Mining area remote sensing change detection method based on IAPM and FDAF

Abstract

The invention relates to the technical field of mining area remote sensing, in particular to a mining area remote sensing change detection method based on IAPM and FDAF, and mainly solves the technical problem that existing mining area remote sensing change detection is low in result accuracy. The method comprises the following steps: S1, acquiring remote sensing images of a mining area under a first time phase and a second time phase; s2, inputting the remote sensing image into the IAPM to obtain a double-time-phase picture after exposure, gamma value, contrast and sharpening degree are self-adapted; s3, inputting the dual-temporal picture into a feature extraction network with ChangeAD for feature extraction to obtain a dual-temporal semantic feature map; s4, inputting the dual-temporal semantic feature map into FDAF to extract change features, and obtaining a segmentation result map; and S5, inputting the double-time-phase image as a guided image and the segmentation result image as an input image into the LGF, and outputting an enhanced segmentation result as a detection result. According to the method, the IAPM, the ChangeAD, the FDAF and the LGF are introduced, so that the accuracy of a change detection result can be effectively improved.

Description

Technical Field

[0001] The invention relates to the field of mining area remote sensing technology, and in particular to a mining area remote sensing change detection method based on IAPM and FDAF. Background Art

[0002] Change detection is an important technology in the field of remote sensing, which is used to identify and quantify the changes in surface cover or objects between different time points. When performing change detection, dual-temporal or multi-temporal remote sensing images are used to compare the before and after states of the same area. Among them, remote sensing change detection in mining areas uses drones to collect remote sensing images in mining areas to compare the before and after states.

[0003] The existing remote sensing change detection in mining areas has the following shortcomings: First, the remote sensing image is not exposed enough, the image is dark overall, and the details are not clear enough; second, the feature interaction is insufficient, and the interaction and combination between different features in the model are less, which is not conducive to capturing complex spatial relationships; third, the feature fusion is insufficient, and features from different sources or different types are not effectively combined, resulting in the model being unable to fully utilize all available information to make judgments; fourth, the segmentation performance of the edge area is insufficient, and it is easy to ignore the important details contained in the area near the boundary of the object in the image; fifth, the alignment of the dual-phase image is insufficient, and the spatial positions of two or more images are inconsistent, resulting in deviations in the originally corresponding positions in the images of different time phases. These shortcomings lead to low accuracy of the change detection results, so there is an urgent need for a remote sensing change detection method for mining areas that can overcome the above shortcomings and improve the accuracy of remote sensing change detection results in mining areas. Summary of the invention

[0004] In order to overcome the technical defect of low result accuracy of existing mining area remote sensing change detection, the present invention provides a mining area remote sensing change detection method based on IAPM and FDAF.

[0005] The mining area remote sensing change detection method based on IAPM and FDAF provided by the present invention comprises the following steps:

[0006] S1. Acquire remote sensing images of the mining area in the first phase and the second phase;

[0007] S2. Inputting the remote sensing image into the IAPM (i.e., image adaptive processing module) to obtain a dual-phase image after the exposure, gamma value, contrast and sharpness are adapted;

[0008] S3. Inputting the bi-phase image into a feature extraction network with ChangeAD (i.e., feature interaction strategy) to extract features and obtain a bi-phase semantic feature map;

[0009] S4. Inputting the bi-temporal semantic feature map into FDAF (i.e., optical flow alignment network) to extract change features and obtain a segmentation result map;

[0010] S5. Using the dual-phase image as a guide map and the segmentation result map as an input map, inputting them into LGF (i.e., a learnable guided filter), and outputting the enhanced segmentation result as a detection result.

[0011] Furthermore, the IAPM includes CNNPP (i.e., a parameter predictor based on a convolutional neural network) and DIF (i.e., a differentiable image filter). The DIF is provided with four parameters, namely, exposure DIF, gamma DIF, contrast DIF, and sharpening DIF. The CNNPP is first used to learn the parameters of the four DIFs from the remote sensing image of the first phase, and then the remote sensing image of the second phase is sequentially input into the exposure DIF, gamma DIF, contrast DIF, and sharpening DIF, and finally the dual-phase image is obtained.

[0012] Furthermore, the CNNPP consists of four convolutional layers, a dropout layer and a fully connected layer.

