Method, device and equipment for remote sensing identification and drift prediction of distressed targets on water

Abstract

The present disclosure relates to a method, device and equipment for remote sensing identification and drift prediction of targets in distress on the water. The present disclosure inputs a remote sensing image into a target intelligent detection model, so that the multiple feature layers contained in the backbone network of the target intelligent detection model extract features of the remote sensing image in turn. Since the multi-scale feature mining module can enhance the feature extraction capability of the target intelligent detection model, and the first deep-shallow feature fusion module and the second deep-shallow feature fusion module can fully interact the information between different feature layers, the remote sensing identification accuracy of targets in distress on the water is improved. Then, the drift prediction is performed in combination with the environmental prediction information corresponding to the preset spatial range around the drift position, and transfer learning is carried out, thereby improving the drift prediction accuracy of targets in distress on the water, and thus timely and accurately searching for targets in distress on the water.

Description

Technical Field

[0001] The present disclosure relates to the field of maritime technology, and in particular to a method, device, and equipment for remote sensing identification and drift prediction of targets in distress on water. Background Art

[0002] Searching for targets in distress at sea requires remote sensing identification (i.e., identifying and locating targets based on remote sensing technology) and drift prediction (i.e., predicting the target's future drift speed, position, trajectory, and search range based on the aquatic environment). However, current remote sensing identification and drift prediction accuracy are low, making it difficult to conduct timely and accurate searches for targets in distress. Summary of the Invention

[0003] In order to solve the above technical problems or at least partially solve the above technical problems, the present disclosure provides a method, device and equipment for remote sensing identification and drift prediction of distressed targets on water, which improves the accuracy of remote sensing identification and drift prediction of distressed targets on water, thereby enabling timely and accurate search for distressed targets on water.

[0004] In a first aspect, an embodiment of the present disclosure provides a method for remote sensing identification and drift prediction of a target in distress on water, the method comprising:

[0005] Acquire remote sensing images taken by remote sensing satellites;

[0006] Inputting the remote sensing image into a target intelligent detection model, the target intelligent detection model includes a backbone network, a multi-scale feature mining module, a first deep-shallow feature fusion module, a second deep-shallow feature fusion module and a detection head, the backbone network includes multiple feature layers that sequentially extract features from the remote sensing image, the multiple feature layers include a first feature layer, a second feature layer, and a third feature layer that are sequentially connected, and the third feature layer is the last feature layer among the multiple feature layers;

[0007] Performing multi-scale feature extraction on the first feature map output by the third feature layer by the multi-scale feature mining module to obtain a second feature map;

[0008] fusing the third feature map output by the first feature layer and the fourth feature map output by the second feature layer through the first deep-shallow feature fusion module to obtain a first fusion result; fusing the second feature map and the first fusion result through the second deep-shallow feature fusion module to obtain a second fusion result;

[0009] Inputting the third feature map, the first fusion result, and the second fusion result into the detection head, the detection head outputting the type of the distressed target on the water in the remote sensing image and the location information of the distressed target on the water in the remote sensing image;

[0010] determining a remote sensing observation position of the target in distress on the water according to position information of the target in distress on the water in the remote sensing image;

[0011] The drift of the target in distress on the water is predicted based on the remote sensing observation position and remote sensing observation time point of the target in distress on the water.

[0012] In a second aspect, an embodiment of the present disclosure provides a device for remote sensing identification and drift prediction of a target in distress on water, the device comprising:

[0013] An acquisition module is used to acquire remote sensing images taken by remote sensing satellites;

[0014] An input module is used to input the remote sensing image into a target intelligent detection model, wherein the target intelligent detection model includes a backbone network, a multi-scale feature mining module, a first deep-shallow feature fusion module, a second deep-shallow feature fusion module, and a detection head, wherein the backbone network includes multiple feature layers that sequentially extract features from the remote sensing image, wherein the multiple feature layers include a first feature layer, a second feature layer, and a third feature layer that are sequentially connected, and the third feature layer is the last feature layer among the multiple feature layers;

[0015] a multi-scale feature extraction module, configured to perform multi-scale feature extraction on the first feature map output by the third feature layer through the multi-scale feature mining module to obtain a second feature map;

[0016] a fusion module, configured to fuse the third feature map output by the first feature layer and the fourth feature map output by the second feature layer through the first deep-shallow feature fusion module to obtain a first fusion result; and fuse the second feature map and the first fusion result through the second deep-shallow feature fusion module to obtain a second fusion result;

[0017] a detection module, configured to input the third feature map, the first fusion result, and the second fusion result into the detection head, wherein the detection head outputs the type of the target in distress on the water in the remote sensing image and the location information of the target in distress on the water in the remote sensing image;

[0018] a determination module, configured to determine the remote sensing observation position of the target in distress on the water according to the position information of the target in distress on the water in the remote sensing image;

[0019] The drift prediction module is used to predict the drift of the target in distress on the water according to the remote sensing observation position and remote sensing observation time point of the target in distress on the water.

[0020] In a third aspect, an embodiment of the present disclosure provides an electronic device, including:

[0021] Memory;

[0022] processor; and

[0023] computer programs;

[0024] The computer program is stored in the memory and is configured to be executed by the processor to implement the method as described in the first aspect.

[0025] In a fourth aspect, an embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the method described in the first aspect.

[0026] In a fifth aspect, an embodiment of the present disclosure provides a computer program product, comprising a computer program, which implements the method described in the first aspect when executed by a processor.

[0027] The methods, devices, and equipment for remote sensing identification and drift prediction of distressed aquatic targets provided by the embodiments of the present disclosure sequentially extract features from the remote sensing image using a remote sensing image input into an intelligent target detection model. The multiple feature layers comprise a first feature layer, a second feature layer, and a third feature layer, connected in sequence, with the third feature layer being the last feature layer. Furthermore, a multi-scale feature mining module performs multi-scale feature extraction on the first feature map output by the third feature layer to obtain a second feature map. A first deep-shallow feature fusion module fuses the third feature map output by the first feature layer with the fourth feature map output by the second feature layer to obtain a first fusion result. A second deep-shallow feature fusion module fuses the second feature map and the first fusion result to obtain a second fusion result. Based on the third feature map, the first fusion result, and the second fusion result, the type of distressed aquatic target and its location in the remote sensing image are identified. Because the multi-scale feature mining module can enhance the feature extraction capabilities of the target intelligent detection model, and the first deep-shallow feature fusion module and the second deep-shallow feature fusion module can fully interact with information between different feature layers, the remote sensing recognition accuracy of distressed targets on the water is improved, specifically the accuracy of the remote sensing observation position of distressed targets on the water. Furthermore, based on the precise remote sensing observation location and remote sensing observation time point, the comprehensive impact of spatially distributed environmental forecast information on the drift of distressed targets on the water, as well as the nonlinear interaction of environmental information such as wind, waves, and currents and their nonlinear impact on the drift of distressed targets on the water, are considered. Furthermore, remote sensing observation information is combined to perform transfer learning and small target drift prediction updates. This allows for accurate drift prediction of distressed targets on the water, thereby enabling timely and accurate search for distressed targets on the water. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0029] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0030] Figure 1 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0031] Figure 2 A schematic diagram of an application scenario provided by an embodiment of the present disclosure;

[0032] Figure 3 A schematic diagram of an intelligent target detection model provided in an embodiment of the present disclosure;

[0033] Figure 4 A schematic diagram of an intelligent target detection model provided in an embodiment of the present disclosure;

[0034] Figure 5 A schematic diagram of a multi-scale feature mining module provided in an embodiment of the present disclosure;

[0035] Figure 6 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0036] Figure 7 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0037] Figure 8 A schematic diagram of a deep-layer and shallow-layer feature fusion module provided in an embodiment of the present disclosure;

[0038] Figure 9 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0039] Figure 10 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0040] Figure 11 A schematic diagram of a preset spatial range provided in an embodiment of the present disclosure;

[0041] Figure 12 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0042] Figure 13 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0043] Figure 14 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0044] Figure 15 A flow chart of the method for remote sensing identification and drift prediction of distressed targets on water provided in an embodiment of the present disclosure;

[0045] Figure 16 A schematic diagram of a small target search range provided by an embodiment of the present disclosure;

[0046] Figure 17 A schematic diagram of the structure of a device for remote sensing identification and drift prediction of distressed targets on water provided by an embodiment of the present disclosure;

[0047] Figure 18 A schematic diagram of the structure of an electronic device embodiment provided by the present disclosure. DETAILED DESCRIPTION

[0048] In order to more clearly understand the above-mentioned objectives, features and advantages of the present disclosure, the scheme of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features therein can be combined with each other in the absence of conflict.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, rather than all of the embodiments.

[0050] When searching for targets in distress on the water, it is necessary to perform remote sensing identification (i.e., identifying and locating the targets in distress on the water based on remote sensing technology) and drift prediction (i.e., predicting the future drift speed, drift position, drift trajectory, and search range of the targets in distress based on the water environment). However, current remote sensing identification accuracy is low, and drift prediction accuracy is also low, resulting in the inability to search for targets in distress on the water in a timely and accurate manner. To address this issue, the present disclosure provides a method for remote sensing identification and drift prediction of targets in distress on the water, which involves the following terminology:

[0051] Targets in distress: people in distress at sea or on water, ships (such as fishing boats, yachts, cargo ships, etc.), life rafts at sea, etc.

[0052] Remote sensing identification: Identification of distressed targets based on remote sensing technology, such as identifying the type and location information of distressed targets.

[0053] Drift prediction: Based on environmental information at sea or on water (such as wind field, wave field, current field), the drift speed, drift position, drift trajectory and search range of distressed targets at sea or on water are predicted over a period of time in the future.

[0054] Search range: The possible distribution range of the distressed target in the future after the distress.

