Target detection method and system based on satellite remote sensing data

Abstract

The invention belongs to the technical field of remote sensing detection, and particularly provides a target detection method and system based on satellite remote sensing data, and the method comprises the steps: receiving a target detection request, and analyzing the received request; acquiring satellite remote sensing original data according to an analysis result; processing the obtained original data to obtain a remote sensing image to be detected; judging whether a corresponding trained target detection model exists or not according to an analysis result; when the trained target detection model does not exist, issuing a model training instruction to each model processing terminal, triggering each model processing terminal to obtain a to-be-trained neural network model based on the training instruction, and performing training for a set number of times to generate a detection model; determining a target model processing terminal based on the accuracy of the detection model, and obtaining the detection model generated by the target model processing terminal as a target detection model; and inputting the remote sensing image to be detected into a target detection model to obtain a detection result output by the target detection model.

Description

Background Art

[0002] With the rapid development of satellite remote sensing technology, satellite remote sensing data plays an increasingly important role in many fields such as earth observation, environmental monitoring, disaster warning, and urban planning. Traditional target detection methods mainly rely on manual visual interpretation, which is not only time-consuming and laborious, but also easily affected by subjective factors, making it difficult to ensure the accuracy and consistency of the detection results. In recent years, the rise of deep learning technology has provided a new solution for target detection in satellite remote sensing images. By training deep learning models, automatic recognition and detection of specific targets in remote sensing images can be achieved, greatly improving detection efficiency and accuracy.

[0003] However, in practical applications, target detection based on satellite remote sensing data still faces many challenges. On the one hand, satellite remote sensing images usually have the characteristics of high resolution and large data volume, which places high demands on computing resources and processing capabilities. On the other hand, the types and characteristics of targets in different application scenarios vary greatly, and corresponding detection models need to be trained for different targets. In addition, with the continuous advancement of satellite remote sensing technology, new data types and target types continue to emerge. How to quickly adapt to these changes and realize rapid training and updating of models is also a problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the above problems faced by target detection based on satellite remote sensing data, the present invention provides a target detection method and system based on satellite remote sensing data.

[0005] In a first aspect, the technical solution of the present invention provides a target detection method based on satellite remote sensing data, comprising the following steps: Receive target detection requests and parse the received requests; According to the analysis results, the original satellite remote sensing data is obtained; Process the acquired raw data to obtain the remote sensing image to be detected; Determine whether the corresponding trained target detection model exists based on the analysis results; When a trained target detection model exists, the remote sensing image to be detected is input into the target detection model to obtain a detection result output by the target detection model; When a trained target detection model does not exist, a model training instruction is issued to each model processing terminal, triggering each model processing terminal to obtain a neural network model to be trained based on the training instruction, and perform a set number of trainings to generate a detection model; Determine the target model processing terminal based on the accuracy of the detection model, and obtain the detection model generated by the target model processing terminal as the target detection model; execute the steps: input the remote sensing image to be detected into the target detection model to obtain the detection result output by the target detection model.

[0006] As a preferred embodiment of the technical solution of the present invention, the steps of receiving a target detection request and parsing the received request include: Receive a target detection request; The target detection request is parsed to obtain target information, including the type of detected target, geographic location information, and detection time range.

[0007] As a preferred embodiment of the technical solution of the present invention, according to the analysis result, the step of obtaining the original satellite remote sensing data includes: Select the satellite data source based on the type and geographic location information of the detection target; Using the time range information, the remote sensing data of the selected satellite data source within the specified time period is filtered out; Download filtered remote sensing data from satellite data sources.

[0008] As a preferred embodiment of the technical solution of the present invention, the step of processing the acquired raw data to obtain the remote sensing image to be detected includes: When receiving remote sensing data, the original data bit stream is parsed layer by layer according to the transmission control protocol to obtain the image data to be detected and the corresponding auxiliary reference data.

[0009] As a preferred embodiment of the technical solution of the present invention, when receiving remote sensing data, the steps of parsing the original data bit stream layer by layer according to the transmission control protocol to obtain the image data to be detected and the corresponding auxiliary reference data include: Receive message segments transmitted in the form of byte streams; Parse the message segments layer by layer, extract the control information from the TCP header, and extract the actual remote sensing image data and auxiliary reference data from the data part; Rearrange the received data according to the sequence number to restore the original data order; After data parsing and reorganization are completed, remote sensing image data and corresponding auxiliary reference data are extracted from the parsed data.

