Noctilucent remote sensing fishing boat intelligent monitoring method based on deep learning

Abstract

The invention discloses a deep learning-based intelligent monitoring method for a noctilucent remote sensing fishing boat, and relates to the technical field of noctilucent remote sensing, and the method comprises the following steps: S1, multi-source data fusion, S2, six-channel tensor determination, S3, three-dimensional physical feature map generation, S4, sample label setting, and S5, multi-task joint output. According to the method, video frames and real-time data are collected through different types of satellites and AI-Buoy buoys, so that minute-level dynamic information pushing is achieved, relevant model parameters are continuously optimized, the analysis accuracy is effectively improved, the omission ratio of complex weather and the false detection rate when monthly illumination is too high are reduced, and the method is suitable for popularization and application. By means of self-supervised light field incremental learning and synthetic cloud and fog confrontation generation, the artificial fine standard requirement is greatly reduced, meanwhile, the illegal fishing boat locking efficiency is improved through the method, important support is provided for sea area fishery resource evaluation, carbon flux research and marine law enforcement, and the modernization level of intelligent fishery supervision and marine governance is integrally improved.

Description

Technical Field

[0001] The present invention relates to the field of night light remote sensing technology, and in particular to a night light remote sensing fishing vessel intelligent monitoring method based on deep learning. Background Art

[0002] Night light remote sensing can capture marine light information at a resolution of hundreds of meters. It is the core technical means for monitoring light fishing vessels. The night light remote sensing fishing vessel intelligent monitoring method based on deep learning is mainly to output real-time visualization of light fishing vessel distribution and light power time series, which is suitable for illegal fishing monitoring and fishery resource management. The invention patent with application number 202310279072.3 discloses "A method for detecting light-induced fishing vessels based on night light remote sensing data based on improved YOLOv5 network, using ship monitoring system data to verify light-induced fishing vessels, and performing light-induced fishing vessel detection on night light remote sensing images. The paper then constructs a fishing vessel target detection dataset containing various complex background noises by labeling the images. It also proposes a multi-scale fusion algorithm based on the attention mechanism of YOLOv5 and uses this algorithm to detect targets in night-light remote sensing imagery data. The paper also discusses a method for producing position information products for light-induced fishing vessels. The proposed light-induced fishing vessel target detection algorithm will provide technical support for applications such as fishing effort estimation in distant-water and near-shore light-induced fisheries, spatiotemporal changes in fishing grounds, and monitoring overfishing.

[0003] The above-mentioned existing technologies solve the problem of being unable to extract effective features under complex conditions. However, when used, since radiation transmission factors such as lunar intensity, cloud-aerosol scattering, atmospheric transmittance and sensor zenith angle are not explicitly modeled, the false detection rate is still high under conditions of thick clouds, strong lunar phases and coastal light interference. In addition, the processing process is usually run offline, which cannot meet the needs of fishery law enforcement for minute-level dynamic intelligence. At the same time, this method fails to use field telemetry data such as AIS, ship-borne video, buoy meteorology and lightning positioning arrays to reverse correct remote sensing results, and the equipment layout requirements are high, making it difficult to deploy at sea. Summary of the Invention

[0004] The purpose of the present invention is to provide a deep learning-based intelligent monitoring method for luminous remote sensing fishing vessels to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning, comprising the following steps:

[0006] S1. Multi-source data fusion: HDF5 / L1B data are collected and converted into radiance. The tilted strips are decompressed along the scan line and reprojected into the WGS-84 grid. The AI-Buoy buoy is used to collect real-time data every five minutes to simultaneously determine the high-altitude cloud water content, mid- and high-level wind field, lunar illumination, Moon-Earth distance, and lunar phase angle.

[0007] S2. Determine a six-channel tensor: After obtaining the lightning frequency, integrate the ship positioning data and use bilinear interpolation and conformal interpolation to map the physical quantities to the WGS-84 grid. Perform Min-Max normalization on the physical quantities and use the processed physical quantities to form a six-channel tensor.

[0008] S3. Generate 3D physical feature maps: A simplified radiation model is established based on physical quantities, and multiple 3D physical feature maps are generated. These feature maps are transferred to Swin-UNet for analysis. After obtaining multi-channel feature vectors, they are fused and analyzed to output the fishing boat mask, optical power, and uncertainty.

