Remote sensing system with pyramid structure and multi-stage feature fusion

Abstract

The invention discloses a remote sensing system with a pyramid structure and multi-stage feature fusion, which relates to the technical field of remote sensing systems and comprises a data preprocessing module, a pyramid feature extraction module, a semantic segmentation module, a feature fusion module, a model evaluation module, an intelligent decision module and a result pushing module. According to the remote sensing system with the pyramid structure and the multi-stage feature fusion, the pyramid structure and the multi-stage feature fusion technology are adopted, and the efficiency of feature extraction and fusion is greatly improved through the methods of Overlap Patch Embedding, Spatial-Reductive Attention and the like, so that the model can efficiently process high-resolution remote sensing images. Meanwhile, the intelligent decision-making module is combined with deep reinforcement learning, so that the system can dynamically adjust a decision-making strategy according to real-time data and historical data, and the accuracy and practicability of decision making are remarkably improved.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing systems, in particular to a remote sensing system with a pyramid structure and multi-stage feature fusion. Background Art

[0002] With the rapid development of remote sensing technology, the processing and analysis of large-scale remote sensing data has become increasingly important. However, existing remote sensing systems still have many deficiencies in feature extraction, data fusion, and decision generation. For example, traditional feature extraction methods often fail to fully utilize multi-scale information, resulting in the loss of detail information of high-resolution remote sensing images. In addition, existing feature fusion technologies mostly rely on a single hierarchical structure and cannot effectively capture the relationship between different features, limiting the performance of the model in complex scenarios. In addition, decision generation is usually based on static models, lacks dynamic adaptability, and cannot respond to environmental changes in real time.

[0003] At the same time, the large remote sensing models currently released by the industry mostly use the Vision Transformer model structure. In computer vision scenarios, for the Vision Transformer, when the input image size increases, the overall time complexity increases by a square factor. Therefore, ViT is naturally not suitable for large-scale image semantic segmentation tasks. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a remote sensing system with a pyramid structure and multi-stage feature fusion to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: a remote sensing system with a pyramid structure and multi-stage feature fusion, comprising a data preprocessing module, a pyramid feature extraction module, a semantic segmentation module, a feature fusion module, a model evaluation module, an intelligent decision-making module and a result push module;

[0006] Data preprocessing module, used to standardize and clean the raw remote sensing data to ensure the accuracy and reliability of subsequent analysis;

[0007] Pyramid feature extraction module, based on the pyramid structure, is used to extract multi-scale features from remote sensing images;

[0008] The semantic segmentation module uses a pre-trained 600 million parameter pyramid model for efficient image segmentation;

[0009] The feature fusion module uses the Cross-Attention mechanism to fuse features extracted at different stages layer by layer to enhance the expressiveness of features;

[0010] The model evaluation module is used to evaluate the performance of the model in real time during training and testing. It uses a variety of evaluation indicators, such as IoU and F1-score, to comprehensively analyze the performance of the model on different data sets.

[0011] Intelligent decision-making module, which uses machine learning algorithms to conduct in-depth mining and intelligent inference of analysis results;

[0012] The result push module is used to push analysis results and decision recommendations to users through multiple channels. Through the API interface, data is pushed to the user's mobile device or Web application to ensure that users can obtain information in real time.

[0013] Further optimizing the technical solution, in the data preprocessing module, the data preprocessing includes the steps of data denoising, geometric correction, and radiation correction;

[0014] Use deep learning-based denoising techniques, such as convolutional neural network (CNN) denoisers, to retain valuable information by learning the characteristic differences between noise and valid signals;

[0015] It also automatically identifies missing values ​​in the data and fills them in using interpolation algorithms to ensure data integrity.

[0016] To further optimize the technical solution, the pyramid feature extraction module divides the image into overlapping small blocks through Overlap Patch Embedding, retaining more spatial information and contextual relationships;

[0017] The pyramid structure enables the model to extract features layer by layer from low-level to high-level, with each layer focusing on details at different scales;

[0018] The Spatial-Reduction Attention mechanism is used to integrate spatial information to improve the processing efficiency of high-resolution remote sensing images.

