CN114821299A

CN114821299A - Remote sensing image change detection method

Info

Publication number: CN114821299A
Application number: CN202210311570.7A
Authority: CN
Inventors: 侍佼; 吴天成; 雷雨; 周德云; 周颖; 何玉亭; 曾丽娜
Original assignee: Northwestern Polytechnical University
Current assignee: Northwestern Polytechnical University
Priority date: 2022-03-28
Filing date: 2022-03-28
Publication date: 2022-07-29
Anticipated expiration: 2042-03-28
Also published as: CN114821299B

Abstract

The invention discloses a method for detecting the change of a remote sensing image, which comprises the following steps: acquiring a first remote sensing image and a second remote sensing image which are shot at the same position at different moments; marking the first remote sensing image and the second remote sensing image respectively to form unlabeled sample data and labeled sample data; respectively carrying out vector format conversion on the unlabeled sample data and the labeled sample data to form unlabeled sample vector data and labeled sample vector data; searching a network structure by using an evolutionary algorithm to obtain a plurality of target step grid models; semi-supervised and unsupervised training is carried out on each target ladder network model to determine a middle optimal ladder network model; and performing semi-supervised and unsupervised training on the intermediate optimal ladder network model to obtain a target optimal ladder network model, wherein an output result corresponding to the target optimal ladder network model is a detection result. The detection method has stronger universality and can obtain better remote sensing image change detection results.

Description

Remote sensing image change detection method

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a remote sensing image change detection method.

Background

Remote sensing image Change Detection (CD for short) is a technology for identifying a Change area of the same area at different times by comparing and processing a pair of remote sensing images. With the development of remote sensing technology, types of remote sensing images are gradually diversified, such as Synthetic Aperture Radar (SAR) images, optical images, heterogeneous images, ultrahigh Resolution (VHR) images, multispectral images, hyperspectral images and the like, so that more abundant remote sensing image information is brought. Nowadays, as more and more remote sensing image information can be fully utilized, applications of remote sensing image change detection are common, such as earth surface change analysis, land cover change, natural disaster assessment, agricultural assessment, city expansion and evolution, environment monitoring and the like.

In recent years, the deep learning technology can extract high-order features of remote sensing images under the support of a large amount of remote sensing data, and therefore the deep learning technology is widely applied to the field of remote sensing image change detection. However, due to the lack of true label exemplars, semi-supervised neural network models that utilize large amounts of unlabeled data to supplement the missing supervised information are of increasing interest. In the case of extreme lack of labels, the semi-supervised method has great potential in the field of remote sensing image change detection. The processing flow of the semi-supervised remote sensing image change detection task is divided into three parts: 1. data processing: the remote sensing data preprocessing module is used for preprocessing the remote sensing data; 2. structural learning of the tag: training the labeled data to generate a structural model thereof; 3. network training: and synthesizing the last module of the untrained processed data, and then outputting a change detection result graph.

However, the existing semi-supervised framework has the following problems:

1. the training of the marked data and the training of the unmarked data are relatively independent, and the marked data and the unmarked data are difficult to train or good remote sensing image information representation structure is difficult to obtain and is spread to the unmarked data, so that the detection precision of the remote sensing image change is influenced;

2. the change of the detection object causes the change of difference information between objects and data scale required to be processed, the adaptability of the fixed network structure is poor, and the detection precision of the change of the remote sensing image is also influenced.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a method for detecting the change of a remote sensing image. The technical problem to be solved by the invention is realized by the following technical scheme:

the embodiment of the invention provides a method for detecting the change of a remote sensing image, which comprises the following steps:

acquiring a first remote sensing image and a second remote sensing image which are shot at the same position at different moments;

respectively marking the first remote sensing image and the second remote sensing image to form unlabeled sample data and labeled sample data; wherein the tagged sample data comprises marked parts in the first remote sensing image and the second remote sensing image; the unlabelled sample data comprises unlabelled parts in the first remote sensing image and the second remote sensing image;

respectively carrying out vector format conversion on the unlabeled sample data and the labeled sample data to form unlabeled sample vector data and labeled sample vector data;

searching a network structure by using an evolutionary algorithm to obtain a plurality of target step grid models;

semi-supervised and unsupervised training is respectively carried out on each target ladder network model to determine an intermediate optimal ladder network model; wherein, the semi-supervised training is training by using the labeled sample vector data; the unsupervised training is to use the labeled sample data and the unlabeled sample vector data for training;

and performing semi-supervised and unsupervised training on the intermediate optimal ladder network model to obtain a target optimal ladder network model, wherein an output result corresponding to the target optimal ladder network model is a change detection result.

In an embodiment of the present invention, the performing vector format conversion on the unlabeled sample data and the labeled sample data to form unlabeled sample vector data and labeled sample vector data respectively includes:

acquiring pixel point pairs at the same position of the first remote sensing image and the second remote sensing image by using a preset neighborhood window to obtain corresponding image block pairs;

for each image block pair, expanding pixel points of each image block in the image block pair by rows to obtain two pixel point vectors;

longitudinally connecting the two pixel point vectors to form a new pixel point vector; forming the label-free sample vector data by all new pixel point vectors formed by the label-free sample data;

for the tagged sample data, further comprising:

meanwhile, judging whether all pixel point pairs change or not;

and longitudinally connecting all the new pixel point vectors and the corresponding judgment results to form the labeled sample vector data.

In one embodiment of the invention, all ladder network models include a first encoder, a second encoder, and a decoder.

In one embodiment of the present invention, the first encoder, the second encoder and the decoder employ the same network structure; the same network structure comprises the network layers and the nodes in the network structure which are respectively corresponding to the same number.

In an embodiment of the present invention, the searching for the network structure by using the evolutionary algorithm to obtain a plurality of target ladder mesh models includes:

selecting a plurality of initial ladder network models as initial populations;

respectively coding the network layer number and the node number of a first coder in each initial ladder network model;

utilizing an evolutionary algorithm to carry out crossing and variation on the coding results of all initial ladder network models in the initial population until a population iteration stop condition is met;

and selecting the coding results of the intersection and the variation which meet the requirements, and decoding to form the plurality of target ladder network models.

