Remote sensing visual system based on supervised training

Abstract

The invention discloses a remote sensing visual system based on supervised training, which relates to the technical field of remote sensing systems and comprises a data enhancement module, a feature extraction module, a target detection module, a semantic segmentation module, a change detection module, a result evaluation module and an intelligent decision module. According to the remote sensing visual system based on supervised training, by introducing an advanced deep learning model and combining multi-module cooperation, efficient processing and analysis of remote sensing images are achieved. According to the system, through cooperative work of the data enhancement module, the feature extraction module, the target detection module, the semantic segmentation module, the change detection module and the like, a target can be accurately recognized and positioned, a pixel-by-pixel classification result is provided, and the accuracy and the real-time performance of remote sensing data analysis are remarkably improved. The method is obtained based on ten-million semantic segmentation remote sensing image slice training, has a good recognition characteristic for most land classes, and can also be used for well recognizing images which are not seen. In downstream tasks, migration can be realized only by a small number of samples.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing systems, and in particular to a remote sensing vision system based on supervised training. Background Art

[0002] In recent years, the rapid development of remote sensing technology has made the acquisition of high-resolution remote sensing images more and more popular. These images are widely used in urban monitoring, land use management, environmental monitoring and other fields. Existing remote sensing vision systems mainly rely on traditional image processing methods and some classic machine learning algorithms, which often cannot fully utilize the potential of deep learning.

[0003] In the prior art, since unlabeled data is used for training, the self-supervised solution needs to construct a supervision task by itself during training. As for the generative self-supervised solution, the model randomly blocks the input image, and then lets the model reconstruct the occluded area based only on the unoccluded part, and uses the original data of the occluded area as the supervision data. Although the model can obtain better pre-training parameters in the case of large amounts of data, the model parameters themselves are not suitable for any downstream tasks. In the remote sensing scenario, even if a self-supervised model is trained using a large amount of remote sensing data, the model itself is not sensitive to remote sensing tasks (such as semantic segmentation, change detection, etc.), and cannot be well migrated to downstream subtasks in small sample scenarios.

[0004] At the same time, in the target detection module, traditional algorithms perform poorly in small target detection and complex backgrounds, and it is difficult to meet high-precision requirements; in the semantic segmentation module, existing methods often have difficulty in achieving ideal segmentation effects when processing remote sensing images with blurred boundaries and rich details. In addition, existing technologies generally lack the ability to fuse multi-layer features, resulting in information loss and low processing efficiency, which affects the final analysis results and decision support. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a remote sensing vision system based on supervised training to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: a remote sensing vision system based on supervised training, comprising a data enhancement module, a feature extraction module, a target detection module, a semantic segmentation module, a change detection module and a result evaluation module; Data enhancement module, which expands remote sensing images based on image transformation, affine transformation, and noise injection technology to increase the diversity of model training samples; The feature extraction module uses deep learning algorithms to extract multi-level features from remote sensing images. It adopts the convolutional neural network (CNN) architecture to extract high-dimensional feature information through layer-by-layer convolution and pooling operations. The target detection module combines YOLO and Faster R-CNN detection algorithms to identify and locate target objects in remote sensing images in real time. It uses multi-scale feature fusion technology to integrate features at different levels to improve the detection capability of small targets. The semantic segmentation module classifies remote sensing images pixel by pixel so that each pixel is labeled as a specific category, such as buildings, roads, and water bodies. It uses a fully convolutional network (FCN) to perform semantic segmentation of images. Through an end-to-end learning method, it takes the entire image as input and outputs the corresponding classification map. The change detection module is used to monitor the changes in the same area between different time points, using time-series remote sensing data and deep learning algorithms to identify changes between image pairs; The result evaluation module is used to comprehensively evaluate the output results of the system, including the calculation of accuracy, precision, and recall indicators, and integrates evaluation tools and visualization tools to help users deeply understand the performance of the model.