[0019] To further optimize the technical solution, the semantic segmentation module labels each pixel in the image as a specific category by performing pixel-level classification on the satellite image;

[0020] The U-Net architecture is adopted, which combines the advantages of the pyramid structure, extracts features at different scales, and better handles complex objects and backgrounds.

[0021] To further optimize the technical solution, in the semantic segmentation module, the pre-trained 600 million parameter pyramid model is constructed by the following process:

[0022] First, the model is pre-trained on the open source dataset ImageNet-1k to obtain the initial 600 million parameter pyramid model parameters;

[0023] Using pre-trained model parameters as a basis, we train semantic segmentation tasks on tens of millions of satellite data.

[0024] The trained model features are transferred to the remote sensing scene to obtain a 600 million parameter pyramid model.

[0025] To further optimize this technical solution, the feature fusion module uses the Cross-Attention mechanism to fuse the features of each stage, including the following processes:

[0026] First, perform Cross-Attention calculation on the first and second stages, and then perform Cross-Attention calculation on the fusion features obtained with the features of the third stage, and so on, considering the interaction between multiple input sources to better capture the correlation between sequences;

[0027] In the first stage, shallow features are extracted, which include low-level information of the image, such as edges and textures;

[0028] Then, the Cross-Attention mechanism is used to fuse the shallow features of the first and second stages. The Cross-Attention mechanism calculates the mutual attention between different features, highlights relevant features, and suppresses irrelevant features, thereby obtaining a richer and more meaningful feature representation.

[0029] Next, the features after fusion of the first and second stages are combined with the features of the third stage to perform another Cross-Attention calculation, and so on;

[0030] Through layer-by-layer fusion, the model comprehensively considers the feature information of the previous layer at each layer, gradually improving the abstraction level and expressiveness of feature representation.

[0031] To further optimize the technical solution, in the model evaluation module, a visual performance curve is drawn based on a visualization tool to help researchers intuitively understand the training progress and potential overfitting of the model. Based on the evaluation results, the module automatically adjusts the hyperparameters of the model to optimize the model performance.

[0032] To further optimize the technical solution, the intelligent decision-making module also combines historical data and real-time analysis results, and the module automatically generates decision-making suggestions, such as land use changes and environmental monitoring;

[0033] It also integrates deep reinforcement learning strategies to adjust decision-making strategies based on feedback, allowing the system to continuously optimize and improve itself.

[0034] To further optimize this technical solution, the intelligent decision-making module has an environmental adaptive decision model (EADM) built in. The model combines historical data, real-time data and deep reinforcement learning strategies to automatically generate decision suggestions based on environmental changes. The model is as follows:

[0035] ;

[0036] in,

[0037] Indicates time Decision output;

[0038] It is the real-time data feature vector at the current time point, such as the segmentation results of remote sensing images and climate data;

[0039] is a feature vector of historical data, such as past records of land use changes;

[0040] The model is in time action strategies;

[0041] When used, the model includes:

[0042] Data collection and preprocessing: collect real-time remote sensing data and historical data, and perform standardization and denoising to ensure data quality;

[0043] Feature extraction and fusion: extracting real-time data using CNN The spatial features of the data and the LSTM are used to process the historical data. The time characteristics of the two feature vectors are combined into a comprehensive feature vector ;

[0044] Decision generation: The comprehensive feature vector Input to the adaptive network , generate preliminary decisions ;

[0045] Reinforcement learning strategy update: observe the decision results, update the Q value function based on feedback information, use the greedy strategy to select the best action, and continuously adjust and optimize the action strategy ;

[0046] Decision output: Generate the final decision based on the optimized strategy and push it to the user, providing suggestions on land use changes and environmental monitoring.

[0047] To further optimize the technical solution, the intelligent decision-making module includes the following process when in use:

[0048] Feature extraction: Through deep learning model and Perform feature extraction, using a combination of convolutional neural networks (CNN) and long short-term memory networks (LSTM) to extract spatial and temporal features and capture key information in complex remote sensing data;

[0049] Decision making: using an adaptive network , based on the extracted features and For decision generation, the network adopts a multi-layer perceptron (MLP) structure and improves the model's expressiveness through non-linear activation functions such as ReLU;

[0050] Reinforcement learning feedback: Through deep reinforcement learning, the results of each decision, success or failure, are fed back to the model to update the strategy , using the Q-learning algorithm, define the Q value function , to assess the status Next Action expected return.

