Remote sensing feature extraction deep reinforcement learning method based on multi-modal knowledge expression

Abstract

The application relates to the technical field of artificial intelligence, and discloses a remote sensing feature extraction deep reinforcement learning method based on multi-modal knowledge expression. The method comprises the following steps: constructing a multi-level feature extraction model library; performing semantic complexity evaluation on input optical, SAR and infrared multi-modal remote sensing data; generating a state vector in combination with an edge device real-time running state and a task priority; dynamically deciding an optimal model and a mode combination through a pre-trained deep reinforcement learning strategy model; and performing adaptive feature extraction. Through dynamic resource allocation, the application can reduce average reasoning delay and energy consumption while guaranteeing high-precision analysis of key areas, and can meet high-timeliness application requirements such as disaster monitoring.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation. Background Technology

[0002] As a core component of Earth observation systems, remote sensing intelligent interpretation has played an increasingly crucial role in fields such as disaster emergency response, environmental monitoring, and land planning in recent years.

[0003] With the widespread application of multi-source sensors such as high-resolution optics, synthetic aperture radar (SAR), and infrared, remote sensing data exhibits multimodal characteristics. Deep reinforcement learning (DRL), due to its adaptive capabilities in complex decision-making tasks, has been introduced into the field of remote sensing feature extraction, enabling dynamic perception and discrimination through the interaction between intelligent agents and the environment.

[0004] However, such methods typically rely on large-scale neural networks to fuse heterogeneous modal information and generate robust policies, resulting in high computational complexity and a large number of parameters in the model.

[0005] Deep reinforcement learning methods for remote sensing feature extraction based on multimodal knowledge representation aim to improve the accuracy of target recognition and change detection in disaster scenarios by jointly modeling multidimensional features such as spectral, texture, geometry, and semantics. These methods typically construct end-to-end policy networks, utilize attention mechanisms or graph neural networks to integrate cross-modal context, and optimize feature selection paths through reward signals. While they have achieved some breakthroughs in accuracy, the model inference process is highly dependent on high-performance computing resources, posing significant challenges when deployed on edge computing devices.

[0006] In existing technologies, multimodal DRL models generally adopt fully connected or densely fused architectures without structural optimization for edge scenarios, leading to increased inference latency. Even when conventional model compression techniques are introduced, the differences in the contribution of different feature paths in the policy network to the final decision are often ignored, resulting in the loss of key remote sensing semantic information and reducing the reliability of disaster identification. Furthermore, static compression strategies cannot adapt to the dynamic changes in target scale, lighting conditions, and disaster types in remote sensing scenarios, making it difficult to achieve a balance between latency constraints and feature fidelity.

[0007] Therefore, achieving low-latency, high-fidelity multimodal remote sensing feature extraction on resource-constrained edge devices has become an urgent technical challenge. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation. This method aims to overcome the inference latency problem caused by limited computing resources on edge devices when processing complex remote sensing tasks. It achieves a dynamic balance between model computational complexity and real-time data processing requirements by constructing a dynamically adaptive inference framework.

[0009] The core technical solution of this invention lies in pre-constructing a feature extraction model library containing multiple models with different computational complexities through a hierarchical knowledge distillation mechanism;

[0010] During the real-time inference phase, the input multimodal remote sensing data is first rapidly analyzed through a lightweight semantic complexity assessment module to generate a data complexity index.

[0011] Simultaneously, the real-time operating status of edge devices is collected;

[0012] Then, the data complexity index and the device operating status are integrated into a state vector, which is input into a dynamic reasoning strategy model trained based on deep reinforcement learning. The strategy model decides to select the most suitable feature extraction model and the combination of data modalities to be activated from the model library based on the current state.

[0013] Finally, the selected model is invoked to process the selected data modality, completing the feature extraction task.

[0014] This method transforms the static, single model inference process into a dynamic, context-adaptive closed-loop control process. While ensuring high-precision analysis capabilities in critical areas, it greatly reduces computational overhead in non-critical or simple scenarios, thereby reducing the average inference latency overall and meeting the needs of time-sensitive applications such as disaster monitoring.