[0007] To further optimize the technical solution, in the data enhancement module, a generative adversarial network (GAN) is combined to perform data enhancement to generate realistic synthetic images and supplement the scarce training data in specific environments; By using GAN-generated samples, the accuracy of object recognition in complex scenarios, such as densely populated urban areas or after natural disasters, can be improved. The module also has the ability to automatically adjust generation parameters, dynamically adjusting the type and quantity of generated samples according to the progress of model training to achieve better training results.

[0008] To further optimize the technical solution, the feature extraction module uses the attention mechanism to automatically weight important areas, highlight key information, and adaptively identify and extract features in different scenes, such as water bodies, urban infrastructure, and vegetation distribution, thereby providing support for subsequent classification and detection.

[0009] To further optimize the technical solution, the target detection module has a built-in instance segmentation model based on image segmentation to identify targets and accurately define target boundaries, thereby obtaining more detailed ground object information, such as land use classification and urban expansion monitoring, to meet higher accuracy requirements; Based on the instance segmentation model of image segmentation, set an objective function To describe the loss of instance segmentation, including category prediction loss, bounding box regression loss and boundary loss, as shown below: ; in, For classification loss, the model predicts the category of each pixel. is the bounding box loss, which represents the difference between the predicted bounding box and the true bounding box; Boundary loss optimizes the smoothness and connectivity of the target boundary; is the weight coefficient used to adjust the relative importance of different loss terms.

[0010] To further optimize the technical solution, the instance segmentation model includes the following process when used: Data preparation: Collect annotated remote sensing image datasets, each image contains the category of the target object and pixel-level boundary information; Network architecture design: Choose a convolutional neural network (CNN) architecture as the feature extractor, such as ResNet or EfficientNet. After the feature extractor, add an upsampling layer to restore the feature map to the resolution of the input image for pixel-level classification. Multi-scale feature fusion: Implement multi-scale feature fusion in the network, fuse low-level and high-level features to better capture the information of small objects, and achieve it through skip connections, helping the model learn richer feature representations; Instance segmentation implementation: In the final output layer, the Softmax function is used to predict the category of each pixel to obtain the category probability of each pixel. At the same time, the regression loss is used to calculate the position of the predicted bounding box. Boundary loss calculation: The contour loss is used to ensure that the boundary of the target is smooth and connected, which is defined as the Euclidean distance between the boundary point and the true boundary point, that is: ; in, is the real boundary point, are the boundary points predicted by the model; Training process: Optimize the above objective function through back propagation , using cross entropy loss to optimize the classification task, using smooth L1 loss to optimize the bounding box, and combining contour loss to improve the accuracy of boundary prediction; Inference and post-processing: In the inference stage, the model classifies the input remote sensing image and predicts the bounding box; then, it uses non-maximum suppression (NMS) to remove overlapping areas and extract the final target instance and its boundary.

[0011] To further optimize the technical solution, the semantic segmentation module has a built-in semantic segmentation model, and the semantic segmentation model sets a loss function To describe the performance of semantic segmentation, including category prediction loss and boundary loss: ; in, It is the semantic segmentation loss, which uses cross entropy loss to calculate the model's category prediction for each pixel; It is a boundary loss, which enhances the smoothness and accuracy of the target boundary; is a weight coefficient used to adjust the relative importance between class loss and boundary loss.