[0010] As a preferred embodiment of the technical solution of the present invention, each model processing terminal is provided with a data set for training different models, and the steps of performing model training at each model processing terminal include: Acquire a data set based on the training instruction; the data set includes a plurality of labeled sample remote sensing images and a plurality of unlabeled sample remote sensing images; The unlabeled sample remote sensing images and the labeled sample remote sensing images are used to train the acquired neural network model to be trained, so as to obtain the target detection model.

[0011] As a preferred embodiment of the technical solution of the present invention, each unlabeled sample remote sensing image and each labeled sample remote sensing image are used to train the acquired neural network model to be trained, and the step of obtaining the target detection model includes: inputting each unlabeled sample remote sensing image and each labeled sample remote sensing image into the neural network model to be trained, and the specific training process is as follows: Performing feature extraction on the unlabeled sample remote sensing image to obtain a first feature image, and performing feature extraction on the labeled sample remote sensing image to obtain a second feature image; Generate a first region image based on the first feature image, and generate a second region image based on the second feature image; extract features of the first region image to generate first region features, and extract features of the second region image to generate second region features; Perform feature enhancement processing on the first region feature and the second region feature respectively; based on the enhanced first region feature and the second region feature, obtain the first detection frame coordinates and the first classification score of each sample target corresponding to the first region feature, and the second detection frame coordinates and the second classification score of each sample target corresponding to the second region feature, so as to optimize the total loss function; When the total loss function value is minimized or converged, the target detection model is determined.

[0012] As a preferred embodiment of the technical solution of the present invention, each model processing terminal is provided with shared data and a local data set for training the model, wherein the shared data is used to characterize the data distribution of the local data of the model processing terminal; the method further comprises: Obtain the shared data and current status of each model processing terminal; A model processing terminal is selected based on the acquired shared data and the current status of each model processing terminal, and a model training instruction is issued to each selected model processing terminal.

[0013] As a preferred embodiment of the technical solution of the present invention, the steps of selecting a model processing terminal according to the acquired shared data and the current state of each model processing terminal, and issuing a model training instruction to each selected model processing terminal include: Analyze the shared data of each model processing terminal; Select a model processing terminal whose data distribution is more similar to the data distribution required by the current model to be trained than a set threshold as the initial model processing terminal; Check the current status of each initial model processing terminal; A model processing terminal whose computing resources are greater than a first set value and whose data delay is less than a second threshold is selected to issue a model training instruction.

[0014] In a second aspect, the technical solution of the present invention further provides a target detection system based on satellite remote sensing data, comprising a detection terminal and a control terminal connected to the detection terminal; the control terminal is connected to a plurality of model processing terminals; The control terminal receives the detection information input by the user and sends a target detection request to the detection terminal based on the detection information input by the user; The detection terminal receives a target detection request and parses the received request; obtains satellite remote sensing raw data according to the parsing result; processes the obtained raw data to obtain a remote sensing image to be detected; determines whether a corresponding trained target detection model exists according to the parsing result; if the trained target detection model exists, inputs the remote sensing image to be detected into the target detection model to obtain a detection result output by the target detection model; When a trained target detection model does not exist, the detection terminal feeds back information to the control terminal, and the control terminal issues a model training instruction to each model processing terminal based on the received feedback information, triggering each model processing terminal to obtain the neural network model to be trained based on the training instruction, and performs a set number of trainings to generate a detection model; The control terminal determines the target model processing terminal based on the accuracy of the detection model, obtains the detection model generated by the target model processing terminal and sends it to the detection terminal as the target detection model.

[0015] As a preferred embodiment of the technical solution of the present invention, the detection terminal receives a target detection request; and parses the target detection request to obtain target information, including the type of the detected target, geographical location information, and a time range for detection.

[0016] As a preferred embodiment of the technical solution of the present invention, the system also includes a satellite data source, and the detection terminal selects the satellite data source according to the type and geographical location information of the detection target; uses the time range information to filter out the remote sensing data of the selected satellite data source within a specified time period; and downloads the filtered remote sensing data from the satellite data source.

[0017] As a preferred embodiment of the technical solution of the present invention, when receiving remote sensing data, the detection terminal parses the original data bit stream layer by layer according to the transmission control protocol to obtain the image data to be detected and the corresponding auxiliary reference data.

[0018] As a preferred embodiment of the technical solution of the present invention, the detection terminal receives a message segment transmitted in the form of a byte stream; parses the message segment layer by layer, extracts control information from the TCP header, and extracts actual remote sensing image data and auxiliary reference data from the data part; rearranges the received data according to the sequence number to restore the original data order; after the data parsing and reorganization are completed, the remote sensing image data and the corresponding auxiliary reference data are extracted from the parsed data.