[0055] Figure 1A flow chart of the method for remote sensing identification and drift prediction of targets in distress on water provided in an embodiment of the present disclosure. Specifically, the method can be executed by a device for remote sensing identification and drift prediction of targets in distress on water, which can be implemented in software and / or hardware, and can be configured on a server, a server cluster, or a terminal device. Among them, a server cluster can be a plurality of servers brought together to perform the same service, which appears to the client as if there is only one server. The server cluster can use multiple computers for parallel computing to obtain a higher computing speed, and can also use multiple computers for backup, so that the entire system can still operate normally after any machine breaks down. Terminal devices are, for example, mobile phones, PDAs, tablet computers, wearable devices with displays, desktop computers, laptop computers, all-in-one computers, smart home devices, etc. Figure 2 The schematic diagram of an application scenario applicable to an embodiment of the present disclosure is shown. After a remote sensing satellite 21 observes the waters where a target 20 in distress is located and obtains a remote sensing image, it transmits the remote sensing image to a server 22. Server 22 processes the remote sensing image to determine whether there is a target in distress. If there is a target in distress in the remote sensing image, server 22 can further identify the type of the target in distress and its location information in the remote sensing image based on the remote sensing image. Based on the location information of the target in distress in the remote sensing image, server 22 can calculate the geographic coordinates of the target in distress, i.e., the remote sensing observation location, thereby achieving remote sensing identification. Alternatively, after a remote sensing satellite 21 observes the waters where the target 20 in distress is located and obtains a remote sensing image, it first determines whether there is a target in distress in the remote sensing image. If there is a target in distress in the remote sensing image, remote sensing satellite 21 transmits the remote sensing image to server 22. The server 22 processes the remote sensing image to determine the type of the distressed target and the location information of the distressed target in the remote sensing image. Based on the location information of the distressed target in the remote sensing image, the server 22 calculates the geographical coordinates of the distressed target, i.e., the remote sensing observation position. The remote sensing observation position can be used as the actual location of the distressed target. The following is an introduction to the remote sensing identification and drift prediction method of the distressed target provided by the present disclosure in conjunction with the illustrations. Figure 1 As shown, the specific steps of this method are as follows:

[0056] S101. Acquire remote sensing images captured by remote sensing satellites.

[0057] For example, the server 22 receives remote sensing images sent by the remote sensing satellite 21 .

[0058] S102. Input the remote sensing image into a target intelligent detection model, wherein the target intelligent detection model includes a backbone network, a multi-scale feature mining module, a first deep-shallow feature fusion module, a second deep-shallow feature fusion module and a detection head, wherein the backbone network includes multiple feature layers for sequentially extracting features from the remote sensing image, wherein the multiple feature layers include a first feature layer, a second feature layer, and a third feature layer connected in sequence, and the third feature layer is the last feature layer among the multiple feature layers.

[0059] For example, a target intelligent detection model is deployed on the server 22, such as Figure 3 The target intelligent detection model 30 shown. Specifically, the target intelligent detection model includes a backbone network, a multi-scale feature mining module, a first deep-shallow feature fusion module, a second deep-shallow feature fusion module and a detection head. Among them, the backbone network can be a high-performance real-time target detector (YOLOX). The server 22 inputs the remote sensing image it receives into the target intelligent detection model. The target intelligent detection model first extracts features from the remote sensing image through the backbone network to obtain a feature map. For example, the backbone network includes multiple feature layers that sequentially extract features from the remote sensing image, and each feature layer outputs a feature map. For example, the backbone network includes 5 feature layers, and the output of the previous feature layer is the input of the next feature layer. As Figure 3 The fifth feature map 31, sixth feature map 32, third feature map 33, fourth feature map 34, and first feature map 35 are, in order, the feature maps output by the five feature layers. The third feature map 33 is recorded as the feature map output by the first feature layer, the fourth feature map 34 is recorded as the feature map output by the second feature layer, and the first feature map 35 is recorded as the feature map output by the third feature layer. In other words, the first, second, and third feature layers are three of the five feature layers, and the first, second, and third feature layers are connected in sequence, with the third feature layer being the last of the five feature layers.

[0060] S103 . Perform multi-scale feature extraction on the first feature map output by the third feature layer through the multi-scale feature mining module to obtain a second feature map.

[0061] For example, the feature map output by the third feature layer is recorded as the first feature map 35, and the multi-scale feature mining module is used to perform multi-scale feature extraction on the first feature map 35 to obtain the second feature map 36.

[0062] S104. Fuse the third feature map output by the first feature layer and the fourth feature map output by the second feature layer through the first deep-shallow feature fusion module to obtain a first fusion result; and fuse the second feature map and the first fusion result through the second deep-shallow feature fusion module to obtain a second fusion result.

[0063] For example, the feature map output by the first feature layer is recorded as the third feature map 33, and the feature map output by the second feature layer is recorded as the fourth feature map 34. The third feature map 33 and the fourth feature map 34 are fused by the first deep-shallow feature fusion module to obtain a first fusion result 37. Then, the second feature map 36 and the first fusion result 37 are fused by the second deep-shallow feature fusion module to obtain a second fusion result 38.

[0064] S105. Input the third feature map, the first fusion result, and the second fusion result into the detection head, and the detection head outputs the type of the distressed target on the water in the remote sensing image, and the location information of the distressed target on the water in the remote sensing image.

[0065] For example, the third feature map 33 is input to the first detection head 39, the first fusion result 37 is input to the second detection head 310, and the second fusion result 38 is input to the third detection head 311. The first detection head 39 outputs the type of the distressed aquatic object in the remote sensing image, such as a ship or cargo ship. The second detection head 310 outputs the location information of the distressed aquatic object in the remote sensing image. The third detection head 311 outputs the confidence level of the first two results. This confidence level can be the confidence level corresponding to the first two results separately, or it can be a total confidence level calculated based on the confidence levels corresponding to the first two results.

[0066] S106 . Determine the remote sensing observation position of the target in distress on the water according to the position information of the target in distress on the water in the remote sensing image.

[0067] Since each pixel in a remote sensing image corresponds to a geographic coordinate, the geographic coordinates of the target in distress on the water, ie, the remote sensing observation position, can be determined based on the position information of the target in distress on the water in the remote sensing image.

[0068] S107: performing drift prediction on the target in distress on the water according to the remote sensing observation position and remote sensing observation time point of the target in distress on the water.

[0069] For example, the moment when the remote sensing satellite 21 observes the distressed target 20 is recorded as the remote sensing observation time point. If the remote sensing image includes the distressed target 20, the remote sensing observation time point may be the moment when the remote sensing satellite 21 captured the remote sensing image. Furthermore, based on the remote sensing observation position and remote sensing observation time point of the distressed target 20, a drift prediction is performed for the distressed target 20. Specifically, the drift prediction for the distressed target 20 includes predicting the drift speed, drift position, drift trajectory, and search range of the distressed target 20 after the remote sensing observation time point.

[0070] The embodiment of the present disclosure inputs a remote sensing image into a target intelligent detection model so that the multiple feature layers contained in the backbone network of the target intelligent detection model sequentially extract features from the remote sensing image. The multiple feature layers include a first feature layer, a second feature layer, and a third feature layer connected in sequence, and the third feature layer is the last feature layer. Furthermore, a multi-scale feature mining module performs multi-scale feature extraction on the first feature map output by the third feature layer to obtain a second feature map. The first deep-shallow feature fusion module fuses the third feature map output by the first feature layer and the fourth feature map output by the second feature layer to obtain a first fusion result. The second deep-shallow feature fusion module fuses the second feature map and the first fusion result to obtain a second fusion result. Then, based on the third feature map, the first fusion result, and the second fusion result, the type of the distressed target on the water and the location information of the distressed target on the water in the remote sensing image are identified. Because the multi-scale feature mining module can enhance the feature extraction capabilities of the target intelligent detection model, and the first deep-shallow feature fusion module and the second deep-shallow feature fusion module can fully interact with information between different feature layers, the remote sensing recognition accuracy of distressed targets on the water is improved, specifically the accuracy of the remote sensing observation position of distressed targets on the water. Furthermore, based on the precise remote sensing observation location and remote sensing observation time point, the comprehensive impact of spatially distributed environmental forecast information on the drift of distressed targets on the water, as well as the nonlinear interaction of environmental information such as wind, waves, and currents and their nonlinear impact on the drift of distressed targets on the water, are considered. Furthermore, remote sensing observation information is combined to perform transfer learning and small target drift prediction updates. This allows for accurate drift prediction of distressed targets on the water, thereby enabling timely and accurate search for distressed targets on the water.

[0071] like Figure 4As shown, the intelligent target detection model described above may also include a spatial-spectral dual attention module. Specifically, the result obtained after processing the second fusion result 38 by the spatial-spectral dual attention module is fused with the first fusion result 37 described above to obtain a seventh feature map 312. The result obtained after processing the seventh feature map 312 by the spatial-spectral dual attention module is fused with the third feature map 33 to obtain an eighth feature map 313. The result obtained after processing the eighth feature map 313 by the spatial-spectral dual attention module is fused with the seventh feature map 312 to obtain a ninth feature map 314. The result obtained after processing the ninth feature map 314 by the spatial-spectral dual attention module is fused with the second fusion result 38 to obtain a tenth feature map 315. Then, the eighth feature map 313 is input to the first detection head 39, the ninth feature map 314 is input to the second detection head 310, and the tenth feature map 315 is input to the third detection head 311. The first detection head 39 outputs the type of the distressed aquatic target in the remote sensing image, such as a ship or cargo ship. The second detection head 310 outputs the location information of the distressed aquatic target in the remote sensing image. The third detection head 311 outputs the confidence levels of the first two results.

[0072] Optionally, the multi-scale feature mining module includes multiple dilation factors of the dilated convolution layer and the maximum pooling layer. Figure 5 The structure 50 of the multi-scale feature mining module shown is Figure 3 or Figure 4 The internal structure of the multi-scale feature mining module is shown in Figure 1. Among them, multiple dilated convolution layers with different expansion factors are as follows: Figure 5 3×3 convolutional layers with Rate=1, 3, 5, and 7 shown.

[0073] Specifically, the multi-scale feature mining module performs multi-scale feature extraction on the first feature map output by the third feature layer to obtain a second feature map, including: Figure 6 The following steps are shown:

[0074] S601 , performing feature extraction on the first feature map output by the third feature layer through the multiple dilation factor-different dilated convolution layers to obtain multi-scale features.

[0075] For example, multi-scale features are extracted from the first feature map 35 output by the third feature layer through 3×3 convolutional layers with Rate=1, 3, 5, and 7, respectively, to obtain multi-scale features.

[0076] S602: Extract key information from the first feature map through the maximum pooling layer.

[0077] For example, the key information in the first feature map 35 is extracted through the Global Max Pooling (GMP) layer.

[0078] S603: Concatenate the multi-scale features and the key information to obtain the second feature map.

[0079] For example, a normalization layer (BatchNorm) normalizes multi-scale features to produce a normalized result. A rectified linear unit (ReLU) is then used to process the normalized result and key information. Furthermore, the outputs of multiple rectified linear units are concatenated to produce a concatenated result. This concatenated result is processed using a 1×1 convolution to produce the second feature map 36.

