Remote sensing image sample selection method based on target scene

Abstract

The present invention discloses a method for optimizing remote sensing image samples based on a target scene. The method comprises: obtaining feature vectors of initial samples and feature vectors of target scene samples; iteratively training a value assessment model and a prediction model in a reinforcement learning decision framework using the feature vectors of the initial samples as model inputs and the feature vectors of the target scene samples as calibration data to obtain a value assessment model and a prediction model based on the target scene; using the target scene-based value assessment model to perform a value assessment on the initial samples to obtain a sample valuation; and selecting samples suitable for the target scene based on the sample valuation. This method applies reinforcement learning technology to optimize remote sensing image samples based on the target scene, achieving accurate sample optimization based on the target scene at a low computational cost.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image processing, and in particular relates to a method for optimizing remote sensing image samples based on a target scene. Background Art

[0002] In recent years, the advent of the information age has provided the remote sensing field with massive amounts of data, driving the rapid development of remote sensing technology. Machine learning technology can effectively apply this massive amount of data, bringing the remote sensing field to a new level.

[0003] It is well known that large amounts of high-quality remote sensing sample data can effectively improve the training performance of machine learning models in the remote sensing field. However, generating such large, high-quality remote sensing sample datasets is costly and challenging. Recent research has shown that, especially when training deep neural networks, different samples have different effects on network training. In some scenarios, removing a large portion of the training sample data can achieve similar or even better training results. In other scenarios, because the large amount of training sample data and test data exists in different domains, it is inevitable that there will be a mismatch in features between the training and test data. Therefore, it is particularly important to select the most relevant sample data to minimize the mismatch between the data.

[0004] Quantifying the value of remote sensing sample data can intuitively demonstrate the differences in value between different sample data. This allows for the selection of high-value samples for target scenarios based on quantified value, minimizing the impact of erroneous and noisy samples and improving sample matching for the target scenario. Furthermore, sample data with quantified value can help guide the development of more optimal data acquisition plans, providing new insights for efficiently constructing very large sample datasets. Currently, common data value quantification methods face two major challenges: poor value quantification effectiveness, such as the leave-one-out (LOO) method, which underestimates the value of nearly identical sample data; and high computational cost, such as the Data Shapley method, whose computational complexity increases exponentially with the number of sample data. Therefore, a remote sensing image sample optimization method based on novel data value quantification methods is needed to improve the efficiency and effectiveness of remote sensing sample optimization. Summary of the Invention

[0005] To solve the above-mentioned problems of the existing technology, reinforcement learning technology is used for remote sensing image sample optimization based on the target scene. By constructing a reinforcement learning decision framework and a small amount of target scene sample data, the value of the initial remote sensing sample is quantified. The initial remote sensing sample is optimized according to the quantified value, and the samples that match the target scene are selected, thereby achieving accurate sample optimization based on the target scene at a low computing cost.

[0006] To achieve the above technical objectives, the present invention provides a method for selecting remote sensing image samples based on a target scene, the method comprising the following steps:

[0007] S1: Obtain the feature vector of the initial sample and the feature vector of the target scene sample;

[0008] S2: Using the feature vector of the initial sample as model input and the feature vector of the target scene sample as calibration data, iteratively training the value assessment model and prediction model in the reinforcement learning decision framework to obtain a value assessment model and prediction model based on the target scene;

[0009] S3: Use the value assessment model based on the target scenario to perform value assessment on the initial sample to obtain a sample valuation, and select a sample suitable for the target scenario based on the sample valuation.

[0010] Optionally, step S2 includes:

[0011] S21: Grouping the feature vectors of the initial samples to obtain feature vectors of the grouped samples;

[0012] S22: Inputting the feature vectors of each grouped sample into the value assessment model in the reinforcement learning decision framework to obtain the valuation of each group of samples;

[0013] S23: Based on the probability distribution of the estimated values ​​of each group of samples, the feature vectors of the initial samples whose estimated values ​​meet the preset requirements are selected and input into the prediction model in the reinforcement learning decision framework;

[0014] S24: Using the feature vector of the target scene sample as calibration data, calculating the loss and reward through the prediction model based on the loss and reward strategy;

[0015] S25: Optimizing the prediction model based on the loss, and feeding the reward back to the value assessment model for model optimization;

[0016] S26: Iterate steps S21-S25 to a preset number of iterations to obtain a value assessment model and a prediction model based on the target scenario.

