A method for detecting building changes in high-resolution remote sensing images

Abstract

The present invention discloses a method for detecting building changes in high-resolution remote sensing images. The method comprises the following steps: S1, obtaining heterogeneous high-resolution remote sensing image datasets of a region at different time phases, and performing registration, fusion, and segmentation on the heterogeneous high-resolution remote sensing image datasets to obtain a unified object set; S2, constructing an evidence set based on the acquired segmentation results; the evidence set includes the non-building index (NBI), the building index (MBI) of the two-phase images, and the differential information between the two-phase images; S3, extracting an evidence confidence index based on the shadow features of the objects in the images at different time phases; S4, performing decision fusion using the Dependency-Structure (D-S) evidence theory, and outputting object-oriented, fine-grained building change detection results. The present invention utilizes the shadow detection results to construct a complete D-S evidence theory change detection model. Compared with the existing technology, the change detection accuracy and Kappa coefficient reach 80% and above 0.7, respectively.

Description

Technical Field

[0001] The present invention relates to a method for detecting building changes, and in particular to a method for detecting building changes using high-resolution remote sensing images. Background Art

[0002] In recent years, with the rapid development of remote sensing technology, acquiring remote sensing imagery has become increasingly convenient. Compared to traditional field surveys, remote sensing imagery offers numerous advantages, including wide coverage and high spatial resolution. It can provide more detailed and rich decision-making support information for urban planning, environmental monitoring, and land change. Change detection of buildings is a key component of urban dynamic monitoring and has become a research hotspot in the remote sensing field.

[0003] Based on the difference in change detection primitives, currently commonly used building change detection methods can be categorized as pixel-level and object-level. Pixel-level change detection methods obtain detection results by directly comparing the differences between matching pixel pairs. However, remote sensing images of different temporal phases can lead to prominent "pseudo-changes" due to differences in scale, image quality, and imaging conditions, and pixel-level methods are often very sensitive to these "pseudo-changes." Furthermore, isolated pixels cannot fully represent the objects and the relationships between them, making it difficult to effectively address the "same object, different spectrum, different objects, same spectrum" phenomenon commonly found in high-resolution remote sensing imagery.

[0004] Compared with pixel-level methods, object-based change detection (OBCD) detects the inherent shape and size of the object for feature extraction and has higher robustness to registration errors, noise, etc. For example, Chen Kuiyi et al. used height analysis to extract building contour information, and based on the real-life three-dimensional model reconstructed from oblique images, made multiple cuts on all elements of the city to obtain the contour features of the objects at different heights, and accurately extracted the buildings based on their contour change rates. After further fine reconstruction of the three-dimensional model of the building based on the contour line, the minimum circumscribed rectangle of the building was used to perform object-level change detection on the building base contour, and the change detection results of medium and high-rise buildings were analyzed; Javed Aisha et al. applied three methods, namely change vector analysis, principal component analysis, and iterative reweighted multivariate alteration detection, to MBI images to expand the pixel-level change detection To object-level change detection, the newly built building area results were obtained; Lu Lichen et al. integrated BMI, CVA, and EM algorithms to perform object-oriented high-resolution remote sensing image building change detection, and output the results as binary building change areas and unchanged areas; Liu Haifei et al. based on the corner points and regular appearance shape characteristics of the building, comprehensively processed the line segments of the two-phase image to make them share line features, based on which the corresponding rectangular search area was generated, and superpixel merging was performed within the area to obtain structural surface objects containing ground feature structure information as basic analysis units. Finally, a feature vector was constructed for each basic analysis unit, and the changed building area was obtained through supervised classification. Despite this, the existing methods mainly focus on whether the building has changed, but cannot reflect other types of changes such as building renovation and renovation. They also have significant limitations in use, which reduces their reference value in practical applications. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a method for detecting building changes in high-resolution remote sensing images that can improve the accuracy of building change detection and refine the types of building changes.

[0006] Technical solution: The detection method of the present invention comprises the following steps:

[0007] S1, obtain heterogeneous high-resolution remote sensing image datasets of a certain area at different time phases, and perform registration, fusion and segmentation on the heterogeneous high-resolution remote sensing image datasets to obtain a unified object set;

[0008] S2, constructing an evidence set based on the obtained segmentation results; the evidence set includes a non-building index NBI, a building index MBI of the two temporal images, and differential information between the two temporal images;

[0009] S3, extracts evidence confidence indicators based on the shadow features of the object in images of different time phases;

[0010] S4 uses DS evidence theory for decision fusion and outputs object-oriented fine-grained building change detection results.

