncertainty of the random error in positioning is set to a preset fixed value; The uncertainty of the position estimation error;

[0030] Among them, for Ships that are navigating with messages sent from both the front and rear. ; This refers to the message interval used when employing linear interpolation.

[0031] for Ships that communicate via messages on one side ; For message time and The interval time, K is a constant;

[0032] For stationary ships with a ground speed (SOG) below a preset speed threshold , where s is the satellite imaging time The standard deviation of AIS data coordinates within a preset time period, where n is the number of valid AIS data.

[0033] As one embodiment of the present invention, for For vessels that communicate via messages on one side, the value of K ranges from 1.0 to 2.0.

[0034] For stationary ships, the preset speed threshold is 3 knots, and the preset time period is 30-50 minutes.

[0035] As one embodiment of the present invention, in step S3, the... The value is 8 meters when the centroid correction has been applied; and 0.25 times the ship length when no correction has been applied.

[0036] In one embodiment of the present invention, in step S4, and The value range is 3 to 5m.

[0037] In one embodiment of the present invention, in step S4, the positional residual An adaptive threshold is set. for The square root,

[0038] like ≥ Then skip other calculations and assign the cost of both. Set it to infinity directly.

[0039] In one embodiment of the present invention, the upper limit (max) is set to 300 m and the lower limit (min) is set to 60 m.

[0040] The beneficial effects of adopting the above technical solution are as follows:

[0041] This invention provides a method for identifying hidden ships by associating satellite imagery with AIS vessels based on position uncertainty. Through gated antenna centroid correction, the antenna position of the AIS message is precisely corrected to the geometric center of the ship, eliminating fixed deviations caused by different antenna installation positions. By modeling position uncertainty through multi-source orthogonal synthesis, a dynamic position uncertainty is calculated for each AIS vessel. Based on this position uncertainty, an inverse variance adaptive weighted multi-feature association cost function is constructed. When AIS position information is highly reliable, strict position alignment is required during matching; conversely, when the position information is unreliable, position deviations are tolerated, and the system relies more on other features such as the ship's size and heading for judgment. The method provided by this invention significantly improves the accuracy and robustness of automatic association between satellite imagery and AIS vessels, and on this basis, achieves efficient and accurate identification of hidden ships. Attached Figure Description

[0042] Figure 1 This is a flowchart of the dark ship identification method of the present invention.

[0043] Figure 2 This is a schematic diagram of the corrected geometric relationship of the antenna centroid.

[0044] Figure 3 This is a schematic diagram of the cross curve of ship position extrapolation error and time. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the invention will be described clearly and completely below in conjunction with specific embodiments.

[0046] This invention proposes a method for identifying hidden ships by associating satellite imagery with AIS vessels driven by position uncertainty. This method uses target position uncertainty to drive dynamic weights and adaptive thresholds instead of fixed weights and thresholds; it introduces gated antenna centroid correction to reduce the spatial system bias of large ships; and it uses median time window positioning for stationary and anchored targets to suppress random position jumps. This significantly improves the accuracy and robustness of automatic association between satellite imagery and AIS vessels, and on this basis, achieves efficient and accurate identification of hidden ships.

[0047] The specific steps and algorithms of the method described in this invention are described in detail below through embodiments. It should be noted that the position coordinates mentioned in this invention are all Cartesian coordinates in meters.

[0048] Example

[0049] like Figure 1 As shown, a method for identifying dark ships by associating satellite imagery with AIS vessels driven by position uncertainty includes the following steps:

[0050] Step S1: AIS data preprocessing and instantaneous position calculation

[0051] The AIS data (messages) are cleaned to remove invalid messages. Specifically, the data is grouped based on MMSI (Marine Mobile Service Identifier) ​​and sorted by time. Illegal MMSIs (valid range 201000000-775999999), invalid latitude and longitude, and zero-value records are removed. The ship size characteristics in the static messages are backfilled into the corresponding dynamic messages according to MMSI.