[0015] This invention provides a deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation, which includes the following steps:

[0016] A multi-level feature extraction model library is constructed, the model library includes at least one first feature extraction model, a second feature extraction model and a third feature extraction model, wherein the computational complexity of the first feature extraction model is greater than that of the second feature extraction model, and the computational complexity of the second feature extraction model is greater than that of the third feature extraction model.

[0017] Acquire multimodal data of the remote sensing image area to be processed, including optical remote sensing data, synthetic aperture radar data, and infrared remote sensing data;

[0018] The multimodal data is subjected to real-time semantic complexity evaluation, and a data complexity scalar value is generated;

[0019] Acquire real-time operating status data of the edge device where the method is deployed, including CPU utilization, graphics processing unit utilization, and memory usage.

[0020] The data complexity scalar value, the real-time running status data, and the preset task priority value are combined together to form a state vector;

[0021] The state vector is input into a pre-trained dynamic inference policy model, which outputs an inference decision instruction containing a model selection index and a modality selection mask.

[0022] Based on the model selection index, a target feature extraction model is selected from the multi-level feature extraction model library;

[0023] Based on the modality selection mask, one or more data modalities are selected from the multimodal data to form a subset of input data;

[0024] The input data subset is fed into the target feature extraction model, and forward inference calculation is performed to generate the final feature representation of the remote sensing image region to be processed.

[0025] As one embodiment of the present invention, the step of constructing a multi-level feature extraction model library specifically includes:

[0026] First, train a teacher model with a complex structure and a large number of parameters on a cloud server or high-performance computing platform until it reaches the preset performance index on a standard remote sensing dataset.

[0027] Next, a hierarchical knowledge distillation method is used to carry out multi-stage knowledge transfer;

[0028] In the first stage, the knowledge of the teacher model is distilled into the first feature extraction model. The first feature extraction model adopts a deep residual network structure. The loss function of the knowledge distillation is composed of a weighted sum of a Kolb-Leibler divergence loss term based on soft labels and a mean squared error loss term based on intermediate layer feature maps.

[0029] In the second stage, the first feature extraction model is used as a new teacher model, and its knowledge is distilled into the second feature extraction model, which adopts a depthwise separable convolutional network structure.

[0030] In the third stage, the second feature extraction model is used as a new teacher model, and its knowledge is distilled into the third feature extraction model. The third feature extraction model adopts an inverted residual and linear bottleneck network structure, thereby completing the construction of the multi-level feature extraction model library.

[0031] As one embodiment of the present invention, the step of performing real-time semantic complexity evaluation on the multimodal data specifically includes:

[0032] The input multimodal data is downsampled using bilinear interpolation in the spatial dimension to reduce data resolution;

[0033] The downsampled multimodal data is input into a shallow convolutional neural network consisting of three convolutional layers and a global average pooling layer;

[0034] The shallow convolutional neural network calculates the mean spatial gradient magnitude and information entropy of the downsampled data, respectively.

[0035] The mean value of the spatial gradient magnitude is linearly weighted and summed with the information entropy. The sum is the scalar value of the data complexity. The range of the scalar value is normalized to a closed interval of 0 to 1.

[0036] As one embodiment of the present invention, the dynamic reasoning strategy model is a deep reinforcement learning policy network obtained by offline training based on the near-end policy optimization algorithm;

[0037] The input to the policy network is the state vector, which has a dimension of 5 and corresponds to the scalar value of data complexity, CPU utilization, graphics processing unit utilization, memory utilization, and task priority value, respectively. All input values ​​have been processed by min-max normalization.

[0038] The policy network is structured as a multilayer perceptron containing three fully connected hidden layers, with each hidden layer followed by a modified linear unit activation function.

[0039] The output layer of the policy network is a classification distribution, which outputs a probability distribution for all predefined combined actions consisting of model selection index and modality selection mask. The inference decision instruction is the action with the highest probability sampled from this distribution.

[0040] In one embodiment of the present invention, the process of training the dynamic reasoning strategy model is carried out in a high-fidelity edge device simulation environment, and its reward function is designed as follows:

[0041] The reward value equals the positive accuracy gain term minus the negative latency penalty term and resource consumption penalty term;

[0042] The accuracy gain term is defined as the difference between the classification accuracy obtained by the model selected in the current decision on the benchmark validation set and the accuracy obtained by always selecting the model with the lowest complexity.