[0011] Furthermore, in step S2, the evidence set constructed includes: non-building indexes of the preceding / following time phases, building indexes of the preceding / following time phases, and multi-time phase differential features; the specific implementation steps are as follows:

[0012] S21, extract the non-building index NBI before and after the phase

[0013] For any object T in the segmentation result i , define the non-building index NBI:

[0014]

[0015] Among them, NDVI is the normalized difference vegetation index, NDWI is the normalized difference water index, and Pr is T i The rectangularity; Pwl is the aspect ratio, Pwl m is the maximum aspect ratio obtained after traversing all objects; S is the area index, let

[0016]

[0017] Among them, s i T i Area: s i =r 2 ×n i , r represents the resolution of the remote sensing image, n i Represents the total number of pixels in the i-th object; s a Defined as the standard value of building area;

[0018] According to the size of the NBI value, the non-building index NBI before and after the extraction phase is N 1i and N 2i ;

[0019] S22, extract the building index MBI before and after the phase

[0020] The steps for extracting the building index MBI of the before / after phase are as follows:

[0021] S221, calculate the brightness value:

[0022]

[0023] Among them, band k (x) is the brightness value of the kth spectral band at pixel x, W is the maximum band number of the visible light spectrum, and the maximum value of each pixel in the visible light band is taken as the brightness value of the pixel;

[0024] S222, Morphological White Hat Reconstruction:

[0025]

[0026] in, is the morphological opening operation of the brightness image b; d and v represent the direction and scale of the linear structure element respectively;

[0027] S223, calculation of differential morphometric profile DMP:

[0028] DMP WTH (d,v)=|WTH(d,(v+Δv))-WTH(d,v)|

[0029] S224, calculate the building index MBI:

[0030]

[0031] Where V=(v max -v min ) / Δv+1, D is the number of directions when calculating the building section; v max , v min are the maximum and minimum scale values of the linear structure element, respectively, and Δv is the scale change step size;

[0032] Fill the holes in the building index MBI results and extract the building index MBI before and after the phases as λ 1i and λ 2i ;

[0033] S23, obtain evidence collection

[0034] Define the differential feature:

[0035]

[0036] Where z is the object T i The total number of pixels in , σ 1k , σ 2k For object T i The corresponding pixel values of the k-th pixel in the two time phases, σ max is the maximum value of the pixel value of the k-th pixel in the two images;

[0037] At this time, the final evidence set R is obtained by combining the two-phase NBI index, the two-phase MBI index and the two-phase differential features. i ={N 1i ,N 2i ,λ 1i ,λ 2i ,Ci}.

[0038] Furthermore, in step S3, based on the evidence set R i , using DS evidence theory, taking shadow as the "pseudo-change" in change detection, the specific steps of extracting evidence confidence indicators are as follows:

[0039] S31, performing shadow detection using a shadow detection method;

[0040] S32, convert the color RGB image into an HSV image, and determine the binarization threshold M by the inter-class variance g value:

[0041] g=ω0(μ0-μ) 2 +ω1(μ1-μ) 2

[0042] Where ω0 is the ratio of the number of foreground pixels with grayscale values less than M to the entire image, and μ0 is its average grayscale; ω1 is the ratio of the number of background pixels with grayscale values greater than M to the entire image, and μ1 is its average grayscale; μ is the total average grayscale of the image;

[0043] When the traversal is completed, g = g max When , the segmentation threshold M is the optimal binary segmentation threshold; then the image is subjected to morphological opening and closing operations to fill holes and filter out isolated points to obtain shadow detection results;

[0044] At this point, the object T is calculated i The shadow area ratios in the dual-phase image are p and q respectively, so T i Confidence of evidence I i :

[0045] I i =(1-p)×(1-q).

