Construction method of semi-supervised remote sensing image change detection model

By optimizing the semi-supervised model using a weighted binary cross-entropy loss function and a contrastive loss function, reliable pseudo-labeled samples are generated, which solves the problem of strong dependence on labeled data in remote sensing image change detection and improves detection accuracy.

CN118918098APending Publication Date: 2024-11-08LIAONING NORMAL UNIVERSITY
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411179585.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing deep learning models rely heavily on labeled data for change detection in remote sensing images. Existing semi-supervised methods have insufficient detection accuracy when using unlabeled data, and the generation of pseudo-labels is inadequate, which affects the accuracy of change detection.

Method used

A deep learning model was trained using a weighted binary cross-entropy loss function. The model state was recorded for three periods to generate a reliable set of pseudo-labeled samples. The semi-supervised detection model was then optimized using a contrastive loss function to construct a semi-supervised remote sensing image change detection model.

Benefits of technology

It improved the quality of pseudo-labels, enhanced the model's generalization ability, and improved the accuracy of remote sensing image change detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918098A_ABST
    Figure CN118918098A_ABST
Patent Text Reader

Abstract

The invention discloses a construction method of a semi-supervised remote sensing image change detection model, which comprises the following steps of: firstly, adjusting and optimizing a pre-trained model by adopting a marked data set and a weighted binary cross entropy loss function, and storing the models in different periods; secondly, the uncertainty of pseudo-labeled samples generated in the pre-training process is evaluated based on a joint measurement standard, the uncertainty of the pseudo-labeled samples is sorted, then the pseudo-labeled samples are divided into a reliable data set and an unreliable data set, a pre-training model is used for obtaining pseudo labels of the reliable data set, and a pseudo-labeled data set is formed; finally, samples of the pseudo-marked data set and the marked data set are mixed, the mixed data set samples are used for carrying out semi-supervised training on the reinitialized detection network, and in the semi-supervised training, changing pixels and unchanged pixels are predicted respectively, so that the detection accuracy is improved. And then a specific CT-SWBCE loss function is designed, and a semi-supervised RS image change detection model is optimized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image recognition, and in particular to a method for constructing a semi-supervised remote sensing image change detection model. Background Art

[0002] At present, deep convolutional neural networks (CNNs) have been successfully applied to RS image change detection. Existing methods are divided into single-stream networks and two-stream networks. In single-stream networks, image pairs are usually directly merged into the network, and then features are extracted through the encoding network and the decoding network to obtain the prediction of the changed area. Alcantarilla et al. used a deconvolution network for change detection. This method provides roughly aligned image pairs to a deep deconvolution network and predicts the changed area; Peng et al. combined low-dimensional features and high-dimensional features extracted from remote sensing images, and then used an attention module to enhance feature recognition capabilities and achieve more accurate predictions. The single-stream network methods mentioned above are generally simple and effective, and neural networks can be used to extract image features to achieve end-to-end change detection, but the biggest problem is the dependence on labeled data.

[0003] In order to alleviate the dependence of deep learning models on labeled data, researchers have proposed some methods. Semi-supervised methods usually require a small labeled data set, and then combine a large amount of unlabeled data during the model training process to obtain a model with significantly improved detection performance; virtual adversarial training (VAT) and semi-supervised learning deep learning methods use consistency regularization to achieve meaningful detection; there are also some methods that use unsupervised algorithms to obtain CD models and use contrast loss and similarity calculations instead of using labeled data, but these methods can easily detect unchanged areas; then, some researchers try to use weak labels to replace pixel-level labels, such as image-level labels and bounding boxes, which can effectively improve the performance of the detection model, but also require additional manual annotation of the data. Some semi-supervised methods apply generative adversarial networks (GANs) to solve the problem of model dependence on annotated data, using generative networks to obtain simulated distribution data and discriminators to distinguish different images. Although these methods improve the robustness of the model, they do not make full use of unlabeled data; another semi-supervised CD method introduces a consistency regularization method for unlabeled data. They add strong perturbations to the unlabeled data or its features, and then use consistency loss to maintain the consistency of different prediction results. However, this method usually sets a threshold for the prediction probability of each pixel to obtain a pseudo label, which only contains part of the original RS image, affecting the feature integrity of the changed object, thereby reducing the accuracy of change detection. Summary of the invention

[0004] In order to solve the above technical problems existing in the prior art, the present invention provides a method for constructing a semi-supervised remote sensing image change detection model.