[0043] The delay penalty term is proportional to the inference time of the selected model, and its scaling factor is a hyperparameter.

[0044] The resource consumption penalty term is the weighted sum of the peak memory usage generated by the selected model during inference and the average utilization rate of the central processing unit and graphics processing unit, and its weight coefficients are a set of hyperparameters.

[0045] By maximizing the expected cumulative reward, the dynamic inference strategy model learns a decision-making strategy that balances accuracy, latency, and resource consumption.

[0046] This invention also provides a deep reinforcement learning system for remote sensing feature extraction based on multimodal knowledge representation, comprising:

[0047] A multi-level feature extraction model library is stored in the non-volatile memory of the edge device. The model library includes at least one first feature extraction model, a second feature extraction model, and a third feature extraction model, with their computational complexity decreasing sequentially.

[0048] The multimodal data acquisition module is used to capture optical remote sensing data, synthetic aperture radar data and infrared remote sensing data of the remote sensing image area to be processed in real time through airborne sensor interface or wireless communication link;

[0049] The semantic complexity evaluation module, which is implemented in hardware as a dedicated digital signal processor, is used to quickly analyze the multimodal data and generate a data complexity scalar value.

[0050] The device status monitoring module is used to periodically query and report the CPU utilization, graphics processing unit utilization, and memory usage of edge devices by calling the operating system kernel interface;

[0051] The state vector generation module connects the semantic complexity evaluation module and the device status monitoring module, and is used to integrate the data complexity scalar value, the real-time operating status data of the device, and the task priority value received from the task management module into a standardized state vector.

[0052] The dynamic inference strategy module has a built-in pre-trained deep reinforcement learning strategy network, which receives the state vector as input and outputs inference decision instructions containing model selection index and modality selection mask.

[0053] The inference execution control module loads the target feature extraction model from the multi-level feature extraction model library into the computing memory according to the model selection index in the inference decision instruction, and configures the data flow channel according to the modality selection mask to transmit the filtered input data subset to the target feature extraction model.

[0054] The feature extraction calculation module, which is an embedded neural network processing unit of the edge device, is used to perform forward inference calculation of the target feature extraction model and output the final feature expression.

[0055] In one embodiment of the present invention, the system further includes a model library construction device, which is deployed on a cloud server and includes:

[0056] Teacher model training unit, used to train high-performance teacher models on large-scale remote sensing datasets;

[0057] The hierarchical knowledge distillation unit contains a loss function calculation module and a model parameter update module, which are used to perform a three-stage hierarchical knowledge distillation process from the teacher model to the first feature extraction model, from the first feature extraction model to the second feature extraction model, and from the second feature extraction model to the third feature extraction model.

[0058] After the hierarchical knowledge distillation process is completed, the model library construction device packages the three trained feature extraction models and sends them to the edge device through a secure transmission protocol to complete the deployment or update of the multi-level feature extraction model library.

[0059] In one embodiment of the present invention, the shallow convolutional neural network in the semantic complexity evaluation module has convolutional kernel sizes of 3×3, 3×3 and 1×1 for its three convolutional layers, and channel numbers of 16, 32 and 8 respectively. The stride of all convolutional operations is 1, and the padding method is zero padding. The global average pooling layer calculates the mean of the feature map output by the last convolutional layer in the spatial dimension to generate an eight-dimensional feature vector. Finally, the eight-dimensional feature vector is mapped to the two output values ​​of the mean spatial gradient magnitude and the information entropy through a fully connected layer.

[0060] In one embodiment of the present invention, the policy network in the dynamic reasoning policy module has 256, 128 and 64 neurons in its three fully connected hidden layers, respectively; the number of nodes in the output layer is equal to the product of the number of models in the multi-level feature extraction model library and the number of all possible modal combinations of the multimodal data, and each node corresponds to a unique reasoning decision action.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] 1. Dynamic adaptive allocation of computing resources is achieved. This invention no longer uses static and unchanging deep learning models for inference. Instead, it dynamically selects the most suitable computing path based on the inherent complexity of the input data and the operating status of the device itself, and accurately invests limited computing resources in the most complex scenarios that need to be processed, thereby achieving optimal resource allocation at the system level.