[0052] Estimate the position of each AIS vessel j (the j-th AIS vessel target) at the satellite imaging time. instantaneous position The specific method is as follows:

[0053] for For vessels that send messages both forward and backward, linear interpolation is used to calculate their... The position coordinates at a given time are calculated using the following formula:

[0054] ;

[0055] in, and They are respectively Most recent message timestamps and Location coordinates; maximum allowable interval between both sides is 30 minutes, if in If no corresponding AIS data is available for this MMSI within 30 minutes before or after, it is considered that the AIS vessel corresponding to this MMSI is in [location missing]. If there is no AIS data before or after this MMSI, the AIS data corresponding to this MMSI is invalid, and no further matching will be performed; if in If only one side has corresponding AIS data within a 30-minute period, then it will be handled according to... Message processing is performed on one side; the linear interpolation method described in this embodiment is based on the following assumption: within the time interval between two consecutive messages ( The ship maintains approximately uniform linear motion. Since this time interval is limited to less than 60 minutes, this assumption holds true in most real-world navigation scenarios. Compared to directly using instantaneous speed and heading for estimation, linear interpolation can precisely anchor the positions of known messages at both ends, avoiding endpoint errors.

[0056] for For vessels with messages sent from only one side, their location can be estimated using a trajectory extrapolation method. Position coordinates at that moment;

[0057] Only When there is a message ahead, the forward trajectory estimation algorithm is as follows:

[0058] extract The last valid message point time latitude and longitude coordinates Ground speed and ground heading ;

[0059] extrapolation time difference is The algorithm assumes that the ship is Maintaining a constant speed and heading over a period of time, calculate forward along the great circle route. Displacement corresponding to time Set the starting point coordinates Displacement S and heading relative to the ground Substituting into the forward algorithm for the great circle route, we obtain... Extrapolation position of time ;

[0060] For only When there is a message behind, the principle of backward trajectory estimation is the same as above, and will not be repeated here;

[0061] The unilateral extrapolation time should not exceed 10 minutes, that is... The time limit should be less than or equal to 10 minutes; otherwise, the corresponding MMSI-related messages will be discarded as invalid messages.

[0062] For stationary ships with a ground speed (SOG) below a preset speed threshold, ±T is the time window. All AIS reporting locations within the window are extracted, and the median is calculated as the median. The location coordinates at any given time are used to suppress single-point jumps; the value of T ranges from 30 to 50 minutes, preferably 40 minutes; in a static state, the AIS message transmission frequency is low, usually 3 minutes, and a total window of 30 to 50 minutes can collect sufficient continuous reporting points, providing enough samples for statistical analysis. In addition, ships are prone to drift due to wind flow and GPS multipath effects when anchored. This window, combined with the median algorithm, can effectively eliminate instantaneous GPS outliers and noise, accurately locking the true position. At the same time, this time span can ensure that a sufficient density of historical reporting points is obtained, while locking the ship within a locally stable static range, avoiding misjudgment of the ship's normal departure or shifting status due to excessive time.

[0063] Step S2: Translation correction of antenna position to the ship's centroid

[0064] Regarding the centroid of the hull The time correction is only performed when the COG (Course of Gait) is valid and the ship length is not less than the preset ship length threshold. The COG value is between 0 and 359° (if the parsed message shows 511, then the field is invalid). The preset ship length threshold ranges from 55 to 65 meters, preferably 60 meters. If the COG is invalid or the ship length is less than the preset ship length threshold, this step is skipped.

[0065] The specific methods for correcting the ship's centroid translation are as follows:

[0066] Extract the dimensional features from the AIS ship message: distance A from the antenna to the bow, distance B to the stern, distance C to the port side, and distance D to the starboard side. Combine this with the heading COG, calculate the eastward correction ΔE and northward correction ΔN of the antenna position relative to the ship's centroid. Then, use the values ​​obtained in step 1 to calculate... By superimposing the aforementioned corrections, the centroid of the hull is obtained at... Corrected instantaneous position coordinates , ;

[0067] The calculation methods for the eastward correction ΔE and the northward correction ΔN are as follows:

[0068] ;

[0069] ;

[0070] The ship's centroid is The corrected instantaneous position coordinates:

[0071] , ,in , for The coordinates; the corrected geometric relationship of the antenna centroid, such as Figure 2 As shown.

[0072] Step S3: Modeling of positional uncertainty

[0073] Based on the dynamic message quality of each AIS vessel j (i.e., the message in...) (Based on the distribution before and after), determine the uncertainty of dynamic estimation error. Based on the implementation of antenna position correction, determine the antenna offset error uncertainty. The dynamic calculation of error uncertainty will be performed. and antenna offset error uncertainty As two independent uncertainty components, the orthogonally combined positional uncertainty The position uncertainty of each AIS ship j is obtained. ;

[0074] The quality of the dynamic message refers to the dynamic message in... Based on the continuity before and after the moment, and according to the broadcast of dynamic messages, ships can be divided into two states: ships in motion and ships at rest (anchored, docked, and at extremely low speeds over land).

