CN-122024066-A - Ground feature change detection method, storage medium and equipment based on double-time-phase remote sensing image

CN122024066ACN 122024066 ACN122024066 ACN 122024066ACN-122024066-A

Abstract

A ground feature change detection method, a storage medium and equipment based on double-time-phase remote sensing images belong to the technical field of remote sensing image detection. The method aims to solve the problem that the existing ground feature change detection based on different time phases is poor in differential feature extraction effect. The invention firstly divides hyperspectral remote sensing images with different time into double-time-phase remote sensing image blocks, then obtains latent variables from the encoding network of an encoder by the double-time-phase remote sensing image blocks and the transposition thereof respectively through variation, calculates the similarity of the double-time-phase remote sensing image blocks, splices the same-time-phase remote sensing image blocks and the transposed latent variables, sends the splicing results of different time phases into a double attention module to obtain interaction enhancement characteristics, obtains time-phase variation characteristics based on the interaction enhancement characteristics, and utilizes the similarity And obtaining the time phase change enhancement characteristic, outputting the change prediction probability through the fully connected classification layer after global maximum pooling, and realizing the ground feature change detection.

Inventors

TANG HAOXIANG
WANG XIAOFEI
LIU SHIXIN
XU ZHIYUAN
WANG LINSEN
ZHU YITING

Assignees

黑龙江大学

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (8)

1. A ground feature change detection method based on double-time-phase remote sensing images is characterized in that S1, hyperspectral remote sensing images with different time are obtained aiming at a region to be detected And Will be Dividing the image blocks to obtain a plurality of image blocks Wherein H and W are the height and width of the image block, C is the spectral band number, and the same way Dividing to obtain multiple image blocks Will (i) be And Inputting a detection network model to detect the change of the ground object, wherein the processing procedure of the detection network model comprises the following steps: s2, will And Respectively inputting a parameter sharing encoder Encoder which adopts a coding network of a variable self-coder and aims at Latent variable is obtained by encoder Encoder Is directed to Obtaining latent variables by encoder Encoder ; Simultaneously calculating similarity of two-time-phase remote sensing image blocks ; S3, pair Transposed to obtain And obtaining characteristics through linear transformation Will be Also through encoder Encoder, latent variables are obtained To (3) pair for the purpose of The same treatment is carried out to obtain latent variable ; S4, will And Splicing to obtain first time phase splicing latent variable Will be And Splicing to obtain a second time phase splicing latent variable ; S5, splicing the first time phase into a latent variable Feeding residual modules to obtain enhanced features Splicing the second time phase with the latent variable The same treatment is carried out to obtain enhanced features ; S6, enhancing the characteristics And Sending the interaction enhancement feature to a dual-attention module, wherein the dual-attention module adopts a two-layer attention mechanism to obtain the interaction enhancement feature And ; S7, based on interaction enhancement features And Obtaining time phase change characteristics and utilizing similarity Enhancing the time phase change characteristic to obtain time phase change enhancement characteristic ; S8, enhancing the phase change And outputting the change prediction probability through the fully connected classification layer after global maximization, so as to realize the ground feature change detection.
2. The method for detecting feature changes based on dual-temporal remote sensing images according to claim 1, wherein in step S4 And Process and apparatus for splicing And And the transverse splicing is adopted in the splicing process.
3. The method for detecting the change of the ground object based on the double-phase remote sensing image according to claim 1, wherein step S5 is to splice the first phase with the latent variable Feeding residual modules to obtain enhanced features The process of (1) comprises: , , Wherein, the Representing a1 x 1 convolutional layer, GN representing a group normalization layer, In order to activate the function, As a matrix of weights that can be learned, Representing a channel-by-channel multiplication.
4. A method for detecting a change in a ground object based on a dual-temporal remote sensing image as claimed in any one of claims 1 to 3, wherein the dual-attention module in step S6 obtains the interaction enhancement feature by using a two-layer attention mechanism And The process of (1) comprises: S601, enhancing features And Respectively sending into a first layer of attention mechanism for processing to obtain features And ; S602, for characteristics And And (3) performing interaction attention mechanism processing: will first Generating Query, key and Value embeddings by linear transformation, respectively denoted as And will be Generating Query, key and Value embeddings by linear transformation, respectively denoted as ; For the purpose of Based on 、 Feature extraction through cross-attention mechanism Utilizing characteristics For a pair of Performing interaction enhancement features ; For the purpose of Based on 、、 Feature extraction through cross-attention mechanism Utilizing characteristics For a pair of Performing interaction enhancement features 。
5. The method for detecting feature changes based on dual-temporal remote sensing images according to claim 4, wherein the first layer of attention mechanism in S601 adopts a channel attention mechanism CAM module.
6. The method for detecting the change of the ground object based on the double-phase remote sensing image according to claim 5, wherein the detection network model is trained in advance, and the training process comprises the following steps: Training based on the training set, the training is performed in step S2 And In the process of respectively inputting a parameter sharing encoder Encoder, the obtained latent variable Also fed into the decoding network of the variable self-encoder to obtain a reconstructed image Based on And Calculating reconstruction errors And KL divergence to obtain ELBO loss Obtaining latent variable in the same way Corresponding reconstructed image And calculate ELBO loss Further obtaining the variation loss of the double-time-phase remote sensing image block ; When the variation prediction probability is obtained in S7, calculating the cross entropy loss of the prediction result ; Bonding of And Calculation of total loss Wherein Is a balance weight coefficient; The detection network model is trained based on the total loss L.
7. A computer storage medium, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement a method for detecting a change in a ground object based on a dual-temporal remote sensing image according to any one of claims 1 to 6.
8. A ground object change detection device based on a dual-temporal remote sensing image, wherein the device comprises a processor and a memory, wherein at least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement a ground object change detection method based on a dual-temporal remote sensing image as set forth in any one of claims 1 to 6.

