A method for constructing high-quality remote sensing datasets based on node entropy

Abstract

In the field of intelligent remote sensing image perception, there are challenges such as difficulty in data acquisition and high annotation costs, necessitating the construction of high-quality datasets. Existing methods mostly take a single approach, resulting in limited sample selection and insufficient guidance for constructing high-quality datasets. To address this need and problem, this invention designs a high-quality remote sensing image construction method based on node entropy. This method comprehensively considers the uncertainty and diversity of samples, enabling a more comprehensive selection of high-quality samples to guide the construction of high-quality remote sensing datasets. During dataset construction, samples with high node entropy are selected, indicating greater uncertainty; connected sub-points are deleted to avoid excessive redundant sample selection, ensuring the diversity of the constructed dataset. The method has been validated on publicly available datasets. This method can be widely applied to the high-quality construction of various intelligent remote sensing image perception datasets.

Description

Technical Field

[0001] This invention relates to the field of intelligent perception of remote sensing images, and designs a method for constructing high-quality datasets of remote sensing images based on node entropy. Background Technology

[0002] With the rapid development of deep learning in computer vision, natural language processing, and other fields, the size of networks and the amount of data required to train these networks have increased significantly. [1] This has led to a significant increase in the time and hardware costs associated with sample collection and labeling, as well as model training. [2] This also limits the application and development of deep learning in areas where data acquisition is difficult. Furthermore, blindly expanding data leads to excessive dataset redundancy; training with a large number of similar samples undoubtedly provides limited gain to neural networks, but the hardware costs are substantial. Therefore, for fields like intelligent perception of remote sensing images, where data acquisition is difficult and annotation costs are high, indiscriminate expansion is not advisable. Instead, it is crucial to strengthen guidance on the sample collection process and develop high-quality dataset construction methods.

[0003] Most existing dataset construction methods rely on active learning algorithms. The core idea of ​​active learning is that the model actively selects high-information-quality samples for annotation and training, which can achieve similar or even better results than training with a large number of randomly selected samples. [3] However, most existing quality assessment methods take a single perspective and have limited sample selection, such as only considering uncertainty or diversity, without integrating multiple methods. Therefore, the improvement in model performance is often unsatisfactory. [4] .

[0004] To address the above issues, we designed a remote sensing image information quality assessment method based on node entropy. This method considers both uncertainty and diversity. Samples located at the intersection of samples from different nodes exhibit greater uncertainty, while samples from different nodes show low similarity and high diversity; such samples are evaluated as high-information-quality samples. When constructing the dataset, we select and annotate high-information-quality samples to achieve a high-quality dataset. This is then applied to two publicly available remote sensing datasets, AID. [5] and RSCN7 [6] This has been verified.

[0005] [1]L.Zhao,W.Luo,Q.Liao,S.Chen,and J.Wu,“Hyperspectral imageclassification with contras tive self-supervised learning under limitedlabeled samples,”IEEE Geoscience and Remote Sensing Letter s,vol.19,pp.1–5,2022.

[0006] [2]P.Zhang,Y.Bai,D.Wang,B.Bai,and Y.Li,“Few-shot classification ofaerial scene images via meta-learning,”Remote Sensing,vol.13,no.1,p.108,2021.

[0007] [3]I.Kalita,R.N.S.Kumar,and M.Roy,“Deep learning-based cross-sensordomain adaptation un der active learning for land cover classification,”IEEEGeoscience and Remote Sensing Letters,vol.19,pp.1–5,2021.

[0008] [4]Y.Li,J.Yang,and J.Wen,“Entropy-based redundancy analysis andinformation screening,”Digital Communications and Networks,2021.

[0009] [5]G.-S.Xia,J.Hu,F.Hu,B.Shi,X.Bai,Y.Zhong,L.Zhang,and

[0010] [6] Q.Zou, L.Ni, T.Zhang, and Q.Wang, "Deep learning based features selection for remote sensing scene classification," IEEE Geoscience and RemoteSensing Letters, vol.12, no.11, pp.2321–2325, 2015. Summary of the Invention

[0011] A high-quality remote sensing dataset construction scheme based on node entropy is proposed, focusing on the selection of node samples and the calculation of sample node entropy. The technical solution is as follows:

[0012] A method for constructing a high-quality remote sensing dataset based on node entropy includes the following steps:

[0013] 1. Extract features from labeled samples.

