Port risk management and control system and method based on low earth orbit satellite

Abstract

The invention discloses a port risk management and control system and method based on a low-orbit satellite, and relates to the technical field of port management and control, and the system comprises a stacking estimation module which is used for carrying out the detection region division, shadow recognition and container cell and shadow cell configuration of a high-resolution satellite image of a port container yard, the vertical stacking layer number and the corresponding reliability of each cell are estimated in sequence, and the container occupancy rate and the congestion level of each detection area are calculated after the low-reliability cells are manually corrected so as to generate storage yard congestion training data; and the congestion level identification module is used for constructing a convolutional neural network model based on the congestion training data output by the stacking estimation module, training by taking the cut image of the detection area as input and the congestion level as output, executing end-to-end reasoning on a new satellite image after the model training is completed, outputting the congestion level of each detection area, and outputting the congestion level of each detection area. The method has the characteristic of quantitative analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of port management and control, in particular to a port risk management and control system and method based on low-orbit satellites. BACKGROUND

[0002] With the popularization of high-resolution optical satellite images and AIS ship identification systems, it is possible to carry out port monitoring based on remote sensing images and ship trajectory data. Existing researches are mostly focused on ship identification, berth occupation inference, port throughput estimation, high-frequency monitoring of port activity intensity, etc. The vertical stacking level in a unit cell cannot be directly obtained by area statistics on the plane scale, and the overall full container level of the container yard is still rough.

[0003] On the other hand, the method of manually interpreting satellite images to count containers is time-consuming and subjective, and is not suitable for systematic analysis in multiple time periods and multiple ports. The method of simply relying on pixel brightness threshold to divide the container yard area cannot distinguish between single-layer stacking and multi-layer stacking areas, and it is also difficult to identify abnormal stacking conditions in local recessed areas. In addition, even if the congestion indicators of the container yard at a certain time can be obtained from the image, the existing methods generally lack a systematic coupling mechanism with the AIS ship arrival and departure time, and cannot quantify the statistical relationship between the congestion level and the ship berthing opportunity, making it difficult to support quantitative analysis of port operation risk and queuing risk. Therefore, it is necessary to design a quantitative analysis of a port risk management and control system and method based on low-orbit satellites. SUMMARY

[0004] The purpose of the present application is to provide a port risk management and control system and method based on low-orbit satellites to solve the problems raised in the background.

[0005] To solve the above technical problems, the present application provides the following technical solution: a port risk management and control system based on low-orbit satellites, comprising a stacking estimation module for detecting and dividing regions, identifying shadows, and configuring container cells and shadow cells on high-resolution satellite images of a port container yard, sequentially estimating the vertical stacking number and corresponding reliability of each cell, and calculating the container occupancy rate and congestion level of each detection region after manually correcting low-reliability cells to generate container yard congestion training data; a congestion level identification module for constructing a convolutional neural network model based on the congestion training data output by the stacking estimation module, training the detection region cropped image as input and the congestion level as output, and performing end-to-end inference on new satellite images after model training to output the congestion level of each detection region; An AIS analysis and evaluation module is configured to acquire AIS ship trajectory data corresponding to a satellite image acquisition date, identify a port arrival event and a port departure event of each berth ship before and after satellite imaging, calculate a time interval between a satellite imaging time and the arrival time or the departure time, obtain a relationship between an average stacking layer number of a yard, a time interval and a ship size based on a congestion level of each detection area, and evaluate congestion evolution characteristics and dynamic operation efficiency of yards of different wharfs.

[0006] According to the technical solution, the stacking estimation module comprises: An image acquisition module is configured to receive high-resolution satellite images covering multiple berths and yard areas from a satellite image acquisition terminal and store the images in a memory at a preset spatial resolution. A detection area configuration module is configured to specify a plurality of rectangular or parallelogram detection areas in a yard range by a user in a display interface, and write pixel coordinate ranges of the detection areas into detection area configuration data. An HSV threshold segmentation module is configured to convert a color image in a detection area to an HSV color space, perform binary judgment on each pixel by a preset threshold, judge shadow pixels and non-shadow pixels, and output a binary image. A cell division module is configured to set a container length direction vector, a width direction vector and a shadow direction vector based on a yard arrangement, a tool starting point, generate a container cell set and a corresponding shadow cell set on a binary image according to a number of container rows and columns and a maximum stacking layer number of each detection area. A pixel statistical stacking module is configured to calculate an average pixel value of each container cell and shadow cell, judge a cell concave state, sequentially estimate a vertical stacking layer number and a reliability of the cell, and mark a cell with a reliability lower than a preset threshold. A cell correction module is configured to highlight a cell with a low reliability in an interface, input a corrected stacking layer number of the cell by an operator, write a correction result into an observation data set, and re-execute stacking estimation. A congestion level generation module is configured to calculate a container occupancy rate of each detection area according to a preset rule after a stacking layer number of a cell is stable, divide the detection area into a plurality of congestion levels, output the congestion levels as yard image labeling results, and store the congestion levels. The congestion level identification module comprises: A training data construction module is configured to combine image blocks obtained by cropping each detection area and corresponding congestion levels into a training sample and a verification sample set. A CNN model construction module is configured to construct a convolutional neural network structure comprising a plurality of convolution layers, pooling layers, dropout layers and fully connected layers in a processor, and set a convolution kernel size, a step length, an activation function and a loss function. The model training module is configured to train the convolutional neural network by using a preset optimization algorithm and taking a training sample set as input, to evaluate accuracy, recall rate and precision rate indexes by using a verification sample set and to adjust network parameters until performance meets requirements; The model inference module is configured to input a detection region image block obtained by cropping a new satellite image into the trained convolutional neural network in an inference stage, to output congestion levels of each detection region and to store the congestion levels; The AIS analysis and evaluation module includes: The AIS data preprocessing module is configured to receive time series data containing ship position, speed, side number, voyage destination and the like from an AIS data acquisition terminal, to clean the data and to divide trajectories according to port berths; The berth event identification module is configured to determine berthing events on each berth based on the relationship between the trajectory stopping position and the geometric range of the berth, to identify in-port ships corresponding to the satellite imaging date and ships that berthed and unberthed most recently before and after imaging; The time interval calculation module is configured to calculate, according to the satellite imaging time, the most recent arrival time and the most recent departure time of each berth, an arrival time interval for export analysis and a departure time interval for import analysis, respectively; The stacking time analysis module is configured to match the congestion level of each detection region with the time interval and ship size information of the corresponding berth, to calculate the average stacking layer number of the yard and to analyze the relationship between the average stacking layer number and the time interval and the ship size; The visualization output module is configured to generate a yard congestion level distribution map, a scatter plot of stacking layer number and time interval and a statistical chart related to ship size, and to output to a display terminal or an external storage medium.