[0019] As a preferred embodiment of the technical solution of the present invention, each model processing terminal is provided with a data set for training different models, and the steps of model training of each model processing terminal include: obtaining a data set based on a training instruction; the data set includes a plurality of labeled sample remote sensing images and a plurality of unlabeled sample remote sensing images; using each unlabeled sample remote sensing image and each labeled sample remote sensing image to train the acquired neural network model to be trained, and obtain a target detection model. The specific training process is as follows: extracting features from the unlabeled sample remote sensing image to obtain a first feature image, and extracting features from the labeled sample remote sensing image to obtain a second feature image; generating a first region image based on the first feature image, and generating a second region image based on the second feature image; extracting features from the first region image to generate a first region feature, and extracting features from the second region image to generate a second region feature; performing feature enhancement processing on the first region feature and the second region feature respectively; obtaining the first detection frame coordinates and the first classification score of each sample target corresponding to the first region feature, and the second detection frame coordinates and the second classification score of each sample target corresponding to the second region feature based on the enhanced first region feature and the second region feature, to optimize the total loss function; when the total loss function value is minimum or converges, the target detection model is determined.

[0020] As a preferred embodiment of the technical solution of the present invention, each model processing terminal is provided with shared data and a local data set for training the model, wherein the shared data is used to characterize the data distribution of the local data of the model processing terminal; the control terminal obtains the shared data and the current status of each model processing terminal; the model processing terminal is selected according to the acquired shared data and the current status of each model processing terminal, and a model training instruction is issued to each selected model processing terminal, specifically for analyzing the shared data of each model processing terminal; a model processing terminal whose data distribution has a similarity with the data distribution required by the current model to be trained greater than a set threshold is selected as the initial model processing terminal; the current status of each initial model processing terminal is checked; and a model processing terminal whose computing resources are greater than a first set value and whose data delay is less than a second threshold is selected to issue a model training instruction.

[0021] It can be seen from the above technical solutions that the present invention has the following advantages: by using the trained target detection model to quickly detect remote sensing images, the detection time can be greatly shortened, while improving the accuracy of the detection results. Compared with the traditional manual visual interpretation method, the present invention can automatically identify and detect targets in remote sensing images, reducing the influence of human intervention and subjective factors.

[0022] The present invention supports the training of corresponding detection models for different targets, and can flexibly respond to target detection requirements in different application scenarios. When a new target type needs to be detected, a new detection model can be quickly generated by issuing model training instructions to each model processing terminal, thereby realizing rapid updating and adaptation of the model. In the model training stage, the present invention can fully utilize computing resources and speed up model training by triggering multiple model processing terminals to perform parallel training. At the same time, by determining the target model processing terminal based on the accuracy of the detection model, it can be ensured that the final detection model has high performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 is a schematic flow chart of a method according to an embodiment of the present invention.

[0025] Figure 2 is a schematic block diagram of a system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0027] like Figure 1 As shown, an embodiment of the present invention provides a target detection method based on satellite remote sensing data, comprising the following steps: Step 1: Receive the target detection request and parse the received request; Step 2: Obtain satellite remote sensing raw data based on the analysis results; Step 3: Process the acquired raw data to obtain the remote sensing image to be detected; Step 4: Determine whether the corresponding trained target detection model exists based on the analysis results; If yes, go to step 5; if no, go to step 6; Step 5: Input the remote sensing image to be detected into the target detection model to obtain the detection result output by the target detection model; Step 6: Send model training instructions to each model processing terminal, trigger each model processing terminal to obtain the neural network model to be trained based on the training instructions, and perform a set number of trainings to generate a detection model; Step 7: Determine the target model processing terminal based on the accuracy of the detection model, obtain the detection model generated by the target model processing terminal as the target detection model; execute step 5.

[0028] When a trained target detection model does not exist, it is necessary to send a neural network model to be trained to each model processing terminal, and trigger these terminals to perform a set number of trainings to generate a detection model. The specific steps of this process can be broken down into the following points: In the central server or management center, a neural network model to be trained is prepared. This model can be pre-designed, and its key parameters such as structure, number of layers, activation function, etc. have been determined. This neural network model to be trained is sent to each model processing terminal through the network or other communication methods. These terminals can be computers or servers distributed in different geographical locations, which have sufficient computing resources and storage capacity to perform the model training task. At the same time as the model is sent, or before it is sent, the parameters required for training, such as learning rate, number of iterations (i.e., a set number of trainings), batch size, etc. are determined. These parameters can be adjusted appropriately according to the complexity of the model and the size of the data. On each model processing terminal, the training process is started. This usually involves using the training data set to update the weights and biases of the model to minimize the loss function. The training process may include multiple iterations, each of which uses a portion of the training data to update the model. During the training process, the training progress and performance of the model, such as the value of the loss function, accuracy, and other indicators, can be monitored in real time. This helps to promptly discover and solve problems that may arise during the training process, such as overfitting and underfitting.

