A dual-temporal high-resolution remote sensing image change detection algorithm

Abstract

The present invention discloses a dual-phase high-resolution remote sensing image change detection algorithm, which relates to the field of remote sensing mapping geographic information technology, and includes the following steps: A. Image acquisition: a satellite camera unit acquires multi-time remote sensing image data and performs registration; B. Image preprocessing: each inertial measurement group source data picture is converted into vector, raster, and PNG formats. The dual-phase high-resolution remote sensing image change detection algorithm can obtain multi-scale features by combining features of regions of different sizes. Driven by this motivation, the image space is evenly divided into sub-regions at a certain scale, and a self-attention mechanism is introduced in each sub-region. By utilizing the spatiotemporal relationship of objects at this scale, the image is divided into multi-scale sub-regions to obtain multi-scale feature representations to better adapt to the scale of the target.

Description

Technical Field

[0001] The invention relates to the field of remote sensing mapping geographic information technology, in particular to a dual-temporal high-resolution remote sensing image change detection algorithm. Background Art

[0002] Change detection is of great significance in remote sensing image interpretation. High-resolution remote sensing images have greatly improved the ability to monitor land use and land cover changes. Multi-temporal high-resolution remote sensing images contain complex geographic information elements. The use of deep learning for fully automatic change detection in remote sensing images has developed rapidly in the past two years, and its accuracy and efficiency have far exceeded traditional methods. For the problem of dual-temporal high-resolution remote sensing change detection, this patent proposes a deep convolutional neural network for target change areas, which realizes end-to-end change detection functions for multi-resolution remote sensing images and achieves 97% change detection accuracy on public remote sensing satellite multi-temporal data sets. Experimental results show that the network has a high accuracy rate and good generalization ability, and can be used in actual production, thereby replacing manual work to more accurately find image change areas.

[0003] Most traditional change detection methods first construct information features from the raw data of the unit. Image pixels and image objects are the two main categories of analysis units. Then, the change category is judged by comparing the feature representations of the analysis units. The general method is to calculate the feature difference map and segment the change area by threshold.

[0004] The change detection method based on deep learning is to determine whether a change has occurred by comparing the parameterized distance of the dual-phase data. The deep fully convolutional network can learn the embedding space. It contains two identical networks that share the same weights. Each network independently generates a feature map for each time image, and then determines whether a change has occurred through the measurement between each pair of feature points. During the training process of the model, different loss functions will be tried in the hope of obtaining better detection results.

[0005] The change category is identified by classifying the extracted bi-temporal data features. The general method is to assign a change score to each position of the image, where the score of the changed position is higher than that of the unchanged position. The category imbalance problem is solved by setting the category weights. The CNN method and RNN method are adopted to capture the sequence relationship and model the time dependency. Some models combining CNN and RNN have difficulty in effectively detecting the changed area of ​​image data of different size resolutions. Summary of the invention

[0006] 1. Technical issues to be resolved

[0007] In view of the shortcomings of the prior art, the present invention provides a dual-phase high-resolution remote sensing image change detection algorithm, the purpose of which is to address the problems existing in dual-phase change detection. The changed object area may have different sizes. By combining the features of areas of different sizes, multi-scale features can be obtained. Driven by this motivation, the image space is evenly divided into sub-areas of a certain scale, and a self-attention mechanism is introduced in each sub-area. By utilizing the spatiotemporal relationship of objects at this scale, by dividing the image into multi-scale sub-areas, multi-scale feature representation can be obtained to better adapt to the scale change of the target.

[0008] (II) Technical solution

[0009] To achieve the above objectives, the technical solution adopted by the present invention is: a dual-phase high-resolution remote sensing image change detection algorithm, the method comprising the following steps:

[0010] A. Image acquisition: The satellite camera unit collects remote sensing image data from multiple times and performs registration;

[0011] B. Image preprocessing: convert each inertial measurement group source data image into vector, raster, and PNG formats;

[0012] C. Dataset division: The preprocessed images are randomly divided into training set, test set and validation set in a ratio of 6:3:1;

[0013] D. Select model evaluation indicators: set overall classification accuracy and Kappa coefficient evaluation indicators for change detection problems;

[0014] E. Model training: deep learning training of neural networks, standard batching, batch size, cropping, and flipping parameter tuning;

[0015] F. Prediction: Select the optimal model, input the first and second phase images of the test, and output the results of change detection.