[0046] Furthermore, in step S4, the specific implementation steps of outputting the building change detection result are as follows:

[0047] S41, define the recognition framework F = [N, D, R, U, O], divide the objects into new construction class N, demolition class D, reconstruction class R, unchanged class U and other class O, then the focal elements include [N], [D], [R], [U], [O], [N, D, R, U, O];

[0048] S42, constructing a total of five mass functions m1, m2, m3, m4, and m5 based on the evidence set extracted in step S2;

[0049] If we assume that the newly created class N is the focal element, The Dempster composition rule for the five mass functions on F is:

[0050]

[0051] Among them, the normalization coefficient N k is the focal element N judged in the k-th mass function;

[0052] S43, for any object T i Establish a probability assignment function BPAF and set the mass function m1 of the NBI index N1 of the previous phase. The expression of the probability assignment function BPAF is as follows:

[0053] m 1i ({N})=0.5×N 1i ×I i

[0054] m 1i ({D})=0.35×(1-N 1i )×I i

[0055] m 1i ({R})=0.3×(1-N 1i )×I i

[0056] m 1i ({U})=0.35×(1-N 1i )×I i

[0057] m 1i ({O})=0.5×N 1i ×I i

[0058] m 1i ({N,D,R,U,O})=1-I i

[0059] Among them, I i is the confidence of the evidence extracted in step S31; R i ={N 1i ,N 2i ,λ 1i ,λ 2i ,C i} The full set of indices provided for each scale evidence, in the rest of the mass functions R i Take the remaining evidence corresponding items in the set; N 1i For object T i The probability of it being a non-building is obtained from the NBI index in the previous phase, N 2i For object T iIn the latter phase, the probability of it being a non-building is obtained from the NBI index; 1i For object T i The probability of it being a building is obtained from MBI in the previous phase, λ 2i The probability that it is a building is obtained from MBI in the next phase; C i Used to describe the object T i The degree of change between the two phases;

[0060] S44, determine the object T according to the probability assignment function value i The category to which it belongs.

[0061] Further, determine the object T i The category to which the is assigned is determined as follows:

[0062] P1, T belonging to N category i satisfy:

[0063] m i ([N])>0.25, or m i ([N])>0.15∧m i ([R])>0.1;

[0064] P2, T belonging to Class D i satisfy:

[0065] m i ([D])>0.25, or m i ([D])>0.15∧m i ([R])>0.1;

[0066] P3, T belonging to R category i satisfy:

[0067] m i ([R])>0.25,

[0068] or {m i ([N])>0.1∨m i ([D])>0.1}∧m i ([R])>0.1∧m i ([O])<0.3;

[0069] P4, T belonging to U category i satisfy:

[0070] m i ([U])>0.3, and m i ([O])<0.3;

[0071] P5, otherwise, T i Belongs to category O;

[0072] Among them, "∧" represents the logical "and" and "∨" represents the logical "or";

[0073] According to the above judgment rules, the fine-grained building change detection results are output, and all objects in the image are finally divided into new construction class, demolition class, reconstruction class, unchanged class and non-building class.

[0074] Compared with the prior art, the present invention has the following significant effects:

[0075] 1. Based on the results of multi-scale image segmentation, this paper first designs a non-building index (NBI) by integrating multiple factors. This evidence set, together with differential information and building index (MBI), can achieve complementary advantages and integrate the advantages of evidence at different scales.

[0076] 2. Using the shadow detection results, an evidence confidence index was proposed, and then a complete DS evidence theory change detection model was constructed, thereby dividing buildings into new construction, demolition and reconstruction categories; compared with the existing technology, the change detection accuracy and Kappa coefficient of the present invention reached 80% and above 0.7 respectively, and the effect was significant in visual analysis and quantitative evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0078] Figure 2(a) is a schematic diagram of the GF-1 image of Yinchuan 1 in dataset 1.

[0079] Figure 2(b) is a schematic diagram of the GF-2 image of Yinchuan 1 in the second phase of dataset 1.

[0080] Figure 2(c) is a reference image of the fine-grained building change detection results of dataset 1.

[0081] Figure 2(d) shows the fine-grained building change detection results of the present invention for dataset 1.

[0082] Figure 2(e) is the binary reference image of dataset 1.

[0083] Figure 2(f) is a binarization result diagram of data set 1 according to the present invention.

[0084] Figure 2(g) shows the result of strategy 1 for dataset 1.

[0085] Figure 2(h) shows the results of Strategy 2 for Dataset 1;

[0086] Figure 2(i) shows the results of Strategy 3 for Dataset 1;

[0087] Figure 3(a) is a schematic diagram of the GF-1 image of Yinchuan 1 in dataset 2.

[0088] Figure 3(b) is a schematic diagram of the GF-2 image of Yinchuan 1 in the second phase of dataset 2.

[0089] Figure 3(c) is a reference image of the fine-grained building change detection results of dataset 2.

[0090] Figure 3(d) shows the fine-grained building change detection results of the present invention for dataset 2.