[0029] After a set number of trainings, each model processing terminal will generate its own detection model. These models may differ slightly in structure and parameters because they are obtained with different training data and training conditions. After obtaining each detection model, their performance needs to be evaluated. This usually involves using a validation data set to test the model's accuracy, recall and other indicators. By comparing these indicators, it can be determined which model processing terminal generates the best detection model. Based on the evaluation results, the model processing terminal with the best performance is selected as the target model processing terminal. The detection model generated by this terminal will be regarded as the target detection model and used for subsequent detection tasks.

[0030] In some embodiments, the steps of receiving a target detection request and parsing the received request include: Step 11: Receive target detection request; Step 12: Parse the target detection request to obtain target information, including the type of detected target, geographic location information, and detection time range.

[0031] Request parsing is performed to obtain target information and detect the type of target, such as buildings, vehicles, crops, or other types of targets. Understanding the target type helps determine the type and resolution of remote sensing data required for subsequent processing and analysis. Geographic location information is key to determining the area where the target is located. This usually includes latitude and longitude coordinates, administrative area names, or specific descriptions of geographic features. This information will be used to locate the target area from satellite remote sensing data. Time range information specifies the time period that needs to be detected. Since satellite remote sensing data is collected in chronological order, understanding the time range helps to filter out remote sensing data taken within the specified time period.

[0032] In some embodiments, according to the analysis result, the step of obtaining the satellite remote sensing raw data includes: Step 21: Select a satellite data source based on the type and geographic location information of the detected target; Step 22: Filter out remote sensing data from the selected satellite data source within a specified time period using the time range information; Step 23: Download the filtered remote sensing data from the satellite data source.

[0033] Select the appropriate satellite data source based on the type of detection target and geographic location information. Different satellites may provide data with different resolutions and spectral bands, so it is necessary to select according to actual needs. Use the time range information to filter out remote sensing data taken within the specified time period. At the same time, further filter the data according to the resolution requirements to ensure that the selected data meets the detection requirements. Download the filtered remote sensing data from the satellite data source and perform necessary preprocessing operations such as correction, cropping, and format conversion. These operations help ensure the accuracy and availability of the data.

[0034] The steps of processing the acquired raw data to obtain the remote sensing image to be detected include: when receiving the remote sensing data, parsing the raw data bit stream layer by layer according to the transmission control protocol to obtain the image data to be detected and the corresponding auxiliary reference data. Specifically include: Step 31: receiving a message segment transmitted in the form of a byte stream; Step 32: Parse the message segment layer by layer, extract the control information from the TCP header, and extract the actual remote sensing image data and auxiliary reference data from the data part; Step 33: Rearrange the received data according to the sequence number to restore the original data order; Step 34: After the data parsing and reorganization are completed, the remote sensing image data and the corresponding auxiliary reference data are extracted from the parsed data.

[0035] In an embodiment of the present invention, at the data receiving end, a TCP connection is first established with the data sending end (such as a satellite or ground data station, i.e., a data source). This usually involves a three-way handshake process, i.e., the sending end sends a SYN message, the receiving end replies with a SYN-ACK message, and finally the sending end sends an ACK message to confirm the establishment of the connection. Once the TCP connection is established, the data begins to be transmitted in the form of a byte stream. The receiving end receives these byte streams and stores them in a buffer for subsequent processing. During the transmission process, TCP data will be divided into multiple segments, each of which contains a data part and a TCP header. The receiving end needs to parse these segments layer by layer, extract control information (such as sequence number, confirmation number, window size, etc.) from the TCP header, and extract actual remote sensing image data and auxiliary reference data from the data part. Since TCP is a byte stream-oriented protocol, the received data may be out of order. Therefore, the receiving end needs to rearrange the received data according to the sequence number to restore the original data order. The TCP protocol includes a checksum mechanism for detecting whether an error occurs during data transmission. When parsing the data, the receiving end calculates the checksum and compares it with the checksum sent by the sending end. If the two do not match, it means that an error occurred during the data transmission process, and the receiving end will request the sending end to retransmit the data. After the data is parsed and reassembled, the receiving end will extract the remote sensing image data and the corresponding auxiliary reference data from the parsed data. These data will be used for subsequent target detection and processing steps.

