CN112577473A - Double-time-phase high-resolution remote sensing image change detection algorithm - Google Patents

Abstract

The invention discloses a double-time-phase high-resolution remote sensing image change detection algorithm, which relates to the technical field of remote sensing mapping geographic information and comprises the following steps: A. acquiring an image: the satellite photographing unit acquires multi-time remote sensing image data and carries out registration; B. preprocessing of the image: and performing vector, grid and PNG format conversion on each IMG source data picture. The double-time-phase high-resolution remote sensing image change detection algorithm can obtain multi-scale features by combining the features of different areas, the image space is averagely divided into sub-areas under a certain scale under the drive of the motivation, a self-attention mechanism is introduced into each sub-area, and the multi-scale feature representation can be obtained by dividing the image into the multi-scale sub-areas by utilizing the space-time relationship of an object under the scale so as to better adapt to the scale of a target.

Description

Double-time-phase high-resolution remote sensing image change detection algorithm

Technical Field

The invention relates to the technical field of remote sensing mapping geographic information, in particular to a double-time-phase high-resolution remote sensing image change detection algorithm.

Background

The change detection has great significance in the interpretation of the remote sensing image, the high-resolution remote sensing image greatly improves the land utilization and the covering change monitoring capability, the multi-temporal high-resolution remote sensing image contains complex geographic information elements, the full-automatic change detection of the remote sensing image is carried out by utilizing deep learning, the two years of rapid development, the precision and the efficiency far exceed the traditional method, the patent provides a deep convolution neural network aiming at a target change area, the end-to-end change detection function of a multiresolution remote sensing image is realized, 97% change detection accuracy is obtained on the multi-temporal data set of the remote sensing satellite, and experimental results show that, the network has high accuracy and good generalization capability, can be used for actual production, and further replaces manual and accurate finding of image change areas.

Most of the conventional change detection methods construct information features from the original data of the cell, where image pixels and image objects are two main categories of the analysis cell, and then judge the change category change identification by comparing the feature representation of the analysis cell, and the general method is to calculate a feature difference map and segment the change area by a threshold.

A change detection method based on deep learning is characterized in that whether change occurs or not is determined by comparing parameterized distances of double time phase data, a deep full convolution network can learn an embedding space and comprises two identical networks sharing the same weight, each network independently generates a feature map of each time image, then whether change occurs or not is judged through measurement between each pair of feature points, and in the training process of a model, different loss functions are tried to obtain a better detection effect.

The general method is that each position of an image is assigned with a change score, wherein the score of the changed position is higher than that of the position without change, the problem of class imbalance is solved through the weight setting of the class, a CNN method and an RNN method are adopted to capture sequence relation and model time dependency, and some models combining the CNN and the RNN are difficult to effectively detect the changed region through photographic data with different size resolutions.

Disclosure of Invention

Technical problem to be solved

Aiming at the defects of the prior art, the invention provides a double-time-phase high-resolution remote sensing image change detection algorithm, aiming at solving the problems in double-time-phase change detection, wherein a change object area can have different sizes, and multi-scale features can be obtained by combining the features of areas with different sizes.

(II) technical scheme

In order to achieve the above purposes, the technical scheme adopted by the invention is as follows: a double-time-phase high-resolution remote sensing image change detection algorithm comprises the following steps:

A. acquiring an image: the satellite photographing unit acquires multi-time remote sensing image data and carries out registration;

B. preprocessing of the image: performing vector, grid and PNG format conversion on each IMG source data picture;

C. data set partitioning: randomly dividing the preprocessed pictures into a training set, a testing set and a verification set according to the ratio of 6: 3: 1;

D. selecting a model evaluation index: setting evaluation indexes such as OA, Kappa and the like aiming at the change detection problem;

E. model training: deep learning and training a neural network, and optimizing parameters such as BN, Batch Size, crop, flit and the like;

F. and (3) prediction: and selecting an optimal model, inputting the first-stage image and the second-stage image of the test, and outputting a change detection result.

Preferably, the specific implementation process of the steps A-F is as follows;

s1, reading the Image data Image and corresponding Label data;

s2, calculating an output size ratio (output stride 16), and the expansion rate of the last stage is 2;

s3, because ASPP has four different rates, a global average pooling is additionally performed;

s4, the result of the encoder is up-sampled by 4 times, then the up-sampled result is convolved by 3x3 together with Conv2 characteristic concat before down-sampling in resnet, and finally the up-sampled result is up-sampled by 4 times to obtain a final result;

and S5, before fusing the low-level information, performing convolution of 1x1 to reduce channels.