[0009] S4. Set sample labels: After obtaining all ship surveillance video frames and constructing positive and negative samples, the CC-GAN model is trained using the fishing boat mask data, the model parameters are optimized, and the corresponding raster cloud and fog multi-channel image is generated based on the set optical thickness as the input condition. This image is then used as a sample and mixed with the real sample.

[0010] S5. Multi-task joint output: Use open-loop Kalman filtering to analyze the observation vector, output the corrected mask and optical power, analyze the connected domain in the fishing vessel mask, output a vectorized target list with confidence, use the central server to roll out the model weights to buoys and law enforcement vessels, and use GIS Dashboard to output real-time visualization of the distribution of lighted fishing vessels, optical power timing and uncertainty.

[0011] Preferably, said S1 comprises the following steps:

[0012] S101. Select and collect the HDF5 / L1B data corresponding to the VIIRS-DNB satellite with an image resolution of 500m, the SDGSAT-1 satellite with an image resolution of 40m, and the Luojia-01 satellite with an image resolution of 130m.

[0013] S102. After correction by NASA SeaDAS data processing software, HDF5 / L1B data are converted to radiance, and tilt strips are extracted along the scan line and reprojected to the WGS-84 / 0.0045° grid;

[0014] S103. Use the AI-Buoy to collect cloud base brightness, wind speed, wind direction, relative humidity, air pressure, and temperature every five minutes, and upload the real-time data in Starlink's "publish-subscribe" mode;

[0015] S104. The high-altitude cloud water content and mid-to-high-level wind field are analyzed using the GFS model. The lunar illumination, moon-to-Earth distance, and lunar phase angle are calculated using the JPL DE430 ephemeris and the IAU lunar nutation model.

[0016] Preferably, said S2 comprises the following steps:

[0017] S201. Generate a five-minute cumulative lightning frequency based on the 1km / 6s data from the China Lightning Positioning Network. Integrate and process the ship positioning data provided by AIS satellites and Beidou satellites, and exclude non-moving targets with valid unique ship identification codes less than 7 digits and speeds less than 0.5 nautical miles per hour. This serves as a reference for post-processing masking.

[0018] S202. Use bilinear interpolation and conformal interpolation to map physical quantities such as radiance, high-altitude cloud water content, lunar illuminance, zenith angle cosine, relative humidity, and lightning frequency to the WGS-84 / 500m grid.

[0019] Preferably, the step S2 further comprises the following steps:

[0020] S203, transform the radiance L to obtain the processed radiance L', where L'=log 10 (L+1), the upper cloud water content, monthly illumination, zenith angle cosine, relative humidity and lightning frequency are normalized by Min-Max;

[0021] S204. Use the processed radiance, high-altitude cloud water content, lunar illumination, zenith angle physical quantity, relative humidity, and lightning frequency to form a six-channel tensor.

[0022] Preferably, the step S3 specifically includes the following steps:

[0023] S301: Obtain cloud optical thickness τ, monthly illumination M, zenith angle θ, and atmospheric extinction coefficient κ, and establish a simplified radiation model based on cloud optical thickness τ, monthly illumination M, zenith angle θ, and atmospheric extinction coefficient κ. The simplified radiation model is specifically:

[0024] L top =L surf τe -κ(τ) +L atm (M,τ,θ)

[0025] Where τ represents the cloud optical thickness, M represents the monthly illumination, θ represents the zenith angle, κ represents the atmospheric extinction coefficient, and L atm(·) represents the atmospheric path radiation function, L top Indicates the received radiation brightness, L surf represents the surface reflectivity;

[0026] S302, after receiving the radiation brightness, generate multiple three-dimensional physical characteristic maps through a table lookup method and a linear differential equation solution method;

[0027] S303: After receiving the three-dimensional physical feature map, it is transferred to Swin-UNet for analysis. After the 64- and 128-channel feature vectors are extracted through a two-layer 3×3 Conv-BN-GELU module, the feature vectors are embedded in the self-attention using the attention formula. All feature vectors are fused and analyzed to output the fishing boat mask, optical power, and uncertainty. The attention formula is specifically:

[0028]

[0029] Among them, Q P represents the query, K V Indicates the key, represents the scaling factor, V P Represents the weighted sum value matrix.