[0007] The port risk management and control method based on low-orbit satellites includes the following steps: S1, obtaining high-resolution satellite images covering multiple berths and yard areas, and demarcating a plurality of detection regions for each yard in the image, and recording the spatial range of each detection region as a basic unit for subsequent processing; S2, performing shadow extraction and cell division on the image in each detection region, performing binaryzation processing based on the difference between the shadow and the container body in the color space, and setting a container cell and a corresponding shadow cell for each container according to a preset geometric parameter, estimating the vertical stacking layer number and the reliability of each cell by sequential calculation, and after manual correction of the cells with low reliability, obtaining the container occupancy rate of each detection region and the congestion level derived from the occupancy rate; S3, sample construction training data set with each detection area cropped image and its corresponding congestion level, build convolutional neural network model, through training get end to end image classification model which can automatically map input detection area image to congestion level, and evaluate model performance with validation data set; S4, for satellite image of the date to be evaluated, crop each detection area and input the trained convolutional neural network model, get the predicted congestion level of each detection area, and summarize the prediction results as the congestion state distribution of the terminal at that time; S5, based on AIS ship trajectory data corresponding to the date to be evaluated, identify the berthing ships, arrival and departure times of each berth before and after satellite imaging, and calculate the time interval between satellite imaging time and the arrival time of the next berthing ship and the time interval between satellite imaging time and the departure time of the last berthing ship, and on this basis, associate the congestion level of each detection area of the terminal with the time interval and the corresponding ship size, and analyze the evolution rule of the average stacking layer number of the terminal with time.

[0008] According to the above technical scheme, the S1 specifically comprises: S1-1, obtain optical satellite images containing multiple berths of the container terminal from the satellite image service platform and other high-resolution image data sources, perform geographic coordinate correction and resolution uniform processing on the images, and store them as two-dimensional raster data available for subsequent image processing; S1-2, load the above satellite images in the user interface, and according to the actual arrangement of the terminal, draw rectangular detection areas along the arrangement direction of the containers within the terminal range by the operating personnel, each detection area covers a certain number of rows and columns of container berths, and each detection area is assigned a unique number; S1-3, write the pixel coordinates of the boundary vertices of each detection area in the image coordinate system into the detection area parameter set, and estimate the maximum arrangement number of containers in the length direction and the width direction and the upper limit of the vertical stacking layer number in each detection area according to the operation characteristics of the terminal, which is used for subsequent cell division.

[0009] According to the above technical scheme, the S2 specifically comprises: S2-1, convert the RGB image in the detection area to HSV color space, for each pixel, when its hue component H≤ preset threshold , saturation component S≤ preset threshold , and brightness component V≤ preset threshold , the pixel is determined as a shadow pixel and is assigned a value of zero in the binary image, and in other cases, it is assigned a value of one, thereby obtaining the binary result of the shadow area and the non-shadow area; S2-2, in the binary image, set a length direction vector Width direction vector Shadow direction vector and the tool starting point for each detection area. The number of containers along the length direction was obtained based on the observation results. Dimensions in the length direction Number of containers in the width direction and maximum vertical stacking layers Based on this, each detection area is classified according to From 0 to , From 0 to Integer indexes generate container cell collections At the same time, based on , , Generate a set of shadow cells arranged along the shadow direction for each container cell. ; S2-3, For each container cell Within the parallelogram of the cell, select a number of pixels and their neighboring pixels along a straight line passing through the intersection of the diagonals and parallel to the vertical axis of the image, and calculate the average pixel value of the container cell. Its value ranges from 0 to 1, for each shaded cell. ,since Start by calculating the average pixel value of the shaded cells one by one. Its value ranges from 0 to 1; S2-4. Use 0.5 as the threshold for shadow detection. Record the first time it exceeds the threshold. Value And calculate the value from the cell according to the following formula. Number of shadow lines produced ,in Indicates the first Line number The number of shadow bars corresponding to the column container cell. This indicates the first index that causes the average pixel value of the shaded cell to exceed 0.5; S2-5, When the average pixel value of the container cell The cell is then classified as a non-recessed cell, and the stacking layer of each cell is determined using the following formula. and reliability Perform sequential estimation: when hour, ,when When this cell is identified as a recessed cell, the number of stacking layers is randomly generated based on the distribution of observed data. The following formula is used to give the number of stacking layers and reliability of the recessed cells. ; S2-6, Reliability Cells with a reliability lower than 0.5 are marked as low-reliability cells and highlighted in the graphical interface. The operator then determines the stacking depth of these cells. Perform manual counting and modification, record the modified results as new observation data, and replace the corresponding cells. Then, repeat the above sequential estimation process until the reliability of all cells is not less than 0.5; S2-7. After obtaining the stable stacking layer number of each cell, calculate the container occupancy rate of each detection area using the following formula. and using natural numbers The detection area is divided into There are several congestion levels, and the width of each level is... ,when Falling in When a certain interval is reached, the detection area is classified as congested. , .