[0080] The disclosed embodiment performs multi-scale feature extraction through dilated convolutions with multiple different dilation factors, which can effectively expand the receptive field range and thus capture more levels of contextual information. Secondly, the most representative features are extracted from the feature map (input) through the maximum pooling operation, so that the network can better focus on the key information in the image. Finally, the features of different scales are cascaded and the number of channels is optimized using a 1×1 convolutional layer, which not only fully mines the correlation between multi-scale features, but also effectively reduces the information loss in the feature combination process. Since the multi-scale feature mining module has significant advantages in feature extraction flexibility and information utilization, it can better adapt to the detection needs of targets of different scales.

[0081] Optionally, the third feature map output by the first feature layer and the fourth feature map output by the second feature layer are fused by the first deep-shallow feature fusion module to obtain a first fusion result, including: Figure 7 The following steps are shown:

[0082] S701. Perform multiple convolution operations of different scales on the third feature map output by the first feature layer to obtain a first convolution operation result.

[0083] Figure 8 The dotted box 80 is a schematic diagram of the structure of the first deep and shallow feature fusion module. The deep feature map is the fourth feature map 34 output by the second feature layer, and the shallow feature map is the third feature map 33 output by the first feature layer. Figure 8 As shown in FIG, a 1×1 convolution layer and a 3×3 convolution layer are used to perform multiple convolution operations of different scales on the shallow feature map in sequence to obtain the result of the first convolution operation.

[0084] S702: Perform multiple convolution operations of different scales on the fourth feature map output by the second feature layer to obtain a second convolution operation result.

[0085] like Figure 8 As shown in FIG, a 1×1 convolution layer and a 3×3 convolution layer are used to perform multiple convolution operations of different scales on the deep feature map in sequence to obtain the result of the second convolution operation.

[0086] S703: Concatenate the first convolution operation result and the second convolution operation result to obtain the first fusion result.

[0087] For example, after splicing the first convolution operation result and the second convolution operation result, a 1×1 convolution layer is used to perform a convolution operation to obtain a first fusion result.

[0088] Optionally, the second feature map and the first fusion result are fused by the second deep-layer shallow-layer feature fusion module to obtain a second fusion result, including: Figure 9 The following steps are shown:

[0089] S901. Perform multiple convolution operations of different scales on the second feature map in sequence to obtain a third convolution operation result.

[0090] For example, Figure 8 The dotted box 80 shown can also be a structural diagram of the second deep and shallow feature fusion module. Among them, the deep feature map is the second feature map 36, and the shallow feature map is the first fusion result. Figure 8 As shown in FIG, a 1×1 convolution layer and a 3×3 convolution layer are used to perform multiple convolution operations of different scales on the deep feature map in sequence to obtain the result of the third convolution operation.

[0091] S902: Perform multiple convolution operations of different scales on the first fusion result to obtain a fourth convolution operation result.

[0092] like Figure 8 As shown in FIG, a 1×1 convolution layer and a 3×3 convolution layer are used to perform multiple convolution operations of different scales on the shallow feature map in sequence to obtain the fourth convolution operation result.

[0093] S903: Concatenate the third convolution operation result and the fourth convolution operation result to obtain the second fusion result.

[0094] For example, after splicing the third convolution operation result and the fourth convolution operation result, a 1×1 convolution layer is used to perform a convolution operation to obtain a second fusion result.

[0095] The disclosed embodiment enables full interaction of information between different feature layers through the first deep-shallow feature fusion module and the second deep-shallow feature fusion module, thereby improving the performance of the target intelligent detection model in target recognition and positioning tasks. In addition, in the first deep-shallow feature fusion module and the second deep-shallow feature fusion module, multiple convolution operations of different scales are performed on the deep feature map and the shallow feature map respectively to integrate the channel attributes of the multi-scale features, and then the fusion features are constructed by channel splicing, and 1×1 convolution is used for dimensionality reduction, and finally the enhanced feature map is output. This design not only retains the unique advantages of features at each level, but also improves the model's adaptability to complex scenes and multi-scale targets through cross-scale information fusion.

[0096] Optionally, based on the remote sensing observation position and remote sensing observation time point of the target in distress on the water, drift prediction of the target in distress on the water is performed, including: Figure 10 The following steps are shown:

[0097] S1001: Obtain environmental prediction information corresponding to the remote sensing observation time point in a preset spatial range around the remote sensing observation location.

[0098] like Figure 2As shown, the position of the distress target 20 may change in real time, resulting in the remote sensing satellite 21 not always being able to observe the distress target 20 in real time. Therefore, the server 22 may not always be able to obtain the actual position of the distress target 20 in real time. Assuming that at a certain moment, such as time t1, the remote sensing satellite 21 observes the distress target 20, time t1 is recorded as the remote sensing observation time point. In this case, the remote sensing observation position of the distress target 20 is used as the starting position for drift prediction of the distress target 20, and the remote sensing observation time point is used as the starting time for drift prediction of the distress target 20. It should be understood that the starting position and starting time for drift prediction of the distress target 20 are not limited to the remote sensing observation position and remote sensing observation time point. For example, if another ship traveling in the same waters observes the distress target 20, the ship may send the time of its observation of the distress target 20, the position of the distress target 20 relative to the ship at the time of observation, and the ship's positioning information at the time of observation to the server 22. The server 22 determines the position of the distress target 20 at the time of observation based on the position of the distress target 20 relative to the ship at the time of observation and the ship's positioning information at the time of observation, and uses this position as the starting position for drift prediction of the distress target 20, and the observation time as the starting time for drift prediction of the distress target 20. Alternatively, the starting position for drift prediction of the distress target 20 may be the position of the distress target 20 at the time of distress, and the starting time for drift prediction of the distress target 20 may be the time when the distress target 20 occurs. That is, the starting position and starting time that can be used for drift prediction of the target 20 in distress on the water are the actual position and time of the target 20 in distress on the water, for example, the actual position and time of the target 20 in distress on the water that the server 22 obtained most recently.

[0099] The embodiment of the present disclosure takes the remote sensing observation position and remote sensing observation time point as the starting position and starting time of drift prediction as an example. Specifically, the server 22 obtains the environmental prediction information corresponding to the remote sensing observation time point in the preset spatial range around the remote sensing observation position. Figure 11As shown, 110 represents the water area where the target 20 in distress on the water is located, 111 represents the remote sensing observation position, and 112 represents the preset spatial range around the remote sensing observation position. The preset spatial range can be a spatial range with a longitude range of lon1 to lon2 and a latitude range of lat1 to lat2 centered on the remote sensing observation position. The environmental prediction information corresponding to the preset spatial range at the remote sensing observation time point can be the wind field forecast result, wave field forecast result, and flow field forecast result at a height of 10 meters above the water surface within the preset spatial range. Among them, the wind field forecast result includes wind speed and wind direction. The wave field forecast result includes wave height, wave direction, wavelength, wave period, etc. The flow field forecast result includes flow velocity and flow direction. For example, if the preset spatial range is divided into multiple high-resolution grids (e.g., with a high resolution of ten meters), the wind field forecast results, wave field forecast results, and flow field forecast results at a height of 10 meters above the water surface within the preset spatial range are, in order, the wind field forecast results, wave field forecast results, and flow field forecast results at a height of 10 meters at each vertex of the high-resolution grid within the preset spatial range. It will be understood that the wind field forecast results, wave field forecast results, and flow field forecast results at a height of 10 meters at the vertex of the same high-resolution grid change over time.

[0100] S1002. Generate multiple initial positions for drift prediction of the distressed target on the water based on the remote sensing observation position, generate multiple initial times for the drift prediction based on the remote sensing observation time point, generate multiple initial environmental information for the drift prediction based on the environmental prediction information corresponding to the remote sensing observation time point, and construct multiple groups of initial conditions based on the multiple initial positions, the multiple initial times and the multiple initial environmental information, each group of initial conditions includes any initial position, any initial time and any initial environmental information, and the drift prediction includes prediction of drift speed, drift position, drift trajectory and search range.

[0101] For example, a random error of normal distribution is added to the remote sensing observation position to obtain multiple initial positions, which constitute a set representing the uncertainty of the spatial position of the distress target on the water. The set is recorded as By adding the random error of normal distribution to the remote sensing observation time point, multiple initial times are obtained. These multiple initial times constitute a set that characterizes the time position uncertainty of the distress target on the water. This set is recorded as . Adding normally distributed random errors to the environmental prediction information corresponding to the remote sensing observation time point, a plurality of initial environmental information is obtained. For example, the environmental prediction information corresponding to the remote sensing observation time point includes wind field forecast results, wave field forecast results, and flow field forecast results. Further, adding normally distributed random errors to the wind field forecast results, the wave field forecast results, and the flow field forecast results, respectively, obtains a plurality of wind field disturbance results (i.e., the result of the wind field forecast result after multiple disturbances), a plurality of wave field disturbance results (i.e., the result of the wave field forecast result after multiple disturbances), and a plurality of flow field disturbance results (i.e., the result of the flow field forecast result after multiple disturbances). Furthermore, the wind field forecast result and the plurality of wind field disturbance results constitute a set that characterizes the uncertainty of the final wind field forecast result, and the set is recorded as The wave field prediction result and the multiple wave field disturbance results constitute a set that characterizes the uncertainty of the final wave field prediction result, and the set is recorded as The flow field prediction result and the multiple flow field disturbance results constitute a set that characterizes the uncertainty of the final flow field prediction result, which is recorded as Each initial environment information as described above includes Any element in (such as the wind field forecast result or any wind field disturbance result), Any element in (such as the wave field forecast result or any wave field disturbance result), and Any element in (such as the flow field prediction result or any flow field disturbance result).

[0102] In getting 、 、 、 、 After these sets are collected, a multivariate random forest sampling method is used to randomly extract an element from each set to form a combination. This combination is recorded as a set of initial conditions. In other words, each set of initial conditions includes any initial position, any initial time, and any initial environment information. For example, N random samplings are performed to obtain N sets of initial conditions.

[0103] S1003. Predict, based on each set of initial conditions, the drift speed and drift position of the target in distress on the water at multiple target time points after any of the initial times. The drift speed of the target in distress on the water at each target time point is obtained based on environmental prediction information corresponding to a preset spatial range around the drift position of the target in distress on the water at the target time point. The drift positions of the target in distress on the water at the multiple target time points constitute a drift trajectory.