[0030] Preferably, the S4 specifically includes the following steps:

[0031] S401: After obtaining all the ship surveillance video frames, the "blank frames" without fishing boats are used as negative samples. The SOL model is used to analyze the ship surveillance video frames in the edge computing environment to generate corresponding high-confidence detection frames. The positive samples are determined based on the high-confidence detection frames, and a self-supervised contrast loss is constructed.

[0032] S402: Using 8,000 finely labeled fishing boat mask data to train the CC-GAN model, calculating the loss value during the training process and optimizing the model parameters;

[0033] S403. After generating the corresponding raster cloud and fog multi-channel image using the CC-GAN model with the set optical thickness as the input condition, the image is mixed with the real sample in a 1:1 ratio as a sample.

[0034] Preferably, the S5 specifically includes the following steps:

[0035] S501: Using an open-loop Kalman filter, an observation vector is constructed using cloud base brightness, lightning frequency, and optical power. Error propagation is performed on the observation vector, and the corrected mask and optical power are output.

[0036] S502: Analyze the connected domain in the fishing boat mask and remove noise points. Compare it with the oil and gas platform vector database and the densely connected lightning areas to eliminate fixed light sources and instantaneous light sources, thereby outputting a vectorized target list with confidence levels.

[0037] Preferably, the S5 further includes the following steps:

[0038] S503. Use the central server to automatically trigger incremental training of the SOL and CC-GAN models every week, and roll out new model weights to buoys and law enforcement vessels.

[0039] S504: Push minute-level alarms via MQTT-TLS and use GIS Dashboard to output real-time visualization of light fishing vessel distribution, optical power timing, and uncertainty.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention collects video frames and real-time data through the VIIRS-DNB satellite with an image resolution of 500m, the SDGSAT-1 satellite with an image resolution of 40m, the Luojia-01 satellite with an image resolution of 130m, and the AI-Buoy buoy, thereby realizing minute-level dynamic intelligence push. The method continuously optimizes relevant model parameters, effectively improves the accuracy of analysis, reduces the missed detection rate in complex weather and the false detection rate when the lunar illumination is too high, and with the help of self-supervised light field incremental learning and synthetic cloud and fog adversarial generation, the demand for manual precision marking is greatly reduced. At the same time, the method improves the efficiency of locking illegal fishing vessels, provides important support for fishery resource assessment, carbon flux research and maritime law enforcement in the sea area, and comprehensively improves the level of smart fishery supervision and marine governance modernization. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Provides an overall method flow chart for an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] See also Figure 1 The present invention provides a technical solution: a method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning, comprising the following steps:

[0045] S1. Multi-source data fusion: HDF5 / L1B data are collected and converted into radiance. The tilted strips are decompressed along the scan line and reprojected into the WGS-84 grid. The AI-Buoy buoy is used to collect real-time data every five minutes to simultaneously determine the high-altitude cloud water content, mid- and high-level wind field, lunar illumination, Moon-Earth distance, and lunar phase angle.

[0046] S2. Determine a six-channel tensor: After obtaining the lightning frequency, integrate the ship positioning data and use bilinear interpolation and conformal interpolation to map the physical quantities to the WGS-84 grid. Perform Min-Max normalization on the physical quantities and use the processed physical quantities to form a six-channel tensor.

[0047] S3. Generate 3D physical feature maps: A simplified radiation model is established based on physical quantities, and multiple 3D physical feature maps are generated. These feature maps are transferred to Swin-UNet for analysis. After obtaining multi-channel feature vectors, they are fused and analyzed to output the fishing boat mask, optical power, and uncertainty.

[0048] S4. Set sample labels: After obtaining all ship surveillance video frames and constructing positive and negative samples, the CC-GAN model is trained using the fishing boat mask data, the model parameters are optimized, and the corresponding raster cloud and fog multi-channel image is generated based on the set optical thickness as the input condition. This image is then used as a sample and mixed with the real sample.