[0063] 2. Reduced average inference latency: By enabling low-power, high-speed lightweight models in a large number of simple or moderately complex remote sensing scenarios, this invention effectively avoids the unnecessary time overhead caused by continuously running heavyweight models, which greatly shortens the average response time of the system and meets the stringent real-time requirements such as rapid disaster response.

[0064] 3. The system energy efficiency is improved without sacrificing the performance of critical tasks. The present invention can reduce the overall computing load through intelligent decision-making while ensuring the ability to extract high-precision features from complex and high-risk target areas. This directly translates into lower power consumption. For edge devices that rely on battery power, this means longer battery life and task execution cycle.

[0065] 4. This invention provides a systematic, end-to-end solution. It not only focuses on the model algorithm itself, but also incorporates multiple dimensions of information such as data characteristics, hardware status, and task priority into a unified decision-making framework, forming a complete closed-loop adaptive inference system. Its robustness and applicability in real and complex environments far exceed those of single model compression or hardware acceleration technologies. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the overall technical architecture of the deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation proposed in this invention.

[0067] Figure 2 This is a schematic diagram of the core principle framework of the dynamic reasoning strategy model in this invention;

[0068] Figure 3 This is a logical flowchart of the construction of a multi-level feature extraction model library and hierarchical knowledge distillation in this invention;

[0069] Figure 4 This is a flowchart illustrating the logical process of multimodal data semantic complexity evaluation and state vector generation in this invention.

[0070] Figure 5 This is a logical flow diagram of dynamic inference execution and feature extraction calculation at the edge device in this invention;

[0071] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between edge devices and cloud servers in the construction and deployment of the model library in this invention. Detailed Implementation

[0072] Please refer to Figures 1 to 6This invention provides a deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation, aiming to solve the technical problem that edge devices have excessively high inference latency when processing multimodal remote sensing data under resource-constrained conditions, and cannot meet the timeliness requirements of tasks such as disaster monitoring.

[0073] This method constructs a feature extraction model library containing multiple models with different computational complexities, and combines a lightweight semantic complexity evaluation module with a device operation status awareness mechanism. It utilizes a pre-trained dynamic inference strategy model to make real-time decisions on the optimal model selection and modality combination, thereby reducing the average inference latency and resource consumption of the system while ensuring the accuracy of feature extraction in key areas.

[0074] The method includes the following steps:

[0075] Construct a multi-level feature extraction model library; acquire multimodal data of the remote sensing image region to be processed;

[0076] The multimodal data is subjected to real-time semantic complexity evaluation, and a data complexity scalar value is generated;

[0077] Obtain real-time operational status data of the edge device where the method is deployed; combine the data complexity scalar value, the real-time operational status data, and the preset task priority value into a state vector;

[0078] The state vector is input into a pre-trained dynamic inference policy model to obtain inference decision instructions containing model selection index and modality selection mask;

[0079] Select a target feature extraction model from the multi-level feature extraction model library according to the model selection index;

[0080] Based on the modality selection mask, one or more data modalities are selected from the multimodal data to form a subset of input data;

[0081] The subset of input data is fed into the target feature extraction model to perform forward inference calculations, generating the final feature representation.

[0082] First, a multi-level feature extraction model library is constructed. This model library contains at least one first feature extraction model, a second feature extraction model, and a third feature extraction model, wherein the computational complexity of the first feature extraction model is greater than that of the second feature extraction model, and the computational complexity of the second feature extraction model is greater than that of the third feature extraction model.

[0083] The model library is built on a cloud server or high-performance computing platform, and includes the following sub-steps:

[0084] A teacher model with a complex structure and a large number of parameters is trained on a large-scale remote sensing dataset until it reaches the preset performance index on a standard remote sensing dataset. This teacher model serves as the source of knowledge distillation.

[0085] A three-stage knowledge transfer method is adopted, using a hierarchical knowledge distillation approach.