[0010] According to the above technical solution, S3 specifically refers to: S3-1. For each detection region, crop an image block containing complete stacking cell and shadow information according to its actual pixel range. Scale the image block to the same size as the input layer of the convolutional neural network and use the corresponding congestion level as the supervision label to form a training sample set and a validation sample set. S3-2. Construct a convolutional neural network model. The convolutional neural network includes four convolutional layers, two pooling layers, three dropout layers, and two fully connected layers connected in sequence. The convolutional layers are used to extract multi-scale texture and shadow features in the local receptive field. The pooling layers are used to reduce the dimension of the feature map while maintaining important features. The dropout layers are used to suppress overfitting. The fully connected layers are used to map the extracted features to the congestion level prediction result. S3-3. Input the training sample set into the convolutional neural network, update the network parameters through forward propagation and backward propagation, and use the validation sample set to calculate the overall accuracy, recall rate and precision evaluation indicators for each congestion level. When the evaluation indicators reach the preset threshold, freeze the network parameters and use the trained network as the yard congestion level identification model.

[0011] According to the above technical solution, S4 specifically refers to: S4-1, the satellite image of the date to be evaluated is cropped according to the same detection area configuration as the training stage to obtain image blocks of each detection area, the pre-processing operation of normalization and size adjustment consistent with the training stage is performed, the processed image blocks are input into the trained convolutional neural network model, and the congestion level prediction results of each detection area are obtained; S4-2, the congestion level results of all detection areas are mapped back to the yard geographical space position to form a yard congestion level distribution map, and the congestion level is statistically summarized according to the berth or wharf when needed to obtain the average yard stacking layer number corresponding to each berth for correlation analysis with AIS data.

[0012] According to the above technical solution, the S5 is specifically: S5-1, according to the relationship between the berthing position of the ship in the AIS trajectory near the wharf and the spatial range of the berth, the nearest berthing process before and after satellite imaging of each berth is identified, the time interval between the satellite imaging time and the next berthing ship arrival time is defined as the arrival time interval , the time interval between the satellite imaging time and the last ship departure time is defined as the departure time interval , and the time interval and the corresponding berth are recorded as analysis samples; S5-2, for at least one wharf, according to the yard layout provided by the wharf and the measured outlet, inlet and empty container quantity at multiple time points, each detection area is divided into outlet detection area and inlet detection area, the outlet or inlet container quantity proportion of each detection area at multiple time points is counted to obtain the outlet proportion and the inlet proportion , wherein is the detection area number, is the berth number; S5-3, the average stacking layer number of the outlet direction and the inlet direction of each wharf is calculated by the following formula using the congestion level of the detection area , the outlet proportion and the inlet proportion : ; S5-4, the and , and ​​And corresponding ship size is analyzed jointly, the average stacking layer and time interval and ship size are drawn, and according to envelope trend, the negative correlation characteristics that average stacking layer decreases with the increase of time interval and the positive correlation characteristics between ship size are judged, quantitative basis is provided for different terminal yard resource deployment strategy and operation efficiency evaluation, and yard congestion control threshold and early warning rule under different risk levels are set, and dynamic control of port operation risk is realized.

[0013] Compared with the prior art, the beneficial effects achieved by the present application are: the present application combines container shadow analysis and convolutional neural network for port container yard congestion analysis, calculates the vertical stacking layer and reliability of each cell through a special labeling tool unit, and generates a congestion level label, and then coupled with AIS data to realize dynamic congestion and berthing time relationship analysis from the third party perspective. By constructing a labeling tool based on container shadows, the yard image is divided into regular container units and shadow units, the vertical stacking layer and its reliability of each unit are automatically calculated, and the congestion level is divided according to the overall occupancy rate, compared with pure manual interpretation, the labeling workload is significantly reduced, and the consistency and reliability of the label are improved.

[0014] Using the above labeling results to construct a multi-layer convolutional neural network congestion recognition model can classify different congestion levels end-to-end under the condition of using only image texture information, the model output is highly consistent with the manual statistical results, and supports automatic processing of large-scale satellite images.