[0013] Furthermore, the expression of the exposure DIF is:

[0014] ;

[0015] in, is the output image of the exposure DIF, is the remote sensing image of the second phase, is the parameter of the exposure DIF;

[0016] The expression of the gamma DIF is:

[0017] ;

[0018] in, is the gamma of the output image of DIF, is the parameter of the gamma DIF;

[0019] The expression of the contrast DIF is:

[0020] ;

[0021] in, is the output image of the contrast DIF, is the parameter of the comparison DIF, is the contrast mapping function, is the enhanced brightness function, is a brightness function based on the sensitivity of the human eye to the three primary colors, , and They are the values ​​of the three color channels: red, green, and blue;

[0022] The expression of the sharpening DIF is:

[0023] ;

[0024] in, is the output image of the sharpened dif, is the parameter of the sharpening DIF, is a Gaussian filter.

[0025] Furthermore, the expression of ChangeAD is:

[0026] ;

[0027] in, For the interaction layer, for or , is the Sigmoid activation function, is the common feature of the bi-temporal semantic feature map, It is a two-layer perceptron. The first layer is the squeeze channel, the second layer is the expansion channel, and GAP is the global average pool. and It is a dual-phase characteristic.

[0028] Furthermore, the expression of FDAF is:

[0029] ;

[0030] in, For FDAF output, and It is a dual-phase feature. enumerate or The position of the pixel in Adjustment or The pixel positions of are used to align the bi-phase features. The offset field.

[0031] Furthermore, the expression of the LGF is:

[0032] ;

[0033] in, is the output graph of the LGF, is the guide map of the LGF, is the input graph of the LGF, For the window, is a pixel, and They are In the window The mean and variance of is the regularization parameter, yes middle The mean of for The number of pixels in .

[0034] Furthermore, all The average value is taken as the filter output.

[0035] The technical solution provided by the present invention has the following advantages compared with the prior art:

[0036] The mining area remote sensing change detection method based on IAPM and FDAF provided by the present invention can make the image adaptive to exposure, gamma value, contrast and sharpness through IAPM, so that the dual-phase image can be aligned and corrected in terms of exposure, gamma value, contrast and sharpness; through FDAF, the perspective alignment and complete correspondence of the dual-phase image can be ensured, and the problem of the difference in acquisition parameters and the mismatch between the acquisition perspective and the acquisition resolution can be solved; through ChangeAD, the interaction effect of the dual-phase image can be improved without introducing additional calculations; through LGF, the significance outside the target can be suppressed, and the downlink detection or segmentation performance can be improved to obtain better edge-preserving segmentation performance. Through the cooperation of these aspects, the accuracy of the change detection result can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 A network structure diagram showing a change detection method according to an embodiment of the present invention;

[0040] Figure 2 A network structure diagram showing an IPAM in an embodiment of the present invention;

[0041] Figure 3 A diagram showing a network structure of ChangeAD in an embodiment of the present invention;

[0042] Figure 4 A network structure diagram of FDAF in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0043] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all of the embodiments.

[0045] Combine the following Figures 1 to 4 The specific embodiments of the present invention are described in detail.

[0046] This embodiment provides a mining area remote sensing change detection method based on IAPM and FDAF, including steps S1 to S5.

[0047] S1. Acquire remote sensing images of the mining area at the first phase and the second phase.

[0048] S2. Input the remote sensing image into IAPM to obtain a dual-phase image after adaptive exposure, gamma value, contrast and sharpness.

[0049] Specifically, the IAPM of this embodiment includes CNNPP and DIF. There are four DIFs, namely exposure DIF, gamma DIF, contrast DIF and sharpening DIF. First, CNNPP is used to learn the parameters of the four DIFs from the remote sensing image of the first phase, and then the remote sensing image of the second phase is input into the exposure DIF, gamma DIF, contrast DIF and sharpening DIF in sequence, and finally a dual-phase image is obtained.

[0050] It is easy to understand that the purpose of CNNPP is to predict the parameters of DIF by understanding the global content of the image, such as brightness, color, hue, and exposure.

[0051] More specifically, the CNNPP of this embodiment consists of four convolutional layers, a dropout layer, and a fully connected layer. This method learns a segmentation network based on CNNPP, first predicts the parameters of four DIFs from a downsampled low-resolution image of size 256×256, and then performs DIF processing on the input high-resolution image to assist the segmentation network in achieving better segmentation performance.

[0052] More specifically, the expression of exposure DIF in this embodiment is:

[0053] ;

[0054] in, is the output image of exposure DIF, is the remote sensing image of the second phase, is the parameter of exposure DIF;

[0055] The expression of gamma DIF is:

[0056] ;

[0057] in, is the output image of gamma DIF, is the parameter of gamma DIF;

[0058] The expression of contrast DIF is:

[0059] ;

[0060] in, To compare the output image of DIF, To compare the parameters of DIF, is the contrast mapping function, is the enhanced brightness function, is a brightness function based on the sensitivity of the human eye to the three primary colors, , and They are the values ​​of the three color channels: red, green, and blue;

[0061] The expression of sharpening DIF is:

[0062] ;

[0063] in, To sharpen the output image of DIF, is the parameter for sharpening DIF, is a Gaussian filter.