[0051] Compared with the prior art, the present invention provides a remote sensing system with a pyramid structure and multi-stage feature fusion, which has the following beneficial effects:

[0052] This remote sensing system with pyramid structure and multi-stage feature fusion has greatly improved the efficiency of feature extraction and fusion by adopting pyramid structure and multi-stage feature fusion technology, through Overlap Patch Embedding and Spatial-Reduction Attention and other methods, so that the model can efficiently process high-resolution remote sensing images. At the same time, the intelligent decision-making module is combined with deep reinforcement learning, so that the system can dynamically adjust the decision-making strategy according to real-time data and historical data, significantly improving the accuracy and practicality of the decision. The system can not only better capture the detailed information in complex scenes, but also provide users with real-time and accurate environmental monitoring and land use change recommendations, thereby effectively supporting various remote sensing applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of the structure of a remote sensing system with a pyramid structure and multi-stage feature fusion proposed by the present invention;

[0054] Figure 2 A flow chart of an environmental adaptability decision model in a remote sensing system with a pyramid structure and multi-stage feature fusion proposed by the present invention;

[0055] Figure 3The present invention proposes a flowchart for constructing a 600 million parameter pyramid model in a remote sensing system with a pyramid structure and multi-stage feature fusion. DETAILED DESCRIPTION

[0056] 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.

[0057] Example:

[0058] See also Figure 1 ,A remote sensing system with a pyramid structure and multi-stage feature fusion, includes a data preprocessing module, a pyramid feature extraction module, a semantic segmentation module, a feature fusion module, a model evaluation module, an intelligent decision making module and a result pushing module.

[0059] The data preprocessing module is used to standardize and clean the raw remote sensing data to ensure the accuracy and reliability of subsequent analysis.

[0060] In this embodiment, in the data preprocessing module, data preprocessing includes the steps of data denoising, geometric correction, and radiation correction;

[0061] Use deep learning-based denoising techniques, such as convolutional neural network (CNN) denoisers, to retain valuable information by learning the characteristic differences between noise and valid signals;

[0062] It also automatically identifies missing values ​​in the data and fills them in using interpolation algorithms to ensure data integrity.

[0063] The pyramid feature extraction module is based on the pyramid structure and is used to extract multi-scale features from remote sensing images.

[0064] In this embodiment, the pyramid feature extraction module divides the image into overlapping small patches through Overlap Patch Embedding, thereby retaining more spatial information and contextual relationships;

[0065] The pyramid structure enables the model to extract features layer by layer from low-level to high-level, with each layer focusing on details at different scales;

[0066] The Spatial-Reduction Attention mechanism is used to integrate spatial information to improve the processing efficiency of high-resolution remote sensing images.

[0067] The semantic segmentation module uses a pre-trained 600 million parameter pyramid model for efficient image segmentation.

[0068] In this embodiment, the semantic segmentation module labels each pixel in the image as a specific category by performing pixel-level classification on the satellite image;

[0069] The U-Net architecture is adopted, which combines the advantages of the pyramid structure, extracts features at different scales, and better handles complex objects and backgrounds.

[0070] like Figure 3 As shown, further, in the semantic segmentation module, the pre-trained 600 million parameter pyramid model includes the following processes when being constructed:

[0071] First, the model is pre-trained on the open source dataset ImageNet-1k to obtain the initial 600 million parameter pyramid model parameters;

[0072] Using pre-trained model parameters as a basis, we train semantic segmentation tasks on tens of millions of satellite data.

[0073] The trained model features are transferred to the remote sensing scene to obtain a 600 million parameter pyramid model.

[0074] Compared with the ViT model with the same number of parameters, this model has faster training and reasoning speed. When the input is a 512x512 image, the training speed is about 2.5 times that of the ViT structure. This can shorten the training time and reduce the development cost.