[0104] For example, for each of the N groups of initial conditions, the group of initial conditions includes any initial position, any initial time, and any initial environmental information. Based on the any initial environmental information, the drift speed of the target 20 in distress on the water at the any initial time is calculated. Assume that what is to be predicted here is the drift trajectory of the target 20 in distress on the water from the any initial time to a certain moment in the future, for example, the drift trajectory of the subsequent M hours starting from the any initial time. There are M target time points after the any initial time, and the first target time point of the M target time points is 1 hour away from the any initial time, and the interval between each two subsequent adjacent target time points is 1 hour. Further, based on the drift speed at the any initial time and the time interval between the first target time point and the any initial time, that is, 1 hour, the drift position of the target 20 in distress on the water at the first target time point can be calculated. Furthermore, the drift speed of the target 20 at the first target time point is calculated based on the environmental prediction information (e.g., wind field forecast results, wave field forecast results, and current field forecast results) for a preset spatial range surrounding the target 20's drift position at the first target time point. Similarly, the drift position of the target 20 at the second target time point can be calculated based on the drift speed of the target 20 at the first target time point and the time interval of one hour between the first and second target time points. Similarly, the drift speed of the target 20 at the second target time point is calculated based on the environmental prediction information for a preset spatial range surrounding the target 20's drift position at the second target time point. In other words, the drift speed of the target 20 at each target time point is obtained based on the environmental prediction information corresponding to that target time point. Furthermore, the drift positions of the target 20 at the M target time points constitute a drift trajectory.

[0105] S1004: Predicting a search range for the target in distress on the water based on an end point of each drift trajectory in the plurality of drift trajectories corresponding to the plurality of sets of initial conditions.

[0106] For example, a drift trajectory can be generated based on each set of initial conditions described above. Thus, N drift trajectories can be generated based on the N sets of initial conditions. Furthermore, based on the endpoints of each of the N drift trajectories, i.e., based on the N endpoints, a search range for the target 20 in distress on the water is predicted. This search range can be a range that includes the N endpoints. For example, a minimum envelope rectangle is determined based on the N endpoints, and this minimum envelope rectangle is recorded as the predicted search range result.

[0107] The embodiment of the present disclosure obtains the starting position, starting time and environmental prediction information for drift prediction of the target in distress on the water, and then generates multiple initial positions based on the starting position, generates multiple initial times based on the starting time, and generates multiple initial environmental information based on the environmental prediction information corresponding to the starting time. Furthermore, multiple groups of initial conditions are constructed based on the multiple initial positions, the multiple initial times and the multiple initial environmental information. For each group of initial conditions in the multiple groups of initial conditions, the drift speed of the target in distress on the water at the target time point is accurately predicted based on the environmental prediction information corresponding to each target time point after any initial time included in the initial conditions. Thus, based on the drift speed of the target in distress on the water at the previous target time point and the time interval between two adjacent target time points, the drift position of the target in distress on the water at each target time point is accurately predicted, so that the drift position of the target in distress on the water at multiple target time points constitutes an accurate drift trajectory. Furthermore, based on the end point of each drift trajectory in the multiple drift trajectories corresponding to the multiple sets of initial conditions, the search range of the distressed target on the water is accurately predicted, thereby improving the drift prediction accuracy of the distressed target on the water, so as to search for the distressed target on the water in a timely and accurate manner.

[0108] Optionally, obtaining environmental prediction information corresponding to a preset spatial range around the remote sensing observation location at the remote sensing observation time point includes: inputting position information corresponding to multiple first-resolution grids contained in the preset spatial range around the remote sensing observation location, environmental prediction information corresponding to the multiple first-resolution grids at the remote sensing observation time point, and position information corresponding to multiple second-resolution grids contained in the preset spatial range into a downscaling deep learning model, and using the environmental prediction information corresponding to the multiple second-resolution grids output by the downscaling deep learning model at the remote sensing observation time point as the environmental prediction information corresponding to the preset spatial range around the remote sensing observation location at the remote sensing observation time point, wherein the second resolution is greater than the first resolution.

[0109] Specifically, the embodiment of the present disclosure uses the first resolution grid as the low resolution grid and the second resolution grid as the high resolution grid. Figure 11Taking the preset spatial range 112 around the remote sensing observation location shown as an example, this preset spatial range 112 can be divided into multiple low-resolution grids or multiple high-resolution grids. The number of high-resolution grids into which this preset spatial range 112 is divided is greater than the number of low-resolution grids. Because the number of low-resolution grids within this preset spatial range 112 is smaller than the number of high-resolution grids, the storage space required to store the location information of the low-resolution grids (e.g., the longitude and latitude of the low-resolution grid vertices) and the environmental prediction information corresponding to the low-resolution grids (e.g., the wind field forecast results, wave field forecast results, and flow field forecast results at a height of 10 meters at the low-resolution grid vertices) is relatively small, and the computation time required to predict the environmental prediction information corresponding to the low-resolution grids is also shorter. The environmental prediction information corresponding to the high-resolution grid (e.g., wind field forecast results, wave field forecast results, and flow field forecast results at a height of 10 meters at the high-resolution grid vertex) is more detailed than the environmental prediction information corresponding to the low-resolution grid. Therefore, when a distress incident occurs on the water, the environmental prediction information corresponding to the high-resolution grid can be predicted based on the environmental prediction information corresponding to the low-resolution grid and the position information of the low-resolution grid. Thus, the drift prediction of the distressed target on the water can be performed based on the environmental prediction information corresponding to the high-resolution grid. Specifically, the position information corresponding to the multiple low-resolution grids within the preset spatial range 112, the environmental prediction information corresponding to the multiple low-resolution grids at the remote sensing observation time point, and the position information corresponding to the multiple high-resolution grids within the preset spatial range 112 (e.g., the latitude and longitude of the high-resolution grid vertices) are input into the downscaling deep learning model. The downscaling deep learning model outputs the environmental prediction information corresponding to the multiple high-resolution grids at the remote sensing observation time point. Here, the environmental prediction information corresponding to the multiple high-resolution grids at the remote sensing observation time point output by the downscaling deep learning model is used as the environmental prediction information corresponding to the preset spatial range 112 at the remote sensing observation time point. That is to say, after a distress accident occurs on the water, when subsequent drift prediction is needed, the downscaling deep learning model is used to quickly and intelligently downscale the low-resolution environmental prediction information within the required time period to obtain high-resolution wind field forecast results, wave field forecast results, and flow field forecast results at a height of 10 meters above the water surface within the required time period.

[0110] Optionally, the environmental prediction information includes wind field forecast results, wave field forecast results, and flow field forecast results; in the embodiment of the present disclosure, the downscaling deep learning model can specifically be a wind-wave-current coupled downscaling deep learning model based on a physical information constrained convolutional long short-term memory network. Specifically, the process of constructing a convolutional long short-term memory network can be the following process: constructing an encoder-decoder network, the encoder extracts features of low-resolution input data, and the decoder generates high-resolution output. Convolutional layers (Conv2D) are used in the encoder and decoder to extract spatial features, and transposed convolutional layers (Conv2DTranspose) are used in the decoder to upsample features and generate high-resolution outputs. A convolutional long short-term memory (ConvLSTM) module is added to the encoder-decoder structure to capture the dependencies between adjacent time solutions in the time series. The downscaling deep learning model is based on the following: Figure 12 The following steps are shown:

[0111] S1201. Calculate environmental prediction information for a plurality of first-resolution grids in a target water area at preset time intervals within a preset future time period based on a wind-wave-current coupling prediction model, wherein the wind-wave-current coupling prediction model is a coupling model of an atmospheric numerical model, a wave numerical model, and a hydrodynamic numerical model.

[0112] In the disclosed embodiments, the atmospheric numerical model can output wind field forecast results, the wave numerical model can output wave field forecast results, and the hydrodynamic numerical model can output flow field forecast results. Specifically, the atmospheric numerical model can be the Weather Research & Forecasting Model (WRF). The hydrodynamic numerical model can be a finite-volume community ocean model (FVCOM) or a regional ocean model (ROMS). The wave numerical model can be a nonlinear random wave numerical model (SWAN) or a wind and wave simulation framework (WAVEWATCH). In the disclosed embodiments, the atmospheric numerical model, the wave numerical model, and the hydrodynamic numerical model can be coupled via a coupler to produce a wind-wave-current coupled forecast model, allowing the atmospheric numerical model, the wave numerical model, and the hydrodynamic numerical model to exchange relevant variables. For example, the atmospheric numerical model provides the wave numerical model and the hydrodynamic numerical model with the wind field and pressure field at a height of 10 meters above the water. The wave numerical model provides the effective wave height, wave direction, wavelength, and spectrum peak period to the hydrodynamic numerical model. The hydrodynamic numerical model provides the flow velocity and water level to the wave numerical model.

[0113] Specifically, according to the wind-wave-current coupled prediction model, simulation is carried out to obtain environmental prediction information of multiple low-resolution grids in the target water area at preset time intervals within the future preset time period. For example, at a fixed time every day, environmental prediction information of each low-resolution grid in the target water area is obtained every hour in the next 7 days. The environmental prediction information is recorded as the first simulation result, and the first simulation result is recorded as The wind field forecast result included in the first simulation result is recorded as The wave field forecast result included in the first simulation result is recorded as The flow field prediction result included in the first simulation result is recorded as .

[0114] S1202: Calculate, based on the wind-wave-current coupling prediction model, environmental prediction information of a plurality of second-resolution grids in the target waters at preset time intervals within the future preset time period.

[0115] For example, the wind-wave-current coupled forecast model can also be used to calculate the environmental forecast information of each high-resolution grid in the target waters every hour in the next 7 days. The environmental forecast information is recorded as the second simulation result, and the second simulation result is recorded as The wind field forecast result included in the second simulation result is recorded as The wave field forecast result included in the second simulation result is recorded as The flow field prediction result included in the second simulation result is recorded as It is understandable that the time granularity corresponding to the second simulation result is the same as that of the first simulation result, except that the second simulation result has more spatial environmental prediction information than the first simulation result.

[0116] S1203. Input the position information corresponding to the multiple first-resolution grids, the environmental prediction information of the multiple first-resolution grids at preset time intervals within the future preset time period, and the position information corresponding to the multiple second-resolution grids into the downscaling deep learning model to be trained, and the downscaling deep learning model outputs the environmental prediction information of the multiple second-resolution grids at preset time intervals within the future preset time period.

[0117] For example, the position information corresponding to the multiple low-resolution grids contained in the target water area, the environmental prediction information of the multiple low-resolution grids every hour in the next 7 days, and the position information corresponding to the multiple high-resolution grids contained in the target water area are input into the downscaling deep learning model to be trained, so that the downscaling deep learning model outputs the environmental prediction information of the multiple high-resolution grids every hour in the next 7 days.