[0049] S5. Multi-task joint output: Use open-loop Kalman filtering to analyze the observation vector, output the corrected mask and optical power, analyze the connected domain in the fishing vessel mask, output a vectorized target list with confidence, use the central server to roll out the model weights to buoys and law enforcement vessels, and use GIS Dashboard to output real-time visualization of the distribution of lighted fishing vessels, optical power timing and uncertainty.

[0050] S1 includes the following steps:

[0051] S101. Select and collect the HDF5 / L1B data corresponding to the VIIRS-DNB satellite with an image resolution of 500m, the SDGSAT-1 satellite with an image resolution of 40m, and the Luojia-01 satellite with an image resolution of 130m.

[0052] S102. After correction by NASA SeaDAS data processing software, HDF5 / L1B data are converted to radiance, and tilt strips are extracted along the scan line and reprojected to the WGS-84 / 0.0045° grid;

[0053] S103. Use the AI-Buoy to collect cloud base brightness, wind speed, wind direction, relative humidity, air pressure, and temperature every five minutes, and upload the real-time data in Starlink's "publish-subscribe" mode;

[0054] S104: Analyze the high-altitude cloud water content and mid-to-high-level wind field using the GFS model, and calculate the lunar illumination, moon-to-Earth distance, and lunar phase using the JPL DE430 ephemeris and the IAU lunar nutation model.

[0055] S2 includes the following steps:

[0056] S201. Generate a five-minute cumulative lightning frequency based on the 1km / 6s data from the China Lightning Positioning Network. Integrate and process the ship positioning data provided by AIS satellites and Beidou satellites, and exclude non-moving targets with valid unique ship identification codes less than 7 digits and speeds less than 0.5 nautical miles per hour. This serves as a reference for post-processing masking.

[0057] S202. Use bilinear interpolation and conformal interpolation to map the physical quantities of radiance, high-altitude cloud water content, lunar illuminance, zenith angle cosine, relative humidity, and lightning frequency to the WGS-84 / 500m grid;

[0058] S2 also includes the following steps:

[0059] S203, transform the radiance L to obtain the processed radiance L', where L'=log 10 (L+1), the upper cloud water content, monthly illumination, zenith angle cosine, relative humidity and lightning frequency are normalized by Min-Max;

[0060] S204. Using the processed radiance, high-altitude cloud water content, lunar illumination, zenith angle physical quantity, relative humidity, and lightning frequency, a six-channel tensor is formed.

[0061] S3 specifically includes the following steps:

[0062] S301: Obtain cloud optical thickness τ, monthly illumination M, zenith angle θ, and atmospheric extinction coefficient κ. Establish a simplified radiation model based on cloud optical thickness τ, monthly illumination M, zenith angle θ, and atmospheric extinction coefficient κ. The simplified radiation model is specifically:

[0063] L top =L surf τe -κ(τ) +L atm (M,τ,θ)

[0064] Where τ represents the cloud optical thickness, M represents the monthly illumination, θ represents the zenith angle, κ represents the atmospheric extinction coefficient, and L atm (·) represents the atmospheric path radiation function, L topIndicates the received radiation brightness, L surf represents the surface reflectivity;

[0065] S302, after receiving the radiation brightness, generate multiple three-dimensional physical characteristic maps through a table lookup method and a linear differential equation solution method;

[0066] S303: After receiving the 3D physical feature map, it is transferred to Swin-UNet for analysis. After the 64- and 128-channel feature vectors are extracted through a 2-layer 3×3 Conv-BN-GELU module, the feature vectors are embedded in the self-attention using the attention formula. All feature vectors are fused and analyzed to output the fishing boat mask, optical power, and uncertainty. The attention formula is as follows:

[0067]

[0068] Among them, Q P represents the query, K V Indicates the key, represents the scaling factor, V P represents the weighted sum value matrix;

[0069] S4 specifically includes the following steps:

[0070] S401. After obtaining all the ship surveillance video frames, the “blank frames” without fishing boats are used as negative samples. The SOL model is used to analyze the ship surveillance video frames in the edge computing environment to generate corresponding high-confidence detection frames. The positive samples are determined based on the high-confidence detection frames, and a self-supervised contrast loss is constructed. The self-supervised contrast loss is specifically:

[0071]