[0015] Combined with the AIS ship arrival and departure time, the relationship between the average stacking layer of export and import containers and the arrival and departure time difference is calculated, providing a technical path for third-party observers to evaluate the port yard congestion risk and berthing time rationality without internal operation data, which is conducive to horizontal comparison and long-term trend analysis. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 It is a schematic diagram of the overall module structure of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0018] Referring to Figure 1 The present application provides a technical solution: a port risk management and control system based on low-orbit satellites, which comprises a stacking estimation module, used for detecting area division, shadow identification, and container cell and shadow cell configuration of high-resolution satellite images of a port container yard, sequentially estimating the vertical stacking layers and corresponding reliability of each cell, and calculating the container occupancy rate and congestion level of each detection area after manual correction of low-reliability cells, for generating yard congestion training data; A congestion level identification module is used for constructing a convolutional neural network model based on the congestion training data output by the stacking estimation module, training by taking the detection area cropped image as input and the congestion level as output, and performing end-to-end inference on new satellite images after model training to output the congestion level of each detection area; An AIS analysis and evaluation module is used for obtaining AIS ship trajectory data corresponding to the satellite image acquisition date, identifying a port ship's one-time arrival event and one-time departure event before and after satellite imaging, calculating the time interval between the satellite imaging time and the above-mentioned arrival time or departure time, combining the congestion level of each detection area to obtain the relationship between the average stacking layers of the yard and the time interval and the ship size, and evaluating the yard congestion evolution characteristics and dynamic operation efficiency of different wharfs; The stacking estimation module comprises: An image acquisition module is used for receiving high-resolution satellite images covering multiple berths and yard areas from a satellite image acquisition terminal, and storing them in a memory at a preset spatial resolution; A detection area configuration module is used for specifying a plurality of rectangular or parallelogram detection areas in the yard range by a user in a display interface, and writing the pixel coordinate range of each detection area into detection area configuration data; An HSV threshold segmentation module is used for converting the color image in the detection area to the HSV color space, performing binary judgment on each pixel by a preset threshold to identify shadow pixels and non-shadow pixels, and outputting a binary image; A cell division module is used for setting container length direction vectors, width direction vectors, and shadow direction vectors, and tool starting points based on yard arrangement, generating a container cell set and a corresponding shadow cell set on the binary image according to the number of container rows and columns and the maximum stacking layers of each detection area; A pixel statistical stacking module is used for calculating the average pixel value of each container cell and shadow cell, judging the cell concave state and sequentially estimating the vertical stacking layers and reliability of the cell, and marking the cells with a reliability lower than a preset threshold; A cell correction module is configured to highlight a cell with low reliability in the interface, input a corrected cell stack layer number by an operator, and write the correction result into an observation data set and then re-perform stack estimation; A congestion level generation module is configured to calculate a container occupancy rate of each detection area according to a preset rule and divide the detection area into a plurality of congestion levels after the cell stack layer number is stable, output the congestion level as a yard image labeling result, and store the congestion level. The congestion level identification module comprises: A training data construction module is configured to combine the image blocks obtained by cutting each detection area and the corresponding congestion levels into a training sample set and a verification sample set. A CNN model construction module is configured to construct a convolutional neural network structure comprising a plurality of convolutional layers, pooling layers, dropout layers, and fully connected layers in the processor, and set the convolution kernel size, step length, activation function, and loss function. A model training module is configured to use a preset optimization algorithm to train the convolutional neural network with the training sample set as input, evaluate the accuracy, recall rate, and precision rate indicators through the verification sample set, and adjust the network parameters until the performance meets the requirements. A model inference module is configured to input the detection area image blocks obtained by cutting the new satellite image into the trained convolutional neural network in the inference stage, output the congestion levels of each detection area, and store the congestion levels. The AIS analysis and evaluation module comprises: An AIS data preprocessing module is configured to receive time series data containing ship position, speed, side number, voyage destination, etc. from an AIS data acquisition terminal, clean the data, and divide the trajectory according to the port berth. A berth event identification module is configured to determine the berthing events on each berth based on the relationship between the trajectory stop position and the geometric range of the berth, identify the in-port ships corresponding to the satellite imaging date, and the ships that berthed and unberthed the nearest time before and after imaging. A time interval calculation module is configured to calculate the arrival time interval for export analysis and the departure time interval for import analysis according to the satellite imaging time, the latest arrival time, and the latest departure time of each berth. A stack time analysis module is configured to match the congestion levels of each detection area with the time interval and ship size information of the corresponding berth, calculate the average stack layer number of the yard, and analyze the relationship between the average stack layer number and the time interval and ship size. A visualization output module is configured to generate a yard congestion level distribution map, a scatter plot of stack layer number and time interval, and a statistical chart related to ship size, and output to a display terminal or an external storage medium. A port risk management and control method based on low-orbit satellites comprises the following steps: S1, acquire high-resolution satellite images covering multiple berths and yard areas, and demarcate a plurality of detection regions for each yard sub-block in the images, and record the spatial range of each detection region as a basic unit for subsequent processing; S2, perform shadow extraction and cell division on the images in each detection region, perform binaryzation processing based on the difference between the shadow and the container body in the color space, and set container cells and corresponding shadow cells for each container according to the preset geometric parameters, estimate the vertical stacking number and reliability of each cell through sequential calculation, and manually correct the cells with low reliability to obtain the container occupancy rate of each detection region and the congestion level derived from the occupancy rate; S3, construct a training data set by using the images of each detection region and the corresponding congestion level as samples, build a convolutional neural network model, obtain an end-to-end image classification model capable of automatically mapping the input detection region image to the congestion level through training, and evaluate the model performance using a validation data set; S4, for the satellite image of the date to be evaluated, crop each detection region and input it into the trained convolutional neural network model to obtain the predicted congestion level of each detection region, and aggregate the prediction results to obtain the congestion state distribution of the yard at that time; S5, based on the AIS ship trajectory data corresponding to the date to be evaluated, identify the berthing ships, arrival and departure times of each berth before and after satellite imaging, and calculate the time interval between the satellite imaging time and the arrival time of the next berthing ship and the time interval between the satellite imaging time and the departure time of the last berthing ship, and on this basis, associate the congestion level of each detection region of the yard with the time interval and the corresponding ship size, and analyze the evolution rule of the average stacking number of the yard over time; S1 specifically includes: S1-1, acquire optical satellite images containing multiple berths of the container terminal from a satellite image service platform and other high-resolution image data sources, perform geographic coordinate correction and resolution unification processing on the images, and store them as two-dimensional raster data for subsequent image processing; S1-2, load the above satellite images in the user interface, and demarcate rectangular detection regions along the container arrangement direction within the yard range according to the actual arrangement of the yard by the operating personnel, each detection region covers a plurality of rows and a plurality of columns of container berths, and each detection region is assigned a unique number; S1-3, write the pixel coordinates of the boundary vertices of each detection region in the image coordinate system into the detection region parameter set, and estimate the maximum arrangement number of containers in the length direction and the width direction and the upper limit of the vertical stacking number in each detection region according to the yard operation characteristics, which is used for subsequent cell division; S2 specifically includes: S2-1. Convert the RGB image within the detection area to the HSV color space. For each pixel, when its hue component H ≤ a preset threshold... Saturation component S ≤ preset threshold The brightness component V ≤ the preset threshold When the pixel is in shadow, it is identified as a shadow pixel and assigned a value of zero in the binarized image; otherwise, it is assigned a value of one, thus obtaining the binarization results of the shadow area and the non-shadow area. S2-2. In the binarized image, a length direction vector is set for each container size. Width direction vector Shadow direction vector and the tool starting point for each detection area. The number of containers along the length direction was obtained based on the observation results. Dimensions in the length direction Number of containers in the width direction and maximum vertical stacking layers Based on this, each detection area is classified according to From 0 to , From 0 to Integer indexes generate container cell collections At the same time, based on , , Generate a set of shadow cells arranged along the shadow direction for each container cell. ; S2-3, For each container cell Within the parallelogram of the cell, select a number of pixels and their neighboring pixels along a straight line passing through the intersection of the diagonals and parallel to the vertical axis of the image, and calculate the average pixel value of the container cell. Its value ranges from 0 to 1, for each shaded cell. ,since Start by calculating the average pixel value of the shaded cells one by one. Its value ranges from 0 to 1; S2-4. Use 0.5 as the threshold for shadow detection. Record the first time it exceeds the threshold. Value And calculate the value from the cell according to the following formula. Number of shadow lines produced ,in Indicates the first Line number The number of shadow bars corresponding to the column container cell. This indicates the first index that causes the average pixel value of the shaded cell to exceed 0.5; S2-5, When the average pixel value of the container cell The cell is then classified as a non-recessed cell, and the stacking layer of each cell is determined using the following formula. and reliability Perform sequential estimation: when hour, ,when When this cell is identified as a recessed cell, the number of stacking layers is randomly generated based on the distribution of observed data. The following formula is used to give the number of stacking layers and reliability of the recessed cells. ; S2-6, Reliability Cells with a reliability lower than 0.5 are marked as low-reliability cells and highlighted in the graphical interface. The operator then determines the stacking depth of these cells. Perform manual counting and modification, record the modified results as new observation data, and replace the corresponding cells. Then, repeat the above sequential estimation process until the reliability of all cells is not less than 0.5; S2-7. After obtaining the stable stacking layer number of each cell, calculate the container occupancy rate of each detection area using the following formula. and using natural numbers The detection area is divided into There are several congestion levels, and the width of each level is... ,when Falling in When a certain interval is reached, the detection area is classified as congested. , ; Most conventional methods only make two-dimensional density estimation and do not distinguish the vertical stacking layers, if the number of layers is to be estimated, either the shadow is completely manually observed, or the length of the shadow is approximately the height to make a simple conversion, there is no row and column physical constraint relationship between each cell, the present application is not a simple independent judgment for each pixel or small block, but according to the real arrangement of the stack, the detection area is accurately divided into container cells Cellc(i,j), and each container cell is configured with a string of shadow cells Cells(i,j,k) according to the shadow direction, the shadow mean Ds(i,j,k), the threshold first out-of-bound index k0 and the row recursive formula N(i,j)=N(i−1,j)+S(i,j) are used to force the stacking layer number of the previous row and the current row to be related, the row continuity + physical shadow constraint is imposed when the stacking layer number is inverted, which is completely different from the conventional method of independently estimating and classifying each block. Under the same image quality: the judgment of the shadow mean Ds(i,j,k) + the row recursive constraint reduces the influence of local noise and occluded cavities on the height estimation of the single cell, and the root mean square error of the overall stacking layer number N(i,j) is significantly smaller than the method of independent estimation of each cell; Especially in the edge area and the place where the light is uneven.