[0017] Optionally, step S24 includes:

[0018] Calculating a loss using a loss function based on an output value of the prediction model and the calibration data;

[0019] Calculate and update the moving average loss based on the said loss;

[0020] A reward is calculated based on the loss and the moving average loss, wherein the loss is used to guide a reduction in reward, and the moving average loss is used to guide an increase in reward.

[0021] Optionally, calculating and updating a moving average loss according to the loss includes:

[0022] Based on the loss, the moving average loss is calculated and updated by the following formula

[0023]

[0024] Among them, L avg is the moving average loss, T is the moving average period, vk is the calibration data, N is the number of calibration data, and Lp is the loss obtained in this calculation.

[0025] Optionally, step S3 includes:

[0026] Inputting the feature vector of the initial sample into the value assessment model based on the target scenario to obtain a sample valuation of the initial sample;

[0027] Based on the distribution of the sample estimates, binarize the estimates of the initial sample to obtain a preferred estimate containing only 0 and 1;

[0028] The sample with the preferred evaluation value of 1 is selected as the sample suitable for the target scene.

[0029] Optionally, step S1 includes:

[0030] Using a convolutional neural network to obtain a feature map of the initial sample and a feature map of the target scene sample, and converting the feature map into a feature vector form to obtain feature vectors of the initial sample and the target scene sample;

[0031] Alternatively, a preset feature extraction operator is used to perform feature extraction on the initial sample and the target scene sample to obtain feature vectors of the initial sample and the target scene sample, wherein the preset feature extraction operator includes SIFT, HOG or LBP.

[0032] The beneficial effects of the present invention are: a remote sensing image sample optimization method based on a target scene is proposed, reinforcement learning technology is used for remote sensing image sample optimization based on a target scene, the value of the initial remote sensing sample is quantified through a constructed reinforcement learning decision framework and a small amount of target scene sample data, the initial remote sensing sample is optimized according to the quantified value, and samples matching the target scene are selected, thereby achieving accurate sample optimization based on the target scene at a low computational cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 Schematic diagram of a flow chart of an embodiment of a method for optimizing remote sensing image samples based on a target scene according to the present invention;

[0035] Figure 2 A structural diagram of the reinforcement learning decision framework constructed for the present invention. DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention are within the scope of protection of the present invention.

[0037] Please refer to Figure 1 , Figure 1 The present invention is a flowchart of an embodiment of a method for optimizing remote sensing image samples based on a target scene, the method comprising the following steps:

[0038] S1: Extract feature vectors of initial samples and target scene samples.

[0039] It should be understood that the extraction of remote sensing sample features has a significant impact on the accuracy of the entire optimization method, and a feature extraction method suitable for the characteristics of remote sensing samples is required. Optionally, a convolutional neural network is used to obtain feature maps of the initial sample and feature maps of a small number of target scene samples, and the feature maps are converted into feature vectors to obtain feature vectors of the initial sample and the small number of target scene samples;

[0040] Alternatively, a preset feature extraction operator is used to perform feature extraction on the initial sample and a small number of target scene samples to obtain feature vectors of the initial sample and the small number of target scene samples, wherein the preset feature extraction operator includes SIFT, HOG or LBP.

[0041] In this embodiment, a convolutional neural network is used to extract features from the initial samples and a small number of target scene samples. In practical applications, Resnet101 can be used as the convolutional neural network model, with the ReLU function as the activation function to prevent network degradation. A batch normalization layer is added to the model to adjust the data distribution and improve the network's generalization ability.

[0042] S2: Using the feature vector of the initial sample as the model input and the feature vector of the target scene sample as the calibration data, the value assessment model and the prediction model in the reinforcement learning decision framework are iteratively trained to obtain the value assessment model and the prediction model based on the target scene.

[0043] The reinforcement learning decision framework described in this embodiment is as follows Figure 2 As shown in Figure 2, the reinforcement learning decision framework includes a sampler, a value assessment model, and a prediction model.