[0036] It should be noted that when the data transmission is completed, the receiving end will send a FIN message to the sending end to request to close the TCP connection. After confirming that all data has been successfully transmitted, the sending end will reply with a FIN-ACK message and close the connection.

[0037] In some embodiments, each model processing terminal is provided with a data set for training different models, and the steps of performing model training at each model processing terminal include: S61: Acquire a data set based on the training instruction; the data set includes a plurality of labeled sample remote sensing images and a plurality of unlabeled sample remote sensing images; S62: Using each unlabeled sample remote sensing image and each labeled sample remote sensing image, the obtained neural network model to be trained is trained to obtain a target detection model.

[0038] In some embodiments, the step of using each unlabeled sample remote sensing image and each labeled sample remote sensing image to train the acquired neural network model to be trained, and obtaining the target detection model includes: inputting each unlabeled sample remote sensing image and each labeled sample remote sensing image into the neural network model to be trained, and the specific training process is as follows: S621: performing feature extraction on the unlabeled sample remote sensing image to obtain a first feature image, and performing feature extraction on the labeled sample remote sensing image to obtain a second feature image; S622: Generate a first region image based on the first feature image, and generate a second region image based on the second feature image; extract features of the first region image to generate first region features, and extract features of the second region image to generate second region features; S623: performing feature enhancement processing on the first region feature and the second region feature respectively; obtaining the first detection frame coordinates and the first classification score of each sample target corresponding to the first region feature, and the second detection frame coordinates and the second classification score of each sample target corresponding to the second region feature based on the enhanced first region feature and the second region feature, so as to optimize the total loss function; S624: When the total loss function value is minimum or converges, determine the target detection model.

[0039] In the embodiment of the present invention, model training includes the following process: 1. Data preparation Unlabeled sample remote sensing images: These images have no annotation information, that is, the specific location and category of the target in the image are unknown.

[0040] Labeled sample remote sensing images: These images contain annotation information of the target, such as the bounding box and category label of the target.

[0041] 2. Feature extraction Feature extraction of unlabeled sample remote sensing images: Use a pre-trained convolutional neural network (CNN) or a custom neural network structure to extract low-level and high-level features of the image and generate a first feature image.

[0042] Feature extraction of labeled sample remote sensing images: The above neural network structure is also used to extract features and generate the second feature image. Since these images have labels, the subsequent steps can directly use these labels for training.

[0043] 3. Region generation and feature extraction Generate a first region image based on the first feature image: Use a region proposal network (RPN) or other methods to generate candidate target regions (proposals) on the first feature image. These regions may contain targets.

[0044] Generate a second region image based on the second feature image: For the feature image of the labeled sample, directly use the annotated bounding box to generate the region image.

[0045] Extracting regional features: For each candidate region in the first region image and the second region image, a fixed-size feature vector is extracted using methods such as ROIPooling or ROI Align to generate the first region features and the second region features.

[0046] 4. Feature enhancement processing: Perform feature enhancement on the first region features and the second region features: Various feature enhancement techniques can be used, such as feature normalization, data enhancement (such as rotation, scaling, flipping, etc.), attention mechanism, etc., to enhance the expressiveness of features.

[0047] 5. Object Detection and Classification Get the detection box coordinates and classification scores: Use a fully connected layer or classifier to perform target detection and classification on the enhanced features, and output the detection box coordinates (i.e., bounding box) and classification score of each candidate region.

[0048] For unlabeled samples, although there is no true label, it can be indirectly evaluated by comparing with other candidate regions or some form of self-supervised learning.

[0049] 6. Optimize the total loss function Define the total loss function: The total loss function usually includes classification loss (such as cross entropy loss) and regression loss (such as Smooth L1 loss) to evaluate the accuracy of model prediction.

[0050] For labeled samples, the loss can be calculated directly; for unlabeled samples, semi-supervised or unsupervised learning methods may be needed to indirectly estimate the loss.

[0051] Optimization process: Using optimization methods such as back-propagation algorithm and gradient descent, the weights of the neural network are iteratively updated to minimize the total loss function.

[0052] 7. Determine the target detection model When the total loss function value reaches the minimum or converges, the training is stopped and the neural network model at this time is saved as the final target detection model.