[0019] The present application maintains the local accuracy of a single cell while significantly suppressing the height estimation jump caused by local shadow loss or noise, so that the overall stacking layer number distribution is closer to the actual stacking state. The effect of the row constraint inversion cannot be expected by the conventional method of independent processing of each cell.

[0020] The conventional method has two ends: either the number of layers is completely manually counted, which is extremely costly, or it is fully automatic, but the estimation in the low-quality area is very poor and there is no feedback loop. The present application first calculates N(i,j) by shadow inversion + row recursive, and then constructs a number that indicates whether the estimation of the cell is reliable or not by P(i,j). Only the cells with P(i,j) below the threshold are manually corrected, and the rest of the cells are completely automatically accepted. The semi-automatic labeling mechanism driven by the reliability P(i,j) of the present application only hands over a small number of unreliable cells to manual judgment. Under the premise that the artificial workload is only a small part of the traditional full manual labeling method, the stacking layer distribution close to the full manual labeling accuracy can still be obtained. The effect of simultaneously achieving low labor cost and high precision cannot be expected by simply increasing or reducing the involvement of manual labor in the field.

[0021] The semi-automatic labeling mechanism driven by the reliability P(i,j) of the present application only hands over a small number of unreliable cells to manual judgment. Under the premise that the artificial workload is only a small part of the traditional full manual labeling method, the stacking layer distribution close to the full manual labeling accuracy can still be obtained. The effect of simultaneously achieving low labor cost and high precision cannot be expected by simply increasing or reducing the involvement of manual labor in the field.