[0044] It should be noted that the above-mentioned value assessment model and the above-mentioned prediction model in this embodiment are both composed of neural networks. This embodiment uses a multilayer perceptron (MLP) to construct a value assessment model. The MLP contains two hidden layers and uses a sigmoid function as an activation function. The MLP outputs a valuation corresponding to each input sample, and the valuation range is 0 to 1. The valuation represents the probability of each input sample being subsequently selected. Subsequently, the valuation is binarized into 0 and 1 through a sampler, and then the sample with a valuation of 1 is selected as the preferred sample and input into the subsequent prediction model.

[0045] It should be noted that the value assessment model outputs sample valuations, not prior sample labels, so common supervised learning methods are not applicable to this scenario. This paper uses reinforcement learning methods, with a prediction model based on the target scenario, and a reward strategy to feed rewards back to the value assessment model to guide model optimization.

[0046] Specifically, step S2 includes:

[0047] S21: Grouping the feature vectors of the initial samples to obtain feature vectors of the grouped samples.

[0048] S22: Input the feature vectors of each grouped sample into the value assessment model in the reinforcement learning decision framework to obtain the estimated value of each group of samples.

[0049] S23: Based on the probability distribution of the estimated values ​​of each group of samples, the feature vectors of the initial samples whose estimated values ​​meet the preset requirements are screened and input into the prediction model in the reinforcement learning decision framework.

[0050] S24: Using the feature vector of the target scene sample as calibration data, the loss and reward are calculated by the prediction model based on the loss and reward strategy.

[0051] S25: Based on the loss, optimize the prediction model, and feed the reward back to the value assessment model for model optimization.

[0052] S26: Iterate steps S21-25 to a preset number of iterations to obtain a value assessment model and a prediction model based on the target scenario.

[0053] In this embodiment, the method of grouping the feature vectors of the initial samples may be random grouping, and the preset number of iterations may be set to 200 times.

[0054] The process of selecting the feature vectors of the initial samples whose estimated values ​​meet the preset requirements is described with a specific embodiment:

[0055] The number of initial samples can be 50,000, and the number of small samples of the target scene can be 200. Both the initial samples and the small samples of the target scene are standardized and normalized. In the actual training process, the initial samples can be divided into 1,000 groups, each with 50 samples. The feature vectors of a group of samples are input at a time and input into the value assessment model one by one to obtain the valuation of the group of samples. The valuation of the group of samples is sampled using a multi-domain distribution sampler to obtain a preferred valuation containing only 0 and 1. The preferred valuation is multiplied by the feature vector of the group of samples using a multiplier to obtain the feature vector of the preferred sample that meets the requirements and has a preferred valuation of 1.

[0056] In this embodiment, the cross entropy loss function is used to calculate the loss of the prediction model.

[0057] Specifically, step S24 includes:

[0058] Calculating a loss using a loss function based on an output value of the prediction model and the calibration data;

[0059] Calculate and update the moving average loss based on the said loss;

[0060] A reward is calculated based on the loss and the moving average loss, wherein the loss is used to guide a reduction in reward, and the moving average loss is used to guide an increase in reward.

[0061] In this embodiment, the prediction model is trained based on loss back propagation, and the moving average loss is calculated and updated based on the loss. The reward is calculated based on the loss and the moving average loss. The loss guides the reduction of the reward, and the moving average loss guides the increase of the reward. The calculation method of the moving average loss is:

[0062]

[0063] Among them, L avg is the moving average loss, T is the moving average period, vk is the calibration data, N is the number of calibration data, and Lp is the loss calculated this time. In this embodiment, T=20 and N=20 can be selected.

[0064] S3: Use the value assessment model based on the target scenario to perform value assessment on the initial sample to obtain a sample valuation, and select a sample suitable for the target scenario based on the sample valuation.

[0065] Specifically, step S3 includes:

[0066] The feature vector of the initial sample is input into the value assessment model based on the target scenario to obtain a sample valuation of the initial sample.

[0067] Based on the distribution of the sample estimates, the estimates of the initial samples are binarized and sampled to obtain preferred estimates containing only 0 and 1.

[0068] The sample with the preferred evaluation value of 1 is selected as the sample suitable for the target scene.

[0069] It should be understood that the aforementioned training process has yielded a value assessment model and prediction model based on the target scenario. Inputting initial samples into this value assessment model yields valuations for all initial samples. At this point, you can choose to continue using the sampler to binarize the valuations based on the distribution and select samples with a value of 1 as preferred samples for the target scenario. Alternatively, you can select a certain proportion of high-valuation samples based on the ranking of the initial sample valuations.

[0070] In this embodiment, samples whose final estimated value is higher than the median of all initial sample estimated values ​​are selected as preferred samples suitable for the target scenario.

[0071] Experimental verification shows that using the selected samples to interpret the target scene achieves significantly better interpretation accuracy than using all initial samples, demonstrating the effectiveness of the selected samples. The overall computation time is significantly shorter than the LOO and Data Shapley methods using the same initial samples and settings.

[0072] The beneficial effects of the present invention are: a remote sensing image sample optimization method based on a target scene is proposed, reinforcement learning technology is used for remote sensing image sample optimization based on a target scene, the value of the initial remote sensing sample is quantified through a constructed reinforcement learning decision framework and a small amount of target scene sample data, the initial remote sensing sample is optimized according to the quantified value, and samples matching the target scene are selected, thereby achieving accurate sample optimization based on the target scene at a low computational cost.

[0073] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for selecting remote sensing image samples based on a target scene, characterized in that: The following steps are involved: S1: Obtain the feature vector of the initial sample and the feature vector of the target scene sample; S2: using the feature vector of the initial sample as a model input and the feature vector of the target scene sample as calibration data to iteratively train the value assessment model and the prediction model in the reinforcement learning decision framework to obtain a value assessment model and a prediction model based on the target scene; S3: Using the value assessment model based on the target scenario to assess the value of the initial sample to obtain a sample valuation, and selecting a sample suitable for the target scenario based on the sample valuation; Step S2 includes: S21: grouping the feature vectors of the initial samples to obtain the feature vectors of the grouped samples; S22: inputting the feature vectors of each grouped sample into the value evaluation model in the reinforcement learning decision framework to obtain the valuation of each group of samples; S23: based on the probability distribution of the valuation of each group of samples, screening the feature vectors of the initial samples whose valuations meet the requirements and inputting them into the prediction model in the reinforcement learning decision framework; S24: using the feature vectors of the target scene samples as calibration data, calculating the loss and reward through the prediction model based on the loss and reward strategy; S25: optimizing the prediction model based on the loss, and feeding the reward back to the value evaluation model for model optimization; S26: iterating steps S21-S25 to a preset number of iterations to obtain a value evaluation model and a prediction model based on the target scene; Step S24 includes: calculating the loss using a loss function based on the output value of the prediction model and the calibration data; calculating and updating the moving average loss according to the loss; and calculating the reward based on the loss and the moving average loss, wherein the loss is used to guide the reduction of the reward and the moving average loss is used to guide the increase of the reward.

2. The method for selecting remote sensing image samples based on target scenes according to claim 1, characterized in that: The calculating and updating of the moving average loss according to the loss includes: Based on the loss, the moving average loss is calculated and updated by the following formula Among them, L avg is the moving average loss, T is the moving average period, v k is the calibration data, N is the number of calibration data, L p is the loss obtained in this calculation.

3. The method for selecting remote sensing image samples based on target scenes according to claim 1, characterized in that: Step S3 includes: Inputting the feature vector of the initial sample into the value assessment model based on the target scenario to obtain a sample valuation of the initial sample; Based on the distribution of the sample estimates, binarize the estimates of the initial sample to obtain a preferred estimate containing only 0 and 1; The sample with the preferred evaluation value of 1 is selected as the sample suitable for the target scene.

4. The method for selecting remote sensing image samples based on target scenes according to claim 1, characterized in that: Step S1 includes: Using a convolutional neural network to obtain a feature map of the initial sample and a feature map of the target scene sample, and converting the feature map into a feature vector form to obtain feature vectors of the initial sample and the target scene sample; Alternatively, a preset feature extraction operator is used to perform feature extraction on the initial sample and the target scene sample to obtain feature vectors of the initial sample and the target scene sample, wherein the preset feature extraction operator includes SIFT, HOG or LBP.

Images