[0012] To further optimize the technical solution, the semantic segmentation model includes the following process when used: Data preparation: Collect remote sensing image datasets with accurate annotations, ensuring that every pixel in each image is labeled as a specific category, such as buildings, roads, and water bodies; Network architecture design: Use convolutional neural networks (CNNs) as feature extractors, such as VGG or ResNet, remove fully connected layers, keep only convolutional layers, and build fully convolutional networks (FCNs); after the convolutional layers, add upsampling layers, such as transposed convolutional layers, to restore the feature maps to the resolution of the original image; Feature fusion: realize multi-layer feature fusion, merge feature maps at different levels, and use skip connections to combine low-level features with high-level features to better capture detailed information; Semantic segmentation output: In the final output layer of the network, the Softmax activation function is used to predict the category of each pixel and obtain the probability distribution of each pixel; Boundary loss calculation: Define boundary loss as the smoothness loss of the boundary points of the target area, using the structural similarity index (SSIM) or gradient-based loss; Training process: Optimize the above loss function through back propagation , combined with cross entropy and boundary loss for joint training, gradually optimizing model performance; Inference and post-processing: In the inference stage, the input remote sensing image is passed through the network to obtain the category prediction of each pixel; thresholding is used to convert the probability map into a discrete category map for subsequent analysis and application.

[0013] To further optimize this technical solution, in the semantic segmentation model: At the semantic segmentation output, the activation function is as follows: ; in, It is The prediction score of the class, is the total number of categories; When calculating the boundary loss, the calculation formula is as follows: ; in, It is the boundary information of the real annotation.

[0014] To further optimize the technical solution, the change detection module is suitable for scenarios of urban expansion and environmental monitoring. It adopts a combination of optical flow method and convolutional neural network to analyze spatial changes while taking into account time series characteristics, and efficiently identifies changes in dynamic environments. It is suitable for application scenarios such as post-natural disaster monitoring and dynamic tracking of resource development.

[0015] To further optimize the technical solution, the system also includes an intelligent decision-making module, which uses a deep reinforcement learning algorithm to provide intelligent decision support based on system output results and real-time environmental information; By studying historical data and analyzing the various factors that affect decision-making, the system provides scenario-based recommendations and solutions. For example, in urban management, the system proposes traffic management optimization plans and environmental protection measures based on real-time monitoring data.

[0016] Compared with the prior art, the present invention provides a remote sensing vision system based on supervised training, which has the following beneficial effects: This remote sensing vision system based on supervised training has achieved efficient processing and analysis of remote sensing images by introducing advanced deep learning models and combining multi-module collaboration. Through the collaborative work of modules such as data enhancement, feature extraction, target detection, semantic segmentation, and change detection, the system can accurately identify and locate targets, provide pixel-by-pixel classification results, and significantly improve the accuracy and real-time performance of remote sensing data analysis. Based on the training of tens of millions of semantically segmented remote sensing image slices, it has good recognition characteristics for most land types and can also better recognize unseen images. In downstream tasks, only a small number of samples are needed to achieve migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of the structure of a remote sensing vision system based on supervised training proposed by the present invention; Figure 2 A schematic diagram of the process of an instance segmentation model in a remote sensing vision system based on supervised training proposed by the present invention; Figure 3 A flowchart of a semantic segmentation model in a remote sensing vision system based on supervised training proposed by the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions 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 are within the scope of protection of the present invention.

[0019] Example: See also Figure 1 ,A remote sensing vision system based on supervised training includes a data enhancement module, a feature extraction module, a target detection module, a semantic segmentation module, a change detection module and a result evaluation module.

[0020] The data enhancement module expands remote sensing images based on image transformation, affine transformation, and noise injection technology to increase the diversity of model training samples.

[0021] In this embodiment, the data enhancement module is combined with a generative adversarial network (GAN) to perform data enhancement, generate realistic synthetic images, and supplement scarce training data in specific environments; By using GAN-generated samples, the accuracy of object recognition in complex scenarios, such as densely populated urban areas or after natural disasters, can be improved. The module also has the ability to automatically adjust generation parameters, dynamically adjusting the type and quantity of generated samples according to the progress of model training to achieve better training results.

[0022] The feature extraction module uses a deep learning algorithm to extract multi-level features from remote sensing images and adopts a convolutional neural network (CNN) architecture to extract high-dimensional feature information through layer-by-layer convolution and pooling operations.