[0118] S1204. Train the downscaling deep learning model based on the environmental prediction information of the multiple second-resolution grids output by the downscaling deep learning model at preset time intervals within the future preset time period, and the environmental prediction information of the multiple second-resolution grids calculated by the wind-wave-current coupling forecast model at preset time intervals within the future preset time period.

[0119] For example, the multiple high-resolution grids output by the downscaling deep learning model are respectively the environmental forecast information and The difference between In addition, the cost function terms of the control equations of the atmospheric numerical model, the hydrodynamic numerical model, and the wave numerical model are respectively recorded as 、 、 . 、 、 It is used to constrain the downscaled wind field forecast results, wave field forecast results, and flow field forecast results to meet physical constraints. Further, the loss function is constructed. , and according to the loss function The downscaling deep learning model is trained. The loss function It is expressed as the following formula (1):

[0120] (1)

[0121] in, 、 、 Represent the coefficients respectively.

[0122] In addition, in some other embodiments, The wind field forecast results, wave field forecast results, and flow field forecast results at a height of 10 meters above the water surface, as well as the longitude and latitude of the low-resolution grid vertices, time information, and the longitude and latitude of the high-resolution grid vertices are used as inputs to the downscaling deep learning model, so that the downscaling deep learning model outputs the environmental prediction information corresponding to the high-resolution grid. Further, based on the environmental prediction information corresponding to the high-resolution grid output by the downscaling deep learning model and the time information corresponding to the high-resolution grid, the downscaling deep learning model is used to calculate the environmental prediction information corresponding to the high-resolution grid. The downscaling deep learning model is trained based on the difference between them.

[0123] Optionally, based on each set of initial conditions, the drift speed and drift position of the target in distress on water at multiple target time points after any initial time are predicted, including: Figure 13 The following steps are shown:

[0124] S1301. Input any of the initial environmental information into a drift speed intelligent prediction model corresponding to the target in distress on the water, and the drift speed intelligent prediction model outputs the drift speed of the target in distress on the water at any of the initial times.

[0125] For example, for each of the N groups of initial conditions described above, the group of initial conditions includes any initial position, any initial time, and any initial environmental information. The any initial environmental information can be regarded as environmental prediction information corresponding to the preset spatial range around the any initial position at the any initial time. The any initial environmental information is input into the drift speed intelligent prediction model corresponding to the distress target 20 on the water, and the drift speed intelligent prediction model outputs the drift speed of the distress target 20 on the water at the any initial time. The drift speed intelligent prediction model corresponding to the distress target 20 on the water can be a drift speed intelligent prediction model corresponding to the type of the distress target 20 on the water. For example, when the distress target 20 on the water is a ship, the drift speed intelligent prediction model corresponding to the distress target 20 on the water is the drift speed intelligent prediction model of the ship.

[0126] S1302: For each target time point after any initial time, calculate the drift position of the target in distress on the water at the target time point based on the drift speed and drift position of the target in distress on the water at the previous target time point.

[0127] For example, there are M target time points after the initial time. The first of the M target time points is one hour away from the initial time, and each subsequent target time point is one hour away from each other. Based on the drift speed of the object 20 in distress at the initial time and the one-hour interval between the first target time point and the initial time, the displacement of the object 20 in distress within the one-hour period can be calculated. Furthermore, based on any initial position included in the set of initial conditions (i.e., the drift position of the object 20 in distress at the initial time) and the displacement, the drift position of the object 20 in distress at the first target time point can be calculated.

[0128] S1303. Input the environmental prediction information corresponding to the preset spatial range around the drift position of the target in distress on the water at the target time point into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed of the target in distress on the water at the target time point.

[0129] In an embodiment of the present disclosure, the environmental prediction information corresponding to the preset spatial range around the drift position of the water distress target 20 at the current moment is input into the drift speed intelligent prediction model corresponding to the water distress target 20, and the drift speed intelligent prediction model can output the drift speed of the water distress target 20 at the current moment.

[0130] For example, the environmental prediction information corresponding to a preset spatial range around the drift position of the target 20 in distress on the water at the first target time point is input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at the first target time point. Further, based on the drift speed and drift position of the target 20 in distress on the water at the first target time point, and the time interval between the first target time point and the second target time point, i.e., one hour, the drift position of the target 20 in distress on the water at the second target time point is calculated. Further, the environmental prediction information corresponding to a preset spatial range around the drift position of the target 20 in distress on the water at the second target time point is input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at the second target time point, and so on. That is to say, for each target time point after any initial time, the environmental prediction information corresponding to the preset spatial range around the drift position of the target 20 in distress on the water at the target time point is input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at the target time point.

[0131] In other embodiments, the environmental prediction information corresponding to the preset spatial range around the drift position of the water distress target 20 at the current moment, and the environmental prediction information corresponding to the preset spatial range around the drift position of the water distress target 20 at historical moments can also be input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model can output the drift speed of the water distress target 20 at the current moment.

[0132] For example, the environmental prediction information corresponding to a preset spatial range around the drift position of the target 20 in distress on the water at the first target time point, as well as any initial environmental information, is input into the drift speed intelligent prediction model. The drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at the first target time point. Furthermore, based on the drift speed and drift position of the target 20 in distress on the water at the first target time point, and the time interval between the first target time point and the second target time point, i.e., one hour, the drift position of the target 20 in distress on the water at the second target time point is calculated. Furthermore, the environmental prediction information corresponding to the preset spatial range around the drift position of the target 20 in distress on the water at the second target time point, the environmental prediction information corresponding to the preset spatial range around the drift position of the target 20 in distress on the water at the first target time point, and any initial environmental information are input into the drift speed intelligent prediction model. The drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at the second target time point, and so on. That is to say, for each target time point after any initial time, the environmental prediction information corresponding to the preset spatial range around the drift position of the water distress target 20 at the target time point, and the environmental prediction information corresponding to the preset spatial range around the drift position of the water distress target 20 before the target time point are input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed of the water distress target 20 at the target time point.

[0133] Optionally, the drift speed intelligent prediction model corresponding to the distressed target on water is based on Figure 14 The following steps are shown:

[0134] S1401. Obtain, through the Beidou positioning system, drift positions of an experimental target representing the target in distress on water corresponding to multiple preset time points during free drifting in the target water area.

[0135] The intelligent drift velocity prediction model for distressed aquatic targets described above is trained using data from aquatic drift experiments. In these experiments, a representative target (e.g., a life raft, lifeboat, dummy, or debris) is dropped into the target waters and allowed to drift freely. If the target is a ship, the ship's power is shut down, allowing it to drift freely. During the experiment, the Beidou positioning system is used to obtain the target's drift position during the free drift process. For example, the drift position corresponding to multiple preset time points during the free drift process is obtained.

[0136] S1402: Calculate the drift speeds of the experimental target corresponding to the plurality of preset time points according to the drift positions corresponding to the plurality of preset time points.

[0137] For example, the drift speeds of the experimental target corresponding to the multiple preset time points are calculated based on the drift positions of the experimental target corresponding to the multiple preset time points and the time interval between every two adjacent preset time points.

[0138] S1403. For each preset time point among the multiple preset time points, input the environmental prediction information corresponding to the preset spatial range around the drift position of the experimental target at the preset time point into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed corresponding to the experimental target at the preset time point.

[0139] In an embodiment of the present disclosure, environmental prediction information corresponding to a preset spatial range around the drift position of the experimental target at the current moment is input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model can output the drift speed of the experimental target at the current moment.

[0140] For example, for each of the multiple preset time points, the environmental prediction information corresponding to the preset spatial range around the drift position corresponding to the preset time point of the experimental target is input into the drift speed intelligent prediction model to be trained, and the drift speed intelligent prediction model outputs the drift speed corresponding to the experimental target at the preset time point.

[0141] In other embodiments, the environmental prediction information corresponding to the preset spatial range around the drift position of the experimental target at the current moment and the environmental prediction information corresponding to the preset spatial range around the drift position of the experimental target at historical moments can also be input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model can output the drift speed of the experimental target at the current moment.

[0142] For example, for each of the multiple preset time points, the environmental prediction information corresponding to the preset spatial range around the drift position of the experimental target at the preset time point, and the environmental prediction information corresponding to the preset spatial range around the drift position of the experimental target before the preset time point are input into the drift speed intelligent prediction model to be trained, and the drift speed intelligent prediction model outputs the drift speed corresponding to the preset time point.

[0143] S1404. Training the drift speed intelligent prediction model according to the calculated drift speed corresponding to the experimental target at the preset time point and the drift speed corresponding to the experimental target at the preset time point output by the drift speed intelligent prediction model.

[0144] For example, for each of the multiple preset time points, the difference between the drift speed of the experimental target at the preset time point output by the drift speed intelligent prediction model and the drift speed of the experimental target at the preset time point calculated in the above S1402 is calculated, and the drift speed intelligent prediction model is trained based on the difference.

[0145] In the disclosed embodiments, for each type of distressed aquatic target, an on-water drift experiment is conducted. Using the on-water drift experiment data, a specific intelligent drift speed prediction model corresponding to that distressed aquatic target is trained. Examples include intelligent drift speed prediction models for ships, people overboard, life rafts, and wreckage. Unless otherwise specified, these models are collectively referred to as intelligent drift speed prediction models corresponding to distressed aquatic targets.

[0146] Since the embodiment of the present disclosure inputs the environmental prediction information within a certain spatial range around the location of the experimental target (i.e., the output of the downscaled deep learning model) into the drift speed intelligent prediction model to be trained, rather than inputting the actual data such as the wind field, wave field, flow field, etc. at the location of the experimental target into the drift speed intelligent prediction model to be trained. Therefore, compared with the traditional water drift experiment, the water drift experiment described in the embodiment of the present disclosure does not require synchronous follow-up observation of the actual data such as the wind field, wave field, flow field, etc. at the location of the experimental target, thereby greatly reducing the experimental cost. In addition, since there is no need to synchronously follow-up observation of the actual data such as the wind field, wave field, flow field, etc. at the location of the experimental target, the number of experiments and experimental data can be increased when conducting water drift experiments in bad weather. In addition, since the embodiment of the present disclosure uses the output of the downscaled deep learning model as part of the input of the drift speed intelligent prediction model to be trained, the system error of the downscaled deep learning model can be optimized during the training of the drift speed intelligent prediction model, so that the trained drift speed intelligent prediction model does not need to consider the impact of the system error of the downscaled deep learning model in the subsequent reasoning process or application process, that is, when predicting the drift of actual water distress targets, thereby improving the accuracy of drift prediction, that is, improving the prediction accuracy of the drift trajectory and search range of the water distress targets.