[0091] Figure 3(e) is the binary reference image of dataset 2.

[0092] FIG3( f ) is a diagram showing the binarization result of the present invention for dataset 2.

[0093] Figure 3(g) shows the result of strategy 1 for dataset 2.

[0094] Figure 3(h) shows the results of Strategy 2 for Dataset 2;

[0095] Figure 3(i) shows the results of Strategy 3 for Dataset 2;

[0096] Figure 4(a) is a schematic diagram of a typical region 1 image in phase 1.

[0097] Figure 4(b) is a schematic diagram of the image of a typical region 1 in phase 2.

[0098] Figure 4(c) is the reference image of region 1.

[0099] Figure 4(d) is the first result diagram of the present invention in region 1.

[0100] Figure 4(e) is a reference to Figure 2 for region 1.

[0101] Figure 4(f) is the second result diagram of the present invention in area 1.

[0102] Figure 4(g) shows the results of strategy 1 in region 1.

[0103] Figure 4(h) shows the results of strategy 2 in region 1;

[0104] Figure 4(i) shows the results of strategy 3 in region 1;

[0105] Figure 5(a) is a schematic diagram of a typical region 2 image in phase 1.

[0106] Figure 5(b) is a schematic diagram of the image of a typical region 2 in phase 2.

[0107] Figure 5(c) is the reference image of region 2.

[0108] Figure 5(d) is the first result diagram of the present invention in area 2.

[0109] Figure 5(e) is the reference figure 2 for region 2.

[0110] FIG5(f) is a second result diagram of the present invention in region 2.

[0111] Figure 5(g) shows the results of strategy 1 in region 2.

[0112] Figure 5(h) shows the results of strategy 2 in region 2;

[0113] Figure 5(i) shows the results of strategy 3 in region 2;

[0114] Note: The color codes of the sub-figures in Figures 2(c) to (i), 3(c) to (i), 4(c) to (i), and 5(c) to (i) are as follows:

[0115]

[0116] Figure 6 This is the overall accuracy change trend of the dataset under different standard values of built-up area. DETAILED DESCRIPTION

[0117] The present invention will be described in further detail below with reference to the accompanying drawings and specific implementations.

[0118] Based on the DS evidence theory, this invention proposes a fine-grained change detection method for buildings. This method not only provides change detection results for buildings and non-buildings, but also further divides building changes into three types: new construction, demolition, and reconstruction. Based on image segmentation, a non-building index (NBI) is designed by integrating the area, rectangularity, aspect ratio and other features of the object; in addition, the corresponding building index MBI, differential information and evidence confidence index based on shadow detection are extracted in combination with the object contour; finally, the above "evidence" is used as the input of the DS decision model to propose a fine-grained building change type recognition model. Multiple groups of experiments show that the overall accuracy of the proposed model can reach more than 85%, which improves the accuracy of building change detection while more finely dividing the types of building changes.

[0119] (1) Implementation process

[0120] like Figure 1 As shown, the method of the present invention comprises the following steps:

[0121] Step S1, data preprocessing

[0122] After obtaining heterogeneous high-resolution remote sensing image datasets of a certain area at different phases, they are registered, fused and segmented to obtain a unified object set.

[0123] The image preprocessing stage specifically includes resampling, registration, and segmentation of multi-temporal images. First, the commercial software ENVI is used for resampling and registration. During the image segmentation stage, based on the popular commercial software eCognition, the fused image is segmented by setting segmentation scale parameters using the weights, compactness, and smoothness of different image bands, thereby obtaining a set of analytical primitives for subsequent object-level change detection. The segmentation parameters used in this paper are set to: a segmentation scale of 100, a shape factor of 0.4, and a compactness of 0.8.

[0124] Step S2: Evidence set construction

[0125] Based on the obtained segmentation results, an evidence set is constructed, which includes the non-building index NBI, the building index MBI of the two-phase images, and the difference information between the two-phase images.

[0126] For the task of building change detection, the evidence set constructed by the present invention mainly consists of five parts: non-building indexes in the previous and next phases, building indexes in the previous and next phases, and multi-phase differential features.

[0127] Step S21: Extracting the non-building index NBI of the front and back phases

[0128] The traditional building index (MBI) describes the possibility of a pixel belonging to a building based on the strong contrast characteristics of the spectrum. On the other hand, there are few quantitative indicators for evaluating whether a pixel belongs to a non-building, which is of great reference value for accurately identifying changes in buildings. Therefore, the present invention uses the segmentation result of any object T to determine the probability of the pixel belonging to a building. i , a non-building index NBI was designed based on the comprehensive rectangularity, aspect ratio, area, NDVI vegetation index and NDWI water index, as shown in formula (1).