[0075] For ships at sea, the uncertainty of dynamic estimation error ,in, To determine the uncertainty of random error, To set a fixed value, a range of 8-12 meters is preferred, with 10 meters being the ideal value. The specific calculation method for the uncertainty of the position estimation error is as follows:

[0076] for Ships that are navigating with messages sent from both the front and rear. ; The message interval when using linear interpolation ( );

[0077] for Ships that communicate via messages on one side ; for The most recent message time before or after The interval time, K is a constant, and its value ranges from 1.0 to 2.0; according to engineering practice, as a preferred embodiment, K is 1.6; in actual applications, the value of K can be manually adjusted according to the application scenario and the accuracy of dark ship identification.

[0078] For stationary ships with a ground speed (SOG) below a preset speed threshold, the dynamic estimation error uncertainty is... , where s is the satellite imaging time The standard deviation of AIS data coordinates within a preset time period (30~50min, preferably 40min), where n is the number of valid AIS data; the preset speed threshold is 3 knots, that is, a ship with a ground speed SOG of less than 3 knots is considered to be stationary.

[0079] The calculation method for s is as follows: ,in For the first time period within the preset time period The planar projection coordinates of a valid AIS data set. This is the average value of the coordinates of these n AIS data points.

[0080] For small boats (with a length less than the preset length threshold), the antenna offset is small and the benefit of centroid correction is limited; the ground heading field of anchored boats is often outdated. Actual test data shows that if the antenna position is corrected, about 15 m of offset will be directed in the wrong direction, which will increase the residual.

[0081] The antenna offset error uncertainty The value is 8 meters when the centroid correction has been applied; and 0.25 times the ship length when no correction has been applied.

[0082] For ships that have undergone centroidal correction The value of 8 meters is chosen because even with centroid correction, some residual random error will still exist due to the ship's movement during navigation, including rolling from wind and waves, changes in draft, and minor errors in the ship's structural parameters. In engineering experience, 8 meters is a reasonable empirical value for the residual error in high-precision differential GPS and remote sensing image registration.

[0083] For the time without correction The 0.25 times ship length is chosen because a ship's antenna typically doesn't extend beyond the hull itself. In the extreme case from bow to stern, the maximum physical distance between the antenna and the centroid is approximately half (0.5 times) the ship's length. Taking half of its standard deviation or statistical mean (i.e., 0.25 times the ship's length) reasonably quantifies the maximum possible deviation caused by the unknown antenna position. Since the upper limit of the ship's length without centroid correction is around 60 meters, 0.25 times the ship's length has its upper limit, constituting a safety truncation mechanism to ensure that the error model for extremely large ships remains within a reasonably controllable range.

[0084] Step S4: Constructing and globally optimal matching of a multi-feature association cost function based on inverse variance adaptive weighting.

[0085] Constructing the association cost function between satellite remote sensing image patch i and AIS ship j , The smaller the value, the more closely the two targets (satellite remote sensing image patch i and AIS ship j) match in space and geometry, and the more likely they are to be identified as the same ship;

[0086] ;in, The square of the positional residual.

[0087] ;

[0088] In the formula , The coordinates of the geometric center of satellite remote sensing image patch i; and The measured length and width of satellite remote sensing image patch i. and The theoretical length and width of the ship carried in the AIS ship j message. and Measurement uncertainty when extracting length and width from remote sensing images;

[0089] Taking the ZY-1 02E satellite as an example, its ground resolution (GSD) of the panchromatic multispectral fusion orthophoto product is approximately 2.5m. In engineering implementation, the standard deviation of 1-2 pixels falls precisely within the 3m-5m range. This effectively absorbs size extraction errors caused by edge pixel blurring, sea foam, or lighting shadows. and They are usually the same value. and The values ​​range from 3 to 5m;

[0090] Based on the aforementioned association cost function, a cost matrix is ​​constructed. The Hungarian algorithm is used to solve the global one-to-one optimal assignment between satellite remote sensing image patches and AIS vessels, and the matching association pairs and unmatched image patches are output. For image patches that fail to match AIS vessels, a list of unmatched vessels is generated.

[0091] Location residual An adaptive threshold is set. for The square root, that is ,

[0092] like ≥ Then skip other calculations and assign the cost of both. The threshold is set directly to infinity; based on the actual size of the ship, the upper limit (max) is set to 300 m to tolerate extrapolation divergence under extreme packet loss scenarios, and the lower limit (min) is set to 60 m to ensure that the small boat does not miss any positioning under the basic positioning error.