[0053] In some embodiments, each model processing terminal is provided with shared data and a local data set for training the model, wherein the shared data is used to characterize the data distribution of the local data of the model processing terminal; the method further includes: S061: Obtaining the shared data and current status of each model processing terminal; S062: Select a model processing terminal according to the acquired shared data and the current status of each model processing terminal, and issue a model training instruction to each selected model processing terminal.

[0054] In some embodiments, the step of selecting a model processing terminal according to the acquired shared data and the current state of each model processing terminal, and issuing a model training instruction to each selected model processing terminal includes: S0621: Analyze the shared data of each model processing terminal; S0622: Select a model processing terminal whose data distribution is more similar to the data distribution required by the current model to be trained than a set threshold as the initial model processing terminal; S0623: Check the current status of each initial model processing terminal; S0624: Select a model processing terminal whose computing resources are greater than a first set value and whose data delay is less than a second threshold to issue a model training instruction.

[0055] It should be noted that when selecting a model processing terminal and issuing model training instructions, the following key factors can be considered. These factors will make decisions based on the acquired shared data and the current status of each model processing terminal: Analyze the shared data and understand the data distribution of the local data set of each model processing terminal.

[0056] Select model processing terminals whose data distribution is most similar to the data distribution required by the current model to be trained or optimized. This can be achieved by calculating the distance between data distributions (such as KL divergence, JS divergence, etc.) or similarity metrics (such as cosine similarity). Check the current status of each model processing terminal, including the occupancy of computing resources such as CPU, GPU, and memory. Give priority to model processing terminals with sufficient computing resources and low load for model training to ensure the efficiency and stability of the training process. Consider the network bandwidth and latency between the model processing terminal and the central control node (or data distribution node). Select a model processing terminal with good network conditions to reduce data transmission latency and bandwidth occupancy.

[0057] You can also select a model processing terminal based on the current system's task priority and load balancing strategy. For example, if a model processing terminal is currently processing a high-priority task, it may not be appropriate to assign a new training task to it.

[0058] After selecting the model processing terminal, the process of issuing model training instructions can be carried out as follows: Construct training tasks based on the model to be trained, training data set, training parameters, etc. Encapsulate the training tasks into instructions or task packages and send them to the selected model processing terminal through the communication network. The central control node or monitoring system can monitor the training progress, resource usage, etc. of each model processing terminal in real time. When the model training is completed, collect the training results of each model processing terminal, including model weights, accuracy, loss values, etc. Evaluate the collected training results and select the optimal model as the final target detection model. Through the above steps, the model processing terminal can be intelligently selected based on the shared data and the current status of each model processing terminal, and the model training instructions can be efficiently issued.

[0059] like Figure 2 As shown, an embodiment of the present invention further provides a target detection system based on satellite remote sensing data, comprising a detection terminal and a control terminal connected to the detection terminal; the control terminal is connected to a plurality of model processing terminals; The control terminal receives the detection information input by the user and sends a target detection request to the detection terminal based on the detection information input by the user; The detection terminal receives a target detection request and parses the received request; obtains satellite remote sensing raw data according to the parsing result; processes the obtained raw data to obtain a remote sensing image to be detected; determines whether a corresponding trained target detection model exists according to the parsing result; if the trained target detection model exists, inputs the remote sensing image to be detected into the target detection model to obtain a detection result output by the target detection model; When a trained target detection model does not exist, the detection terminal feeds back information to the control terminal, and the control terminal issues a model training instruction to each model processing terminal based on the received feedback information, triggering each model processing terminal to obtain the neural network model to be trained based on the training instruction, and performs a set number of trainings to generate a detection model; The control terminal determines the target model processing terminal based on the accuracy of the detection model, obtains the detection model generated by the target model processing terminal and sends it to the detection terminal as the target detection model.

[0060] In some embodiments, the detection terminal receives a target detection request; and parses the target detection request to obtain target information, including the type of the detected target, geographic location information, and a time range for detection.

[0061] In some embodiments, the system also includes a satellite data source, and the detection terminal selects a satellite data source based on the type and geographic location information of the detection target; uses the time range information to filter out remote sensing data from the selected satellite data source within a specified time period; and downloads the filtered remote sensing data from the satellite data source.

[0062] In some embodiments, when receiving remote sensing data, the detection terminal parses the original data bit stream layer by layer according to the transmission control protocol to obtain the image data to be detected and the corresponding auxiliary reference data.

[0063] In some embodiments, the detection terminal receives a message segment transmitted in the form of a byte stream; parses the message segment layer by layer, extracts control information from the TCP header, and extracts actual remote sensing image data and auxiliary reference data from the data part; rearranges the received data according to the sequence number to restore the original data order; after the data parsing and reorganization are completed, extracts the remote sensing image data and corresponding auxiliary reference data from the parsed data.