[0086] In the first stage, the knowledge of the teacher model is distilled into the first feature extraction model, which adopts a deep residual network structure. Its loss function is composed of a weighted loss term based on soft labels, Kohlbek-Leibler divergence, and a loss term based on intermediate layer feature maps, with weight coefficients of 0.7 and 0.3, respectively.

[0087] In the second stage, the first feature extraction model is used as a new teacher model, and its knowledge is distilled into the second feature extraction model. This model adopts a deep separable convolutional network structure, and its loss function also includes a soft label KL divergence term and an intermediate feature map MSE term, with the weight coefficients adjusted to 0.6 and 0.4.

[0088] In the third stage, the second feature extraction model is used as a new teacher model, and its knowledge is distilled into the third feature extraction model. This model adopts an inverted residual and linear bottleneck network structure. The weight of the soft label KL divergence term in its loss function is 0.5, and the weight of the intermediate feature map MSE term is 0.5.

[0089] After the above three-stage distillation, the three student models reduce the number of parameters and computational load while maintaining high accuracy. Finally, they are packaged and sent to the non-volatile memory of the edge device through a secure transmission protocol, forming a multi-level feature extraction model library for solidified deployment.

[0090] Next, multimodal data of the remote sensing image area to be processed is acquired. The multimodal data includes optical remote sensing data, synthetic aperture radar data, and infrared remote sensing data.

[0091] Optical remote sensing data is acquired through visible light and near-infrared band sensors, with a spatial resolution of 0.5 meters to 2 meters, including four channels: red, green, blue, and near-infrared.

[0092] Synthetic aperture radar data is acquired through C-band or X-band radar, and amplitude and phase information are recorded in complex form. After multi-view processing, a single-channel intensity image is generated.

[0093] Infrared remote sensing data is acquired through thermal infrared sensors, which record the surface radiation temperature, with a spatial resolution typically ranging from 10 to 30 meters.

[0094] The three types of data were collected synchronously in time and aligned at the pixel level in space through georegistration, forming a three-dimensional multimodal tensor with a size of [size missing]. ,in and These are the image height and width, respectively. The total number of channels is 4 plus 1 plus 1, which is 6.

[0095] Next, the semantic complexity of the multimodal data is evaluated in real time, and a data complexity scalar value is generated.

[0096] The evaluation process is performed by a dedicated shallow convolutional neural network, which is hardware-implemented in a digital signal processor and features low latency and low power consumption.

[0097] The specific process is as follows:

[0098] First, the input multimodal data is downsampled by bilinear interpolation in the spatial dimension to reduce the original resolution to 64×64 pixels, thereby reducing the subsequent computational load.

[0099] The downsampled data is then input into a shallow convolutional neural network consisting of three convolutional layers and a global average pooling layer.

[0100] The first convolutional layer uses a 3×3 convolutional kernel with 16 channels, a stride of 1, and zero padding.

[0101] The second convolutional layer also uses a 3×3 convolutional kernel, with 32 channels, a stride of 1, and zero padding.

[0102] The third convolutional layer uses a 1×1 convolutional kernel with 8 channels, a stride of 1, and is padded with zeros.

[0103] Each convolutional layer is followed by batch normalization and modified linear unit activation functions.

[0104] Subsequently, the global average pooling layer calculates the mean of the eight-channel feature map output by the last convolutional layer in the spatial dimension, generating an eight-dimensional feature vector.

[0105] This eight-dimensional vector is mapped to two output values ​​through a fully connected layer:

[0106] Spatial gradient magnitude mean and information entropy.

[0107] The mean spatial gradient magnitude is obtained by averaging the Sobel operator gradient magnitudes on the downsampled data across the entire image, reflecting the complexity of the image texture.

[0108] Information entropy is obtained by calculating the Shannon entropy from the histogram of pixel values ​​in each channel and then averaging the channels, reflecting the richness of information.

[0109] Finally, the mean magnitude of the spatial gradient and the information entropy are linearly weighted and summed with weight coefficients of 0.6 and 0.4 respectively. The sum is the scalar value of the data complexity, and it is constrained to a closed interval between 0 and 1 by max-min normalization.

[0110] Simultaneously, real-time operational status data of edge devices is acquired. This data is periodically collected by the device status monitoring module, with a sampling period of 100 milliseconds.