In one embodiment of the present invention, a process for semi-supervised and unsupervised training of a ladder network model comprises:

carrying out noise adding compression on the unlabeled sample vector data and the labeled sample vector data by utilizing a first encoder in the ladder network model to obtain a noise sample;

calculating the difference between the output result of the first encoder corresponding to the labeled sample vector data in the noise sample and the judgment result in the labeled sample vector data to obtain a supervision loss result;

denoising an output result of a first encoder in the ladder network model through a decoder in the ladder network model to obtain a denoised sample;

carrying out pure compression on the unlabeled sample vector data and the labeled sample vector data by utilizing a second encoder in the ladder network model to obtain a plurality of noiseless samples;

reconstructing the denoised samples and each noiseless sample by using a decoder in the ladder network model, and calculating the difference between a reconstruction result and each noiseless sample to obtain a plurality of unsupervised loss results;

and calculating a network loss result corresponding to the ladder network model according to the supervision loss result and all unsupervised loss results.

In an embodiment of the present invention, before performing semi-supervised and unsupervised training on the ladder network model, the method further includes:

dividing the unlabeled sample vector data into a plurality of batches; wherein the size of the unlabeled sample vector data of each batch is close to that of the labeled sample vector data;

respectively enabling the unlabeled sample vector data and the labeled sample vector data of each batch to form training data, and performing semi-supervised and unsupervised training on the ladder network model by using the training data to obtain a corresponding network loss result of the ladder network model;

determining a corresponding optimized ladder network model according to all network loss results;

the optimized ladder network model determined by training each target ladder network model is a plurality of target ladder network models, and the intermediate optimal ladder network model is determined from the target ladder network models; training the intermediate optimal ladder network model to determine an optimized ladder network model which is the target optimal ladder network model; and the iteration threshold adopted when each target ladder network model is trained is smaller than the iteration threshold adopted when the intermediate optimal ladder network model is trained.

In one embodiment of the present invention, the process of determining the intermediate optimal ladder network model from the plurality of target ladder network models comprises:

calculating the average difference level corresponding to all target ladder network models;

and determining the intermediate optimal ladder network model according to the average difference level, the network loss results of all target ladder network models and the unsupervised loss result of the target ladder network model.

In an embodiment of the present invention, the output result corresponding to the target optimal ladder network model is a change detection result, which includes:

classifying the output result of the first encoder in the target optimal ladder network model to obtain the change detection result; and when the change detection result is obtained by classification, the first encoder in the corresponding target optimal ladder network model does not add noise interference.

In one embodiment of the present invention, further comprising:

classifying the output result of the first encoder in the target optimal ladder network model to obtain a semi-supervised classification result;

classifying output results of each layer except the first layer of network layer of the second encoder in the target optimal ladder network model to obtain a plurality of unsupervised classification results;

fusing the semi-supervised classification result and the plurality of unsupervised classification results;

counting the total scores of all classification results according to the semi-supervised classification result and the unsupervised classification results, judging whether the classification result of each pixel point is reliable or not according to the total scores, selecting all reliable unlabelled sample data in a fusion result as pseudo-labeled sample data, and adding the pseudo-labeled sample data into the labeled sample data to form new labeled sample data;

and performing semi-supervised and unsupervised training on the target optimal ladder network model according to the unlabelled sample data and the new labeled sample data to obtain an updated target optimal ladder network model, wherein the output corresponding to the updated target optimal ladder network model is a change detection result.

The invention has the beneficial effects that:

the invention provides a remote sensing image change detection method, which is a new strategy for combining labeled sample data and non-labeled sample data, combines the advantages of a semi-supervised and non-supervised change detection method, utilizes the labeled sample data in the semi-supervised change detection, utilizes the labeled sample data and the non-labeled sample data in the non-supervised change detection, trains the labeled sample data and the non-labeled sample data to obtain a good remote sensing image information representation structure, improves the remote sensing image change detection precision, directly carries out generalized learning on the labeled sample data in a network, does not need to additionally learn a labeled structured model, and simplifies a network model; the invention adopts an evolutionary algorithm to search a network to obtain a plurality of target ladder network models, then trains each target ladder network model by adopting semi-supervised and unsupervised training modes, in each network cycle iteration, the structure and parameters of the target ladder network model can be automatically adjusted to adapt to the most effective ladder network to obtain an intermediate optimal ladder network model, and the intermediate optimal ladder network model is again subjected to semi-supervised and unsupervised training to realize the final fine tuning of the network to obtain the target optimal ladder network model, and the change detection of remote sensing images is realized through the target optimal ladder network model, so that the automatic network adjustment avoids the manual network configuration and can also adapt to the conditions that the change of a detection object causes the change of difference information among objects and the scale of data to be processed, and the change detection method has stronger universality, the semi-supervised and unsupervised double prediction results are beneficial to supplementing classification information of different angles, and better remote sensing image change detection results can be obtained.

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Drawings

FIG. 1 is a schematic flow chart of a method for detecting changes in remote sensing images according to an embodiment of the present invention;

FIG. 2 is a schematic flow chart of a network structure search using an evolutionary algorithm according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of encoding the number of network layers and the number of nodes of the first encoder in each initial ladder network model according to an embodiment of the present invention;

fig. 4 is a schematic diagram illustrating the crossing and variation of the encoding results of all the initial ladder network models in the initial population according to the embodiment of the present invention;

FIG. 5 is a schematic flow chart of semi-supervised and unsupervised training of the ladder network model according to the embodiment of the present invention;

FIG. 6 is a schematic flow chart of determining an intermediate optimal ladder network model from a plurality of target ladder network models according to an embodiment of the present invention;

fig. 7 is a schematic diagram of a detection process for implementing a circular ladder network by using pseudo tag sample data according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of a training process for implementing a circular ladder network using pseudo tag sample data according to an embodiment of the present invention;

FIG. 9 is a complete diagram of the output change detection result of the optimal ladder network model according to the present invention;

fig. 10(a) to fig. 10(c) are schematic diagrams of multi-temporal remote sensing images of a test heterogeneous image Wuhan according to an embodiment of the present invention;

fig. 11(a) to fig. 11(c) are schematic diagrams of a multi-temporal remote sensing image of a test multispectral image Hongqi provided by an embodiment of the present invention;

fig. 12(a) to 12(d) are schematic diagrams illustrating comparison of change detection results of different algorithms in testing heterogeneous image Wuhan multi-temporal remote sensing images according to an embodiment of the present invention;

fig. 13(a) to fig. 13(d) are schematic diagrams showing comparison of change detection results of different algorithms in testing a multi-spectral image Hongqi multi-temporal remote sensing image according to an embodiment of the present invention.

Detailed Description

The present invention will be described in further detail with reference to specific examples, but the embodiments of the present invention are not limited thereto.