[0129]

[0130] Among them, NDVI is the normalized difference vegetation index, NDWI is the normalized difference water index, and Pr is T i The rectangularity; Pwl is the aspect ratio, Pwl m is the maximum aspect ratio obtained after traversing all objects. S is the area index. Since buildings usually have a certain area but not too large or too small, let

[0131]

[0132] Among them, s i T i Area: s i =r 2 ×n i, r represents the resolution of the remote sensing image, n i Represents the total number of pixels in the i-th object. a Defined as the standard value of building area, the standard value of building area s taken by this invention is a The value is 2000m 2 The larger the NBI value, the more likely the object is a non-building object; conversely, the more likely the object is a building. On this basis, the non-building index NBI is extracted for the front and back phases, and N is obtained respectively. 1i and N 2i .

[0133] Step S22: Extracting the MBI of the preceding and following phases

[0134] The Morphological Building Index (MBI) mainly considers the spectral structural characteristics of buildings, including brightness, local contrast, shape, size, and directionality, and performs a series of morphological operations. The specific extraction steps are as follows:

[0135] Step S221: Calculate the brightness value.

[0136]

[0137] Among them, band k (x) is the brightness value of the kth spectral band at pixel x, and W is the maximum number of bands in the visible light spectrum. Since the visible light band has a significant impact on the spectral information of buildings, the maximum value of each pixel in the visible light band is used as the brightness value of the pixel.

[0138] Step S222: Morphological white-hat reconstruction.

[0139]

[0140] in, is the morphological opening operation of the brightness image b, while d and v represent the direction and scale of the linear structure element, respectively.

[0141] Step S223: Calculate differential morphological profiles (DMP).

[0142] DMP WTH (d,v)=|WTH(d,(v+Δv))-WTH(d,v)| (5)

[0143] Step S224: Calculate the MBI index.

[0144]

[0145] Where V=(vmax -v min ) / Δv+1, D is the number of directions when calculating the building section. The v used in this invention max =105,v min =5, Δv=25. However, the interior of the larger area building objects extracted by MBI often leads to incomplete extraction results due to the presence of heterogeneous interference factors. The present invention fills the holes in the MBI results to improve the accuracy of building area extraction. On this basis, the building index MBI before and after extraction is λ 1i and λ 2i .

[0146] Step S23: Two-phase differential features

[0147] On the basis of extracting the unified object primitives, we further extract the spectral difference features between the two temporal images. The specific steps are as follows: define the difference features:

[0148]

[0149] Where z is the object T i The total number of pixels in , σ 1k , σ 2k For object T i The corresponding pixel values of the k-th pixel in the two time phases, σ max is the larger pixel value of the k-th pixel in the two images. At this time, combining the two-phase NBI index, the two-phase MBI index and the two-phase differential features, the final evidence set R can be obtained. i ={N 1i ,N 2i ,λ 1i ,λ 2i ,C i Step S3: Extracting evidence confidence index

[0150] Evidence confidence indicators are extracted based on the shadow features of the object in images of different time phases.

[0151] Based on the evidence set R i , the present invention adopts DS evidence theory (Dempster-Shafer envidence theory) for decision fusion, thereby achieving fine-grained building change degree classification. Compared with traditional probabilistic reasoning theory, the prior data required by DS evidence theory is more intuitive, easy, does not require excessive processing, and has the advantage of being able to integrate heterogeneous information as evidence support. In addition, considering that shadows are the main source of "pseudo-changes" in change detection, the present invention designs an evidence confidence index I that takes shadow factors into account i , as a basis for measuring the credibility of different evidence.

[0152] Shadows are one of the main sources of "false changes" in change detection using high-resolution remote sensing images. Therefore, this paper first employs a shadow detection method to detect shadows. By converting the color RGB image into an HSV image, the binarization threshold M is determined based on the inter-class variance g calculated using formula (8).