[0005] The technical solution of the present invention is: a method for constructing a semi-supervised remote sensing image change detection model, which is carried out according to the following steps:

[0006] Step 1. Take a labeled dataset D l Input into the deep learning model, train the deep learning model and record three periods in the training process, the three periods are model T1, model T2, and model T3, and the loss function in the training process is a weighted binary cross entropy loss function;

[0007] The weighted binary cross entropy loss function expression is as follows:

[0008]

[0009] Among them, L SWBCE represents the weighted binary cross entropy loss function, i, j represents the spatial position index of the image pixel, y (i,j) Represents the label value at pixel (i, j), y (i,j) The value of is 0 or 1, and y (i,j) =1 means there is a change, y (i,j) =0 means no change, p (i,j) represents the change graph value at point (i, j), M and N represent the change graph dimensions, Indicates the sum of M-dimensional data, represents the sum of N-dimensional data, log represents the logarithmic function, TP represents the number of accurately identified changed pixels, TN represents the number of correctly detected unchanged pixels, FP represents the number of unchanged pixels that are mistakenly detected as changed pixels, and FN represents the number of changed pixels that are mistakenly detected as unchanged pixels;

[0010] Step 2. Based on the three periods of model T1, model T2, and model T3, the unlabeled dataset D ul Perform uncertainty calculations, sort the uncertainty calculation results, and then generate a reliable pseudo-labeled sample set based on the sorting results and unreliable pseudo-labeled sample sets

[0011] The expression for the uncertainty calculation is as follows:

[0012]

[0013] in, Indicates that through T JThe model is the pseudo-labeled result predicted by the I-th unlabeled sample pair, and J = 1, 2, Represents the pre-trained model T K is the pseudo-labeling result predicted by the Ith unlabeled sample pair, where K = 3; n represents the number of unlabeled image pairs; δ I represents the I-th weight coefficient; meanIoU(·) represents the mean intersection-over-union calculation function; OA(·) represents the overall accuracy calculation function, It means to sum K-1 data. It means summing the elements in the unlabeled sample set;

[0014] The Ith weight coefficient δ I The expression is as follows:

[0015]

[0016] Among them, mean(·) represents the mean calculation function;

[0017] Step 3. Use a reliable pseudo-labeled sample set And the labeled dataset D l Train the semi-supervised detection model and establish a loss function to build a semi-supervised detection model;

[0018] The loss function expression is as follows:

[0019] L CT-SWBCE =αL CT +(1-α)L SWBCE

[0020] Among them, L CT-SWBCE represents the loss function for training the semi-supervised detection model, L CT represents the contrast loss function, L SWBCE represents the SWBCE loss function; α represents the combined weight coefficient.

[0021] The preferred technical solution is that the reliable pseudo-labeled sample set The number of samples in accounts for 50% of the total number of pseudo-labeled samples.

[0022] The present invention firstly uses a labeled data set and a weighted binary cross entropy loss function to tune the pre-trained model, and saves the models of different periods; secondly, based on the joint metric, the uncertainty of the pseudo-labeled samples generated in the pre-training process is evaluated, and the uncertainty of the pseudo-labeled samples is sorted, and then the pseudo-labeled samples are divided into reliable data sets and unreliable data sets, and then the pseudo-labeled data sets are obtained by using the pre-training model to form a pseudo-labeled data set; finally, the pseudo-labeled data set and the labeled data set samples are mixed, and the reinitialized detection network is semi-supervised trained with these mixed data set samples. In the semi-supervised training, the present invention predicts the changed pixels and the unchanged pixels respectively, and then designs a specific CT-SWBCE loss function and optimizes the semi-supervised RS image change detection model. The present invention improves the quality of pseudo-labels, and the constructed semi-supervised change detection model has strong generalization ability, which can effectively improve the accuracy of remote sensing image change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the overall framework of an embodiment of the present invention. DETAILED DESCRIPTION

[0024] A method for constructing a semi-supervised remote sensing image change detection model of the present invention is as follows Figure 1 As shown, follow these steps:

[0025] Step 1. Take a labeled dataset D l Input into the deep learning model, train the deep learning model and record three periods in the training process, the three periods are model T1, model T2, and model T3, and the loss function in the training process is a weighted binary cross entropy loss function;

[0026] The weighted binary cross entropy loss function expression is as follows:

[0027]

[0028] Among them, L SWBCE represents the weighted binary cross entropy loss function, i, j represents the spatial position index of the image pixel, y (i,j) Represents the label value at pixel (i, j), y (i,j) The value of is 0 or 1, and y (i,j) =1 means there is a change, y (i,j) =0 means no change, p (i,j) represents the change graph value at point (i, j), M and N represent the change graph dimensions, Indicates the sum of M-dimensional data, represents the sum of N-dimensional data, log represents the logarithmic function, TP represents the number of accurately identified changed pixels, TN represents the number of correctly detected unchanged pixels, FP represents the number of unchanged pixels that are mistakenly detected as changed pixels, and FN represents the number of changed pixels that are mistakenly detected as unchanged pixels;