Example one

In order to improve the accuracy of detecting changes in remote sensing images, referring to fig. 1, an embodiment of the present invention provides a method for detecting changes in remote sensing images, including the following steps:

s10, acquiring a first remote sensing image and a second remote sensing image which are shot at the same position at different moments;

s20, marking the first remote sensing image and the second remote sensing image respectively to form unlabeled sample data and labeled sample data; the labeled sample data comprises marked parts in the first remote sensing image and the second remote sensing image; the unlabeled sample data includes unlabeled parts in the first remote sensing image and the second remote sensing image.

Specifically, S10 and S20 are remote sensing data acquisition stages, the acquired first remote sensing image and second remote sensing image may be SAR images, or may also be multispectral images, and the form of the specifically acquired first remote sensing image and second remote sensing image is not limited. And then, marking the first remote sensing image and the second remote sensing image, specifically marking the same position of the two remote sensing images, thereby forming unlabeled sample data and labeled sample data, wherein both the first remote sensing image and the second remote sensing image contain labeled data and unlabeled data.

And S30, respectively carrying out vector format conversion on the unlabeled sample data and the labeled sample data to form unlabeled sample vector data and labeled sample vector data.

Specifically, the embodiment of the present invention adopts a neighborhood window acquisition mode, and performs vector format conversion on unlabeled sample data and labeled sample data to form unlabeled sample vector data and labeled sample vector data, which specifically includes:

acquiring pixel point pairs at the same positions of the first remote sensing image and the second remote sensing image by using a preset neighborhood window to obtain corresponding image block pairs; for each image block pair, expanding the pixel points of each image block in the image block pair according to rows to obtain two pixel point vectors; longitudinally connecting the two pixel point vectors to form a new pixel point vector; forming a new pixel point vector by using the unlabeled sample data to form unlabeled sample vector data;

for tagged sample data, further comprising:

meanwhile, judging whether all pixel point pairs change or not;

and longitudinally connecting all the new pixel point vectors and the corresponding judgment results to form labeled sample vector data.

For example, the remote sensing data set D ═ { I ═ I ₁ ,I ₂ In (I), I ₁ Representing a first remote sensing image, I ₂ Representing the second remote sensing image, respectively for the first remote sensing image I ₁ And a second remote sensing image I ₂ Selecting pixel points at the same position to form a pixel point pair, respectively recording the pixel point pair as p '═ i', j 'and p' ═ i ', j', respectively taking (i ', j') and (i ', j') as pixel centers, carrying out pixel point acquisition on each pixel point in the pixel point pair by using a window with the size of a neighborhood window being w multiplied by w to obtain a corresponding image block pair, wherein the size of each image block in the image block pair is w multiplied by w, and expanding the pixel points in each image block according to rows to obtain two w multiplied by w ² Vector of pixel points of dimension, and then two w ² The pixel point vectors of the dimension are longitudinally connected to form 2w ² A new pixel point vector of dimension. For the first remote sensing image I ₁ And a second remote sensing image I ₂ All the pixel points at the same position in the system are subjected to vector format conversion in the neighborhood window acquisition mode to correspondingly form 2w ² A new pixel point vector of dimension.

For the first remote sensing image I ₁ And a second remote sensing image I ₂ Of 2w, which is formed from all 2w ² And forming unlabeled sample vector data by using the new pixel point vectors of the dimension.

For the first remote sensing image I ₁ And a second remote sensing image I ₂ The labeled sample data is also considered, the change condition of the pixel point at the position is also considered, whether the pixel point pair at the selected current position is changed or not is judged, the change condition is recorded regardless of whether the pixel point pair is changed or not, for example, the pixel point pair is marked as 1 when being changed and is marked as 0 when not being changed, and the judgment result and the 2w formed according to the neighborhood window acquisition mode are recorded ² The pixel point vectors of the dimension are longitudinally connected to form 2w ² A new pixel point vector of +1 dimension. And the judgment result is used for calculating the monitoring loss result in the subsequent semi-monitoring training process.

For the first remote sensing image I ₁ And a second remote sensing image I ₂ With the labeled sample data of all 2w formed ² And forming labeled sample vector data by the + 1-dimensional new pixel point vectors.

It should be noted that, if the first remote sensing image I is ₁ And a second remote sensing image I ₂ As a multi-spectral mapLike having a plurality of ordinary segments, the above-described vector format conversion process is also applicable. Such as the first remote sensing image I ₁ And a second remote sensing image I ₂ All have n ₁ Each common segment corresponds to a remote sensing image, and the first remote sensing image I ₁ And a second remote sensing image I ₂ Corresponding to the respective sum of n ₁ A remote sensing image, for the first remote sensing image I ₁ And a second remote sensing image I ₂ Each remote sensing image is processed by adopting the neighborhood window acquisition mode, namely, the first remote sensing image I is processed ₁ And a second remote sensing image I ₂ All the pixel points at the same position of each remote sensing image are subjected to the neighborhood window expansion to respectively and correspondingly form 2w ² New pixel point vector of dimension, 2w corresponding to all ordinary segments ² The dimensional pixel point vectors are longitudinally connected to form 2n ₁ w ² And (5) dimension pixel point vectors. For the first remote sensing image I ₁ And a second remote sensing image I ₂ The sample data with the label also needs to consider the change condition of the pixel point pair at the same position, and the judgment result and the 2n formed according to the neighborhood window acquisition mode ₁ w ² The pixel point vectors of dimension are longitudinally connected to form 2n ₁ w ² A new pixel point vector of +1 dimension.

Similarly, for the multispectral image, for the first remote sensing image I ₁ And a second remote sensing image I ₂ Of (4) unlabeled sample data consisting of all 2n formed ₁ w ² Forming unlabeled sample vector data by using the new pixel point vectors of the dimension; for the first remote sensing image I ₁ And a second remote sensing image I ₂ With tagged sample data of all 2n formed ₁ w ² And forming labeled sample vector data by the + 1-dimensional new pixel point vectors.

S40, searching a network structure by using an evolutionary algorithm to obtain a plurality of target step grid models;

specifically, all ladder network models related to the embodiments of the present invention include a first encoder, a second encoder, and a decoder, such as an initial ladder network model, a target ladder network model, an intermediate optimal ladder network model, and a target optimal ladder network model, which are mentioned later, are all the network structures. The network structures of the first encoder, the second encoder and the decoder can adopt the same network structure, and also can adopt different network structures, and only if different network structures are adopted, the complexity of network search is increased, because the first encoder, the second encoder and the decoder adopt the same network structure preferentially in the embodiment of the invention; the same network structure comprises the network layers and the nodes in the network structure which are respectively corresponding to the same number.