[0072] Among them, L ssl represents the loss value, Represents the positive sample features output by the PILNet encoder, represents the negative sample features output by the PILNet encoder, f(x i ′) represents the anchor sample features output by the PILNet encoder, and γ represents the hyperparameter;

[0073] S402: Using 8,000 finely labeled fishing boat mask data to train the CC-GAN model, calculating the loss value during the training process and optimizing the model parameters;

[0074] S403, after generating the corresponding raster cloud and fog multi-channel image with the set optical thickness as the input condition through the CC-GAN model, the image is mixed with the real sample in a 1:1 ratio as a sample;

[0075] S5 specifically includes the following steps:

[0076] S501: Use open-loop Kalman filtering to construct an observation vector using cloud base brightness, lightning frequency, and optical power. Perform error propagation on the observation vector and output the corrected mask and optical power. The error propagation is specifically as follows:

[0077]

[0078] Among them, K k represents the gain, z k represents the observation vector, H represents the parameter, represents the open-loop prediction state, Indicates the true state;

[0079] S502: Analyze the connected domain in the fishing boat mask, remove noise points, and compare it with the oil and gas platform vector database and the dense lightning contiguous area to eliminate fixed light sources and instantaneous light sources, thereby outputting a vectorized target list with confidence levels.

[0080] S5 specifically further includes the following steps:

[0081] S503. Use the central server to automatically trigger incremental training of the SOL and CC-GAN models every week, and roll out new model weights to buoys and law enforcement vessels.

[0082] S504: Push minute-level alarms via MQTT-TLS and use GIS Dashboard to output real-time visualization of light fishing vessel distribution, optical power timing, and uncertainty.

[0083] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0084] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The deep learning-based intelligent monitoring method for night-light remote sensing fishing vessels is characterized by: The method comprises the following steps: S1. Multi-source data fusion: HDF5 / L1B data are collected and converted into radiance. The tilted strips are decompressed along the scan line and reprojected into the WGS-84 grid. The AI-Buoy buoy is used to collect real-time data every five minutes to simultaneously determine the high-altitude cloud water content, mid- and high-level wind field, lunar illumination, Moon-Earth distance, and lunar phase angle. S2. Determine a six-channel tensor: After obtaining the lightning frequency, integrate the ship positioning data and use bilinear interpolation and conformal interpolation to map the physical quantities to the WGS-84 grid. Perform Min-Max normalization on the physical quantities and use the processed physical quantities to form a six-channel tensor. S3. Generate 3D physical feature maps: A simplified radiation model is established based on physical quantities, and multiple 3D physical feature maps are generated. These feature maps are transferred to Swin-UNet for analysis. After obtaining multi-channel feature vectors, they are fused and analyzed to output the fishing boat mask, optical power, and uncertainty. S4. Set sample labels: After obtaining all ship surveillance video frames and constructing positive and negative samples, the CC-GAN model is trained using the fishing boat mask data, the model parameters are optimized, and the corresponding raster cloud and fog multi-channel image is generated based on the set optical thickness as the input condition. This image is then used as a sample and mixed with the real sample. S5. Multi-task joint output: Use open-loop Kalman filtering to analyze the observation vector, output the corrected mask and optical power, analyze the connected domain in the fishing vessel mask, output a vectorized target list with confidence, use the central server to roll out the model weights to buoys and law enforcement vessels, and use GIS Dashboard to output real-time visualization of the distribution of lighted fishing vessels, optical power timing and uncertainty.

2. The method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning according to claim 1 is characterized in that: Said S1 comprises the following steps: S101. Select and collect the HDF5 / L1B data corresponding to the VIIRS-DNB satellite with an image resolution of 500m, the SDGSAT-1 satellite with an image resolution of 40m, and the Luojia-01 satellite with an image resolution of 130m. S102. After correction by NASA SeaDAS data processing software, HDF5 / L1B data are converted to radiance, and tilt strips are extracted along the scan line and reprojected to the WGS-84 / 0.0045° grid; S103. Use the AI-Buoy to collect cloud base brightness, wind speed, wind direction, relative humidity, air pressure, and temperature every five minutes, and upload the real-time data via Starlink's "publish-subscribe" model. S104. The high-altitude cloud water content and mid-to-high-level wind field are analyzed using the GFS model. The lunar illumination, moon-to-Earth distance, and lunar phase angle are calculated using the JPL DE430 ephemeris and the IAU lunar nutation model.