[0022] S3 specifically is: S3-1, for each detection area, according to its actual pixel range, the image block containing complete yard cell and shadow information is obtained by cropping, the image block is scaled to the size consistent with the input layer of convolutional neural network, and the corresponding congestion level is taken as the supervision label to form the training sample set and the verification sample set; S3-2, a convolutional neural network model is constructed, the convolutional neural network includes four convolutional layers, two pooling layers, three dropout layers and two fully connected layers connected in sequence, the convolutional layer is used to extract multi-scale texture and shadow features in the local receptive field, the pooling layer is used to reduce the feature map dimension while keeping important features, the dropout layer is used to suppress overfitting, and the fully connected layer is used to map the extracted features to congestion level prediction results; S3-3, input the training sample set into the convolutional neural network, update the network parameters through forward propagation and back propagation, calculate the overall accuracy, recall rate and precision rate of each congestion level using the verification sample set, when the evaluation index reaches the preset threshold, freeze the network parameters and take the trained network as the yard congestion level identification model; S4 specifically is: S4-1, the satellite image of the date to be evaluated is cropped according to the same detection area configuration as in the training stage to obtain image blocks of each detection area, the preprocessing operations of normalization and size adjustment consistent with the training stage are performed, the processed image blocks are input into the trained convolutional neural network model, and the congestion level prediction results of each detection area are obtained; S4-2, map the congestion level results of all detection areas back to the yard geographical space position to form the yard congestion level distribution map, and when needed, count and summarize the congestion level according to the berth or wharf to obtain the average yard stacking layer corresponding to each berth for correlation analysis with AIS data; S5 specifically is: S5-1, according to the relationship between the berthing position of the ship in the AIS trajectory near the wharf and the spatial range of the berth, the nearest berthing process of each berth before and after satellite imaging is identified, the time interval between satellite imaging time and the arrival time of the next berthing ship is defined as the arrival time interval , the time interval between satellite imaging time and the departure time of the last departing ship is defined as the departure time interval , and the time interval and the corresponding berth are recorded as analysis samples; The smaller: the closer the next ship berths, the more nervous the export operation time, The larger, the last ship has been away for a long time.

[0023] S5-2, for at least one terminal, according to the yard layout provided by the terminal and the number of export, import and empty containers measured at multiple time points, dividing each detection area into an export detection area and an import detection area, and calculating the proportion of the number of containers exported or imported at multiple time points in each detection area to obtain an export proportion and an import proportion , wherein is the detection area number, is the berth number; S5-3, using the congestion level of the detection area and the export proportion and the import proportion , the average stacking level of the export direction and the import direction of each terminal is calculated by the following formula and : , ; The export direction: the capacity risk + the risk of loading delay, The larger the value is, the higher the average stacking level of the detection area of the export direction of the yard is, the more full the yard is, and the less space is available for new container stacking. In the same time interval , if is larger, the amount of containers that can be cleared before the arrival of the next ship is limited, and the export containers are difficult to collect and place in time, and there is a risk of loading delay. In the same stacking level , if is smaller, the ship arrives faster, and the time available for loading and clearing the yard is shortened, and the risk of delay increases under the same load.

[0024] The import direction: the risk of insufficient distribution + the risk of long-term congestion, The larger the value is, the higher the import containers are stacked in the yard, and the more containers are not distributed, in the same time interval , if is larger, it means that the time from the departure of the last ship is the same, but there are more import containers left in the yard, reflecting the inefficiency of land-side distribution and the existence of bottlenecks in the inland channel. In the same stacking level , if is larger, the ship departs later, and the longer the time is, the higher the yard remains , indicating that the import containers are long-term stranded, forming a long-term congestion risk of the yard.

[0025] By combining the congestion level estimated based on satellite images with AIS time intervals and ship sizes, the application can construct a monotonic relationship between the average stacking layer number and the time interval of arrival and departure in both export and import directions, thereby predicting the stacking yard capacity risk and operation delay risk under a given stacking state, realizing the upgrade from static stacking detection to dynamic port operation risk assessment. This technical route of evolving stacking parameters into risk indicators is not possessed by traditional satellite stacking detection methods.

[0026] S5-4, the average stacking layer number and the time interval of arrival and departure of each wharf under different time interval grouping conditions are combined with the corresponding ship size for joint analysis, a relationship diagram of the average stacking layer number and the time interval and the ship size is drawn, and the negative correlation characteristics of the average stacking layer number decreasing with the increase of the time interval and the positive correlation characteristics between the average stacking layer number and the ship size are judged according to the envelope trend, so as to provide quantitative basis for different wharf yard resource allocation strategies and operation efficiency evaluation, and set the yard congestion control threshold and early warning rules under different risk levels, realize the dynamic management and control of the port operation risk. With , With and the corresponding ship size are combined for joint analysis, a relationship diagram of the average stacking layer number and the time interval and the ship size is drawn, and the negative correlation characteristics of the average stacking layer number decreasing with the increase of the time interval and the positive correlation characteristics between the average stacking layer number and the ship size are judged according to the envelope trend, so as to provide quantitative basis for different wharf yard resource allocation strategies and operation efficiency evaluation, and set the yard congestion control threshold and early warning rules under different risk levels, realize the dynamic management and control of the port operation risk.

[0027] It should be noted that, in this text, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0028] Finally, it should be noted that: the above only describes the preferred embodiments of the application, and is not used to limit the application, although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A port risk management system based on low-Earth orbit satellites, characterized in that: It includes a stacking estimation module, which is used to divide the detection area, identify shadows, and configure container cells and shadow cells in high-resolution satellite images of port container yards. It sequentially estimates the vertical stacking layers and corresponding reliability of each cell, and calculates the container occupancy rate and congestion level of each detection area after manually correcting low-reliability cells, in order to generate yard congestion training data. The congestion level identification module is used to build a convolutional neural network model based on the congestion training data output by the stacking estimation module. It takes the cropped image of the detection area as input and the congestion level as output for training. After the model training is completed, it performs end-to-end inference on new satellite images and outputs the congestion level of each detection area. The AIS analysis and evaluation module is used to acquire AIS vessel trajectory data corresponding to the satellite image acquisition date, identify one arrival event and one departure event of vessels at each berth before and after satellite imaging, calculate the time interval between the satellite imaging time and the aforementioned arrival or departure time, combine the congestion level of each detection area to obtain the relationship between the average number of stacking layers in the yard and the time interval and vessel size, and evaluate the yard congestion evolution characteristics and dynamic operational efficiency of different terminals.