[0147] Optionally, the method further includes: when the remote sensing satellite observes the distressed target on the water again, performing transfer learning on the drift speed intelligent prediction model corresponding to the distressed target on the water based on the current observation position and current observation time point of the distressed target on the water, and the historical observation position and historical observation time point of the distressed target on the water.

[0148] For example, when the remote sensing satellite observes the distress target 20 on the water again, the current observation position and the current observation time point are recorded. Furthermore, based on the current observation position and the current observation time point, as well as the historical observation position and the historical observation time point, the drift speed intelligent prediction model corresponding to the distress target 20 on the water is transferred and learned. The historical observation position includes the position of the distress target 20 on the water when it was in distress, and / or the position of the distress target 20 on the water when it was observed before the current observation time point. The historical observation time point includes the time when the distress target 20 on the water was in distress, and / or the time when the distress target 20 on the water was observed before the current observation time point.

[0149] For example, the historical observed locations of a target 20 in distress on water include loc1 and loc2. The historical observed time points include T1 and T2. Loc1 corresponds to T1, and loc2 corresponds to T2. The current observed location of the target 20 in distress on water is recorded as loc3, and the current observed time point is recorded as T3. The displacement is calculated based on loc1 and loc2, and the time interval is calculated based on T1 and T2. The drift velocity calculated based on this displacement and time interval can be used as the drift velocity of the target 20 in distress on water at time T2, and this drift velocity is recorded as V2. Similarly, the displacement is calculated based on loc2 and loc3, and the time interval is calculated based on T2 and T3. The drift velocity calculated based on this displacement and time interval can be used as the drift velocity of the target 20 in distress on water at time T3, and this drift velocity is recorded as V3.

[0150] In the disclosed embodiment, the drift speed intelligent prediction model corresponding to the distressed target 20 is a model trained based on the above-described drift experimental data. Transfer learning of the drift speed intelligent prediction model corresponding to the distressed target 20 is to further and more accurately train the already trained drift speed intelligent prediction model.

[0151] In one feasible implementation, the environmental prediction information corresponding to the preset spatial range around loc2 at time T2 is input into the drift speed intelligent prediction model corresponding to the target 20 in distress on the water. The drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at time T2. Further, based on the difference between the drift speed of the target 20 in distress on the water at time T2 output by the drift speed intelligent prediction model and V2 as described above, the drift speed intelligent prediction model is retrained, i.e., transfer learning. Similarly, the environmental prediction information corresponding to the preset spatial range around loc3 at time T3 is input into the drift speed intelligent prediction model corresponding to the target 20 in distress on the water. The drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at time T3. Further, based on the difference between the drift speed of the target 20 in distress on the water at time T3 output by the drift speed intelligent prediction model and V3 as described above, the drift speed intelligent prediction model is retrained, i.e., transfer learning, and so on.

[0152] In another feasible implementation, the environmental prediction information corresponding to the preset spatial range around loc2 at time T2, and the environmental prediction information corresponding to the preset spatial range around loc1 at time T1, are input into the drift speed intelligent prediction model corresponding to the target 20 in distress on the water. The drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at time T2. Further, based on the difference between the drift speed of the target 20 in distress on the water at time T2 output by the drift speed intelligent prediction model and V2 as described above, the drift speed intelligent prediction model is retrained, i.e., transfer learning. Similarly, the environmental prediction information corresponding to the preset spatial range around loc3 at time T3, the environmental prediction information corresponding to the preset spatial range around loc2 at time T2, and the environmental prediction information corresponding to the preset spatial range around loc1 at time T1 are input into the drift speed intelligent prediction model corresponding to the target 20 in distress on the water. The drift speed intelligent prediction model outputs the drift speed of the target 20 in distress on the water at time T3. Furthermore, based on the difference between the drift speed of the target in distress 20 on water at time T3 output by the drift speed intelligent prediction model and V3 as described above, the drift speed intelligent prediction model is retrained, i.e., transfer learning, and so on.

[0153] During the transfer learning process, the convolutional layers, activation functions, and pooling layers in the drift velocity intelligent prediction model remain fixed, and only the parameters of the fully connected layer are further fine-tuned to achieve transfer learning based on real-time remote sensing observation data. The drift velocity intelligent prediction model after transfer learning is then used to carry out subsequent drift predictions. It is understandable that since medium and large targets such as ships and life rafts may be observed by remote sensing satellites, the drift velocity intelligent prediction model corresponding to these targets can be further transfer-learned. However, since small targets such as people and debris cannot be observed by remote sensing satellites, the drift velocity intelligent prediction model corresponding to these targets cannot be transfer-learned.

[0154] Optionally, after performing transfer learning on the drift speed intelligent prediction model corresponding to the distressed target on the water, the method further includes: performing drift prediction on the distressed target on the water again based on the drift speed intelligent prediction model after transfer learning and the multiple sets of initial conditions to obtain a new search range for the distressed target on the water.

[0155] For example, after transfer learning of the drift speed intelligent prediction model corresponding to the distressed target 20, the drift prediction of the distressed target 20 is performed again based on the transferred drift speed intelligent prediction model and the N sets of initial conditions described above to obtain a new search range for the distressed target 20. The specific drift prediction process can be referred to the drift prediction process described in the above embodiment and will not be repeated here.

[0156] In addition, based on the drift speed intelligent prediction model after the transfer learning and the N groups of initial conditions as described above, the drift prediction of the distress target 20 on the water is performed again to obtain a new search range of the distress target 20 on the water. The method further includes the following steps: Figure 15 The following steps are shown:

[0157] S1501. When the remote sensing satellite observes the distress target on the water within the new search range, a preset number of end points are selected around the distress target on the water within the new search range.

[0158] For example, there are 6 groups of initial conditions for drift prediction of the distress target 20 on the water, which are sequentially recorded as initial condition 1, initial condition 2, ..., initial condition 6. The initial positions included in initial condition 1, initial condition 2, ..., initial condition 6 are as follows: Figure 16 The initial position 1, initial position 2, ..., initial position 6 are shown. Based on the 6 sets of initial conditions and the drift speed intelligent prediction model after transfer learning corresponding to the distress target 20 on water, 6 new drift trajectories can be generated, thereby obtaining 6 new end points. The 6 new end points are as follows: Figure 16The end point 7-end point 12 shown. According to the end point 7-end point 12, a new search range of the distress target 20 on the water is obtained, and the new search range is as follows Figure 16 As shown in the first search range 160, assume that the remote sensing satellite observes a distress target 20 within the first search range 160. Furthermore, a predetermined number of endpoints are selected around the distress target 20 within the first search range 160. For example, three endpoints are selected around the distress target 20, namely, endpoint 7 through endpoint 9. Alternatively, endpoints within a predetermined range 161 around the distress target 20 are selected, such as endpoints 7 through endpoint 9.

[0159] S1502. Predict the drift speed, drift position, drift trajectory, and search range of other targets in distress based on the initial conditions corresponding to the preset number of endpoints and the drift speed intelligent prediction model corresponding to other targets in distress at the same time as the target in distress on the water, where the size of the other targets in distress is smaller than the size of the target in distress on the water.

[0160] For example Figure 16 As shown, based on endpoints 7, 8, and 9, the initial conditions corresponding to endpoints 7, 8, and 9 are traced back, such as initial conditions 1, initial conditions 2, and initial conditions 3. Furthermore, based on initial conditions 1, initial conditions 2, and initial conditions 3, as well as the drift velocity intelligent prediction model corresponding to other distressed targets simultaneously with the distressed target 20 on the water, the drift trajectory and search range of the other distressed targets are predicted. The specific prediction process refers to the prediction process described in the above embodiment and will not be repeated here. For example, the distressed target 20 on the water is a large-sized distressed target such as a fishing boat, yacht, or cargo ship that is easily observed by remote sensing satellites. Other distressed targets are smaller-sized distressed targets such as people and wreckage that are not easily observed by remote sensing satellites.

[0161] For example, the predicted drift trajectory of the other distressed target based on initial condition 1 is b, the predicted drift trajectory of the other distressed target based on initial condition 2 is a, and the predicted drift trajectory of the other distressed target based on initial condition 3 is c. End point 14 is the end point of drift trajectory a, end point 13 is the end point of drift trajectory b, and end point 15 is the end point of drift trajectory c. Furthermore, based on end points 13, 14, and 15, a search range for other distressed targets is obtained, such as second search range 162.

[0162] In water distress incidents, medium-to-large targets such as ships and life rafts, as well as smaller targets such as people and debris, are often present simultaneously. However, medium-to-large targets may be identified and located by remote sensing satellites, while small targets may not. The disclosed embodiments not only predict the search range for larger targets in distress, but also, based on the remote sensing observation position of the target within the search range, infer the drift trajectory and search range of smaller targets that were in distress at the same time as the target. This overcomes the problem of small targets being unable to be identified by remote sensing satellites and improves the timeliness and accuracy of small target searches.

[0163] Figure 17 The schematic diagram of the structure of the remote sensing identification and drift prediction device for water distress targets provided by the embodiment of the present disclosure. The remote sensing identification and drift prediction device for water distress targets provided by the embodiment of the present disclosure can execute the processing flow provided by the remote sensing identification and drift prediction method for water distress targets, such as Figure 17 As shown, the remote sensing identification and drift prediction device 170 for distressed targets on water includes:

[0164] An acquisition module 171 is used to acquire remote sensing images captured by remote sensing satellites;

[0165] An input module 172 is configured to input the remote sensing image into a target intelligent detection model, wherein the target intelligent detection model includes a backbone network, a multi-scale feature mining module, a first deep-shallow feature fusion module, a second deep-shallow feature fusion module, and a detection head. The backbone network includes multiple feature layers that sequentially extract features from the remote sensing image. The multiple feature layers include a first feature layer, a second feature layer, and a third feature layer that are sequentially connected. The third feature layer is the last feature layer among the multiple feature layers.

[0166] A multi-scale feature extraction module 173 is configured to perform multi-scale feature extraction on the first feature map output by the third feature layer through the multi-scale feature mining module to obtain a second feature map;

[0167] a fusion module 174 configured to fuse the third feature map output by the first feature layer and the fourth feature map output by the second feature layer through the first deep-shallow feature fusion module to obtain a first fusion result; and fuse the second feature map and the first fusion result through the second deep-shallow feature fusion module to obtain a second fusion result;

[0168] a detection module 175 configured to input the third feature map, the first fusion result, and the second fusion result into the detection head, and the detection head outputting the type of the distressed water target in the remote sensing image and the location information of the distressed water target in the remote sensing image;

[0169] a determination module 176 for determining a remote sensing observation position of the target in distress on the water based on the position information of the target in distress on the water in the remote sensing image;

[0170] The drift prediction module 177 is used to perform drift prediction on the target in distress on the water according to the remote sensing observation position and remote sensing observation time point of the target in distress on the water.