[0023] In this embodiment, the feature extraction module not only focuses on traditional low-level features such as texture and shape, but also uses the attention mechanism to automatically weight important areas, highlight key information, and adaptively identify and extract features in different scenes, such as water bodies, urban infrastructure, and vegetation distribution, thereby providing support for subsequent classification and detection.

[0024] The target detection module combines YOLO and Faster R-CNN detection algorithms to identify and locate target objects in remote sensing images in real time. It uses multi-scale feature fusion technology to integrate features at different levels to improve the detection capability of small targets.

[0025] In this embodiment, the target detection module has a built-in instance segmentation model based on image segmentation to identify targets and accurately define target boundaries, thereby obtaining more detailed ground object information, such as land use classification and urban expansion monitoring, to meet higher accuracy requirements; Based on the instance segmentation model of image segmentation, set an objective function To describe the loss of instance segmentation, including category prediction loss, bounding box regression loss and boundary loss, as shown below: ; in, For classification loss, the model predicts the category of each pixel. is the bounding box loss, which represents the difference between the predicted bounding box and the true bounding box; Boundary loss optimizes the smoothness and connectivity of the target boundary; is the weight coefficient used to adjust the relative importance of different loss terms.

[0026] like Figure 2 As shown, the instance segmentation model, when used, includes the following processes: Data preparation: Collect annotated remote sensing image datasets, each image contains the category of the target object and pixel-level boundary information; Network architecture design: Choose a convolutional neural network (CNN) architecture as the feature extractor, such as ResNet or EfficientNet. After the feature extractor, add an upsampling layer to restore the feature map to the resolution of the input image for pixel-level classification. Multi-scale feature fusion: Implement multi-scale feature fusion in the network, fuse low-level and high-level features to better capture the information of small objects, and achieve it through skip connections, helping the model learn richer feature representations; Instance segmentation implementation: In the final output layer, the Softmax function is used to predict the category of each pixel to obtain the category probability of each pixel. At the same time, the regression loss is used to calculate the position of the predicted bounding box. Boundary loss calculation: The contour loss is used to ensure that the boundary of the target is smooth and connected, which is defined as the Euclidean distance between the boundary point and the true boundary point, that is: ; in, is the real boundary point, are the boundary points predicted by the model; Training process: Optimize the above objective function through back propagation , using cross entropy loss to optimize the classification task, using smooth L1 loss to optimize the bounding box, and combining contour loss to improve the accuracy of boundary prediction; Inference and post-processing: In the inference stage, the model classifies the input remote sensing image and predicts the bounding box; then, it uses non-maximum suppression (NMS) to remove overlapping areas and extract the final target instance and its boundary.

[0027] The system can achieve high-precision detection and segmentation of targets in remote sensing images, especially in scenarios such as urban expansion and environmental monitoring. The processing of small targets and complex boundaries can significantly improve the detection effect. Each loss term in the model can be adjusted according to task requirements to ensure a balance between target detection and boundary recognition to achieve the best segmentation effect.

[0028] The semantic segmentation module classifies remote sensing images pixel by pixel so that each pixel is labeled as a specific category, such as buildings, roads, and water bodies. It uses a fully convolutional network (FCN) to perform semantic segmentation of images. Through an end-to-end learning method, it takes the entire image as input and outputs the corresponding classification map. It integrates a multi-task learning framework, combines the tasks of target detection and semantic segmentation, and optimizes the overall performance of the model by sharing a feature extraction network. This mechanism not only improves segmentation accuracy, but also reduces the amount of data and computing resources required for training.

[0029] In this embodiment, the semantic segmentation module has a built-in semantic segmentation model, and the semantic segmentation model sets a loss function To describe the performance of semantic segmentation, including category prediction loss and boundary loss: ; in, It is the semantic segmentation loss, which uses cross entropy loss to calculate the model's category prediction for each pixel; It is a boundary loss, which enhances the smoothness and accuracy of the target boundary; is a weight coefficient used to adjust the relative importance between class loss and boundary loss.