[0016] The specific implementation process of steps AF is as follows:

[0017] S1, read image data Image and corresponding marking data;

[0018] S2, calculate the ratio of input size to output size (output step size = 16), and the dilation rate of the last kernel parameter of the dilated convolution is 2;

[0019] S3: Because the atrous spatial pyramid module has four different atrous convolutions, an additional global average pooling is performed;

[0020] S4, first upsample the encoder result by 4 times, then align it with the feature row of the 2D matrix convolution operation before downsampling in the residual network, then perform 3x3 convolution, and finally upsample by 4 times to get the final result;

[0021] S5. Before fusing low-level information, a 1x1 convolution is performed to reduce the number of channels.

[0022] (III) Beneficial effects

[0023] The beneficial effects of the present invention are:

[0024] This dual-phase high-resolution remote sensing image change detection algorithm can obtain multi-scale features by combining the features of regions of different sizes. Driven by this motivation, the image space is evenly divided into sub-regions of a certain scale, and a self-attention mechanism is introduced in each sub-region. The spatiotemporal relationship of objects at this scale is utilized. By dividing the image into multi-scale sub-regions, a multi-scale feature representation can be obtained to better adapt to the scale changes of the target. This can significantly solve the problem of change detection in dual-phase multi-resolution images, and the model OA index is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the work flow chart of the present invention;

[0026] Figure 2 It is the Kappa coefficient formula diagram of the present invention;

[0027] Figure 3 It is the Tversky Loss formula diagram of the present invention;

[0028] Figure 4 A first phase of non-building detection diagram of the building change detection result of the present invention;

[0029] Figure 5 This is the second phase new building detection map of the building change detection result of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] like Figure 1-5 As shown, the present invention provides a technical solution: a dual-phase high-resolution remote sensing image change detection algorithm, the method comprising the following steps:

[0032] A. Image acquisition: The satellite camera unit collects remote sensing image data from multiple times and performs registration;

[0033] B. Image preprocessing: convert each inertial measurement group source data image into vector, raster, and PNG formats;

[0034] C. Dataset division: The preprocessed images are randomly divided into training set, test set and validation set in a ratio of 6:3:1;

[0035] D. Select model evaluation indicators: set overall classification accuracy and Kappa coefficient evaluation indicators for change detection problems;

[0036] E. Model training: deep learning training of neural networks, standard batching, batch size, cropping, and flipping parameter tuning;

[0037] F. Prediction: Select the optimal model, input the first and second phase images of the test, and output the results of change detection.

[0038] The specific implementation process of step AF is as follows:

[0039] S1, read image data Image and corresponding marking data;

[0040] S2, calculate the ratio of input size to output size (output step size = 16), and the dilation rate of the last kernel parameter of the dilated convolution is 2;

[0041] S3: Because the atrous spatial pyramid module has four different atrous convolutions, an additional global average pooling is performed;

[0042] S4, first upsample the encoder result by 4 times, then align it with the feature row of the 2D matrix convolution operation before downsampling in the residual network, then perform 3x3 convolution, and finally upsample by 4 times to get the final result;

[0043] S5. Before fusing low-level information, a 1x1 convolution is performed to reduce the number of channels. For example, there are 512 channels, but the encoder result has only 256 channels.

[0044] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A dual-temporal high-resolution remote sensing image change detection algorithm, the method comprising the following steps: A. Image acquisition: The satellite camera unit collects remote sensing image data from multiple times and performs registration; B. Image preprocessing: convert each inertial measurement group source data image into vector, raster, and PNG formats; C. Dataset division: The preprocessed images are randomly divided into training set, test set and validation set in a ratio of 6:3:1; D. Select model evaluation indicators: set overall classification accuracy and Kappa coefficient evaluation indicators for change detection problems; E. Model training: deep learning training of neural networks, standard batching, batch size, cropping, and flipping parameter tuning; F. Prediction: Select the optimal model, input the first and second phase images of the test, and output the change detection results; The specific implementation process of steps AF is as follows: S1, read image data Image and corresponding marking data; S2, calculate the ratio of input size to output size (output step size = 16), and the dilation rate of the last kernel parameter of the dilated convolution is 2; S3: Because the atrous spatial pyramid module has four different atrous convolutions, an additional global average pooling is performed; S4, first upsample the encoder result by 4 times, then align it with the feature row of the 2D matrix convolution operation before downsampling in the residual network, then perform 3x3 convolution, and finally upsample by 4 times to get the final result; S5. Before fusing low-level information, a 1x1 convolution is performed to reduce the number of channels.

Images