[0171] Optionally, the multi-scale feature mining module includes multiple dilated convolution layers and maximum pooling layers with different dilation factors; the multi-scale feature extraction module 173 performs multi-scale feature extraction on the first feature map output by the third feature layer through the multi-scale feature mining module to obtain a second feature map, specifically for:

[0172] Extracting features from the first feature map output by the third feature layer through the plurality of dilation factor-different dilated convolutional layers to obtain multi-scale features;

[0173] Extracting key information from the first feature map through the maximum pooling layer;

[0174] The multi-scale features and the key information are concatenated to obtain the second feature map.

[0175] Optionally, the fusion module 174 fuses the third feature map output by the first feature layer and the fourth feature map output by the second feature layer through the first deep-layer and shallow-layer feature fusion module to obtain a first fusion result, which is specifically used to:

[0176] Performing multiple convolution operations of different scales on the third feature map output by the first feature layer to obtain a first convolution operation result;

[0177] Performing multiple convolution operations of different scales on the fourth feature map output by the second feature layer in sequence to obtain a second convolution operation result;

[0178] The first convolution operation result and the second convolution operation result are concatenated to obtain the first fusion result.

[0179] Optionally, the fusion module 174 fuses the second feature map and the first fusion result through the second deep-layer shallow-layer feature fusion module to obtain a second fusion result, which is specifically used to:

[0180] Performing multiple convolution operations of different scales on the second feature map in sequence to obtain a third convolution operation result;

[0181] Performing multiple convolution operations of different scales on the first fusion result to obtain a fourth convolution operation result;

[0182] The third convolution operation result and the fourth convolution operation result are concatenated to obtain the second fusion result.

[0183] Optionally, when the drift prediction module 177 performs drift prediction on the target in distress on the water based on the remote sensing observation position and remote sensing observation time point of the target in distress on the water, it is specifically used to:

[0184] Obtaining environmental prediction information corresponding to the remote sensing observation time point in a preset spatial range around the remote sensing observation location;

[0185] Generating multiple initial positions for drift prediction of the target in distress on the water based on the remote sensing observation position, generating multiple initial times for the drift prediction based on the remote sensing observation time point, generating multiple initial environmental information for the drift prediction based on environmental prediction information corresponding to the remote sensing observation time point, and constructing multiple groups of initial conditions based on the multiple initial positions, the multiple initial times, and the multiple initial environmental information, each group of initial conditions including any initial position, any initial time, and any initial environmental information, wherein the drift prediction includes prediction of drift speed, drift position, drift trajectory, and search range;

[0186] predicting, based on each set of initial conditions, a drift speed and a drift position of the target in distress at a plurality of target time points after any of the initial times, wherein the drift speed of the target in distress at each target time point is obtained based on environmental prediction information corresponding to a preset spatial range around the drift position of the target in distress at the target time point, and the drift positions of the target in distress at the plurality of target time points constitute a drift trajectory;

[0187] The search range of the target in distress on the water is predicted according to the end point of each drift trajectory in the multiple drift trajectories corresponding to the multiple groups of initial conditions.

[0188] Optionally, when the drift prediction module 177 predicts the drift speed and drift position of the target in distress on water at multiple target time points after any initial time according to each set of initial conditions, it is specifically configured to:

[0189] Inputting any of the initial environmental information into a drift speed intelligent prediction model corresponding to the target in distress on the water, and the drift speed intelligent prediction model outputting the drift speed of the target in distress on the water at any of the initial times;

[0190] For each target time point after any of the initial times, calculating the drift position of the target in distress on the water at the target time point based on the drift speed and drift position of the target in distress on the water at the previous target time point;

[0191] The environmental prediction information corresponding to the preset spatial range around the drift position of the target in distress on the water at the target time point is input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed of the target in distress on the water at the target time point.

[0192] Optionally, the intelligent prediction model for the drift speed of the target in distress on water is trained according to the following steps:

[0193] Obtaining, by means of a BeiDou positioning system, drift positions of an experimental target representing the target in distress on water at a plurality of preset time points during its free drift in the target waters;

[0194] Calculating the drift speeds of the experimental target corresponding to the plurality of preset time points according to the drift positions corresponding to the plurality of preset time points;

[0195] For each of the plurality of preset time points, inputting environmental prediction information corresponding to a preset spatial range around the drift position of the experimental target at the preset time point into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputting the drift speed of the experimental target corresponding to the preset time point;

[0196] The drift speed intelligent prediction model is trained according to the calculated drift speed corresponding to the experimental target at the preset time point and the drift speed corresponding to the experimental target at the preset time point output by the drift speed intelligent prediction model.

[0197] Optionally, the remote sensing identification and drift prediction device 170 for distressed targets on water also includes: a transfer learning module 178, which is used to perform transfer learning on the drift speed intelligent prediction model corresponding to the distressed target on water according to the current observation position and current observation time point of the distressed target on water, and the historical observation position and historical observation time point of the distressed target on water when the remote sensing satellite observes the distressed target on water again.

[0198] Optionally, after the transfer learning module 178 performs transfer learning on the drift speed intelligent prediction model corresponding to the distressed target on the water, the drift prediction module 177 is also used to: perform drift prediction on the distressed target on the water again according to the drift speed intelligent prediction model after transfer learning and the multiple sets of initial conditions to obtain a new search range for the distressed target on the water.

[0199] Optionally, the apparatus 170 for remote sensing identification and drift prediction of distressed targets on water further includes:

[0200] A selection module 179 is configured to select a preset number of endpoints around the water distress target within the new search range when the remote sensing satellite observes the water distress target within the new search range;

[0201] The drift prediction module 177 is also used to: predict the drift speed, drift position, drift trajectory and search range of other distress targets based on the initial conditions corresponding to the preset number of end points and the drift speed intelligent prediction model corresponding to other distress targets that are in distress at the same time as the distress target on the water, and the size of the other distress targets is smaller than the size of the distress target on the water.

[0202] Optionally, when the drift prediction module 177 obtains the environmental prediction information corresponding to the remote sensing observation time point in a preset spatial range around the remote sensing observation location, it is specifically used to:

[0203] The position information corresponding to multiple first-resolution grids contained in a preset spatial range around the remote sensing observation position, the environmental prediction information corresponding to the multiple first-resolution grids at the remote sensing observation time point, and the position information corresponding to multiple second-resolution grids contained in the preset spatial range are input into the downscaling deep learning model, and the environmental prediction information corresponding to the multiple second-resolution grids at the remote sensing observation time point output by the downscaling deep learning model is used as the environmental prediction information corresponding to the preset spatial range around the remote sensing observation position at the remote sensing observation time point, wherein the second resolution is greater than the first resolution.

[0204] Optionally, the environmental prediction information includes wind field forecast results, wave field forecast results, and flow field forecast results;

[0205] The downscaling deep learning model is trained according to the following steps:

[0206] Calculating environmental prediction information for a plurality of first-resolution grids in the target waters at preset time intervals within a preset future time period according to a wind-wave-current coupled prediction model, wherein the wind-wave-current coupled prediction model is a coupled model of an atmospheric numerical model, a wave numerical model, and a hydrodynamic numerical model;

[0207] Calculating, according to the wind-wave-current coupled prediction model, environmental prediction information of a plurality of second-resolution grids in the target waters at preset time intervals within the future preset time period;

[0208] Inputting the position information corresponding to the plurality of first-resolution grids, the environmental prediction information of the plurality of first-resolution grids at preset time intervals within the future preset time period, and the position information corresponding to the plurality of second-resolution grids into a downscaling deep learning model to be trained, the downscaling deep learning model outputting the environmental prediction information of the plurality of second-resolution grids at preset time intervals within the future preset time period;

[0209] The downscaling deep learning model is trained based on the environmental prediction information of the multiple second-resolution grids output by the downscaling deep learning model at preset time intervals within the future preset time period, and the environmental prediction information of the multiple second-resolution grids calculated by the wind-wave-current coupling forecast model at preset time intervals within the future preset time period.

[0210] Figure 17 The remote sensing identification and drift prediction device for distressed targets on water in the illustrated embodiment can be used to implement the technical solution of the above-mentioned method embodiment. Its implementation principle and technical effects are similar and will not be repeated here.

[0211] The above describes the internal functions and structure of the device for remote sensing identification and drift prediction of targets in distress on water. The device can be implemented as an electronic device. Figure 18 This is a schematic diagram of the structure of an electronic device embodiment provided by the present disclosure. Figure 18 As shown, the electronic device includes a memory 181 and a processor 182 .

[0212] The memory 181 is used to store programs. In addition to the aforementioned programs, the memory 181 may also be configured to store various other data to support operations on the electronic device. Examples of such data include instructions for any application or method operating on the electronic device, contact data, phone book data, messages, images, videos, and the like.

[0213] Memory 181 can be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0214] The processor 182 is coupled to the memory 181 and executes the program stored in the memory 181 to implement the technical solution of the above method embodiment.

[0215] Further, if Figure 18As shown, the electronic device may further include: a communication component 183, a power component 184, an audio component 185, a display 186 and other components. Figure 18 Only some components are shown schematically, which does not mean that the electronic device only includes Figure 18 Components shown.

[0216] The communication component 183 is configured to facilitate wired or wireless communication between the electronic device and other devices. The electronic device can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 183 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 183 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0217] The power supply component 184 provides power to various components of the electronic device. The power supply component 184 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device.

[0218] The audio component 185 is configured to output and / or input audio signals. For example, the audio component 185 includes a microphone (MIC), which is configured to receive external audio signals when the electronic device is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signal can be further stored in the memory 181 or transmitted via the communication component 183. In some embodiments, the audio component 185 also includes a speaker for outputting audio signals.

[0219] The display 186 includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, slides, and gestures on the touch panel. The touch sensor can not only sense the boundaries of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0220] In addition, an embodiment of the present disclosure further provides a computer-readable storage medium on which a computer program is stored. The computer program is executed by a processor to implement the method described in the above embodiment.

[0221] The exemplary embodiments of the present disclosure further provide a computer program product, including a computer program, wherein when the computer program is executed by a processor of a computer, the computer is configured to enable the computer to implement the method described in the above embodiment.