[0064] It should be noted that is a proportional factor, the degree of sharpening can be optimized To adjust to get better segmentation performance.

[0065] S3. Input the bi-temporal image into the feature extraction network with ChangeAD for feature extraction to obtain the bi-temporal semantic feature map. During calculation, the bi-temporal image is first projected into the feature space to obtain the global common features, and then the distributed attention map is used to adaptively re-weight each channel of the bi-temporal image to improve the interaction effect of the bi-temporal image without introducing additional calculations.

[0066] Specifically, the expression of ChangeAD in this embodiment is:

[0067] ;

[0068] in, For the interaction layer, for or , is the Sigmoid activation function, is the common feature of the bi-temporal semantic feature map, It is a two-layer perceptron. The first layer is the squeeze channel, the second layer is the expansion channel, and GAP is the global average pool. and It is a dual-phase characteristic.

[0069] During the calculation, we first aggregate the features from the Siamese branch by element-wise summation, then use global average pooling to generate global information, use MLP to extract common features, and use the Sigmoid activation function to obtain the bi-temporal semantic feature map.

[0070] S4. Input the bi-temporal semantic feature map into FDAF to extract the change features and obtain the segmentation result map.

[0071] It should be noted that FDAF will apply the estimated offset to the front and back images respectively, and then extract the change features by aligning and subtracting. Optical flow alignment mainly solves the problems of perspective offset and incomplete correspondence of dual-phase images.

[0072] It is easy to understand that FDAF comes with an explicit bidirectional alignment of optical flow. This module allows interactive alignment and feature fusion, introducing a priori task for change detection, namely finding the difference between images. Specifically, the bi-temporal feature maps are resampled through a deformable field to obtain their respective corrected features, and then the distance to the other original feature map is taken respectively and input into the subsequent forward propagation. In other words, FDAF implements an explicit task conversion function to convert the object extraction task into a change detection task. This solves the problem that the difference in acquisition parameters does not correspond to the acquisition perspective and acquisition resolution.

[0073] Specifically, the expression of FDAF in this embodiment is:

[0074] ;

[0075] in, For FDAF output, and It is a dual-phase feature. enumerate or The position of the pixel in Adjustment or The pixel positions of are used to align the bi-temporal features. The offset field.

[0076] S5. Use the dual-phase image as the guide map and the segmentation result map as the input map, input them into the LGF, and output the enhanced segmentation result as the detection result.

[0077] It is easy to understand that guided filtering is an edge-preserving and gradient-preserving image operation that uses the object boundary in the guided image to detect the saliency of the object. It can suppress the saliency outside the object and improve the downstream detection or segmentation performance. The LGF module is trained in an end-to-end manner with other modules to ensure that LGF performs adaptive processing on each image to obtain better edge-preserving segmentation performance.

[0078] Specifically, the expression of LGF in this embodiment is:

[0079] ;

[0080] in, is the output graph of LGF, This is the mind map of LGF. is the input graph of LGF, For the window, is a pixel, and They are In the window The mean and variance of is the regularization parameter, yes middle The mean of for The number of pixels in .

[0081] It is easy to understand that yes In pixels Centered window The linear transformation in .

[0082] More specifically, in this embodiment, all The average value is taken as the filter output.

[0083] The mining area remote sensing change detection method based on IAPM and FDAF provided in this embodiment can be applied to but not limited to the following four aspects:

[0084] First, monitoring of terrain changes in mining areas. This method can quickly identify areas of surface subsidence caused by mining activities, as well as areas of terrain recovery due to reclamation measures. Compared with traditional change detection methods, this method has improved the accuracy of identifying terrain changes in mining areas by about 20% and reduced the false alarm rate by 15%. In particular, the accuracy of detecting terrain changes in remote sensing images under low exposure conditions, such as settlement in shadow areas or vegetation recovery, has been significantly improved, effectively supporting mineral resource management and environmental protection work;

[0085] Second, analysis of vegetation coverage changes. This method can more accurately distinguish increases or decreases in vegetation density and improve detection accuracy. Compared with traditional change detection methods, this method has increased the accuracy of distinguishing vegetation changes caused by natural growth and human intervention by 10%, which provides more reliable data support for evaluating the ecological restoration effect of mining areas and formulating corresponding environmental protection measures;