2. The port risk management system based on low-Earth orbit satellites according to claim 1, characterized in that: The stack estimation module includes: The image acquisition module is used to receive high-resolution satellite images covering multiple berths and yard areas from the satellite image acquisition terminal and store them in the memory at a preset spatial resolution; The detection area configuration module is used to allow users to specify several rectangular or parallelogram detection areas within the yard area in the display interface, and to write the pixel coordinate range of each detection area into the detection area configuration data. The HSV threshold segmentation module is used to convert the color image within the detection area to the HSV color space, and to binarize each pixel to determine shadow pixels and non-shadow pixels by using a preset threshold, and output a binarized image. The cell segmentation module is used to set the container length direction vector, width direction vector, and shadow direction vector, as well as the tool starting point, based on the yard layout. It generates a set of container cells and a corresponding set of shadow cells on the binary image according to the number of container rows and columns and the maximum stacking layer in each detection area. The pixel statistics stacking module is used to calculate the average pixel value for each container cell and shadow cell, determine the cell's concavity state, and sequentially estimate the vertical stacking layer and reliability of the cell. Cells with reliability below a preset threshold are then marked. The cell correction module is used to highlight cells with low reliability in the interface. The operator inputs the corrected cell stacking layer number, and the correction result is written into the observation data set before stacking estimation is re-executed. The congestion level generation module is used to calculate the container occupancy rate of each detection area according to preset rules after the number of stacking layers in the cell has stabilized, and divide it into multiple congestion levels accordingly. The congestion level is output as the yard image annotation result and stored. The congestion level identification module includes: The training data construction module is used to combine the image patches cropped from each detection region with the corresponding congestion level to form training samples and validation sample sets. The CNN model building module is used to build a convolutional neural network structure in the processor, which includes multiple convolutional layers, pooling layers, dropout layers, and fully connected layers, and to set the kernel size, stride, activation function, and loss function. The model training module is used to train the convolutional neural network with a training sample set as input and a preset optimization algorithm. It evaluates the accuracy, recall and precision metrics by verifying the sample set and adjusts the network parameters until the performance meets the requirements. The model inference module is used to input the detection area image blocks obtained by cropping new satellite images into the trained convolutional neural network during the inference stage, output the congestion level of each detection area and store it. The AIS analysis and evaluation module includes: The AIS data preprocessing module is used to receive time-series data containing information such as ship position, speed, hull number, and voyage destination from the AIS data acquisition terminal, clean the data, and divide the trajectory according to port berths. The berth event identification module is used to determine berthing events at each berth based on the relationship between the trajectory berthing position and the geometric range of the berth, and to identify vessels in port corresponding to the satellite imaging date and vessels that have most recently berthed and departed before and after the imaging. The time interval calculation module is used to calculate the arrival time interval for export analysis and the departure time interval for import analysis based on the satellite imaging time, the most recent arrival time and the most recent departure time of each berth. The stacking time analysis module is used to match the congestion level of each detection area with the time interval and ship size information of the corresponding berth, calculate the average stacking layer in the yard and analyze its relationship with the time interval and ship size. The visualization output module is used to generate a distribution map of yard congestion levels, a scatter plot of stacking layers and time intervals, and statistical charts related to ship size, and output them to a display terminal or external storage medium.

3. A port risk management method based on low-Earth orbit satellites, characterized by: Includes the following steps: S1. Acquire high-resolution satellite images covering multiple berths and storage yards, and delineate several detection areas for each storage yard in the images. Record the spatial range of each detection area as the basic unit for subsequent processing. S2. Perform shadow extraction and cell division on the images in each detection area. Perform binarization processing based on the difference between the shadow and the container body in the color space. Set container cells and corresponding shadow cells for each container according to preset geometric parameters. Estimate the vertical stacking layer and reliability of each cell by sequential calculation. After manually correcting cells with low reliability, obtain the container occupancy rate of each detection area and the congestion level derived from the occupancy rate. S3. Construct a training dataset using cropped images of each detection region and their corresponding congestion levels as samples, build a convolutional neural network model, and obtain an end-to-end image classification model that can automatically map the input detection region image to the congestion level through training. Then, evaluate the model performance using the validation dataset. S4. For satellite images of the date to be evaluated, crop each detection area and input it into the trained convolutional neural network model to obtain the predicted congestion level of each detection area. Then, summarize the prediction results into the congestion status distribution of the storage yard at that time. S5. Based on the AIS vessel trajectory data corresponding to the date to be evaluated, identify the berthed vessels at each berth before and after satellite imaging, their arrival and departure times, and calculate the time interval between the satellite imaging time and the arrival time of the next berthed vessel, as well as the time interval between the satellite imaging time and the departure time of the previous departing vessel. On this basis, correlate the congestion level of each detection area in the yard with the time interval and the corresponding vessel size, and analyze the evolution of the average stacking layer number in the yard over time.

4. The port risk management method based on low-orbit satellites according to claim 3, characterized in that: S1 specifically includes: S1-1. Obtain optical satellite images containing multiple berths of the container terminal from satellite image service platforms and other high-resolution image data sources, perform geographic coordinate correction and resolution unification processing on the images, and store them as two-dimensional raster data for subsequent image processing. S1-2. Load the above satellite imagery into the user interface. The operator shall, according to the actual layout of the yard, delineate a rectangular inspection area within the yard along the container arrangement direction. Each inspection area shall cover several rows and columns of container stacks and assign a unique number to each inspection area. S1-3. Write the pixel coordinates of the boundary vertices of each detection area in the image coordinate system into the detection area parameter set, and estimate the maximum number of containers arranged in the length and width directions and the upper limit of the vertical stacking layers in each detection area according to the characteristics of yard operations, for subsequent cell division.