[0064] In some embodiments, each model processing terminal is provided with a data set for training different models, and the steps of model training performed by each model processing terminal include: obtaining a data set based on a training instruction; the data set includes a plurality of labeled sample remote sensing images and a plurality of unlabeled sample remote sensing images; using each unlabeled sample remote sensing image and each labeled sample remote sensing image to train the acquired neural network model to be trained, and obtain a target detection model. The specific training process is as follows: extracting features from the unlabeled sample remote sensing image to obtain a first feature image, and extracting features from the labeled sample remote sensing image to obtain a second feature image; generating a first region image based on the first feature image, and generating a second region image based on the second feature image; extracting features from the first region image to generate a first region feature, and extracting features from the second region image to generate a second region feature; performing feature enhancement processing on the first region feature and the second region feature, respectively; obtaining the first detection frame coordinates and the first classification score of each sample target corresponding to the first region feature, and the second detection frame coordinates and the second classification score of each sample target corresponding to the second region feature based on the enhanced first region feature and the second region feature, to optimize the total loss function; when the total loss function value is minimum or converges, determining the target detection model.

[0065] In some embodiments, each model processing terminal is provided with shared data and a local data set for training the model, wherein the shared data is used to characterize the data distribution of the local data of the model processing terminal; the control terminal obtains the shared data and the current status of each model processing terminal; the model processing terminal is selected according to the obtained shared data and the current status of each model processing terminal, and a model training instruction is issued to each selected model processing terminal, specifically for analyzing the shared data of each model processing terminal; a model processing terminal whose data distribution has a similarity with the data distribution required by the current model to be trained greater than a set threshold is selected as the initial model processing terminal; the current status of each initial model processing terminal is checked; and a model processing terminal whose computing resources are greater than a first set value and whose data delay is less than a second threshold is selected to issue a model training instruction.

[0066] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0067] An embodiment of the target detection system based on satellite remote sensing data provided in an embodiment of the present invention belongs to the same inventive concept as the target detection method based on satellite remote sensing data in the above-mentioned embodiments. For details not described in detail in the embodiment of the target detection system based on satellite remote sensing data, reference can be made to the above-mentioned embodiment of the target detection method based on satellite remote sensing data.

[0068] The target detection system based on satellite remote sensing data is a unit and algorithm step of each example described in combination with the embodiments disclosed herein, which can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0069] Those skilled in the art will appreciate that various aspects of the target detection method based on satellite remote sensing data can be implemented as a system, method or program product. Therefore, various aspects of the present disclosure can be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, which can be collectively referred to as "circuit", "module" or "system" here.

[0070] Although the present invention has been described in detail with reference to the accompanying drawings and in combination with the preferred embodiments, the present invention is not limited thereto. Without departing from the spirit and essence of the present invention, a person of ordinary skill in the art may make various equivalent modifications or substitutions to the embodiments of the present invention, and these modifications or substitutions shall be within the scope of the present invention. Any person of ordinary skill in the art may easily think of changes or substitutions within the technical scope disclosed by the present invention, and these shall be within the scope of protection of the present invention.

Claims

1. A target detection method based on satellite remote sensing data, characterized in that: The steps include: Receive target detection requests and parse the received requests; According to the analysis results, the original satellite remote sensing data is obtained; Process the acquired raw data to obtain the remote sensing image to be detected; Determine whether the corresponding trained target detection model exists based on the analysis results; When a trained target detection model exists, the remote sensing image to be detected is input into the target detection model to obtain a detection result output by the target detection model; When a trained target detection model does not exist, a model training instruction is issued to each model processing terminal, triggering each model processing terminal to obtain a neural network model to be trained based on the training instruction, and perform a set number of trainings to generate a detection model; Determine the target model processing terminal based on the accuracy of the detection model, and obtain the detection model generated by the target model processing terminal as the target detection model; Execution steps: input the remote sensing image to be detected into the target detection model to obtain the detection result output by the target detection model.

2. The target detection method based on satellite remote sensing data according to claim 1, characterized in that: The steps of receiving a target detection request and parsing the received request include: Receive a target detection request; The target detection request is parsed to obtain target information, including the type of detected target, geographic location information, and detection time range.

3. The target detection method based on satellite remote sensing data according to claim 2, characterized in that: According to the analysis results, the steps to obtain satellite remote sensing raw data include: Select the satellite data source based on the type and geographic location information of the detection target; Using the time range information, the remote sensing data of the selected satellite data source within the specified time period is filtered out; Download filtered remote sensing data from satellite data sources.