[0030] like Figure 3 As shown, the semantic segmentation model, when used, includes the following processes: Data preparation: Collect remote sensing image datasets with accurate annotations, ensuring that every pixel in each image is labeled as a specific category, such as building, road, water body.

[0031] Network architecture design: Use convolutional neural network (CNN) as feature extractor, such as VGG or ResNet, remove the fully connected layer, keep only the convolution layer, and build a fully convolutional network (FCN); after the convolution layer, add upsampling layers, such as transposed convolution layers to restore the feature map to the resolution of the original image.

[0032] Feature fusion: Implement multi-layer feature fusion, merge feature maps at different levels, and use skip connections to combine low-level features with high-level features to better capture detailed information.

[0033] Semantic segmentation output: In the final output layer of the network, the Softmax activation function is used to predict the category of each pixel and obtain the probability distribution of each pixel; in the semantic segmentation output, the activation function is as follows: ; in, It is The prediction score of the class, is the total number of categories.

[0034] Boundary loss calculation: Define boundary loss as the smoothness loss of the boundary points of the target area, using the structural similarity index (SSIM) or gradient-based loss. When calculating boundary loss, the calculation formula is as follows: ; in, It is the boundary information of the real annotation.

[0035] Training process: Optimize the above loss function through back propagation , combined with cross entropy and boundary loss for joint training, gradually optimizing model performance.

[0036] Inference and post-processing: In the inference stage, the input remote sensing image is passed through the network to obtain the category prediction of each pixel; thresholding is used to convert the probability map into a discrete category map for subsequent analysis and application.

[0037] The fully convolutional network realizes pixel-by-pixel classification of remote sensing images and effectively identifies various types of objects in the image. The boundary loss term in the model can effectively improve the accuracy of the target boundary, making the segmentation result more delicate and suitable for processing complex remote sensing scenes. By flexibly adjusting the weight coefficient of the loss term , it can balance category recognition and boundary accuracy according to specific task requirements to meet the needs of different application scenarios.

[0038] The change detection module is used to monitor the changes in the same area between different time points. It uses time-series remote sensing data and deep learning algorithms to identify changes between image pairs.

[0039] In this embodiment, the change detection module is suitable for scenarios of urban expansion and environmental monitoring. It adopts a combination of optical flow method and convolutional neural network to analyze spatial changes while taking into account time series characteristics. It can efficiently identify changes in a dynamic environment and is suitable for application scenarios such as post-natural disaster monitoring and dynamic tracking of resource development.

[0040] The result evaluation module is used to comprehensively evaluate the output results of the system, including the calculation of accuracy, precision, and recall indicators, and integrates evaluation tools and visualization tools to help users deeply understand the performance of the model.

[0041] In this embodiment, the evaluation process is made more comprehensive and detailed by introducing multi-dimensional evaluation criteria, such as F1-score, IoU, etc. This module not only provides feedback for model optimization, but also supports comparison with industry standards, providing a scientific basis for user decision-making.

[0042] In this embodiment, the system further includes an intelligent decision-making module, which uses a deep reinforcement learning algorithm to perform intelligent decision support based on system output results and real-time environmental information; By learning historical data and analyzing various factors that affect decision-making, the system provides scenario-based suggestions and solutions. For example, in urban management, the system proposes traffic management optimization solutions and environmental protection measures based on real-time monitoring data. The introduction of an adaptive learning mechanism enables the system to continuously update and optimize decision-making strategies based on new data, thereby improving the accuracy and timeliness of decision-making.

[0043] The training data of this system is as high as tens of millions of slice samples, including 18 common remote sensing land scenes, which conform to the common category rule system of remote sensing. The model trained on this large-scale data itself has strong generalization and special semantics of remote sensing scenes, that is, the model itself focuses on the remote sensing mission objectives.