Description

Ground feature change detection method, storage medium and equipment based on double-time-phase remote sensing image Technical Field The invention belongs to the technical field of remote sensing image detection, and relates to a ground feature change detection method, a storage medium and equipment based on double-time-phase remote sensing images. Background With the development of remote sensing image detection technology, the current remote sensing image detection technology has been applied to a plurality of fields, and the detection of ground features by using remote sensing image detection has become mature. Early detection of remote sensing image features mainly depends on manually designed feature operators and shallow machine learning models. Researchers use priori knowledge of the spectral features, texture statistics, geometric shapes, shadow relations and the like of the ground features and combine methods such as a support vector machine and the like to realize pixel-level classification. Such methods are not particularly ideal and have poor generalization ability and poor feature detection ability for complex backgrounds. Along with the development of the deep learning model, the method gradually becomes a main scheme for remote sensing image analysis, and various schemes have evolved and have good effects. For ground object change detection, at present, remote sensing images of different time phases are commonly adopted for feature extraction, then difference images based on feature images are detected, in such schemes, the network model with shared remote sensing image input parameters of different time phases is used for feature extraction, so that identical space transformation and feature extraction can be guaranteed for different time phases, the unification of feature space is guaranteed to a certain extent for the corresponding extracted features, but the remote sensing image features of different time phases are mutually independent to be extracted, and time and space information of the remote sensing images of different time phases are not fully utilized, so that targets with insignificant change ranges and the change detection capability of small targets are not ideal, and further optimization and improvement are needed. Disclosure of Invention The method and the device aim to solve the problems that the traditional ground feature change detection based on different time phases has poor differential feature extraction effect and the network type cannot fully utilize time and space information. A ground feature change detection method based on double-time-phase remote sensing images comprises the following steps of S1, aiming at a region to be detected, obtaining hyperspectral remote sensing images at different timesAndWill beDividing the image blocks to obtain a plurality of image blocksWherein H and W are the height and width of the image block, C is the spectral band number, and the same wayDividing to obtain multiple image blocksWill (i) beAndInputting a detection network model to detect the change of the ground object, wherein the processing procedure of the detection network model comprises the following steps: s2, will AndRespectively inputting a parameter sharing encoder Encoder which adopts a coding network of a variable self-coder and aims atLatent variable is obtained by encoder EncoderIs directed toObtaining latent variables by encoder Encoder; Simultaneously calculating similarity of two-time-phase remote sensing image blocks; S3, pairTransposed to obtainAnd obtaining characteristics through linear transformationWill beAlso through encoder Encoder, latent variables are obtainedTo (3) pair for the purpose ofThe same treatment is carried out to obtain latent variable; S4, willAndSplicing to obtain first time phase splicing latent variableWill beAndSplicing to obtain a second time phase splicing latent variable; S5, splicing the first time phase into a latent variableFeeding residual modules to obtain enhanced featuresSplicing the second time phase with the latent variableThe same treatment is carried out to obtain enhanced features; S6, enhancing the characteristicsAndSending the interaction enhancement feature to a dual-attention module, wherein the dual-attention module adopts a two-layer attention mechanism to obtain the interaction enhancement featureAnd; S7, based on interaction enhancement featuresAndObtaining time phase change characteristics and utilizing similarityEnhancing the time phase change characteristic to obtain time phase change enhancement characteristic; S8, enhancing the phase changeAnd outputting the change prediction probability through the fully connected classification layer after global maximization, so as to realize the ground feature change detection. Further, in step S4AndProcess and apparatus for splicingAndAnd the transverse splicing is adopted in the splicing process. Further, step S5 is to splice the first time phase latent variablesFeeding