In a traditional change detection algorithm, a fixed network model is mostly considered, but in a multi-spectrum remote sensing image, more spectrums bring abundant change information and different input data sizes, so that the adoption of the fixed network model can cause greater detection difficulty. However, in the way of manually adjusting the network model to improve the change detection accuracy, an optimal network model cannot be achieved after many times of manual adjustments, and the problem of low detection accuracy still exists. Therefore, the embodiment of the invention provides a mode for searching a network structure by using an evolutionary algorithm to determine an optimal network model.

In the selection process of the evolutionary algorithm, only the number of network layers, i.e. Layer _ size, and the number of nodes of a single network Layer are searched, so that the whole evolutionary network search space is reduced. In order to prevent the training difficulty of the network caused by excessive mutation, the step network structure in the embodiment of the invention cannot select a too large and too complex network structure, and preferably selects a network structure with a small Layer number of nodes in each Layer of the network and a large Layer size of the network, so that the subsequent unsupervised feature extraction is deeper, and the generalization capability of the labeled sample data can be better improved by using the unlabeled sample data. The embodiment of the present invention limits the number of maximum network layers maxLayerSize and the maximum number of nodes maxNeuSzie of a single network layer, for example, the number of maximum network layers maxLayerSize is limited to 6, and the number of maximum nodes maxNeuSzie is limited to 150, but is not limited to this limiting manner. Because the first encoder, the second encoder and the decoder are preferably selected to adopt the same network structure, in the embodiment of the invention, when the network search is carried out by the evolutionary algorithm, only one of the first encoder, the second encoder and the decoder needs to be selected for carrying out the network search, and after the network structure of the first encoder is determined by the network search, the network structure of the second encoder and the decoding layer adopts the network structure completely consistent with that of the first encoder, so that an additional network structure search process is not needed, and the complexity of the network search is reduced.

Taking the network search of the first encoder as an example, the corresponding evolutionary algorithm is utilized to perform the network structure search to obtain a plurality of target ladder grid models, please refer to fig. 2, which includes the following steps:

s401, selecting a plurality of initial ladder network models as initial populations.

Specifically, for the initial ladder network model selection, the above requirements of the maximum number of network layers maxLayerSize and the maximum number of nodes maxNeuSzie of a single network layer need to be satisfied. In the evolutionary algorithm, an initial ladder network model is an individual, and an initial population includes a plurality of such individuals, each individual having a different network structure. For example, the size popSize of the initial population of an embodiment of the invention is 30.

S402, respectively coding the network layer number and the node number of the first coder in the initial ladder network model.

Specifically, referring to fig. 3, the embodiment of the present invention takes a network net1 and a network net2 as examples, and describes a case of encoding the number of network layers and the number of nodes of the first encoder in the initial ladder network model. As can be seen from fig. 3: from left to right, the encoding method for the network layer from the top to the bottom of the network is performed, the encoding value corresponding to each network layer is the number of nodes of the network layer, for example, the network net1 is a 4-layer network, the number of nodes corresponding to each network layer is 6,4,3,2, the encoding result of the first encoder in the initial ladder network model net1 is [6,4,3,2], and similarly, the network net2 is a 3-layer network, the number of nodes corresponding to each network layer is 4,3,2, and the encoding result of the first encoder in the initial ladder network model net2 is [4,3,2 ].

According to the coding mode, the real network structure of each initial ladder network model in the initial population is coded, the coding result is shown in the upper left corner of fig. 4, each row is the coding result of the initial ladder network model, each row contains a number of numbers representing the number of layers of the network, the number from the top to the bottom of the network is sequentially represented from left to right, specifically, each number corresponds to the number of nodes of the network, the first row is taken as an example to represent that the number of the first coder in the initial ladder network model is 4, and the number of the network nodes corresponding to the top to the bottom is 18, 47, 27 and 2.

And S403, crossing and varying the coding results of all the initial ladder network models in the initial population by using an evolutionary algorithm until a population iteration stop condition is met.

Specifically, using S402, the encoding of all initial ladder network models in the initial population is completed, and the intersection and variation shown in fig. 4 are performed on all encoding results. For the case of cross variation, both the number of network layers and the number of nodes of the network layers can be guaranteed to always meet the requirements of the maximum number of network layers maxLayerSize and the maximum number of nodes maxNeuSzie of a single network layer, but for the case of mutation variation, the number of network layers may no longer meet the requirement of the maximum number of network layers maxLayerSize, and the number of nodes of the network layers may also no longer meet the requirement of the maximum number of nodes maxNeuSzie of a single network layer, and for the mutation individuals which do not meet the requirements, the mutation individuals are deleted in the population iteration process so as not to influence the evolution effect. In the embodiment of the invention, the mating operation for generating offspring exists in the network structure search, the mating pool matchPoolSize is limited to 20, the cross probability pc is 0.3, the variation probability pm is 0.3, the population iteration frequency maxIter is 30, and all the stepped network structures after cross and variation are output after 30 population iterations.

S404, selecting the coding results of the intersection and the variation meeting the requirements, and decoding to form a plurality of target ladder network models.

Specifically, after the population iteration is stopped, if the cross and mutation coding results also possibly have the condition that the number of the network layers does not satisfy the requirement of the maximum number of the network layers maxLayerSize, and the number of the nodes of the network layers does not satisfy the requirement of the maximum number of the nodes maxNeuSzie of a single network layer, the cross and mutation coding results satisfying the requirement are selected, and a plurality of target ladder network models are formed by decoding. Wherein, the decoding is the inverse process of the S402 coding, for example, if the coding result of the intersection and the variation is [18,132,60,27,2], the target ladder network model is 5 layers, the number of nodes of each layer network layer corresponding to the top layer of the network to the bottom layer of the network is 18,132,60,27,2, respectively, and the target ladder network model corresponding to the coding result of the intersection and the variation is formed by decoding.

In order to prevent meaningless network evolution, the embodiment of the invention adopts the principle of 'no priority and stop'. This means that if the classification of the future three generations of network offspring does not perform as well as the worst network individuals in the parent, the cycle will directly stop. At this time, all the encoding results of the crossover and mutation are decoded to form a plurality of target ladder network models.