4. The target detection method based on satellite remote sensing data according to claim 3 is characterized in that: The steps of processing the acquired raw data to obtain the remote sensing image to be detected include: When receiving remote sensing data, the original data bit stream is parsed layer by layer according to the transmission control protocol to obtain the image data to be detected and the corresponding auxiliary reference data.

5. The target detection method based on satellite remote sensing data according to claim 4, characterized in that: When receiving remote sensing data, the steps of parsing the original data bit stream layer by layer according to the transmission control protocol to obtain the image data to be detected and the corresponding auxiliary reference data include: Receive message segments transmitted in the form of byte streams; Parse the message segments layer by layer, extract the control information from the TCP header, and extract the actual remote sensing image data and auxiliary reference data from the data part; Rearrange the received data according to the sequence number to restore the original data order; After data parsing and reorganization are completed, remote sensing image data and corresponding auxiliary reference data are extracted from the parsed data.

6. The target detection method based on satellite remote sensing data according to claim 5, characterized in that: Each model processing terminal is provided with a data set for training different models. The steps of model training performed by each model processing terminal include: Acquire a data set based on the training instruction; the data set includes a plurality of labeled sample remote sensing images and a plurality of unlabeled sample remote sensing images; The unlabeled sample remote sensing images and the labeled sample remote sensing images are used to train the acquired neural network model to be trained, so as to obtain the target detection model.

7. The target detection method based on satellite remote sensing data according to claim 6, characterized in that: The steps of training the obtained neural network model to be trained by using each unlabeled sample remote sensing image and each labeled sample remote sensing image to obtain the target detection model include: inputting each unlabeled sample remote sensing image and each labeled sample remote sensing image into the neural network model to be trained, and the specific training process is as follows: Performing feature extraction on the unlabeled sample remote sensing image to obtain a first feature image, and performing feature extraction on the labeled sample remote sensing image to obtain a second feature image; Generate a first region image based on the first feature image, and generate a second region image based on the second feature image; extract features of the first region image to generate first region features, and extract features of the second region image to generate second region features; Perform feature enhancement processing on the first region feature and the second region feature respectively; based on the enhanced first region feature and the second region feature, obtain the first detection frame coordinates and the first classification score of each sample target corresponding to the first region feature, and the second detection frame coordinates and the second classification score of each sample target corresponding to the second region feature, so as to optimize the total loss function; When the total loss function value is minimized or converged, the target detection model is determined.

8. The target detection method based on satellite remote sensing data according to claim 6, characterized in that: Each model processing terminal is provided with shared data and a local data set for training the model, wherein the shared data is used to characterize the data distribution of the local data of the model processing terminal; the method further includes: Obtain the shared data and current status of each model processing terminal; A model processing terminal is selected based on the acquired shared data and the current status of each model processing terminal, and a model training instruction is issued to each selected model processing terminal.

9. The target detection method based on satellite remote sensing data according to claim 7, characterized in that: The steps of selecting a model processing terminal according to the acquired shared data and the current status of each model processing terminal and issuing a model training instruction to each selected model processing terminal include: Analyze the shared data of each model processing terminal; Select a model processing terminal whose data distribution is more similar to the data distribution required by the current model to be trained than a set threshold as the initial model processing terminal; Check the current status of each initial model processing terminal; A model processing terminal whose computing resources are greater than a first set value and whose data delay is less than a second threshold is selected to issue a model training instruction.

10. A target detection system based on satellite remote sensing data, characterized in that: It includes a detection terminal and a control terminal connected to the detection terminal; the control terminal is connected to a plurality of model processing terminals; The control terminal receives the detection information input by the user and sends a target detection request to the detection terminal based on the detection information input by the user; The detection terminal receives the target detection request and parses the received request; obtains the original satellite remote sensing data according to the parsing result; processes the obtained original data to obtain the remote sensing image to be detected; Determine whether a corresponding trained target detection model exists according to the analysis result; if the trained target detection model exists, input the remote sensing image to be detected into the target detection model to obtain the detection result output by the target detection model; When a trained target detection model does not exist, the detection terminal feeds back information to the control terminal, and the control terminal issues a model training instruction to each model processing terminal based on the received feedback information, triggering each model processing terminal to obtain the neural network model to be trained based on the training instruction, and performs a set number of trainings to generate a detection model; The control terminal determines the target model processing terminal based on the accuracy of the detection model, obtains the detection model generated by the target model processing terminal and sends it to the detection terminal as the target detection model.

Images