5. The port risk management method based on low-Earth orbit satellites according to claim 4, characterized in that: Specifically, S2 is: S2-1. Convert the RGB image within the detection area to the HSV color space. For each pixel, when its hue component H ≤ a preset threshold... Saturation component S ≤ preset threshold The brightness component V ≤ the preset threshold When the pixel is in shadow, it is identified as a shadow pixel and assigned a value of zero in the binarized image; otherwise, it is assigned a value of one, thus obtaining the binarization results of the shadow area and the non-shadow area. S2-2. In the binarized image, a length direction vector is set for each container size. Width direction vector Shadow direction vector and the tool starting point for each detection area. The number of containers along the length direction was obtained based on the observation results. Dimensions in the length direction Number of containers in the width direction and maximum vertical stacking layers Based on this, each detection area is classified according to From 0 to , From 0 to Integer indexes generate container cell collections At the same time, based on , , Generate a set of shadow cells arranged along the shadow direction for each container cell. ; S2-3, For each container cell Within the parallelogram of the cell, select a number of pixels and their neighboring pixels along a straight line passing through the intersection of the diagonals and parallel to the vertical axis of the image, and calculate the average pixel value of the container cell. Its value ranges from 0 to 1, for each shaded cell. ,since Start by calculating the average pixel value of the shaded cells one by one. Its value ranges from 0 to 1; S2-4. Use 0.5 as the threshold for shadow detection. Record the first time it exceeds the threshold. Value And calculate the value from the cell according to the following formula. Number of shadow lines produced ,in Indicates the first Line 1 The number of shadow bars corresponding to the column container cell. This indicates the first index that causes the average pixel value of the shaded cell to exceed 0.5; S2-5, When the average pixel value of the container cell The cell is then classified as a non-recessed cell, and the stacking layer of each cell is determined using the following formula. and reliability Perform sequential estimation: when hour, ,when When this cell is identified as a recessed cell, the number of stacking layers is randomly generated based on the distribution of observed data. The following formula is used to give the number of stacking layers and reliability of the recessed cells. ; S2-6, Reliability Cells with a reliability lower than 0.5 are marked as low-reliability cells and highlighted in the graphical interface. The operator then determines the stacking depth of these cells. Perform manual counting and modification, record the modified results as new observation data, and replace the corresponding cells. Then, repeat the above sequential estimation process until the reliability of all cells is not less than 0.5; S2-7. After obtaining the stable stacking layer number of each cell, calculate the container occupancy rate of each detection area using the following formula. and using natural numbers The detection area is divided into There are several congestion levels, and the width of each level is... ,when Falling in When a certain interval is reached, the detection area is classified as congested. , .

6. The port risk management method based on low-Earth orbit satellites according to claim 5, characterized in that: Specifically, S3 is: S3-1. For each detection region, crop an image block containing complete stacking cell and shadow information according to its actual pixel range. Scale the image block to the same size as the input layer of the convolutional neural network and use the corresponding congestion level as the supervision label to form a training sample set and a validation sample set. S3-2. Construct a convolutional neural network model. The convolutional neural network includes four convolutional layers, two pooling layers, three dropout layers, and two fully connected layers connected in sequence. The convolutional layers are used to extract multi-scale texture and shadow features in the local receptive field. The pooling layers are used to reduce the dimension of the feature map while maintaining important features. The dropout layers are used to suppress overfitting. The fully connected layers are used to map the extracted features to the congestion level prediction result. S3-3. Input the training sample set into the convolutional neural network, update the network parameters through forward propagation and backward propagation, and use the validation sample set to calculate the overall accuracy, recall rate and precision evaluation indicators for each congestion level. When the evaluation indicators reach the preset threshold, freeze the network parameters and use the trained network as the yard congestion level identification model.

7. The port risk management method based on low-Earth orbit satellites according to claim 6, characterized in that: Specifically, S4 is: S4-1. The satellite images of the date to be evaluated are cropped into image blocks of each detection region according to the same detection region configuration as in the training phase. The same normalization and size adjustment preprocessing operations as in the training phase are performed. The processed image blocks are input into the trained convolutional neural network model to obtain the congestion level prediction results of each detection region. S4-2. Map the congestion level results of all detection areas back to the geospatial location of the yard to form a yard congestion level distribution map. When necessary, summarize the congestion levels according to berths or wharves to obtain the average number of stacking layers in the yard corresponding to each berth, which can be used for correlation analysis with AIS data.

8. The port risk management method based on low-Earth orbit satellites according to claim 7, characterized in that: Specifically, S5 is: S5-1. Based on the relationship between the berthing position of the vessel near the dock and the spatial range of the berth in the AIS trajectory, identify the most recent berthing process of each berth before and after satellite imaging, and define the time interval between the satellite imaging time and the arrival time of the next berthing vessel as the arrival time interval. The time interval between the satellite imaging time and the departure time of the previous departing vessel is defined as the departure time interval. The time interval and the corresponding berth are recorded as the analysis sample; S5-2. For at least one terminal, based on the yard layout provided by the terminal and the measured export, import, and empty container quantities at multiple times, divide each inspection area into an export inspection area and an import inspection area. Calculate the export or import container volume ratio for each inspection area at multiple time points to obtain the export ratio. Import ratio ,in The detection area is numbered. Number the berth; S5-3, Utilizing the congestion level of the detection area and export ratio Import ratio The average number of stacking layers in the export and import directions of each terminal can be calculated using the following formula. and : , ; S5-4, obtained by grouping each wharf under different time intervals. and , and In addition, a joint analysis was conducted on the corresponding ship size to draw a graph showing the relationship between the average number of stacking layers, the time interval, and the ship size. Based on the envelope trend, the negative correlation between the average number of stacking layers and the time interval and the positive correlation between the average number of stacking layers and the ship size were determined. This provides a quantitative basis for resource allocation strategies and operational efficiency assessments for different terminal yards. Based on this, yard congestion control thresholds and early warning rules were set for different risk levels to achieve dynamic management and control of port operation risks.

Images