[0014] (1) Split the dataset.

[0015] The dataset was divided into a pool dataset and a test set in a 5:1 ratio, and 10% of the pool dataset was selected as the initial labeled dataset.

[0016] (2) Train the feature extraction network.

[0017] The feature extraction backbone network was selected as ResNet-18. The backbone network was optimized using the initial labeled dataset and the test set to obtain the feature extraction network.

[0018] (3) Obtain sample features.

[0019] The samples from the initial labeled dataset are input into a pre-trained feature extraction network, which then performs 512-dimensional feature extraction on the initial labeled data. Similarly, the candidate samples are input into the pre-trained feature extraction network, which then performs 512-dimensional feature extraction on the candidate samples.

[0020] 2. Select node samples from the initial labeled dataset.

[0021] (1) Calculate the distance between features of each sample in the initial labeled dataset.

[0022] d ij =||f (i) -f (j) ||2

[0023] Among them, f (i) f represents the features of the i-th sample in the initial labeled dataset. (j) Let d be the feature of the j-th sample in the initial labeled dataset. ij Let be the feature distance between the i-th and j-th samples.

[0024] (2) Combine the distances between samples to form a numerical ascending sequence.

[0025] D = {d1, d2, ..., d} n*n}

[0026] Where n is the number of samples in the initial labeled dataset.

[0027] (3) Use the quartile method to calculate the threshold determined by the nodes in the sample set.

[0028]

[0029] (4) Calculate the connectivity between samples in the initial labeled dataset.

[0030] For each sample, if the distance between it and other samples is less than the threshold β, then these two sub-points are considered to be connected, and the number of connections between these two samples is:

[0031]

[0032] (5) Calculate the total connectivity of each sample in the initial labeled dataset.

[0033] The number of connections for each sample is the sum of its connections to all other samples:

[0034]

[0035] (6) Select node samples from the initial labeled dataset.

[0036] The sample with the highest number of connections in the initial labeled dataset is the node sample. When selecting multiple nodes, after selecting a previous node, its connected child nodes should be deleted before selecting the next node. This avoids redundancy in multi-node selection. Finally, m initial labeled node samples are selected.

[0037] 3. Calculate the node entropy of the candidate samples.

[0038] (1) Calculate the distance from the candidate sample to each initial labeled node sample.

[0039]

[0040] Among them, f (k) For the features of the k-th sample in the dataset to be evaluated, The features of the i-th initial labeled node sample.

[0041] (2) Normalize the distance between the candidate samples and the initial labeled node samples.

[0042]

[0043] (3) Calculate the node entropy of the candidate samples.

[0044]

[0045] 4. Calculate the connectivity of the candidate samples.

[0046] (1) Calculate the connectivity between the candidate samples.

[0047] For each sample, if the distance between it and other samples is less than the threshold β, then these two sub-points are considered to be connected, and the number of connections between these two samples is:

[0048]

[0049] Where β is the threshold calculated using the quartile method when selecting node samples from the initial labeled dataset.

[0050] (2) Obtain the connection matrix of the candidate samples.

[0051] Val = Val p*p =(val(x) (i) ,x (j) )) p*p

[0052] Where p is the total number of candidate samples.

[0053] 5. Select data to complete the overall construction of the dataset.

[0054] (1) Select the sample x with the largest node entropy from the candidate sample set. (q) The data is labeled and added to the initial labeled dataset to form a new constructed dataset.

[0055] (2) With this sample x (q) Connected candidate samples are deleted to form a new candidate sample set.

[0056] (3) Repeat the above two steps, select the sample with the largest node entropy and delete the connected sub-points until the labeling budget is exhausted.

[0057] The proposed method for constructing a high-quality remote sensing dataset based on node entropy involves the following steps: First, a feature extraction network is trained using a small number of existing labeled samples to extract features from labeled samples and candidate samples. Second, using the features of various types of samples in the labeled dataset, the node entropy of each candidate sample is obtained through a node sample selection module and a node entropy calculation module; samples with higher node entropy have higher information quality. Third, samples with high node entropy are labeled and added to the dataset, and their connected sub-points are deleted. Finally, the steps of selecting samples with high node entropy and deleting connected sub-points are repeated based on the labeling and addition budget to achieve the construction of a high-quality remote sensing dataset. Attached Figure Description

[0058] Figure 1 A schematic diagram illustrating methods for constructing high-quality remote sensing datasets;

[0059] Figure 2 Node sample display diagram;

[0060] Figure 3 A comparison chart of experimental results for constructing high-quality datasets and random datasets in this invention. Detailed Implementation

[0061] This example uses a high-quality deep learning remote sensing dataset as the research object to describe in detail the implementation of the present invention.

[0062] To make the technical solution of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings. A schematic diagram of the method for constructing a high-quality remote sensing dataset is shown below. Figure 1 As shown. Details are as follows:

[0063] The first step is to extract features from the labeled samples.

[0064] (1) Split the dataset.

[0065] The dataset was divided into a pool dataset and a test set in a 5:1 ratio, and 10% of the pool dataset was selected as the initial labeled dataset.

[0066] (2) Train the feature extraction network.

[0067] The feature extraction backbone network was selected as ResNet-18. The backbone network was optimized using the initial labeled dataset and the test set to obtain the feature extraction network.

[0068] (3) Obtain sample features.

[0069] The samples from the initial labeled dataset are input into a pre-trained feature extraction network, which then performs 512-dimensional feature extraction on the initial labeled data. Similarly, the candidate samples are input into the pre-trained feature extraction network, which then performs 512-dimensional feature extraction on the candidate samples.

[0070] The second step is to select node samples from the initial labeled dataset.

[0071] (1) Calculate the distance between features of each sample in the initial labeled dataset.

[0072] d ij =||f (i) -f (j) ||2

[0073] Among them, f (i) f represents the features of the i-th sample in the initial labeled dataset. (j) Let d be the feature of the j-th sample in the initial labeled dataset. ij Let be the feature distance between the i-th and j-th samples.

[0074] (2) Combine the distances between samples to form a numerical ascending sequence.

[0075] D = {d1, d2, ..., d} n*n}

[0076] Where n is the number of samples in the initial labeled dataset.

[0077] (3) Use the quartile method to calculate the threshold determined by the nodes in the sample set.

[0078]

[0079] (4) Calculate the connectivity between samples in the initial labeled dataset.

[0080] For each sample, if the distance between it and other samples is less than the threshold β, then these two sub-points are considered to be connected, and the number of connections between these two samples is:

[0081]

[0082] (5) Calculate the total connectivity of each sample in the initial labeled dataset.

[0083] The number of connections for each sample is the sum of its connections to all other samples:

[0084]

[0085] (6) Select node samples from the initial labeled dataset.

[0086] The sample with the highest number of connections in the initial labeled dataset is the node sample. When selecting multiple nodes, after selecting a previous node, its connected child nodes should be deleted before selecting the next node. This avoids redundancy in multi-node selection. Finally, m initial labeled node samples are selected. Some sub-points and nodes in the initial labeled dataset, such as Figure 2 As shown, there are clear differences between nodes, and corresponding nodes and their sub-nodes have obvious similarities.

[0087] The third step is to calculate the entropy of the candidate sample nodes.

[0088] (1) Calculate the distance from the candidate sample to each initial labeled node sample.

[0089]

[0090] Among them, f (k) For the features of the k-th sample in the dataset to be evaluated, The features of the i-th initial labeled node sample.

[0091] (2) Normalize the distance between the candidate samples and the initial labeled node samples.

[0092]

[0093] (3) Calculate the node entropy of the candidate samples.

[0094]

[0095] The fourth step is to calculate the connectivity of the candidate samples.

[0096] (1) Calculate the connectivity between the candidate samples.

[0097] For each sample, if the distance between it and other samples is less than the threshold β, then these two sub-points are considered to be connected, and the number of connections between these two samples is:

[0098]

[0099] Where β is the threshold calculated using the quartile method when selecting node samples from the initial labeled dataset.

[0100] (2) Obtain the connection matrix of the candidate samples.

[0101] Val = Val p*p =(val(x) (i) ,x (j) )) p*p

[0102] Where p is the total number of candidate samples.

[0103] The fifth step is to select data to complete the overall construction of the dataset.

[0104] (1) Select the sample x with the largest node entropy from the candidate sample set. (q) The data is labeled and added to the initial labeled dataset to form a new constructed dataset.

[0105] (2) With this sample x (q) Connected candidate samples are deleted to form a new candidate sample set.