(III) advantageous effects

The invention has the beneficial effects that:

the double-time-phase high-resolution remote sensing image change detection algorithm can obtain multi-scale features by combining the features of different areas, the image space is averagely divided into sub-areas under a certain scale under the drive of the motivation, a self-attention mechanism is introduced into each sub-area, the space-time relationship of an object under the scale is utilized, the multi-scale feature representation can be obtained by dividing the image into the multi-scale sub-areas, the target scale change can be better adapted, the problem of two-time-phase multi-resolution image change detection can be remarkably solved, and the OA index of the model is remarkably improved.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a Kappa coefficient formula diagram according to the present invention;

FIG. 3 is a Tverseky Loss formula diagram of the present invention;

FIG. 4 is a first no building detection diagram of the building change detection results of the present invention;

fig. 5 is a second-stage newly-built building detection diagram of the building change detection result of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

As shown in fig. 1 to 5, the present invention provides a technical solution: a double-time-phase high-resolution remote sensing image change detection algorithm comprises the following steps:

A. acquiring an image: the satellite photographing unit acquires multi-time remote sensing image data and carries out registration;

B. preprocessing of the image: performing vector, grid and PNG format conversion on each IMG source data picture;

C. data set partitioning: randomly dividing the preprocessed pictures into a training set, a testing set and a verification set according to the ratio of 6: 3: 1;

D. selecting a model evaluation index: setting evaluation indexes such as OA, Kappa and the like aiming at the change detection problem;

E. model training: deep learning and training a neural network, and optimizing parameters such as BN, Batch Size, crop, flit and the like;

F. and (3) prediction: and selecting an optimal model, inputting the first-stage image and the second-stage image of the test, and outputting a change detection result.

The specific implementation process of the steps A-F is as follows;

s1, reading the Image data Image and corresponding Label data;

s2, calculating an output size ratio (output stride 16), and the expansion rate of the last stage is 2;

s3, because ASPP has four different rates, a global average pooling is additionally performed;

s4, the result of the encoder is up-sampled by 4 times, then the up-sampled result is convolved by 3x3 together with Conv2 characteristic concat before down-sampling in resnet, and finally the up-sampled result is up-sampled by 4 times to obtain a final result;

s5, before fusing the low-level information, performing convolution of 1x1 to reduce channels, for example, 512 channels, and the encoder result only has 256 channels.

The above-mentioned embodiments, objects, technical solutions and advantages of the present invention are further described in detail, it should be understood that the above-mentioned embodiments are only illustrative of the present invention and are not intended to limit the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (2)

1. A double-time-phase high-resolution remote sensing image change detection algorithm comprises the following steps:

A. acquiring an image: the satellite photographing unit acquires multi-time remote sensing image data and carries out registration;

B. preprocessing of the image: performing vector, grid and PNG format conversion on each IMG source data picture;

C. data set partitioning: randomly dividing the preprocessed pictures into a training set, a testing set and a verification set according to the ratio of 6: 3: 1;

D. selecting a model evaluation index: setting evaluation indexes such as OA, Kappa and the like aiming at the change detection problem;

E. model training: deep learning and training a neural network, and optimizing parameters such as BN, Batch Size, crop, flit and the like;

F. and (3) prediction: and selecting an optimal model, inputting the first-stage image and the second-stage image of the test, and outputting a change detection result.

2. The dual-temporal high-resolution remote sensing image change detection algorithm according to claim 1, characterized in that: the specific implementation process of the steps A-F is as follows;

s1, reading the Image data Image and corresponding Label data;

s2, calculating an output size ratio (output stride 16), and the expansion rate of the last stage is 2;

s3, because ASPP has four different rates, a global average pooling is additionally performed;

s4, the result of the encoder is up-sampled by 4 times, then the up-sampled result is convolved by 3x3 together with Conv2 characteristic concat before down-sampling in resnet, and finally the up-sampled result is up-sampled by 4 times to obtain a final result;

and S5, before fusing the low-level information, performing convolution of 1x1 to reduce channels.

Images