S50, respectively carrying out semi-supervised and unsupervised training on each target ladder network model to determine an intermediate optimal ladder network model; wherein, the semi-supervised training is training by using labeled sample vector data; the unsupervised training is training by using labeled sample data and unlabeled sample vector data;

specifically, a plurality of target ladder network models are formed by S40, and an optimal ladder network model is identified from these target ladder network models. Due to the complexity of the self change of the remote sensing image and the limitation of the label, the network model has serious overfitting to the change detection result, and the limitation of limited label information is difficult to break through by a single semi-supervised change detection structure. Therefore, the invention proposes to adopt a semi-supervised and unsupervised simultaneous training mode. In practice, the data size of the vector data of the labeled sample is not large, but the data size of the vector data of the unlabeled sample is large, if the vector data of the unlabeled sample is input as the training sample at one time, although the accuracy of the change detection result of the method is not influenced, the calculation amount is large, and the training speed is seriously influenced. Therefore, before performing semi-supervised and unsupervised training on the ladder network model, the method further comprises the following steps:

dividing the unlabeled sample vector data into a plurality of batches; wherein the size of the unlabeled sample vector data of each batch is close to that of the labeled sample vector data; respectively combining the unlabeled sample vector data and the labeled sample vector data of each batch into training data, and performing semi-supervised and unsupervised training on the ladder network model by using the training data to obtain a network loss result of the corresponding ladder network model; determining a corresponding optimized ladder network model according to all network loss results; the optimized ladder network model determined by training each target ladder network model is a plurality of target ladder network models, and a middle optimal ladder network model is determined from the plurality of target ladder network models; training the intermediate optimal ladder network model to determine an optimized ladder network model as a target optimal ladder network model; and the iteration threshold adopted when each target ladder network model is trained is smaller than the iteration threshold adopted when the intermediate optimal ladder network model is trained.

Therefore, the unlabeled sample vector data are divided into a plurality of batches, the unlabeled sample vector data and the labeled sample vector data of each batch are used as training data to perform semi-supervised and unsupervised training, the data amount is small during training, and the training speed is high. When training is performed by using training data composed of unlabeled sample vector data and labeled sample vector data of each batch, semi-supervised and unsupervised training is performed on the ladder network model according to the unlabeled sample vector data and the labeled sample vector data, please refer to fig. 5, which includes the following steps:

s501, a first encoder in the ladder network model is used for conducting noise adding compression on the unlabeled sample vector data and the labeled sample vector data to obtain noise samples.

S502, calculating the difference between the output result of the first encoder corresponding to the labeled sample vector data in the noise sample and the judgment result in the labeled sample vector data to obtain a supervision loss result.

S503, denoising an output result of the first encoder in the ladder network model through a decoder in the ladder network model to obtain a denoised sample.

S504, pure compression is carried out on the unlabeled sample vector data and the labeled sample vector data by utilizing a second encoder in the ladder network model to obtain a plurality of noiseless samples.

S505, reconstructing the denoised sample and each noiseless sample by using a decoder in the ladder network model, and calculating the difference between a reconstruction result and each noiseless sample to obtain a plurality of unsupervised loss results;

and S506, calculating a network loss result corresponding to the ladder network model according to the supervision loss result and all unsupervised loss results.

S501 and S504 can respectively adopt the existing noise adding and denoising modes to carry out noise adding compression and pure compression on the unlabeled sample vector data and the labeled sample vector data, noise is added on the last network layer of the first encoder to obtain a noise added sample, and other network layers except the first layer in the second encoder are denoised to obtain a plurality of noise free samples; s503 may perform denoising processing on the noise sample by using a corresponding decoding manner, where the noise sample is the noisy data including the unlabeled sample vector data and the labeled sample vector data.

In S502, the output result of the first encoder is the output corresponding to the last network layer, and the result of the calculated supervision loss is recorded as L in the input data corresponding to the batch in the embodiment of the present invention _c The result of the supervision loss L _c The formula is expressed as:

L _c (x _label )＝-logP(g(x _label )|x _label ) (1)

wherein L is _c Representing the supervision loss result, x, corresponding to the output result of the first encoder in the ladder network model _label Indicates the judgment result, g (x), in the labeled sample vector data _label ) Denotes x _label The output result of the first encoder in the input post-staircase network model, P (-) represents a priori x _label Lower g (x) _label ) Average negative log probability of.

In S505, the output result of the second encoder is the output corresponding to the other network layers except the first layer, and the result of the unsupervised loss calculated in the embodiment of the present invention under the input data corresponding to the batch is recorded as the unsupervised loss resultL _u The unsupervised loss result L _u The formula is expressed as:

wherein λ is _i Representing an unsupervised penalty factor, m, for the i-layer network layer of the second encoder in the ladder network model _i Representing the number of nodes of the i-layer network layer of the second encoder in the ladder network model, x representing the combined data of the labeled sample vector data and the unlabeled sample vector data, (x-N) _noise ) _i Representing a noise-free sample, N, obtained after x is input through an i-layer network layer of a second encoder in the ladder network model _noise Representing Gaussian noise following a 0-1 Normal distribution, f (x) _i And the output result of the i-th network layer corresponding to the second encoder in the ladder network model after x is reconstructed by the decoder in the ladder network model is represented.

Correspondingly, in S506, under the input data corresponding to the batch, according to the result L of monitoring the loss _c And all unsupervised loss results L _u And calculating a network loss result corresponding to the ladder network model, wherein the formula is as follows:

performing the semi-supervised and unsupervised training of the above-mentioned S501-S506 on the unlabelled sample vector data and the labeled sample vector data of all batches to obtain the network loss result of the corresponding ladder network model, and determining the corresponding optimized ladder network model according to all the network loss results, wherein all the network loss result formulas are expressed as:

wherein L (-) represents the net loss result of the ladder net model, L _c Representing a first code in a ladder network modelThe monitor outputs a monitoring loss result corresponding to the result, l represents the network layer number of the second encoder in the ladder network model,

representing the output result of the i-layer network layer except the first layer of the second encoder in the ladder network model, n representing the batch into which the unlabeled sample vector data is divided, x ^j _label Indicates the judgment result, g (x), in the labeled sample vector data corresponding to the jth batch ^j _label ) Denotes x ^j _label The output result of the first encoder in the post-staircase network model is input, _P (. represents a priori x) ^j _label Lower g (x) ^j _label ) Average negative logarithmic probability of λ _i Representing an unsupervised penalty factor, m, for the i-layer network layer of the second encoder in the ladder network model _i Representing the number of nodes of the i-layer network layer of the second encoder in the ladder network model,

representing a noise-free sample obtained by the training data corresponding to the jth batch through the ith network layer of the second encoder in the ladder network model, N _noise Representing Gaussian noise that follows a normal distribution of 0 to 1,

and representing the output result of the i-layer network layer of the second encoder in the corresponding ladder network model after the training data corresponding to the jth batch is reconstructed by the decoder in the ladder network model.