[0222] It should be noted that, in this document, relational terms such as "first" and "second" 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, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0223] The foregoing description is intended only to provide specific embodiments of the present disclosure, intended to enable those skilled in the art to understand and implement the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure is not intended to be limited to the embodiments described herein, but rather to be construed in the broadest manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for remote sensing identification and drift prediction of distressed targets on water, characterized in that: The method comprises: Acquire remote sensing images taken by remote sensing satellites; Inputting the remote sensing image into a target intelligent detection model, the target intelligent detection model includes a backbone network, a multi-scale feature mining module, a first deep-shallow feature fusion module, a second deep-shallow feature fusion module and a detection head, the backbone network includes multiple feature layers that sequentially extract features from the remote sensing image, the multiple feature layers include a first feature layer, a second feature layer, and a third feature layer that are sequentially connected, and the third feature layer is the last feature layer among the multiple feature layers; Performing multi-scale feature extraction on the first feature map output by the third feature layer by the multi-scale feature mining module to obtain a second feature map; fusing the third feature map output by the first feature layer and the fourth feature map output by the second feature layer through the first deep-shallow feature fusion module to obtain a first fusion result; fusing the second feature map and the first fusion result through the second deep-shallow feature fusion module to obtain a second fusion result; Inputting the third feature map, the first fusion result, and the second fusion result into the detection head, the detection head outputting the type of the distressed target on the water in the remote sensing image and the location information of the distressed target on the water in the remote sensing image; determining a remote sensing observation position of the target in distress on the water according to position information of the target in distress on the water in the remote sensing image; The drift of the target in distress on the water is predicted based on the remote sensing observation position and remote sensing observation time point of the target in distress on the water.

2. The method according to claim 1, characterized in that The multi-scale feature mining module includes multiple dilation factors of the hollow convolution layer and the maximum pooling layer; Performing multi-scale feature extraction on the first feature map output by the third feature layer by the multi-scale feature mining module to obtain a second feature map includes: Extracting features from the first feature map output by the third feature layer through the plurality of dilation factor-different dilated convolutional layers to obtain multi-scale features; Extracting key information from the first feature map through the maximum pooling layer; The multi-scale features and the key information are concatenated to obtain the second feature map.

3. The method according to claim 1, characterized in that The first deep-layer and shallow-layer feature fusion module fuses the third feature map output by the first feature layer and the fourth feature map output by the second feature layer to obtain a first fusion result, including: Performing multiple convolution operations of different scales on the third feature map output by the first feature layer to obtain a first convolution operation result; Performing multiple convolution operations of different scales on the fourth feature map output by the second feature layer in sequence to obtain a second convolution operation result; The first convolution operation result and the second convolution operation result are concatenated to obtain the first fusion result.

4. The method according to claim 1, wherein Fusing the second feature map and the first fusion result through the second deep-layer shallow-layer feature fusion module to obtain a second fusion result, including: Performing multiple convolution operations of different scales on the second feature map in sequence to obtain a third convolution operation result; Performing multiple convolution operations of different scales on the first fusion result to obtain a fourth convolution operation result; The third convolution operation result and the fourth convolution operation result are concatenated to obtain the second fusion result.

5. The method according to claim 1, wherein Predicting the drift of the target in distress on the water according to the remote sensing observation position and remote sensing observation time point of the target in distress on the water includes: Obtaining environmental prediction information corresponding to the remote sensing observation time point in a preset spatial range around the remote sensing observation location; Generating multiple initial positions for drift prediction of the target in distress on the water based on the remote sensing observation position, generating multiple initial times for the drift prediction based on the remote sensing observation time point, generating multiple initial environmental information for the drift prediction based on environmental prediction information corresponding to the remote sensing observation time point, and constructing multiple groups of initial conditions based on the multiple initial positions, the multiple initial times, and the multiple initial environmental information, each group of initial conditions including any initial position, any initial time, and any initial environmental information, wherein the drift prediction includes prediction of drift speed, drift position, drift trajectory, and search range; predicting, based on each set of initial conditions, a drift speed and a drift position of the target in distress at a plurality of target time points after any of the initial times, wherein the drift speed of the target in distress at each target time point is obtained based on environmental prediction information corresponding to a preset spatial range around the drift position of the target in distress at the target time point, and the drift positions of the target in distress at the plurality of target time points constitute a drift trajectory; The search range of the target in distress on the water is predicted according to the end point of each drift trajectory in the multiple drift trajectories corresponding to the multiple groups of initial conditions.

6. The method according to claim 5, characterized in that Predicting, based on each set of initial conditions, the drift speed and drift position of the target in distress on water at a plurality of target time points after any of the initial times, comprises: Inputting any of the initial environmental information into a drift speed intelligent prediction model corresponding to the target in distress on the water, and the drift speed intelligent prediction model outputting the drift speed of the target in distress on the water at any of the initial times; For each target time point after any of the initial times, calculating the drift position of the target in distress on the water at the target time point based on the drift speed and drift position of the target in distress on the water at the previous target time point; The environmental prediction information corresponding to the preset spatial range around the drift position of the target in distress on the water at the target time point is input into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputs the drift speed of the target in distress on the water at the target time point.

7. The method according to claim 6, characterized in that The intelligent prediction model for the drift speed of the target in distress on water is trained according to the following steps: Obtaining, by means of a BeiDou positioning system, drift positions of an experimental target representing the target in distress on water at a plurality of preset time points during its free drift in the target waters; Calculating the drift speeds of the experimental target corresponding to the plurality of preset time points according to the drift positions corresponding to the plurality of preset time points; For each of the plurality of preset time points, inputting environmental prediction information corresponding to a preset spatial range around the drift position of the experimental target at the preset time point into the drift speed intelligent prediction model, and the drift speed intelligent prediction model outputting the drift speed of the experimental target corresponding to the preset time point; The drift speed intelligent prediction model is trained according to the calculated drift speed corresponding to the experimental target at the preset time point and the drift speed corresponding to the experimental target at the preset time point output by the drift speed intelligent prediction model.

8. The method according to claim 7, characterized in that The method further comprises: When the remote sensing satellite observes the distressed target on the water again, transfer learning is performed on the drift speed intelligent prediction model corresponding to the distressed target on the water based on the current observation position and current observation time point of the distressed target on the water, as well as the historical observation position and historical observation time points of the distressed target on the water.

9. The method according to claim 8, characterized in that After performing transfer learning on the drift speed intelligent prediction model corresponding to the target in distress on water, the method further includes: According to the drift speed intelligent prediction model after the transfer learning and the multiple sets of initial conditions, the drift prediction of the target in distress on the water is performed again to obtain a new search range of the target in distress on the water.

10. The method according to claim 9, characterized in that The method further comprises: When the remote sensing satellite observes the distress target on the water within the new search range, a preset number of end points around the distress target on the water are selected within the new search range; Based on the initial conditions corresponding to the preset number of end points and the drift speed intelligent prediction model corresponding to other distressed targets that are in distress at the same time as the distressed target on the water, the drift speed, drift position, drift trajectory and search range of the other distressed targets are predicted, and the size of the other distressed targets is smaller than the size of the distressed target on the water.

11. The method according to claim 5, characterized in that Obtaining environmental prediction information corresponding to a preset spatial range around the remote sensing observation location at the remote sensing observation time point, including: The position information corresponding to multiple first-resolution grids contained in a preset spatial range around the remote sensing observation position, the environmental prediction information corresponding to the multiple first-resolution grids at the remote sensing observation time point, and the position information corresponding to multiple second-resolution grids contained in the preset spatial range are input into the downscaling deep learning model, and the environmental prediction information corresponding to the multiple second-resolution grids at the remote sensing observation time point output by the downscaling deep learning model is used as the environmental prediction information corresponding to the preset spatial range around the remote sensing observation position at the remote sensing observation time point, wherein the second resolution is greater than the first resolution.

12. The method according to claim 11, characterized in that The environmental prediction information includes wind field forecast results, wave field forecast results, and flow field forecast results; The downscaling deep learning model is trained according to the following steps: Calculating environmental prediction information for a plurality of first-resolution grids in the target waters at preset time intervals within a preset future time period according to a wind-wave-current coupled prediction model, wherein the wind-wave-current coupled prediction model is a coupled model of an atmospheric numerical model, a wave numerical model, and a hydrodynamic numerical model; Calculating, according to the wind-wave-current coupled prediction model, environmental prediction information of a plurality of second-resolution grids in the target waters at preset time intervals within the future preset time period; Inputting the position information corresponding to the plurality of first-resolution grids, the environmental prediction information of the plurality of first-resolution grids at preset time intervals within the future preset time period, and the position information corresponding to the plurality of second-resolution grids into a downscaling deep learning model to be trained, the downscaling deep learning model outputting the environmental prediction information of the plurality of second-resolution grids at preset time intervals within the future preset time period; The downscaling deep learning model is trained based on the environmental prediction information of the multiple second-resolution grids output by the downscaling deep learning model at preset time intervals within the future preset time period, and the environmental prediction information of the multiple second-resolution grids calculated by the wind-wave-current coupling forecast model at preset time intervals within the future preset time period.

13. A device for remote sensing identification and drift prediction of distressed targets on water, characterized in that: The device comprises: An acquisition module is used to acquire remote sensing images taken by remote sensing satellites; An input module is used to input the remote sensing image into a target intelligent detection model, wherein the target intelligent detection model includes a backbone network, a multi-scale feature mining module, a first deep-shallow feature fusion module, a second deep-shallow feature fusion module, and a detection head, wherein the backbone network includes multiple feature layers that sequentially extract features from the remote sensing image, wherein the multiple feature layers include a first feature layer, a second feature layer, and a third feature layer that are sequentially connected, and the third feature layer is the last feature layer among the multiple feature layers; a multi-scale feature extraction module, configured to perform multi-scale feature extraction on the first feature map output by the third feature layer through the multi-scale feature mining module to obtain a second feature map; a fusion module, configured to fuse the third feature map output by the first feature layer and the fourth feature map output by the second feature layer through the first deep-shallow feature fusion module to obtain a first fusion result; and fuse the second feature map and the first fusion result through the second deep-shallow feature fusion module to obtain a second fusion result; a detection module, configured to input the third feature map, the first fusion result, and the second fusion result into the detection head, wherein the detection head outputs the type of the target in distress on the water in the remote sensing image and the location information of the target in distress on the water in the remote sensing image; a determination module, configured to determine the remote sensing observation position of the target in distress on the water according to the position information of the target in distress on the water in the remote sensing image; The drift prediction module is used to predict the drift of the target in distress on the water according to the remote sensing observation position and remote sensing observation time point of the target in distress on the water.

14. An electronic device, characterized in that: include: Memory; processor; as well as computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the method according to any one of claims 1 to 12.

Images