[0075] The pyramid network extracts features from different scales through a multi-layer structure, which is very important for semantic segmentation tasks, helps improve segmentation accuracy, is more suitable for semantic segmentation tasks, and reduces model training time. Based on the remote sensing large model, by coordinating different decoder structures, the model can be migrated to various downstream tasks, such as multispectral segmentation, vertical semantic segmentation, and change detection.

[0076] The pyramid structure plays a key role in this model. Operations such as Overlap Patch Embedding and Spatial-Reduction Attention are used to significantly reduce the computational complexity, making the model more suitable for processing large-scale semantic segmentation tasks. Overlap Patch Embedding divides the image into small blocks in an overlapping manner, which helps to retain more spatial information and contextual relationships. Spatial-Reduction Attention improves the efficiency of the model in processing high-resolution remote sensing images by effectively integrating and simplifying spatial information.

[0077] The feature fusion module uses the Cross-Attention mechanism to fuse the features extracted at different stages layer by layer to enhance the expressiveness of the features.

[0078] In this embodiment, the feature fusion module uses the Cross-Attention mechanism to fuse the features of each stage, including the following process:

[0079] First, perform Cross-Attention calculation on the first and second stages, and then perform Cross-Attention calculation on the fusion features obtained with the features of the third stage, and so on, considering the interaction between multiple input sources to better capture the correlation between sequences;

[0080] In the first stage, shallow features are extracted, which include low-level information of the image, such as edges and textures;

[0081] Then, the Cross-Attention mechanism is used to fuse the shallow features of the first and second stages. The Cross-Attention mechanism calculates the mutual attention between different features, highlights relevant features, and suppresses irrelevant features, thereby obtaining a richer and more meaningful feature representation.

[0082] Next, the features after fusion of the first and second stages are combined with the features of the third stage to perform another Cross-Attention calculation, and so on;

[0083] Through layer-by-layer fusion, the model comprehensively considers the feature information of the previous layer at each layer, gradually improving the abstraction level and expressiveness of feature representation.

[0084] The introduction of the Cross-Attention mechanism enables the model to fully consider the interaction between different input sources during the feature fusion process. This is particularly important for processing multi-source remote sensing data, because remote sensing images usually contain data from different sensors, different times, and different spatial resolutions. Through Cross-Attention, the model can better understand the relationship between these diverse data, thereby mining potential and valuable information to a greater extent.

[0085] The model evaluation module is used to evaluate the performance of the model in real time during training and testing. It uses a variety of evaluation indicators, such as IoU and F1-score, to comprehensively analyze the performance of the model on different data sets.

[0086] In this embodiment, in the model evaluation module, a visual performance curve is drawn based on a visualization tool to help researchers intuitively understand the training progress of the model and potential overfitting. Based on the evaluation results, the module automatically adjusts the hyperparameters of the model to optimize the model performance.

[0087] The intelligent decision-making module uses machine learning algorithms to conduct in-depth mining and intelligent inference of analysis results.

[0088] In this embodiment, the intelligent decision-making module also combines historical data and real-time analysis results, and the module automatically generates decision-making suggestions, such as land use changes and environmental monitoring;

[0089] It also integrates deep reinforcement learning strategies to adjust decision-making strategies based on feedback, allowing the system to continuously optimize and improve itself.

[0090] Furthermore, the intelligent decision-making module has a built-in environmental adaptive decision model (EADM). The model combines historical data, real-time data and deep reinforcement learning strategies to automatically generate decision suggestions based on environmental changes. The model is as follows:

[0091] ;

[0092] in,

[0093] Indicates time Decision output;

[0094] It is the real-time data feature vector at the current time point, such as the segmentation results of remote sensing images and climate data;

[0095] is a feature vector of historical data, such as past records of land use changes;

[0096] The model is in time action strategies;

[0097] like Figure 2 As shown, the model, when used, includes:

[0098] Data collection and preprocessing: collect real-time remote sensing data and historical data, and perform standardization and denoising to ensure data quality;

[0099] Feature extraction and fusion: extracting real-time data using CNN The spatial features of the data and the LSTM are used to process the historical data. The time characteristics of the two feature vectors are combined into a comprehensive feature vector ;

[0100] Decision generation: The comprehensive feature vector Input to the adaptive network , generate preliminary decisions ;

[0101] Reinforcement learning strategy update: observe the decision results, update the Q value function based on feedback information, use the greedy strategy to select the best action, and continuously adjust and optimize the action strategy ;

[0102] Decision output: Generate the final decision based on the optimized strategy and push it to the user, providing suggestions on land use changes and environmental monitoring.

[0103] The environmentally adaptive decision model combines real-time and historical data with deep reinforcement learning to form a self-optimizing decision-making mechanism. It builds a joint model of feature extraction and decision generation, which can flexibly respond to dynamic environmental changes and provide users with practical decision support.

[0104] Furthermore, when the intelligent decision-making module is used, it includes the following process:

[0105] Feature extraction: Through deep learning model and Perform feature extraction, using a combination of convolutional neural networks (CNN) and long short-term memory networks (LSTM) to extract spatial and temporal features and capture key information in complex remote sensing data;

[0106] Decision making: using an adaptive network , based on the extracted features and For decision generation, the network adopts a multi-layer perceptron (MLP) structure and improves the model's expressiveness through non-linear activation functions such as ReLU;

[0107] Reinforcement learning feedback: Through deep reinforcement learning, the results of each decision, success or failure, are fed back to the model to update the strategy , using the Q-learning algorithm, define the Q value function , to assess the status Next Action expected return.

[0108] The result push module is used to push analysis results and decision recommendations to users through multiple channels. Through the API interface, data is pushed to the user's mobile device or Web application to ensure that users can obtain information in real time.

[0109] In this embodiment, the module also implements intelligent report generation technology, which automatically generates customized reports based on user needs, including data visualization, charts, and key information summaries, to facilitate user decision-making.

[0110] The beneficial effects of the present invention are:

[0111] This remote sensing system with pyramid structure and multi-stage feature fusion has greatly improved the efficiency of feature extraction and fusion by adopting pyramid structure and multi-stage feature fusion technology, through Overlap Patch Embedding and Spatial-Reduction Attention and other methods, so that the model can efficiently process high-resolution remote sensing images. At the same time, the intelligent decision-making module is combined with deep reinforcement learning, so that the system can dynamically adjust the decision-making strategy according to real-time data and historical data, significantly improving the accuracy and practicality of the decision. The system can not only better capture the detailed information in complex scenes, but also provide users with real-time and accurate environmental monitoring and land use change recommendations, thereby effectively supporting various remote sensing applications.

[0112] 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.

[0113] 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 system with a pyramid structure and multi-stage feature fusion, characterized in that: It includes data preprocessing module, pyramid feature extraction module, semantic segmentation module, feature fusion module, model evaluation module, intelligent decision-making module and result push module; Data preprocessing module, used to standardize and clean the raw remote sensing data to ensure the accuracy and reliability of subsequent analysis; Pyramid feature extraction module, based on the pyramid structure, is used to extract multi-scale features from remote sensing images; The semantic segmentation module uses a pre-trained 600 million parameter pyramid model for efficient image segmentation; The feature fusion module uses the Cross-Attention mechanism to fuse features extracted at different stages layer by layer to enhance the expressiveness of features; The model evaluation module is used to evaluate the performance of the model in real time during training and testing. It uses a variety of evaluation indicators, such as IoU and F1-score, to comprehensively analyze the performance of the model on different data sets. Intelligent decision-making module, which uses machine learning algorithms to conduct in-depth mining and intelligent inference of analysis results; The result push module is used to push analysis results and decision recommendations to users through multiple channels. Through the API interface, data is pushed to the user's mobile device or Web application to ensure that users can obtain information in real time.

2. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: In the data preprocessing module, data preprocessing includes the steps of data denoising, geometric correction, and radiation correction; Use deep learning-based denoising techniques, such as convolutional neural network (CNN) denoisers, to retain valuable information by learning the characteristic differences between noise and valid signals; It also automatically identifies missing values ​​in the data and fills them in using interpolation algorithms to ensure data integrity.

3. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: The pyramid feature extraction module divides the image into overlapping small patches through Overlap Patch Embedding, retaining more spatial information and contextual relationships; The pyramid structure enables the model to extract features layer by layer from low-level to high-level, with each layer focusing on details at different scales; The Spatial-Reduction Attention mechanism is used to integrate spatial information to improve the processing efficiency of high-resolution remote sensing images.

4. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: The semantic segmentation module labels each pixel in the image as a specific category by performing pixel-level classification on the satellite image; The U-Net architecture is adopted, which combines the advantages of the pyramid structure, extracts features at different scales, and better handles complex objects and backgrounds.

5. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: In the semantic segmentation module, the pre-trained 600 million parameter pyramid model is constructed by the following process: First, the model is pre-trained on the open source dataset ImageNet-1k to obtain the initial 600 million parameter pyramid model parameters; Using pre-trained model parameters as a basis, we train semantic segmentation tasks on tens of millions of satellite data. The trained model features are transferred to the remote sensing scene to obtain a 600 million parameter pyramid model.

6. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: The feature fusion module uses the Cross-Attention mechanism to fuse the features of each stage. The process includes: First, perform Cross-Attention calculation on the first and second stages, and then perform Cross-Attention calculation on the fusion features obtained with the features of the third stage, and so on, considering the interaction between multiple input sources to better capture the correlation between sequences; In the first stage, shallow features are extracted, which include low-level information of the image, such as edges and textures; Then, the Cross-Attention mechanism is used to fuse the shallow features of the first and second stages. The Cross-Attention mechanism calculates the mutual attention between different features, highlights relevant features, and suppresses irrelevant features, thereby obtaining a richer and more meaningful feature representation. Next, the features after fusion of the first and second stages are combined with the features of the third stage to perform another Cross-Attention calculation, and so on; Through layer-by-layer fusion, the model comprehensively considers the feature information of the previous layer at each layer, gradually improving the abstraction level and expressiveness of feature representation.

7. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: In the model evaluation module, a visual performance curve is drawn based on a visualization tool to help researchers intuitively understand the model's training progress and potential overfitting. Based on the evaluation results, the module automatically adjusts the model's hyperparameters to optimize model performance.

8. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: The intelligent decision-making module also combines historical data and real-time analysis results, and the module automatically generates decision-making suggestions, such as land use changes and environmental monitoring; It also integrates deep reinforcement learning strategies to adjust decision-making strategies based on feedback, allowing the system to continuously optimize and improve itself.

9. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 1, characterized in that: The intelligent decision-making module has an environmental adaptive decision model (EADM) built in. The model combines historical data, real-time data, and deep reinforcement learning strategies to automatically generate decision suggestions based on environmental changes. The model is as follows: ； in, Indicates time Decision output; It is the real-time data feature vector at the current time point, such as the segmentation results of remote sensing images and climate data; is a feature vector of historical data, such as past records of land use changes; The model is in time action strategies; When used, the model includes: Data collection and preprocessing: collect real-time remote sensing data and historical data, and perform standardization and denoising to ensure data quality; Feature extraction and fusion: extracting real-time data using CNN The spatial features of the data and the LSTM are used to process the historical data. The time characteristics of the two feature vectors are combined into a comprehensive feature vector ; Decision generation: The comprehensive feature vector Input to the adaptive network , generate preliminary decisions ; Reinforcement learning strategy update: observe the decision results, update the Q value function based on feedback information, use the greedy strategy to select the best action, and continuously adjust and optimize the action strategy ; Decision output: Generate the final decision based on the optimized strategy and push it to the user, providing suggestions on land use changes and environmental monitoring.

10. A remote sensing system with a pyramid structure and multi-stage feature fusion according to claim 9, characterized in that: When the intelligent decision-making module is used, it includes the following process: Feature extraction: Through deep learning model and Perform feature extraction, using a combination of convolutional neural networks (CNN) and long short-term memory networks (LSTM) to extract spatial and temporal features and capture key information in complex remote sensing data; Decision making: using an adaptive network , based on the extracted features and For decision generation, the network adopts a multi-layer perceptron (MLP) structure and improves the model's expressiveness through non-linear activation functions such as ReLU; Reinforcement learning feedback: Through deep reinforcement learning, the results of each decision, success or failure, are fed back to the model to update the strategy , using the Q-learning algorithm, define the Q value function , to assess the status Next Action expected return.

Images