By using the above manner, semi-supervised and unsupervised training is performed on all target ladder network models obtained by searching of the evolutionary algorithm to obtain network loss results of the corresponding ladder network models, and a middle optimal ladder network model is determined according to all network loss results, that is, the middle optimal ladder network model is determined from a plurality of target ladder network models, and the specific process please refer to fig. 6 and includes the following steps:

s601, calculating average difference levels corresponding to all target ladder network models.

Specifically, the maximum value and the minimum value of the network loss result are selected from the network loss results corresponding to all the target ladder network models, and the average difference level corresponding to all the target ladder network models is calculated according to the maximum value and the minimum value, wherein the average difference level formula is represented as follows:

α＝(L(x；x _label ) _max -L(x；x _label ) _min )/N (5)

wherein alpha represents the average difference level, L (x; x) _label ) _max Representing the maximum value of the net loss result, L (x; x _label ) _min And N represents the number of the ladder network models searched by the evolutionary algorithm network. The two target ladder network model basic difference indices are evaluated by using the average difference level α, which ensures that each selected target ladder network model is above the average difference level.

S602, determining the intermediate optimal ladder network model according to the average difference level, the network loss results of all target ladder network models and the unsupervised loss result of the target ladder network model.

Specifically, all target ladder network models are denoted as A ₁ 、A ₁ 、……、A _N And the network loss result corresponding to each target ladder network model is recorded as

Unsupervised loss results are reported

Determining the formula of the intermediate optimal ladder network model according to the average difference level alpha, the network loss results of all target ladder network models and the unsupervised loss results of the target ladder network models as follows:

wherein, (L) _B 、(L) _A Is from A ₁ 、A ₁ 、……、A _N The network loss results corresponding to the two selected target ladder network models in (L) _u ) _B 、(L _u ) _A For the corresponding unsupervised loss results, it can be seen that: the embodiment of the invention calculates the difference (L) of the network loss results corresponding to the two target ladder network models _B -(L) _A Judging the difference value (L) _B -(L) _A From the average difference level a, if the difference (L) _B -(L) _A If the average difference level alpha is larger than or equal to the average difference level alpha, the target ladder network model A corresponding to the smaller network loss result is taken as the current intermediate optimal ladder network model, otherwise, the unsupervised loss results corresponding to the two target supervised network models are judged, and the smaller unsupervised loss result (L) is taken as the current intermediate optimal ladder network model _u ) _A And selecting the corresponding target ladder network model A as the current intermediate optimal ladder network model, and selecting the target ladder network model B as the current intermediate optimal ladder network model under other conditions.

Next to it, from A ₁ 、A ₁ 、……、A _N Continuously selecting a target ladder network model, continuously comparing the target ladder network model with the current intermediate optimal ladder network model by using a formula (6) to determine the intermediate optimal ladder network model in the round, and repeating the process until the comparison A is finished ₁ 、A ₁ 、……、A _N And (4) obtaining the final intermediate optimal ladder network model by using all the target ladder network models.

And S60, performing semi-supervised and unsupervised training on the intermediate optimal ladder network model to obtain a target optimal ladder network model, wherein an output result corresponding to the target optimal ladder network model is a change detection result.

Specifically, the optimal target ladder network model is obtained by training in the same supervised and unsupervised training mode for each target ladder network model, and the specific training process is not repeated here. In order to prevent the phenomenon of overfitting of change detection of the remote sensing image after model training, the iteration frequency of each target step network model during training is lower than that of the intermediate optimal step network model during network training.

And finally, classifying the output result of the first encoder in the target optimal ladder network model to obtain a final change detection result. In the final change detection process, the output end of the first encoder is connected with a classification layer to realize the classification of the output result of the first encoder, and the first encoder in the target optimal ladder network model does not have noise interference.

Further, in order to improve the change detection performance of the network, the invention provides a new strategy of circular self-training, which utilizes more pseudo label sample data to guide the training of the network again, and the network training introduces richer prior knowledge, thereby achieving the purpose of optimizing the network structure. And, referring to fig. 7, the process of guiding the network training again by using more pseudo label sample data includes the following steps:

s701, classifying the output result of the first encoder in the target optimal ladder network model to obtain a semi-supervised classification result.

Specifically, in the embodiment of the present invention, the output end of the first encoder in the target optimal ladder network model is connected to a classification layer, so as to classify the output result of the first encoder, obtain a semi-supervised classification result, and record the semi-supervised classification result as CM ₀ . The semi-supervised process extracts high-order features which are more prone to being classified under the guidance of labeled sample data.

S702, classifying the output results of each layer except the first layer of network layer of the second encoder in the target optimal ladder network model to obtain a plurality of unsupervised classification results.

Specifically, the unsupervised process is applied to extract high-order features from each layer other than the first layer, thereby suppressing the influence of redundant information and noise and being more suitable for classification. Therefore, in the unsupervised process, the embodiment of the present invention adopts, but is not limited to, a Compressed change vector analysis (C for short) ² VA) algorithm, compressing pixel space toAnd (3) in a polar coordinate space, classifying the categories through a phase angle range, and finally extracting high-order features from each layer except the first layer to form a plurality of unsupervised classification results which are recorded as: CM (compact message processor) ₁ 、CM ₂ 、……、CM _l-1 And l is the number of layers of the second encoder. For example, in fig. 8, if the number of network layers of the second encoder is 4, the unsupervised classification result CM obtained by classification is obtained ₁ 、CM ₂ 、CM ₃ 。

And S703, fusing the semi-supervised classification result and the plurality of unsupervised classification results.

Specifically, the semi-supervised classification result and the plurality of unsupervised classification results are fused, and the fusion is to select the voting for each pixel point at the same position on the semi-supervised classification result and the plurality of unsupervised classification result graphs, so that more reliable detection classification results are screened. The mode of obtaining the classification result through different angles can enrich the difference information quantity, the semi-supervised classification result of the network is fused with a plurality of unsupervised classification results obtained by an unsupervised method, the relatively reliable classification results are screened by adopting the principle that a small number of classification results obey a majority, and the fused classification results have higher change detection precision.