[0106] (3) Repeat the above two steps, select the sample with the largest node entropy and delete the connected sub-points until the labeling budget is exhausted.

[0107] The sixth step is to test the effectiveness of constructing a high-quality remote sensing dataset.

[0108] To demonstrate the effectiveness of the proposed imbalanced dataset construction, we used two publicly available remote sensing datasets, AID and RSCN7. During testing, data from the test set was input into classifier models trained on the different datasets. The corresponding class predictions were then compared with the true labels in the test set to calculate the model's prediction accuracy.

[0109] The test results of classification models trained on the high-quality remote sensing dataset constructed according to the method of this invention and on randomly constructed datasets were compared. Each dataset was trained three times, and the test result was the average of the three test results. Experimental results show that, under the same budget, the training accuracy of the high-quality remote sensing dataset model constructed according to the method of this invention is higher than that of the randomly constructed dataset model. This indicates that the high-quality remote sensing dataset constructed using the method of this invention has higher sample information quality than the randomly constructed dataset, as shown in the experimental results. Figure 3 As shown.

[0110] The above-described embodiments are merely preferred embodiments. The description of specific implementations in this paper is intended to help understand the method and core ideas of the present invention. Modifications to the technical solutions described in the foregoing embodiments or equivalent substitutions of some indicators should be included within the protection scope of the present invention.

Claims

1. A method for constructing a high-quality remote sensing image dataset based on node entropy, characterized in that, Includes the following steps: (1) Extracting features from labeled samples The remote sensing image dataset was divided into a pooling dataset and a test set in a 5:1 ratio. 10% of the pooling dataset was selected as the initial labeled dataset. The feature extraction backbone network was selected as ResNet-18. The backbone network was optimized using the initial labeled dataset and the test set to obtain the feature extraction network. The samples in the initial labeled dataset are input into the trained feature extraction network, which is then used to extract 512-dimensional features from the initial labeled data. The candidate samples are input into a trained feature extraction network, which is then used to extract 512-dimensional features from the candidate samples. (2) Select node samples from the initial labeled dataset First, calculate the distance between the features of each sample in the initial labeled dataset. ,in, The features of the i-th sample in the initial labeled dataset, For the features of the j-th sample in the initial labeled dataset, The feature distance between the i-th and j-th samples; Secondly, the distances between the samples are combined to form a numerical ascending sequence. Where n is the number of samples in the initial labeled dataset; the threshold determined by the nodes in the sample set is calculated using the quartile method. , Take the [th]th element in sequence D The value of the ] bit; Next, calculate the connectivity between samples in the initial labeled dataset. For each sample, the distance to other samples must be less than a threshold. If these two sub-points are considered to be connected, then the number of connections between these two samples is: Next, calculate the total connectivity of each sample in the initial labeled dataset, and the number of connections of each sample itself, which is the sum of its connections with all other samples: Finally, node samples from the initial labeled dataset are selected; the sample with the highest number of connections in the initial labeled dataset is chosen as the node sample. When selecting multiple nodes, after selecting a previous node, its connected child nodes are deleted before selecting the next node. This avoids redundancy in multi-node selection, ultimately selecting m initial labeled node samples. ; (3) Calculate the entropy of the candidate sample nodes Calculate the distance from the candidate sample to each initial labeled node sample. ,in, For the features of the k-th sample in the dataset to be evaluated, The features of the i-th initially labeled node sample; The distance between the candidate samples and the initial labeled node samples is normalized: Calculate the node entropy of the candidate samples: (4) Calculate the connectivity of candidate samples Calculate the connectivity between candidate samples. For each sample, the distance to other samples is less than a threshold. If these two sub-points are considered to be connected, then the connection number between these two samples is: Where β is the threshold calculated using the quartile method when selecting node samples from the initial labeled dataset; subsequently, the connection matrix of the candidate samples is obtained. Where p is the total number of candidate samples; (5) Select data to complete the overall construction of the dataset. (5.1) Select the sample with the largest node entropy from the candidate sample set. The data is labeled and added to the initial labeled dataset to form a new constructed dataset; (5.2) will be compared with this sample Interconnected candidate samples are deleted, forming a new candidate sample set; (5.3) Repeat the above two steps, select the sample with the largest node entropy and delete the connected sub-points until the labeling budget is exhausted.

Images