3. The method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning according to claim 1, characterized in that: The S2 comprises the following steps: S201. Generate a five-minute cumulative lightning frequency based on the 1km / 6s data from the China Lightning Positioning Network. Integrate and process the ship positioning data provided by AIS satellites and Beidou satellites, and exclude non-moving targets with valid unique ship identification codes less than 7 digits and speeds less than 0.5 nautical miles per hour. This serves as a reference for post-processing masking. S202. Use bilinear interpolation and conformal interpolation to map physical quantities such as radiance, high-altitude cloud water content, lunar illuminance, zenith angle cosine, relative humidity, and lightning frequency to the WGS-84 / 500m grid.

4. The method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning according to claim 3 is characterized in that: The S2 further comprises the following steps: S203, transform the radiance L to obtain the processed radiance L', where L'=log 10 (L+1), the upper cloud water content, monthly illumination, zenith angle cosine, relative humidity and lightning frequency are normalized by Min-Max; S204. Use the processed radiance, high-altitude cloud water content, lunar illumination, zenith angle physical quantity, relative humidity, and lightning frequency to form a six-channel tensor.

5. The method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning according to claim 1, characterized in that: The S3 specifically includes the following steps: S301: Obtain cloud optical thickness τ, monthly illumination M, zenith angle θ, and atmospheric extinction coefficient κ, and establish a simplified radiation model based on cloud optical thickness τ, monthly illumination M, zenith angle θ, and atmospheric extinction coefficient κ. The simplified radiation model is specifically: L top =L surf the -κ(τ) +L atm (M,t,i) Where τ represents the cloud optical thickness, M represents the monthly illumination, θ represents the zenith angle, κ represents the atmospheric extinction coefficient, and L atm (·) represents the atmospheric path radiation function, L top Indicates the received radiation brightness, L surf represents the surface reflectivity; S302, after receiving the radiation brightness, generate multiple three-dimensional physical characteristic maps through a table lookup method and a linear differential equation solution method; S303. After receiving the three-dimensional physical feature map, it is transferred to Swin-UNet for analysis. After the 64- and 128-channel feature vectors are extracted through a two-layer 3×3 Conv-BN-GELU module, the feature vectors are embedded in the self-attention using the attention formula. All feature vectors are fused and analyzed to output the fishing boat mask, optical power, and uncertainty.

6. The method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning according to claim 1, characterized in that: The S4 specifically includes the following steps: S401: After acquiring all the ship surveillance video frames, the "blank frames" without fishing boats are used as negative samples. The SOL model is used to analyze the ship surveillance video frames in the edge computing environment to generate corresponding high-confidence detection frames. The positive samples are determined based on the high-confidence detection frames, and a self-supervised contrast loss is constructed. S402: Using 8,000 finely labeled fishing boat mask data to train the CC-GAN model, calculating the loss value during the training process and optimizing the model parameters; S403. After generating the corresponding raster cloud and fog multi-channel image using the CC-GAN model with the set optical thickness as the input condition, the image is mixed with the real sample in a 1:1 ratio as a sample.

7. The method for intelligent monitoring of fishing vessels using night-light remote sensing technology based on deep learning according to claim 1, characterized in that: The S5 specifically includes the following steps: S501: Using an open-loop Kalman filter, an observation vector is constructed using cloud base brightness, lightning frequency, and optical power. Error propagation is performed on the observation vector, and the corrected mask and optical power are output. S502: Analyze the connected domain in the fishing boat mask and remove noise points. Compare it with the oil and gas platform vector database and the densely connected lightning areas to eliminate fixed light sources and instantaneous light sources, thereby outputting a vectorized target list with confidence levels.

8. The deep learning-based luminous remote sensing fishing vessel intelligent monitoring method according to claim 7 is characterized by: The S5 specifically further includes the following steps: S503. Use the central server to automatically trigger incremental training of the SOL and CC-GAN models every week, and roll out new model weights to buoys and law enforcement vessels. S504: Push minute-level alarms via MQTT-TLS and use GIS Dashboard to output real-time visualization of light fishing vessel distribution, optical power timing, and uncertainty.

Images