S704, counting the total scores of all classification results, judging whether the classification result of each pixel point is reliable or not according to the total scores, selecting all reliable label-free sample data in the fusion result as pseudo label sample data, and adding the pseudo label sample data into the labeled sample data to form new labeled sample data.

Specifically, the total scores of all classification results are counted, and the purpose of the statistics is only to judge whether each pixel point is changed or not, so that the score values assigned to all classification results are not limited specifically, and only for simplicity, the same score assignment is performed on the semi-supervised classification result and the unsupervised classification result, for example, the score is assigned to be 1 uniformly. The statistical total score NL can be formulated as:

where l is the number of network layers of the second encoder, nl _i A score value, nl, representing the value assigned to the unsupervised classification result corresponding to the i-th layer network layer in the second encoder _semi And the score value assigned to the semi-supervised classification result is shown.

And for the scoring condition of the changed region and the unchanged region, the formula is expressed as:

as can be seen from the formula (8), for any pixel point, when the pixel point changes, the statistical score is 1, and when the pixel point does not change, the statistical score is 0. The following rules are used to filter reliable varying and invariant pseudo-label sample data, the formula is expressed as:

the embodiment of the invention combines a semi-supervised classification result and an unsupervised classification result to select pseudo label sample data, when the total score NL counted in the formula (9) is l, the pixel point is reliably changed, when the total score NL counted is 0, the pixel point is not changed, and when the total score NL counted is 1-l-1, the pixel point is uncertain. In the embodiment of the invention, when the statistical total score NL is l and serves as a condition for generating the pseudo-label sample data, all reliable non-label sample data in the classification result obtained by fusion are marked to form the pseudo-label sample data and are marked as the pseudo-label sample data when the statistical total score NL is met

Sample data of pseudo label

Added to a labeled sampleThis data X _labeled In (2), new tagged sample data is formed

At this time, the corresponding unlabeled sample data is still X _unlabeled 。

S705, semi-supervised and unsupervised training is carried out on the target optimal ladder network model according to the unlabelled sample data and the new labeled sample data to obtain an updated target optimal ladder network model, and the output corresponding to the updated target optimal ladder network model is a detection result.

Specifically, referring to fig. 8, the target optimal ladder network model is semi-supervised and unsupervised trained by using the new labeled sample data and the unlabeled sample data obtained in S704, and the specific training process refers to the above-mentioned training process for each target ladder network model, which is not described herein again. At this time, the number of iterations used in the training still needs to be greater than the number of iterations used in the training of each target ladder network model.

Referring to fig. 9, a remote sensing image change detection process provided by the embodiment of the invention is completely shown, unlabeled sample vector data and labeled sample vector data are formed through S10 to S30, and a corresponding change detection result is output according to a target optimal ladder network model determined through S40 to S60, so that remote sensing image change detection is realized. And generating pseudo label sample data by using unsupervised classification and semi-supervised classification, and guiding the optimal target ladder network model training again by using the pseudo label sample data in combination with the graph 8 so as to realize more accurate remote sensing image change detection.

It should be noted that, in order to ensure that the target optimal ladder network model determined each time can better meet the change detection requirement of the current scene, the evolutionary algorithm can be adopted to preferentially select a plurality of suitable target ladder network models each time change detection is performed, then the final target optimal ladder network model is determined through the semi-supervised and unsupervised training processes, and high-precision change detection is realized through the target optimal ladder network model.

In order to verify the effectiveness of the remote sensing image change detection method provided by the embodiment of the invention, the verification is performed through the following experiment.

1. Simulation conditions

Example of the invention simulation experiments were conducted in an Intel (R) core (TM) i7-4790 CPU @3.60GHz Windows 10, Python3.6.2, Tensorflow1.3.2 environment.

2. Evaluation index

For the simulation experiment, qualitative and quantitative analysis are used to evaluate the performance of the algorithm, and the main evaluation indexes used in the quantitative analysis are as follows:

(ii) a correct classification (PCC) formula:

wherein, the True rate (TP) represents the number of pixels for correctly detecting the change area; a True Negative rate (TN) indicating the number of pixels for correctly detecting an unchanged area; a False Positive (FP) indicating that an originally unchanged area is detected as a change class; the False Negative (FN) indicates that an originally changing area is not detected.

Measuring the Kappa coefficient of the consistency of the simulation experiment result graph and the change reference graph, wherein the formula is as follows:

wherein M is _c And M _u The number of pixels representing the actual changed and unchanged areas, respectively; PRE represents the desired coherency ratio.

3. Content of simulation experiment

The simulation experiment is carried out on different types of remote sensing image data sets by using the existing method, and the change detection method provided by the invention is suitable for various remote sensing image types, wherein heterogeneous images and multispectral images are selected as examples to carry out the simulation experiment.

The first group used in the simulation is heterogeneous images as shown in fig. 10(a) to 10(c), fig. 10(a) is an SAR remote sensing image captured in the region of wuhan city in 6 months 2008, and the size of the image is 503 × 495; fig. 10(b) is an RGB three-channel optical image of the same size captured in the same region of 9 months in 2012; fig. 10(c) is a variation reference diagram between fig. 10(a) and 10(b), and a white area indicates a variation region and a black area indicates an unchanged region.

The second set used in the simulation is the multispectral images shown in fig. 11(a) to 11(c), fig. 11(a) is the multispectral image of the red flag canal area shot in 12, 9 and 2013, and the size of the image is 539 × 543; fig. 11(b) is a multispectral image of the same area and the same size taken 10, 16 days 2015; fig. 11(c) is a variation reference diagram of fig. 11(a) and 11(b), and a white area indicates a variation region and a black area indicates an unchanged region.

First simulation, change detection simulation is performed on a group of remote sensing images shown in fig. 10(a) to 10(c) by using the method of the present invention and two existing methods, and change detection results are shown in fig. 12(a) to 12(d), where fig. 12(a) is a change detection result of a Residual Network algorithm (Resnet for short), fig. 12(b) is a change detection result of a Semi-Supervised generated countermeasure Network algorithm (Semi-Supervised general adaptive Network for short), fig. 12(c) is a change detection result of the method of the present invention, and fig. 12(d) is a change detection reference diagram.

The results of quantitative evaluation analysis obtained by analyzing the change detection simulation experiment data of the heterogeneous images shown in fig. 10(a) to 10(c) on the Wuhan heterogeneous image dataset are shown in table 1.