[0044] The beneficial effects of the present invention are: This remote sensing vision system based on supervised training has achieved efficient processing and analysis of remote sensing images by introducing advanced deep learning models and combining multi-module collaboration. Through the collaborative work of modules such as data enhancement, feature extraction, target detection, semantic segmentation, and change detection, the system can accurately identify and locate targets, provide pixel-by-pixel classification results, and significantly improve the accuracy and real-time performance of remote sensing data analysis. Based on the training of tens of millions of semantically segmented remote sensing image slices, it has good recognition characteristics for most land types and can also better recognize unseen images. In downstream tasks, only a small number of samples are needed to achieve migration.

[0045] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0046] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A remote sensing vision system based on supervised training, characterized in that: It includes data enhancement module, feature extraction module, target detection module, semantic segmentation module, change detection module and result evaluation module; Data enhancement module, which expands remote sensing images based on image transformation, affine transformation, and noise injection technology to increase the diversity of model training samples; The feature extraction module uses deep learning algorithms to extract multi-level features from remote sensing images. It adopts the convolutional neural network (CNN) architecture to extract high-dimensional feature information through layer-by-layer convolution and pooling operations. The target detection module combines YOLO and Faster R-CNN detection algorithms to identify and locate target objects in remote sensing images in real time. It uses multi-scale feature fusion technology to integrate features at different levels to improve the detection capability of small targets. The semantic segmentation module classifies remote sensing images pixel by pixel so that each pixel is labeled as a specific category, such as buildings, roads, and water bodies; Use a fully convolutional network (FCN) to perform semantic segmentation of images. Through an end-to-end learning method, the entire image is taken as input and the corresponding classification map is output; The change detection module is used to monitor the changes in the same area between different time points, using time-series remote sensing data and deep learning algorithms to identify changes between image pairs; The result evaluation module is used to comprehensively evaluate the output results of the system, including the calculation of accuracy, precision, and recall indicators, and integrates evaluation tools and visualization tools to help users deeply understand the performance of the model.

2. A remote sensing vision system based on supervised training according to claim 1, characterized in that: In the data enhancement module, a generative adversarial network (GAN) is combined to perform data enhancement to generate realistic synthetic images and supplement the scarce training data in a specific environment; By using GAN-generated samples, the accuracy of object recognition in complex scenarios, such as densely populated urban areas or after natural disasters, can be improved. The module also has the ability to automatically adjust generation parameters, dynamically adjusting the type and quantity of generated samples according to the progress of model training to achieve better training results.

3. A remote sensing vision system based on supervised training according to claim 1, characterized in that: The feature extraction module uses the attention mechanism to automatically weight important areas, highlight key information, and adaptively identify and extract features in different scenes, such as water bodies, urban infrastructure, and vegetation distribution, thereby providing support for subsequent classification and detection.

4. A remote sensing vision system based on supervised training according to claim 1, characterized in that: The target detection module has a built-in instance segmentation model based on image segmentation to identify targets and accurately define target boundaries, thereby obtaining more detailed ground object information, such as land use classification and urban expansion monitoring, to meet higher accuracy requirements; Based on the instance segmentation model of image segmentation, set an objective function To describe the loss of instance segmentation, including category prediction loss, bounding box regression loss and boundary loss, as shown below: ； in, For classification loss, the model predicts the category of each pixel. is the bounding box loss, which represents the difference between the predicted bounding box and the true bounding box; Boundary loss optimizes the smoothness and connectivity of the target boundary; is the weight coefficient used to adjust the relative importance of different loss terms.