[0111] Specifically, this includes CPU utilization, graphics processing unit utilization, and memory usage.

[0112] The CPU utilization rate is obtained by reading the operating system kernel's / proc / stat interface, calculating the percentage of idle time, and then working backwards to determine the CPU utilization rate.

[0113] The graphics processing unit utilization rate is obtained by calling the performance counter interface provided by the device driver.

[0114] Memory usage can be obtained by querying the ratio of used memory to total memory in the / proc / meminfo interface.

[0115] All three indicators are expressed as percentages and have undergone max-min normalization to map to the 0-1 range.

[0116] The data complexity scalar value, the three device operating status data, and the preset task priority value are then combined into a five-dimensional state vector.

[0117] The task priority value is set by the task management module according to the current task type. For example, disaster emergency response tasks are set to 1, routine inspection tasks are set to 0.3, and low-priority tasks at night are set to 0.1.

[0118] This value is also normalized to the range of 0 to 1.

[0119] The five dimensions of the state vector are as follows:

[0120] Data complexity scalar value, CPU utilization, graphics processing unit utilization, memory utilization, and task priority value.

[0121] The five-dimensional state vector is input into a pre-trained dynamic inference strategy model.

[0122] This model is a deep reinforcement learning policy network trained offline based on a proximal policy optimization algorithm.

[0123] Its network structure is a multilayer perceptron containing three fully connected hidden layers. The first hidden layer has 256 neurons, the second hidden layer has 128 neurons, and the third hidden layer has 64 neurons. Each layer is followed by a modified linear unit activation function.

[0124] The number of output layer nodes is equal to the product of the number of models in the model library and the number of all possible modal combinations of the multimodal data.

[0125] In this embodiment, the number of models is 3, and the modal combinations include optical only, SAR only, infrared only, optical + SAR, optical + infrared, SAR + infrared, and optical + SAR + infrared, for a total of 7 combinations. Therefore, the number of output layer nodes is 21.

[0126] Each node corresponds to a unique inference decision action, which is a specific combination of model selection index and modality selection mask.

[0127] The policy network outputs a 21-dimensional probability distribution, and the inference decision instruction is the model index and modal mask corresponding to the action with the highest probability in this distribution.

[0128] The training process of the dynamic reasoning strategy model is carried out in a high-fidelity edge device simulation environment.

[0129] The simulation environment accurately simulates the computational latency, memory bandwidth, and power consumption characteristics of real edge devices. The reward function is designed as follows:

[0130] The reward value equals the accuracy gain term minus the latency penalty term and the resource consumption penalty term.

[0131] The accuracy gain term is defined as the difference between the classification accuracy obtained by the model selected in the current decision on the benchmark validation set and the accuracy obtained by always selecting the third feature extraction model;

[0132] The delay penalty term is equal to the measured inference time of the selected model under the current device conditions multiplied by the hyperparameters. , Set it to 0.01;

[0133] The resource consumption penalty term is the weighted sum of the peak memory usage (after normalization) during the inference process of the selected model and the average utilization rate of the central processing unit and graphics processing unit, with weight coefficients of 0.4, 0.3 and 0.1, respectively.

[0134] By maximizing the expected cumulative reward, the policy network learns how to balance accuracy, latency, and resource consumption under different state combinations, thereby generating the optimal decision.

[0135] After receiving the inference decision instruction, the corresponding target feature extraction model is loaded into the computing memory from the multi-level feature extraction model library according to the model selection index.

[0136] If the first feature extraction model is selected, the parameters of the deep residual network are loaded; if the second feature extraction model is selected, the parameters of the deep separable convolutional network are loaded; if the third feature extraction model is selected, the parameters of the inverted residual and linear bottleneck network are loaded.

[0137] Simultaneously, the data stream channels are configured based on the mode selection mask. The mode selection mask is a three-bit binary vector, corresponding to the optical, SAR, and infrared modes respectively. A value of 1 enables the mode, and a value of 0 disables it.

[0138] For example, if the mask is 110, only optical and SAR data will be enabled, forming a subset of input data with 4 plus 1 equals 5 channels.

[0139] Finally, the subset of input data is fed into the target feature extraction model to perform forward inference computation.

[0140] The target feature extraction model runs on an embedded neural network processing unit on an edge device, which supports INT8 quantization inference to further accelerate computation.

[0141] The forward inference process includes standard operations such as convolution, activation, and pooling, ultimately outputting a fixed-dimensional feature vector, for example, 1280 dimensions, which serves as the final feature representation of the remote sensing image region to be processed. This feature representation can be used for subsequent high-level tasks such as object detection, change detection, or semantic segmentation.

[0142] At the system level, the present invention also provides a corresponding deep reinforcement learning system.

[0143] The system includes a multi-level feature extraction model library, a multimodal data acquisition module, a semantic complexity evaluation module, a device status monitoring module, a state vector generation module, a dynamic inference strategy module, an inference execution control module, and a feature extraction calculation module.

[0144] The multi-level feature extraction model library is permanently stored in the flash memory of the edge device;

[0145] The multimodal data acquisition module receives optical, SAR, and infrared data in real time via an airborne sensor interface or wireless communication link;

[0146] The semantic complexity evaluation module is implemented using a dedicated digital signal processor to ensure that the evaluation latency is less than 10 milliseconds.

[0147] The device status monitoring module periodically collects resource usage data through the operating system kernel interface; the status vector generation module standardizes the data from various sources and concatenates them into a five-dimensional vector.

[0148] The dynamic inference strategy module incorporates the aforementioned pre-trained strategy network, keeping inference latency within 5 milliseconds.

[0149] The inference execution control module is responsible for model loading and data routing;

[0150] The feature extraction computation module is a neural network accelerator that supports parallel computing and memory-optimized scheduling.

[0151] The entire system forms a closed-loop adaptive inference architecture, which can complete the entire process from data input to feature output in milliseconds, meeting the stringent real-time requirements of disaster monitoring scenarios.

[0152] In summary, this embodiment effectively solves the real-time challenge of multimodal remote sensing feature extraction in resource-constrained environments by constructing a hierarchical distillation model library, designing a lightweight complexity evaluation mechanism, introducing device state awareness and task priority fusion, training a PPO-based dynamic policy network, and implementing a complete inference control closed loop at the edge. This method reduces average inference latency and memory usage without sacrificing the accuracy of critical tasks, thus enhancing the practicality and deployment value of edge intelligent systems.

[0153] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish an entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0154] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation, characterized in that, include: A multi-level feature extraction model library is constructed, the model library includes at least one first feature extraction model, a second feature extraction model and a third feature extraction model, wherein the computational complexity of the first feature extraction model is greater than that of the second feature extraction model, and the computational complexity of the second feature extraction model is greater than that of the third feature extraction model. Acquire multimodal data of the remote sensing image area to be processed, including optical remote sensing data, synthetic aperture radar data, and infrared remote sensing data; The multimodal data is subjected to real-time semantic complexity evaluation, and a data complexity scalar value is generated; Acquire real-time operating status data of the edge device where the method is deployed, including CPU utilization, graphics processing unit utilization, and memory usage. The data complexity scalar value, the real-time running status data, and the preset task priority value are combined together to form a state vector. The state vector is input into a pre-trained dynamic inference policy model, which outputs an inference decision instruction containing a model selection index and a modality selection mask. Based on the model selection index, a target feature extraction model is selected from the multi-level feature extraction model library; Based on the modality selection mask, one or more data modalities are selected from the multimodal data to form a subset of input data; The subset of input data is fed into the target feature extraction model, and forward inference calculation is performed to generate the final feature representation of the remote sensing image region to be processed. Constructing a multi-level feature extraction model library includes: training a complex, parameter-intensive teacher model on a cloud server or high-performance computing platform until it reaches a preset performance index on a standard remote sensing dataset; employing a hierarchical knowledge distillation method for three-stage knowledge transfer; in the first stage, distilling the knowledge of the teacher model into the first feature extraction model, which uses a deep residual network structure, and the loss function of the knowledge distillation is a weighted sum of a soft-label-based Kolb-Leibler divergence loss term and a mean squared error loss term based on intermediate layer feature maps; in the second stage, using the first feature extraction model as a new teacher model, distilling its knowledge into the second feature extraction model, which uses a deep separable convolutional network structure; and in the third stage, using the second feature extraction model as a new teacher model, distilling its knowledge into the third feature extraction model, which uses an inverted residual and linear bottleneck network structure, thereby completing the construction of the multi-level feature extraction model library. Real-time semantic complexity evaluation of the multimodal data to generate a data complexity scalar value includes: performing bilinear interpolation downsampling on the input multimodal data in the spatial dimension to reduce data resolution; inputting the downsampled multimodal data into a shallow convolutional neural network consisting of three convolutional layers and a global average pooling layer; the shallow convolutional neural network calculating the mean spatial gradient magnitude and information entropy of the downsampled data respectively; and linearly weighting the mean spatial gradient magnitude and the information entropy to obtain the sum, which is the data complexity scalar value, and the range of the scalar value is normalized to a closed interval of 0 to 1. The dynamic inference policy model is a deep reinforcement learning policy network trained offline based on a near-end policy optimization algorithm. The input of the policy network is the state vector, which has a dimension of 5 and corresponds to the scalar value of data complexity, CPU utilization, GPU utilization, memory utilization, and task priority. All input values ​​have undergone min-max normalization. The structure of the policy network is a multilayer perceptron containing three fully connected hidden layers, with each hidden layer followed by a modified linear unit activation function. The output layer of the policy network is a classification distribution, which outputs a probability distribution for all predefined combinations of actions consisting of a model selection index and a modality selection mask. The inference decision instruction is the action with the highest probability sampled from this distribution.