[0093] Finally, the cost matrix that satisfies the adaptive threshold constraint The algorithm utilizes the Hungarian algorithm to perform one-to-one optimization from a global perspective, outputting matching pairs containing MMSI, ship name, and length and width dimensions. For the list of unlisted ships, further screening can be conducted by relevant personnel or algorithms.

[0094] The dark ship identification method described in this invention uses an association cost function. China passed Dynamic transfer and adaptive fusion of the right to determine ship position and size were achieved. If the target ship's AIS data points are dense, Very small; if the target vessel is extrapolated for a long time due to data packet loss, then It automatically increases in size, thereby mathematically avoiding and tolerating extrapolation errors on the order of hundreds of meters, and preventing omissions.

[0095] Experimental Example

[0096] To verify the feasibility and reliability of the hidden ship identification method provided by this invention, a certain nearshore sea area was selected as the study area. This sea area has a high density of small and medium-sized targets such as fishing boats and a low AIS broadcast rate, which is a typical scenario with a high hidden ship rate, thus fully testing the adaptability of the association method.

[0097] (1) Source of experimental verification data

[0098] The remote sensing image data is a scene from the panchromatic multispectral fusion orthophoto product of the ZY-1 02E satellite, with a ground resolution of approximately 2.5 m, a coordinate system of CGCS2000, and a nominal imaging time. The timeframe is XXXX-XX-XX XX:XX:XX (Beijing time). Target detection yielded 663 image patches, each containing the center latitude and longitude coordinates, the long side and short side of the smallest bounding rectangle, and the north azimuth angle. The target detection results from the satellite remote sensing imagery were used as the associated input for this experimental example.

[0099] The AIS data (messages) consist of approximately 4.67 million daily imaging messages and approximately 15,000 unique MMSIs; additional AIS data from the same region for one week was used for error calibration and control experiments. Antenna installation parameters were taken from the A, B, C, and D fields of the AIS static messages, covering a total of 31,659 ships, of which approximately 58% were valid records, with a median antenna offset of approximately 15 m. A summary of the main data is shown in Table 1.

[0100] Regarding coordinate references, CGCS2000 and WGS84 used by AIS are consistent in definition at the centimeter level. Due to the difference in reference epochs, there is a system offset of about 0.5-0.8 m, which is much smaller than other error sources. Push-broom imaging causes the actual imaging time of different latitude rows to differ by the second level, which is also a minor error, and both are ignored.

[0101] Table 1 Overview of the main datasets

[0102]

[0103] (2) Overall correlation results

[0104] Of the 663 satellite remote sensing image patches, 385 were successfully associated, a correlation rate of 58.1%. For the 278 suspected vessels on the list that were not successfully associated, manual screening initially removed 10 false alarms related to land-based structures, 9 false alarms related to offshore wind power facilities, and 4 false alarms related to size anomalies (significant differences in length, width, or aspect ratio compared to the vessel). The median position residual of the matching pairs was 16.0 m, and the normalized residual... The median is approximately 1.0. The normalized residuals being close to 1 indicates that the uncertainty model gives... The method for identifying dark ships provided by this invention is consistent with the magnitude of the measured residuals, meaning that the error budget is neither underestimated nor significantly overestimated. Two features that did not enter the association cost provide independent cross-validation: the median length ratio of the matched set patches to the AIS data is approximately 1.08, and the median heading (COG) difference is approximately 5°, both consistent with the expectation of correct association.

[0105] (3) Ablation analysis of antenna centroid correction

[0106] For 349 matched targets with effective antenna parameters, the positional residuals before and after correction were statistically analyzed according to ship length, and the results are shown in Table 2. The correction reduced the median overall residual from 45.7m to 15.3m (−66%), with large ships longer than 100m reducing the median from 66.5m to 10.3m (−84%). Approximately 71% of the targets experienced a reduction in residual after correction. Gain was concentrated on large ships; for small ships, due to their small antenna offset and unreliable anchoring bow direction, correction yielded almost no benefit or even slightly worsened the situation. This is precisely the effect of the gating strategy. The basis for absorbing residual errors.