[0086] Third, building and infrastructure monitoring. This method can accurately detect the construction of new buildings, the demolition of old buildings, and the changes in infrastructure such as roads and bridges. Compared with traditional change detection methods, this method successfully identified 98% of building change cases, which increased the detection accuracy by 10% compared with traditional methods. In addition, through in-depth analysis of optical flow consistency scores, potential safety hazards such as illegal buildings or infrastructure damage can be discovered in a timely manner, providing an important basis for mine safety management;

[0087] Fourth, geological disaster monitoring in mining areas. This method can accurately capture the tiny displacements of the ground surface caused by geological activities, which is the key to identifying the dynamic development of landslides or cracks. Compared with traditional change detection methods, this method successfully identified the precursor changes in potential landslide areas several weeks in advance, and the accuracy was 15% higher than that of traditional detection methods based on spectral changes. Especially in the landslide areas under the shadow, this method showed extremely high sensitivity. Through continuous monitoring and analysis, this method effectively assisted the geological disaster early warning system in the mining area, shortened the response time, provided a valuable time window for emergency evacuation and disaster prevention measures, and significantly improved the safety of mining operations. In addition, for the follow-up monitoring of landslide events that have occurred, this method can also accurately assess the terrain changes after the disaster and provide a scientific basis for post-disaster recovery work.

[0088] The above is only a specific implementation of the present invention, which enables those skilled in the art to understand or implement the present invention. Although detailed descriptions are given with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A mining area remote sensing change detection method based on IAPM and FDAF, characterized in that: The steps include: S1. Acquire remote sensing images of the mining area in the first phase and the second phase; S2. Inputting the remote sensing image into the IAPM to obtain a dual-phase image after the exposure, gamma value, contrast and sharpness are adapted; S3. Inputting the dual-phase image into a feature extraction network with ChangeAD to extract features and obtain a dual-phase semantic feature map; S4. Inputting the bi-phase semantic feature map into FDAF to extract the change features and obtain the segmentation result map; S5. Use the dual-phase image as a guide map, use the segmentation result map as an input map, input them into the LGF, and output the enhanced segmentation result as a detection result.

2. The mining area remote sensing change detection method based on IAPM and FDAF according to claim 1 is characterized in that: The IAPM includes CNNPP and DIF, and the DIF is set to four, namely exposure DIF, gamma DIF, contrast DIF and sharpening DIF. First, the CNNPP is used to learn the parameters of the four DIFs from the remote sensing image of the first phase, and then the remote sensing image of the second phase is input into the exposure DIF, gamma DIF, contrast DIF and sharpening DIF in sequence, and finally the dual-phase image is obtained.

3. According to the mining area remote sensing change detection method based on IAPM and FDAF according to claim 2, the CNNPP consists of four convolutional layers, a dropout layer and a fully connected layer.

4. The mining area remote sensing change detection method based on IAPM and FDAF according to claim 2 is characterized in that: The expression of the exposure DIF is: ； in, is the output image of the exposure DIF, is the remote sensing image of the second phase, is the parameter of the exposure DIF; The expression of the gamma DIF is: ； in, is the gamma of the output image of the dif, is the parameter of the gamma DIF; The expression of the contrast DIF is: ； in, is the output image of the contrast DIF, is the parameter of the comparison DIF, is the contrast mapping function, is the enhanced brightness function, is a brightness function based on the sensitivity of the human eye to the three primary colors, , and They are the values ​​of the three color channels: red, green, and blue; The expression of the sharpening DIF is: ； in, is the output image of the sharpened dif, is the parameter of the sharpening DIF, is a Gaussian filter.

5. The mining area remote sensing change detection method based on IAPM and FDAF according to claim 1 is characterized in that: The expression of ChangeAD is: ； in, For the interaction layer, for or , is the Sigmoid activation function, is the common feature of the bi-temporal semantic feature map, It is a two-layer perceptron. The first layer is the squeeze channel, the second layer is the expansion channel, and GAP is the global average pool. and It is a dual-phase characteristic.

6. The mining area remote sensing change detection method based on IAPM and FDAF according to claim 1 is characterized in that: The expression of FDAF is: ； in, For FDAF output, and It is a dual-phase feature. enumerate or The position of the pixel in Adjustment or The pixel positions of are used to align the bi-temporal features. The offset field.

7. The mining area remote sensing change detection method based on IAPM and FDAF according to claim 1 is characterized in that: The expression of the LGF is: ； in, is the output graph of the LGF, is the guide map of the LGF, is the input graph of the LGF, For the window, is a pixel, and They are In the window The mean and variance of is the regularization parameter, yes middle The mean of for The number of pixels in .

8. The mining area remote sensing change detection method based on IAPM and FDAF according to claim 7 is characterized in that: All Take the average value as the filter output.

Images