2. The deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation according to claim 1, characterized in that, The convolutional kernel sizes of the three convolutional layers of the shallow convolutional neural network are 3×3, 3×3 and 1×1, respectively, and the number of channels are 16, 32 and 8, respectively. The stride of all convolutional operations is 1, and the padding method is zero padding. The global average pooling layer calculates the average value of the feature map output by the last convolutional layer in the spatial dimension, generating an eight-dimensional feature vector. The eight-dimensional feature vector is mapped to two output values, the mean of the spatial gradient magnitude and the information entropy, through a fully connected layer.

3. The deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation according to claim 2, characterized in that, The number of neurons in the three fully connected hidden layers are 256, 128, and 64, respectively; the number of nodes in the output layer is equal to the product of the number of models in the multi-level feature extraction model library and the number of all modal combinations in the multimodal data, and each node corresponds to a unique inference decision action.

4. The deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation according to claim 3, characterized in that, The training of the dynamic inference strategy model is carried out in a high-fidelity edge device simulation environment, and its reward function is designed as follows: The reward value equals the positive accuracy gain term minus the negative latency penalty term and resource consumption penalty term; The accuracy gain term is defined as the difference between the classification accuracy obtained by the model selected in the current decision on the benchmark validation set and the accuracy obtained by always selecting the model with the lowest complexity. The delay penalty term is proportional to the inference time of the selected model, and its scaling factor is a hyperparameter. The resource consumption penalty term is the weighted sum of the peak memory usage generated by the selected model during inference and the average utilization rate of the central processing unit and graphics processing unit, and its weight coefficients are a set of hyperparameters. By maximizing the expected cumulative reward, the dynamic inference strategy model learns a decision-making strategy that balances accuracy, latency, and resource consumption.

5. The deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation according to claim 4, characterized in that, The mode selection mask is a three-bit binary vector, corresponding to optical remote sensing data, synthetic aperture radar data, and infrared remote sensing data, respectively. A value of one enables the corresponding mode, and a value of 0 disables the corresponding mode.

6. The deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation according to claim 5, characterized in that, The multimodal data is spatially aligned at the pixel level through georegistration, forming a three-dimensional multimodal tensor with a size of [size missing]. ,in and These are the image height and width, respectively. The total number of channels is 6.

7. The deep reinforcement learning method for remote sensing feature extraction based on multimodal knowledge representation according to claim 6, characterized in that, The target feature extraction model runs on the embedded neural network processing unit of the edge device and supports INT8 quantization inference to accelerate computation; the forward inference computation outputs a fixed-dimensional feature vector as the final feature expression.

Images