[0029] Step 2. Based on the three periods of model T1, model T2, and model T3, the unlabeled dataset D ul Perform uncertainty calculations, sort the uncertainty calculation results, and then generate a reliable pseudo-labeled sample set based on the sorting results and unreliable pseudo-labeled sample sets

[0030] The expression for the uncertainty calculation is as follows:

[0031]

[0032] in, Indicates that through T J The model is the pseudo-labeled result predicted by the I-th unlabeled sample pair, and J = 1, 2, Represents the pre-trained model T K is the pseudo-labeling result predicted by the Ith unlabeled sample pair, where K = 3; n represents the number of unlabeled image pairs; δ I represents the I-th weight coefficient; meanIoU(·) represents the mean intersection-over-union calculation function; OA(·) represents the overall accuracy calculation function, It means to sum K-1 data. It means summing the elements in the unlabeled sample set;

[0033] The Ith weight coefficient δ I The expression is as follows:

[0034]

[0035] Among them, mean(·) represents the mean calculation function;

[0036] Furthermore, the unlabeled sample set is divided into a reliable pseudo-labeled sample set and unreliable pseudo-labeled sample sets Selecting a reliable pseudo-labeled sample set and the labeled sample set D l The subsequent semi-supervised detection model is constructed, in which a reliable pseudo-labeled sample set The number of samples in accounts for 50% of the total number of pseudo-labeled samples;

[0037] Step 3. Use a reliable pseudo-labeled sample set And the labeled dataset D l Train the semi-supervised detection model and establish a loss function to build a semi-supervised detection model;

[0038] The loss function expression is as follows:

[0039] L CT-SWBCE =αL CT +(1-α)L SWBCE

[0040] Among them, L CT-SWBCE represents the loss function for training the semi-supervised detection model, L CT represents the contrast loss function, L SWBCE represents the SWBCE loss function; α represents the combined weight coefficient.

Claims

1. A method for constructing a semi-supervised remote sensing image change detection model. Features Follow these steps: Step 1. Take a labeled dataset D l Input into the deep learning model, train the deep learning model and record three periods in the training process, the three periods are model T1, model T2, and model T3, and the loss function in the training process is a weighted binary cross entropy loss function; The weighted binary cross entropy loss function expression is as follows: Among them, L SWBCE represents the weighted binary cross entropy loss function, i, j represents the spatial position index of the image pixel, y (i,j) Represents the label value at pixel (i, j), y (i,j) The value of is 0 or 1, and y (i,j) =1 means there is a change, y (i,j) =0 means no change, P (i,j) represents the change graph value at point (i, j), M and N represent the change graph dimensions, Indicates the sum of M-dimensional data, represents the sum of N-dimensional data, log represents the logarithmic function, TP represents the number of accurately identified changed pixels, TN represents the number of correctly detected unchanged pixels, FP represents the number of unchanged pixels that are mistakenly detected as changed pixels, and FN represents the number of changed pixels that are mistakenly detected as unchanged pixels; Step 2. Based on the three periods modelT1, modelT2, and modelT3, the unlabeled dataset D ul Perform uncertainty calculations, sort the uncertainty calculation results, and then generate a reliable pseudo-labeled sample set based on the sorting results and unreliable pseudo-labeled sample sets The expression for the uncertainty calculation is as follows: in, Indicates that through T J The model is the pseudo-labeled result predicted by the I-th unlabeled sample pair, and J = 1, 2, Represents the pre-trained model T K is the pseudo-labeling result predicted by the I-th unlabeled sample pair, where K = 3; n represents the number of unlabeled image pairs; δ represents the I-th weight coefficient; meanIoU(·) represents the mean intersection-over-union calculation function; OA(·) represents the overall accuracy calculation function, Indicates the sum of K-1 data. It means summing the elements in the unlabeled sample set; The expression of the I-th weight coefficient δ, is as follows: Among them, mean(·) represents the mean calculation function; Step 3. Use a reliable pseudo-labeled sample set And the labeled dataset D l Train the semi-supervised detection model and establish a loss function to build a semi-supervised detection model; The loss function expression is as follows: L CT-SWBCE =αL CT +(1-α)L SWBCE Among them, L CT-SWBCE represents the loss function for training the semi-supervised detection model, L CT represents the contrast loss function, L SWBCE represents the SWBCE loss function; α represents the combined weight coefficient.

2. According to claim 1, a method for constructing a semi-supervised remote sensing image change detection model, Features: The reliable pseudo-labeled sample set The number of samples in accounts for 50% of the total number of pseudo-labeled samples.

Citation Information

Cited By

  • Change detection pattern spot optimization method, electronic equipment and computer product

    CN119887833A

  • Change detection patch optimization method, electronic device, and computer product

    CN119887833B

  • Remote sensing forest land change detection method based on time sequence consistency judgment

    CN121191004A