5. A remote sensing vision system based on supervised training according to claim 4, characterized in that: When the instance segmentation model is used, it includes the following process: Data preparation: Collect annotated remote sensing image datasets, each image contains the category of the target object and pixel-level boundary information; Network architecture design: Choose a convolutional neural network (CNN) architecture as the feature extractor, such as ResNet or EfficientNet; After the feature extractor, an upsampling layer is added to restore the feature map to the resolution of the input image for pixel-level classification. Multi-scale feature fusion: Implement multi-scale feature fusion in the network, fuse low-level and high-level features to better capture the information of small objects, and achieve it through skip connections, helping the model learn richer feature representations; Instance segmentation implementation: In the final output layer, the Softmax function is used to predict the category of each pixel to obtain the category probability of each pixel. At the same time, the regression loss is used to calculate the position of the predicted bounding box. Boundary loss calculation: The contour loss is used to ensure that the boundary of the target is smooth and connected, which is defined as the Euclidean distance between the boundary point and the true boundary point, that is: ； in, is the real boundary point, are the boundary points predicted by the model; Training process: Optimize the above objective function through back propagation , using cross entropy loss to optimize the classification task, using smooth L1 loss to optimize the bounding box, and combining contour loss to improve the accuracy of boundary prediction; Inference and post-processing: In the inference stage, the model classifies the input remote sensing image and predicts the bounding box; then, it uses non-maximum suppression (NMS) to remove overlapping areas and extract the final target instance and its boundary.

6. A remote sensing vision system based on supervised training according to claim 1, characterized in that: The semantic segmentation module has a built-in semantic segmentation model, which sets a loss function To describe the performance of semantic segmentation, including category prediction loss and boundary loss: ； in, It is the semantic segmentation loss, which uses cross entropy loss to calculate the model's category prediction for each pixel; It is a boundary loss, which enhances the smoothness and accuracy of the target boundary; is a weight coefficient used to adjust the relative importance between class loss and boundary loss.

7. A remote sensing vision system based on supervised training according to claim 6, characterized in that: When the semantic segmentation model is used, it includes the following processes: Data preparation: Collect remote sensing image datasets with accurate annotations, ensuring that every pixel in each image is labeled as a specific category, such as buildings, roads, and water bodies; Network architecture design: Use convolutional neural networks (CNNs) as feature extractors, such as VGG or ResNet, remove fully connected layers, keep only convolutional layers, and build fully convolutional networks (FCNs); after the convolutional layers, add upsampling layers, such as transposed convolutional layers, to restore the feature maps to the resolution of the original image; Feature fusion: realize multi-layer feature fusion, merge feature maps at different levels, and use skip connections to combine low-level features with high-level features to better capture detailed information; Semantic segmentation output: In the final output layer of the network, the Softmax activation function is used to predict the category of each pixel and obtain the probability distribution of each pixel; Boundary loss calculation: Define boundary loss as the smoothness loss of the boundary points of the target area, using the structural similarity index (SSIM) or gradient-based loss; Training process: Optimize the above loss function through back propagation , combined with cross entropy and boundary loss for joint training, gradually optimizing model performance; Inference and post-processing: In the inference phase, the input remote sensing image is passed through the network to obtain the category prediction of each pixel; Use thresholding to convert the probability map into a discrete category map for subsequent analysis and application.

8. A remote sensing vision system based on supervised training according to claim 7, characterized in that: In the semantic segmentation model: At the semantic segmentation output, the activation function is as follows: ； in, It is The prediction score of the class, is the total number of categories; When calculating the boundary loss, the calculation formula is as follows: ； in, It is the boundary information of the real annotation.

9. The remote sensing vision system based on supervised training according to claim 1, characterized in that: The change detection module is suitable for scenarios such as urban expansion and environmental monitoring. It combines the optical flow method with a convolutional neural network to analyze spatial changes while taking into account time series characteristics. It can efficiently identify changes in a dynamic environment and is suitable for application scenarios such as post-natural disaster monitoring and dynamic tracking of resource development.

10. A remote sensing vision system based on supervised training according to claim 1, characterized in that: The system also includes an intelligent decision-making module, which uses deep reinforcement learning algorithms to provide intelligent decision support based on system output results and real-time environmental information; By studying historical data and analyzing the various factors that affect decision-making, the system provides scenario-based recommendations and solutions. For example, in urban management, the system proposes traffic management optimization plans and environmental protection measures based on real-time monitoring data.

Images