TABLE 1 determination, evaluation and analysis results of Wuhan heterogeneous image data under different algorithms

Meanwhile, please refer to fig. 12(a) to fig. 12(d), with reference to fig. 12 (d): both the Resnet and Semi-GAN methods shown in FIGS. 12(a) to 12(b) inevitably cause noise problems, and the detection performance of the unchanged area and the changed area is seriously reduced, so that the Kappa coefficient scores of the methods are lower; the method of the present invention shown in fig. 12(c) suppresses noise interference to a certain extent, and the detection performance of the unchanged region and the changed region is improved compared with other algorithms, and the corresponding scores of the PCC and Kappa coefficients are also improved. In general, the method of the present invention has the best detection performance.

And secondly, performing change detection simulation on a group of remote sensing images shown in fig. 11(a) to 11(c) by using the method of the invention and two conventional methods, wherein the change detection results are shown in fig. 13(a) to 13(d), wherein fig. 13(a) is the change detection result of a residual error network algorithm Resnet, fig. 13(b) is the change detection result of a Semi-supervised generation antagonistic network algorithm Semi-GAN, fig. 13(c) is the change detection result of the method of the invention, and fig. 13(d) is a change detection reference diagram.

The results of quantitative evaluation analysis obtained by analyzing the data of the change detection simulation experiment of the heterogeneous images shown in fig. 11(a) to 11(c) on the Hongqi multispectral image dataset are shown in table 2.

TABLE 2 Hougqi multispectral image dataset theoretic evaluation analysis results under different algorithms

Meanwhile, referring to fig. 13(a) to 13(d), referring to fig. 13(d), it can be seen that the difficulty in detecting the Hongqi multispectral image dataset is that it has a large number of linear change regions (four elongated lines in the middle region), which greatly test the edge generalization ability of the change detection method: both the Resnet and Semi-GAN approaches shown in fig. 13(a) -13 (b) inevitably suffer from noise problems, especially in the unchanged area, and details still need to be enhanced, so both PCC and Kappa coefficients are low; the method of the invention shown in fig. 13(c) can keep the best details in the whole remote sensing image, and well inhibit noise in both the changed area and the unchanged area, thereby achieving the best PCC and Kappa coefficients.

Through the simulation experiment analysis, it can be seen that the method has better classification performance and higher change detection precision for the problem of remote sensing image change detection no matter the Wuhan heterogeneous image data set or the Hongqi multispectral image data set, and is superior to the method widely used at present.

In summary, the embodiment of the present invention provides a method for detecting a change in a remote sensing image, which is a new strategy that combines labeled sample data and unlabeled sample data, and the strategy combines the advantages of semi-supervised and unsupervised change detection methods, wherein the labeled sample data is used in the semi-supervised change detection, the labeled sample data and the unlabeled sample data are used in the unsupervised change detection, and the labeled sample data and the unlabeled sample data are trained to obtain a good structure for representing information of the remote sensing image, so that the accuracy of detecting a change in a remote sensing image is improved, and the labeled sample data is directly subjected to generalized learning in a network, so that an additional learning of a labeled sample data structured model is not required, and a network model is simplified; the embodiment of the invention adopts an evolutionary algorithm to search a network to obtain a plurality of target ladder network models, then trains each target ladder network model by adopting semi-supervised and unsupervised training modes, in each network cycle iteration, the structure and parameters of the target ladder network models can be automatically adjusted to adapt to the most effective ladder network to obtain an intermediate optimal ladder network model, the intermediate optimal ladder network model is again subjected to semi-supervised and unsupervised training to realize the final fine tuning of the network to obtain the target optimal ladder network model, the change detection of remote sensing images is realized through the target optimal ladder network model, the automatic network adjustment avoids the manual network configuration, and can also adapt to the condition that the change of a detection object causes the change of difference information between objects and the scale change of data to be processed, and the change detection method has stronger universality, the semi-supervised and unsupervised double prediction results are beneficial to supplementing classification information of different angles, and better remote sensing image change detection results can be obtained.

Meanwhile, the embodiment of the invention also provides that more pseudo label sample data are utilized to guide the training of the ladder network model again in the cyclic detection, and the self-training mode can improve the classification performance of the ladder network model and obtain a better remote sensing image change detection result due to the fact that richer prior knowledge is brought.

In the description of the present invention, it is to be understood that the terms "first", "second" and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implying any number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the present invention, "a plurality" means two or more unless specifically defined otherwise.

While the present application has been described in connection with various embodiments, other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed application, from a review of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.

Claims

1. A method for detecting changes of remote sensing images is characterized by comprising the following steps:

2. The method for detecting remote sensing image change according to claim 1, wherein the performing vector format conversion on the unlabeled sample data and the labeled sample data to form unlabeled sample vector data and labeled sample vector data respectively comprises:

aiming at each image block pair, expanding pixel points of each image block in the image block pair according to rows to obtain two pixel point vectors;

longitudinally connecting the two pixel point vectors to form a new pixel point vector; all new pixel point vectors formed by the unlabeled sample data form the unlabeled sample vector data;

for the tagged sample data, further comprising:

meanwhile, judging whether all pixel point pairs change or not;

3. The remote sensing image change detection method according to claim 2, wherein all ladder network models include a first encoder, a second encoder, and a decoder.

4. The remote sensing image change detection method according to claim 3, wherein the first encoder, the second encoder, and the decoder employ the same network structure; the same network structure comprises the network layers and the nodes in the network structure which are respectively corresponding to the same number.

5. The method for detecting changes in remote sensing images of claim 4, wherein the step of searching the network structure using the evolutionary algorithm to obtain a plurality of target ladder mesh models comprises:

selecting a plurality of initial ladder network models as initial populations;

6. The remote sensing image change detection method according to claim 3, wherein the process of semi-supervised and unsupervised training of the ladder network model comprises:

7. The method for detecting changes in remote sensing images as claimed in claim 6, wherein prior to semi-supervised and unsupervised training of the ladder network model, further comprising:

respectively combining the unlabelled sample vector data and the labeled sample vector data of each batch into training data, and performing semi-supervised and unsupervised training on the ladder network model by using the training data to obtain a corresponding network loss result of the ladder network model;

8. The method for detecting changes in remote sensing images of claim 7, wherein the process of determining the intermediate optimal ladder network model from the plurality of target ladder network models comprises:

9. The remote sensing image change detection method according to claim 3, wherein the output result corresponding to the target optimal ladder network model is a change detection result, and comprises:

10. The remote sensing image change detection method according to claim 3, further comprising: