Remote sensing image change detection method based on feature inverse information entropy network

By using an iterative training method based on feature inversion information entropy networks, the problems of low accuracy and class imbalance in remote sensing image change detection are solved, achieving higher detection accuracy and lower false negative rate.

CN118334529BActive Publication Date: 2026-11-10XIDIAN UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410365425.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-28
Publication Date
2026-11-10
Estimated Expiration
2044-03-28

AI Technical Summary

Technical Problem

Existing technologies for detecting changes in remote sensing images have low accuracy and suffer from difficulties in filtering out false change areas and class imbalance, resulting in high rates of missed detections and false detections.

Method used

We employ a feature-reversed information entropy network-based approach, which iteratively trains a twin encoder, a feature fusion module, a decoder, and a feature reversal module. We utilize a deformable information entropy convolution module to filter out pseudo-change regions, deeply supervised features to enhance the features of change regions, and the feature reversal module to enhance modeling capabilities.

Benefits of technology

It effectively improved the accuracy of remote sensing image change detection, reduced the false detection rate and false negative rate of pseudo-change areas, and improved the balance of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118334529B_ABST
    Figure CN118334529B_ABST
Patent Text Reader

Abstract

This invention proposes a remote sensing image change detection method based on a feature-reversed entropy network. The steps are as follows: acquiring training and testing sample sets; constructing a remote sensing image change detection network model based on a feature-reversed entropy network; iteratively training the remote sensing image change detection network model; and obtaining the remote sensing image change detection results. The deformable entropy convolution module in this invention reduces the impact of the limited available information of pseudo-changes on the false detection rate by calculating the entropy of the probability distribution of pixel changes within the deformable region of each feature map. Simultaneously, the deep-supervised feature enhancement module alleviates the class imbalance problem in remote sensing image change detection through selective deep supervision of the change region, reducing the model's false negative rate. Furthermore, the feature reversal module enhances the modeling capability of the feature reversal module by utilizing the invertible relationship between convolution and deconvolution, thereby improving the detection accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a method for detecting changes in remote sensing images. Specifically, it relates to a supervised method for detecting changes in remote sensing images based on a feature inverse information entropy network, which can be applied to fields such as urban expansion planning and farmland area monitoring. Background Technology

[0002] The objective of remote sensing image change detection is to distinguish changes of interest (ROI) in a pair of bitemporal images of the same area taken at different times. The definition of ROI varies depending on the task, such as changes in buildings, farmland, or forests. Because not all changes in a bitemporal image pair are ROI, a large number of pseudo-change regions exist in remote sensing image change detection. Some of these pseudo-change regions do not actually belong to the change category but appear different due to differences in season, humidity, and lighting when the two images were taken; others do belong to the change category but are not the changes of interest. Therefore, how to reasonably screen and filter pseudo-change regions in bitemporal image pairs is a key issue in remote sensing image change detection. Furthermore, because changed regions often constitute only a small proportion of a bitemporal image pair, with the remaining large background areas classified as invariant, change detection suffers from a severe class imbalance problem. Balancing the large difference in the number of samples between the two classes to reduce the false negative rate of the network is also a crucial research focus in change detection.

[0003] For example, patent application CN 117494765 A, entitled "A Twin Network and Method for Change Detection in Ultra-High Spatial Resolution Remote Sensing Images," discloses a method for detecting changes in remote sensing images. This invention employs asymmetric convolutional residual blocks to extract deeper features that provide more discriminative information about changes. A spatiotemporal attention mechanism is used to enhance the spatiotemporal domain of the fused multi-scale difference features, thereby enhancing areas of real change while suppressing areas of non-real change. Simultaneously, correlations are established between channels of difference feature maps at different scales to improve the representational ability of difference features at different scales. This invention is designed for change detection in high spatial resolution remote sensing images, effectively extracting real changes in ground features and eliminating the influence of external factors such as illumination and season, thus reducing the false detection rate of spurious changes. However, because spurious changes themselves have limited usable information and no solution is found to alleviate the severe class imbalance problem in remote sensing image change detection, the detection accuracy remains low. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a remote sensing image change detection method based on feature inverse information entropy network to solve the technical problem of low accuracy in the existing technology.

[0005] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0006] (1) Obtain the training sample set and the test sample set:

[0007] The changed regions in each of the M acquired bitemporal remote sensing image pairs are labeled, and K bitemporal remote sensing image pairs and their labels are used to form a training sample set R. The remaining MK bitemporal remote sensing image pairs are used as a test sample set E, where M > 1000.

[0008] (2) Construct a remote sensing image change detection network model O based on feature inverse information entropy network:

[0009] A remote sensing image change detection network model O is constructed, comprising a cascaded twin encoder (Encoder), a feature fusion module (FB), a decoder (Decoder), and a feature inversion module (FRB). The Encoder extracts features at D scales from the two temporal remote sensing images in a dual-temporal image pair. The FB fuses the feature maps at multiple scales extracted by the Encoder from the two temporal remote sensing images. The Decoder decodes the fusion result from the FB and includes a cascaded dual residual convolutional layer, a deformable information entropy convolution module (DIEC), and a deep supervised feature enhancement module (DSFE). The FRB inverts the classification features decoded by the Decoder, where D > 3.

[0010] (3) Iteratively train the remote sensing image change detection network model:

[0011] The training sample set R is used as input to the remote sensing image change detection network model O for iterative training, resulting in the trained remote sensing image change detection network model O. * ;

[0012] (4) Obtain the change detection results of remote sensing images:

[0013] The test sample set E is used as the trained remote sensing image change detection network model O. * The input is propagated forward to obtain the change detection and segmentation result map of MK test samples.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] 1. In the process of training the remote sensing image change detection network model, the deformable information entropy convolution module of this invention selects the region to be calculated for information entropy by the offset of each feature map, and calculates the information entropy of the pixel change probability distribution within the selected region. This achieves the filtering of pseudo-change regions in each feature map, avoiding the impact of the limited amount of available information of pseudo-changes on the false detection rate of existing technologies, and effectively improving the detection accuracy.

[0016] 2. In the process of training the remote sensing image change detection network model, the depth supervision feature enhancement module in the decoder is based on depth supervision. By selectively performing depth supervision on the changed areas, the feature intensity of the changed areas is enhanced, thereby alleviating the class imbalance problem in remote sensing image change detection, reducing the false negative rate of the model, and further improving the detection accuracy.

[0017] 3. In the process of training the remote sensing image change detection network model, the feature inversion module of this invention utilizes the invertible relationship between convolution and deconvolution to allow labels to participate more deeply in the model training. This enhances the modeling capability of the feature inversion module without increasing the number of parameters and computational load, thereby further improving the detection accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0019] Figure 2 This is a diagram showing the overall structure of the remote sensing image change detection network model of this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0021] Reference Figure 1 The present invention includes the following steps:

[0022] Step 1) Obtain the training sample set and the test sample set:

[0023] The changed regions in each of the M acquired bitemporal remote sensing image pairs are labeled, and K bitemporal remote sensing image pairs and their labels are used to form a training sample set R. The remaining MK bitemporal remote sensing image pairs are used as a test sample set E, where M > 1000. In this embodiment, M = 10000 and K = 7000.

[0024] Step 2) Construct a remote sensing image change detection network model O based on a feature inverse information entropy network:

[0025] Build as Figure 2The remote sensing image change detection network model O shown includes a cascaded twin encoder (Encoder), feature fusion module (FB), decoder (Decoder), and feature inversion module (FRB). The Encoder extracts features at D scales from the two temporal remote sensing images in the dual-temporal remote sensing image pair. The FB fuses the feature maps at multiple scales extracted from the two temporal remote sensing images by the Encoder. The Decoder decodes the fusion result from the FB and includes a cascaded dual residual convolutional layer, deformable information entropy convolution module (DIEC), and deep supervised feature enhancement module (DSFE). The FRB inverts the classification features decoded by the Decoder, where D > 3; in this embodiment, D = 5.

[0026] The twin encoder consists of two feature extraction networks, MobileNetV2, arranged in parallel.

[0027] The decoder, which contains a deformable information entropy convolution module DIEC, consists of a cascaded convolution kernel offset module CKO and an information entropy calculation module CIE.

[0028] The Feature Inversion Module (FRB) consists of a label feature extraction module composed of multiple convolutional layers arranged in parallel with the same number of layers, and an inversion module composed of multiple deconvolutional layers.

[0029] Step 3) Iteratively train the remote sensing image change detection network model:

[0030] The training sample set R is used as input to the remote sensing image change detection network model O for iterative training, resulting in the trained remote sensing image change detection network model O. * ;

[0031] (3a) Initialize the number of iterations to i, the maximum number of iterations to I, I>10000, and the change detection network model for remote sensing images in the i-th iteration is O i The weights and bias parameters in the equation are w. i b i And let i = 0;

[0032] (3b) The twin encoder encodes the two temporal remote sensing images in each pair of dual-temporal remote sensing images. Feature extraction is performed at D scales respectively; the feature fusion module FB processes the features extracted by the Encoder. The d-th feature map The fusion process is performed; the decoder processes the D-map fused feature maps obtained from the FB fusion. Decoding yields D segmentation maps for supervision. and a classification feature map F k* The Feature Inversion Module (FRB) decodes the classification feature map F obtained from the Decoder. k* Feature inversion is performed to obtain the change detection and segmentation result P of the k-th dual-temporal remote sensing image pair. k* ;

[0033] The feature fusion module FB is used for... When merging, firstly The feature map is obtained by concatenating the data along the channel dimension. Then, The input is passed through several convolutional layers for channel reduction to obtain a fused feature map.

[0034] The double residual convolutional layer connects adjacent feature maps in the first D-1 scales. and Perform stitching, and integrate all stitching results with the feature map at the D-th scale to form a feature map set.

[0035] The kernel offset module CKO in the deformable information entropy convolution module DIEC uses each feature map The offset is used to select the region where the information entropy will be calculated; the information entropy calculation module CIE calculates the information entropy of the probability distribution of pixel changes within the selected region, thereby filtering out pseudo-change regions in each feature map;

[0036] Specifically, the input feature map First, input CKO to obtain the offsets Δu = (Δx, Δy) of each pixel in the horizontal and vertical directions. Then, select the region where the information entropy IE(·) will be calculated based on Δu = (Δx, Δy). Next, CIE calculates the information entropy within the selected region using the information entropy formula. The calculation process is as follows:

[0037]

[0038]

[0039]

[0040] Where u represents the feature map The pixels in, u n Indicates the range of the 3x3 grid containing u. Nine pixels inside.

[0041] The Deep Supervised Feature Enhancement (DSFE) module applies the filtered feature maps to each feature map. The classification process yields segmentation maps at D scales for supervision. During supervision, only for The changes in the region are monitored, and the segmentation map at each scale is generated. Feature maps of corresponding scales The product as The feature maps are then augmented, and the augmented D feature maps are concatenated along the channel dimension to obtain the classification feature map F. k* .

[0042] The label feature extraction module in the Feature Reversal Module (FRB) extracts the feature representation LF for each label GT; the reversal module then compares the LF with the classification feature map F. k* The loss between L mse (F k* Optimization of F(LF) to achieve F k* Fitting to LF and using the fitted feature map F k* Obtain the change detection and segmentation result P of the k-th dual-temporal remote sensing image pair. k* .

[0043] (3c) Segmentation results P based on label GT and change detection k* Calculate the loss value Loss of O, and then apply Loss to w. i b i The remote sensing image change detection network model O is updated to obtain the model for this iteration. i ;

[0044] Loss = L DSFE +L FRB

[0045]

[0046] L FRB =L cls (P k* ,GT)+L mse (F k* ,LF)

[0047] L cls (P k* ,GT)=L bce (P k* ,GT)+L dice (P k* ,GT)

[0048]

[0049]

[0050] Among them, L DSFE and L FRB L represents the loss of the Deep Supervised Feature Enhancement (DSFE) module and the Feature Reversal (FRB) module, respectively;cls L represents the classification loss function; mse L represents the MSE loss function; bce and L dice Let w′ represent the binary cross-entropy loss function and the DICE loss function, respectively. i b′ i They represent w respectively i b i The update result, where lr represents the learning rate. and They represent Loss for w respectively i b i The partial derivative of .

[0051] (3d) Determine whether i≥I holds true. If yes, obtain the trained remote sensing image change detection network model O. * , otherwise, let i=i+1, O=O i And perform step (3b).

[0052] The deformable information entropy convolution module in this invention selects the region to be calculated for information entropy by the offset of each feature map, and calculates the information entropy of the probability distribution of pixel changes within the selected region. This filters out pseudo-change regions in each feature map, avoiding the impact of the limited amount of usable information in pseudo-changes on the false detection rate in existing technologies, and effectively improving the detection accuracy.

[0053] The deep-supervised feature enhancement module in the decoder of this invention is based on deep supervision. By selectively performing deep supervision on the changed regions, it enhances the feature intensity of the changed regions, thereby alleviating the class imbalance problem in remote sensing image change detection and reducing the false negative rate of the model. Furthermore, the feature inversion module utilizes the invertible relationship between convolution and deconvolution to allow the labels to participate more deeply in the model training. This enhances the modeling capability of the feature inversion module without increasing the number of parameters and computational cost, further improving the detection accuracy.

[0054] Step 4) Obtain the change detection results of the remote sensing image:

[0055] The test sample set E is used as the trained remote sensing image change detection network model O. * The input is propagated forward to obtain the change detection and segmentation result map of MK test samples.

Claims

1. A method for detecting changes in remote sensing images based on feature inverse information entropy networks, characterized in that, Includes the following steps: (1) Obtain the training sample set and the test sample set: For the acquired The changed regions in each of the two temporal remote sensing image pairs are labeled, and the changes are... The training sample set consists of two pairs of bitemporal remote sensing images and their labels. , the remaining A pair of bitemporal remote sensing images were used as the test sample set. ,in, , ; (2) Construct a remote sensing image change detection network model based on feature inverse information entropy network. : Constructing a twin encoder including cascades Feature fusion module decoder and feature inversion module Remote sensing image change detection network model ;in, Used to perform separate processing on the two temporal remote sensing images in a dual-temporal remote sensing image pair. Feature extraction at each scale; Used for The feature maps at multiple scales extracted from the two temporal remote sensing images are fused; Used for The fusion results are decoded, including cascaded dual residual convolutional layers and deformable information entropy convolutional modules. and deep supervision feature enhancement module , Concatenated kernel offset modules and information entropy calculation module composition, This is used to enhance the feature intensity of variable regions based on depth supervision by selectively performing depth supervision on variable regions. It includes a label feature extraction module consisting of multiple convolutional layers arranged in parallel with the same number of layers, and an inversion module consisting of multiple deconvolutional layers, used for... The decoded classification features are inverted, where... ; (3) Iteratively train the remote sensing image change detection network model: training sample set As a network model for detecting changes in remote sensing images The input is used to iteratively train the model to obtain a trained remote sensing image change detection network model. ; (4) Obtain the results of remote sensing image change detection: Test sample set As a trained remote sensing image change detection network model The input is propagated forward to obtain The result of segmentation detection of changes in each test sample.

2. The method according to claim 1, characterized in that, The remote sensing image change detection network model described in step (2) ,in: Twin encoder It includes two feature extraction networks, MobileNetV2, arranged in parallel.

3. The method according to claim 2, characterized in that, The iterative training of the remote sensing image change detection network model described in step (3) is implemented as follows: (3a) Initialize the number of iterations to be The maximum number of iterations is , , No. Sub-iteration Transformation Remote Sensing Image Change Detection Network Model The weights and bias parameters in the data are respectively , and order ; (3b) Twin encoder For each pair of two temporal remote sensing images , Perform separately Feature extraction at each scale; features Fusion Module right Extracted , The Feature map , Perform fusion; decoder right The result of fusion Aspect-Fusion Feature Map Decode to obtain A segmented image used for supervision and a classification feature map Feature reversal module right Decoded classification feature map Perform feature reversal to obtain the first Change detection and segmentation results of two pairs of remote sensing images. ; (3c) via label And change detection segmentation result image calculate loss value and through right , The remote sensing image change detection network model is updated to obtain the model for this iteration. ; (3d) Judgment If true, then a well-trained remote sensing image change detection network model is obtained. Otherwise, let , Then proceed with step (3b).

4. The method according to claim 3, characterized in that, The decoder described in step (3b) right The result of fusion Aspect-Fusion Feature Map Decoding is performed by following these steps: (3b1) Double residual convolutional layer for the previous Adjacent feature maps at each scale and Perform splicing, and combine all splicing results with the first... Feature maps at various scales are integrated into a feature map set. ; (3b2) Deformable Information Entropy Convolution Module Convolution kernel offset module Through each feature map The offset is used to select the region where information entropy will be calculated; Information entropy calculation module Calculate the information entropy of the probability distribution of pixel changes within the selected region to filter out pseudo-change regions in each feature map; (3b3) Deep Supervision Feature Enhancement Module For each feature map filtered Classify to obtain data for supervision. Segmentation map at various scales During supervision, only for The changes in the region are monitored, and the segmentation map at each scale is generated. Feature maps of corresponding scales The product as The feature enhancement map, and then the enhanced... The amplitude feature maps are concatenated along the channel dimension to obtain the classification feature map. .

5. The method according to claim 3, characterized in that, The feature inversion module described in step (3b) right Decoded classification feature map The steps to perform feature reversal are as follows: Feature inversion module The label feature extraction module extracts each label. Feature representation The reversal module, through... With classification feature map Losses between Optimization to achieve Towards The fitting was performed, and the feature map after fitting was used to obtain the feature map. Get the Change detection and segmentation results of two pairs of remote sensing images. .

6. The method according to claim 3, characterized in that, The steps described in step (3c) loss value The calculation formula is: ; ; ; ; in, and These represent the deep supervision feature enhancement modules. and feature inversion module The loss; Represents the classification loss function; Represents the MSE loss function; and Let represent the binary cross-entropy loss function and the DICE loss function, respectively.

7. The method according to claim 3, characterized in that, In step (3c), the weights and bias parameters are... , The update is performed using the following formulas: ; ; in, , They represent , The update results Indicates the learning rate. and They represent for , The partial derivative of .

Citation Information

Patent Citations

  • Ultrahigh spatial resolution remote sensing image change detection twin network and method

    CN117494765A

  • Semi-supervised high-resolution remote sensing image scene classification method based on generative adversarial network

    CN110689086A

  • Small sample change detection method based on multi-scale feature extraction

    CN112668494A