[0107] Table 2. Residual position of antenna centroid before and after correction (by ship length)

[0108]

[0109] (4) Verification of imaging time alignment: value boundary of simple inference and learning model

[0110] To examine whether linear interpolation for aligning imaging moments sacrifices accuracy due to its simplicity, this experiment specifically trained a masked spatiotemporal Transformer trajectory model (RMST-Former) as a control: the model uses a 60-second step size and a 50-step sequence, randomly occluding a 30-minute segment in the middle, and reconstructs the occluded trajectory based on the context on both sides, i.e., compared with... The task involves isomorphic bilateral interpolation; the input includes features such as progressive displacement, speed, heading, turning rate, ship length, beam, and hull type. The network is a 4-layer encoder-decoder Transformer (d_model=128). Training data consists of messages from all days except for the imaging day, and testing is conducted on the imaging day. To eliminate the possibility of "remembering specific ships," the training and test sets are re-partitioned according to non-overlapping MMSI (test ships do not appear at all in training) for verification.

[0111] The results are shown in Table 3: Under both classifications, linear interpolation outperforms the learning model in all indicators. This is also true for each difficulty level after stratification by chord deviation (Table 4), including the strongest turning level. Chord deviation is defined as the maximum vertical distance from the actual trajectory point of the occluded segment to the line connecting the two ends of the gap (chord). A larger value indicates that the trajectory deviates more from a straight line and the turning is stronger, which is more detrimental to linear interpolation; therefore, it is used as the basis for stratifying interpolation difficulty. The mechanism is that linear interpolation in bilateral interpolation naturally anchors to both ends of the gap, with minimal endpoint error; while the learning model of free regression decoding drifts at the end of the gap, with endpoint errors reaching the hundreds of meters level. This conclusion is robust across data classifications, indicating that under the condition of bilateral clamping of reporting points, linear interpolation is the optimal choice for imaging time alignment in an experimental sense.

[0112] Table 3 Comparison of linear interpolation and learned model on two-sided interpolation task (occlusion for 30 min)

[0113]

[0114] Table 4. Median of sample-by-sample average displacement error stratified by trajectory bending strength (by day)

[0115]

[0116] For another design where unilateral extrapolation does not exceed 10 minutes, this experimental example further verifies the results using minute-by-minute error curves: placing the occlusion at the end of the sequence to form a pure extrapolation task, and statistically analyzing the median per-sample position error per minute within a prediction duration of 1–30 minutes, such as… Figure 3 As shown (this indicator characterizes the error level of a typical target, which differs from the average displacement error calculated by averaging all occlusion steps in Table 3, as the latter is affected by a few strong turning samples).

[0117] The results show a clear intersection of error durations between the two methods: within 10 minutes of extrapolation, the error of constant-speed extrapolation is smaller (nearly 0 at 1 minute, approximately 3.9 m at 10 minutes), while the learning model is slightly worse due to the inherent jitter of the regression output; after about 9-10 minutes, the two lines intersect (intersecting at 9 minutes when divided by day, and at 10 minutes when divided by non-overlapping MMSI, with robustness across divisions), after which the error of constant-speed extrapolation increases linearly with time (reaching 16.7 m at 30 minutes), while the learning model's increase is significantly slower (approximately 6.9 m at 30 minutes). Therefore, setting the upper limit of unilateral extrapolation to 10 minutes in the correlation process is not an empirical value, but rather the boundary where constant-speed extrapolation maintains its advantage; in the experimental data, targets requiring unilateral extrapolation account for only about 0.6% of moving targets, and the reporting intervals are all within 2-3 minutes, well within this boundary.

[0118] Combine Tables 3 and 4 with Figure 3 As shown, the value boundary of the simple extrapolation and learning model can be summarized into three conclusions: for interpolation with bilateral observations, linear interpolation is optimal; for unilateral extrapolation within 10 minutes, constant-speed extrapolation is optimal; for unilateral extrapolation exceeding 10 minutes, the learning model is significantly superior. The imaging time alignment falls into the first two cases, and the method of this invention is experimentally supported in all branches.

[0119] This invention addresses the problem of ship association between single-temporal satellite imagery and AIS data, proposing an association method centered on per-target position uncertainty, namely multi-source orthogonal synthesis. The model-driven inverse variance weighted cost and adaptive threshold, combined with gated antenna centroid correction and a robust median positioning strategy for anchored ships, demonstrate through real-data experiments that: the correlation rate is 58.1%, the median residual is 16.0 m, and the median normalized residual is approximately 1.0; antenna centroid correction reduces the median residual of large ships by 84%; tests using the learned model as a control define that simple extrapolation of the value boundary is linearly optimal for bilateral interpolation, optimal for one-sided extrapolation within 10 minutes at constant speed, and the learned model dominates after approximately 10 minutes, and is robust across partitions.

[0120] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying dark ships by associating satellite imagery with AIS vessels driven by position uncertainty, characterized in that, Includes the following steps: Step S1: Clean the AIS data to remove invalid messages, and backfill the ship size characteristics in the static messages into the corresponding dynamic messages based on MMSI; Estimate the satellite imaging time for each AIS ship j instantaneous position ; Step S2: Extract the dimensional features from the AIS ship message: distance A from the antenna to the bow, distance B to the stern, distance C to the port side, and distance D to the starboard side. Combine this with the heading COG, calculate the eastward correction ΔE and northward correction ΔN of the antenna position relative to the ship's centroid. By superimposing the aforementioned corrections, the centroid of the hull is obtained at... Corrected instantaneous position coordinates , ; Step S3: Determine the uncertainty of dynamic estimation error based on the dynamic message quality of each AIS vessel j. Based on the execution status of antenna position correction, determine the uncertainty of antenna offset error. ,Will and As two independent uncertainty components, the orthogonally combined positional uncertainty ; Step S4: Construct the association cost function between satellite remote sensing image patch i and AIS ship j , ,in, The square of the positional residual. ， In the formula, , The coordinates of the geometric center of satellite remote sensing image patch i; and The measured length and width of satellite remote sensing image patch i. and The theoretical length and width of the ship carried in the AIS ship j message. and Measurement uncertainty when extracting length and width from remote sensing images; Based on the aforementioned association cost function, a cost matrix is ​​constructed. The Hungarian algorithm is used to solve the global one-to-one optimal assignment between satellite remote sensing image patches and AIS vessels, and the matching association pairs and unmatched image patches are output. For image patches that fail to match AIS vessels, a list of unmatched vessels is generated.

2. The method according to claim 1, characterized in that, In step S1, the AIS ship j is at the satellite imaging time instantaneous position The calculation method is as follows: for For vessels that send messages both forward and backward, linear interpolation is used to calculate their... Position coordinates at that moment; for For vessels with messages sent from only one side, their location can be estimated using a trajectory extrapolation method. Position coordinates at that moment; For stationary ships with a ground speed (SOG) below a preset speed threshold, ±T is the time window. All AIS reporting locations within the window are extracted, and the median is calculated as the median. The location coordinates at a given time, where T is 30~50 minutes.

3. The method according to claim 1, characterized in that, In step S2, the calculation methods for the eastward correction ΔE and the northward correction ΔN are as follows: ； ； The ship's centroid is The corrected instantaneous position coordinates: , ,in , for The coordinates.

4. The method according to claim 1, characterized in that, In step S2, the centroid of the hull is in The time correction is only performed when the heading to the ground is valid and the ship length is not less than a preset ship length threshold, wherein the preset ship length threshold ranges from 55 to 65 meters.

5. The method according to claim 2, characterized in that, for For vessels with messages on one side, the extrapolation time should not exceed 10 minutes; otherwise, the corresponding MMSI-related messages will be removed. In step S3, for a ship at sea, the error uncertainty is dynamically calculated. ,in, The uncertainty of the random error in positioning is set to a preset fixed value; The uncertainty of the position estimation error; Among them, for Ships that are navigating with messages sent from both the front and rear. ; This refers to the message interval used when employing linear interpolation. for Ships that communicate via messages on one side ; For message time and The interval time, K is a constant; For stationary ships with a ground speed (SOG) below a preset speed threshold , where s is the satellite imaging time The standard deviation of AIS data coordinates within a preset time period, where n is the number of valid AIS data.

6. The method according to claim 5, characterized in that, for For vessels that communicate via messages on one side, the value of K ranges from 1.0 to 2.

0. For stationary ships, the preset speed threshold is 3 knots, and the preset time period is 30-50 minutes.

7. The method according to claim 4, characterized in that, In step S3, the The value is 8 meters when the centroid correction has been applied; and 0.25 times the ship length when no correction has been applied.

8. The method according to claim 1, characterized in that, In step S4 and The value range is 3 to 5m.

9. The method according to claim 1, characterized in that, In step S4, the position residual An adaptive threshold is set. for The square root, like ≥ Then skip other calculations and assign the cost of both. Set it to infinity directly.

10. The method according to claim 9, characterized in that, The upper limit (max) of the threshold is set to 300 m, and the lower limit (min) is set to 60 m.

Images