[0153] g=ω0(μ0-μ) 2 +ω1(μ1-μ) 2 (8)

[0154] Among them, ω0 is the ratio of the number of foreground pixels with grayscale values less than M to the entire image, μ0 is its average grayscale; ω1 is the ratio of the number of background pixels with grayscale values greater than M to the entire image, μ1 is its average grayscale; μ is the total average grayscale of the image. When the traversal is completed, g=g max When , the segmentation threshold M used is the optimal binary segmentation threshold. On this basis, the image is opened and closed by morphological operation to fill the holes and filter out isolated points, thereby obtaining the shadow detection result. At this point, the object T can be calculated. i The shadow area ratios in the dual-phase image are p and q respectively, so T i Confidence in evidence:

[0155] I i =(1-p)×(1-q) (9)

[0156] S4, outputs the building change detection results; uses DS evidence theory for decision fusion and outputs object-oriented fine-grained building change detection results.

[0157] First, define the recognition framework F = [N, D, R, U, O], and divide the objects into the newly built class N, the demolished class D, the rebuilt class R, the unchanged class U, and the other class O. Then the focal element includes [N], [D], [R], [U], [O], [N, D, R, U, O]. Based on the evidence set extracted in step S2, a total of 5 mass functions m1, m2, m3, m4, and m5 are constructed. If the newly built class N is the focal element The Dempster composition rule for the five mass functions on F is:

[0158]

[0159] Among them, the normalization coefficient N k is the focal element N in the kth mass function. Therefore, for any object T i The basic probability assignment function BPAF is established. Taking the mass function m1 of the NBI index N1 of the previous phase as an example, the BPAF formula is as follows:

[0160] m 1i ({N})=0.5×N 1i ×I i (11)

[0161] m 1i ({D})=0.35×(1-N 1i )×I i (12)

[0162] m 1i ({R})=0.3×(1-N 1i )×I i (13)

[0163] m 1i ({U})=0.35×(1-N 1i )×I i (14)

[0164] m 1i ({O})=0.5×N 1i ×I i (15)

[0165] m 1i ({N,D,R,U,O})=1-I i (16)

[0166] Among them, I i is the confidence of the evidence extracted in step S31. i ={N 1i ,N 2i ,λ 1i ,λ 2i ,C i} The full set of indices provided for each scale evidence, in the rest of the mass functions R i Take the remaining evidence corresponding items in the set. N 1i For object T i The probability of it being a non-building is obtained from the NBI index in the previous phase, N 2i For object T i In the latter phase, the probability of it being a non-building is obtained from the NBI index; 1i For object T i The probability of it being a building is obtained from MBI in the previous phase, λ 2i The probability that it is a building is obtained from MBI in the next phase; C i Used to describe the object T i The degree of change between the two phases.

[0167] The basic basis for fine-grained classification of building changes using the following discrimination rules is that the object T i The greater the possibility of belonging to a certain class, the greater the probability distribution function value corresponding to that class should be:

[0168] ① T belonging to N category i Should meet: m i ([N])>0.25, or m i ([N])>0.15∧m i ([R])>0.1;

[0169] ②T belonging to Class D i Should meet: m i ([D])>0.25, or m i ([D])>0.15∧m i ([R])>0.1;

[0170] ③T belonging to R category i Should meet: m i ([R])>0.25, or {m i ([N])>0.1∨m i ([D])>0.1}∧m i ([R])>

[0171] 0.1∧m i ([O])<0.3;

[0172] ④T belonging to U category i Should meet: m i ([U])>0.3, and m i ([O])<0.3;

[0173] ⑤ Otherwise, T i Belongs to category O.

[0174] Among them, "∧" represents logical "and" and "∨" represents logical "or".

[0175] According to the above judgment rules, the fine-grained building change detection results are output, and all objects in the image are finally divided into five categories: new construction, demolition, reconstruction, unchanged, and non-building.

[0176] (2) Experiment and analysis

[0177] To verify the effectiveness and accuracy of the proposed method, two sets of heterogeneous high-resolution remote sensing image data were used in the experiment. The two sets of data used heterogeneous R, G, and B three-band images from the Gaofen-1 (GF-1) and Gaofen-2 (GF-2) satellites, covering different areas of Yinchuan City, Ningxia Hui Autonomous Region, China, with spatial resolutions of 2 meters and 0.81 meters, respectively. The images were acquired in 2018 and 2020, respectively. After resampling and registration, the two sets of images were both 1024 pixels × 1024 pixels in size, with a resolution of 2 meters.

[0178] In order to verify the effectiveness of the method proposed in this invention, three advanced methods are used to conduct comparative experiments.

[0179] Strategy 1: Based on an object-oriented approach, the MBI algorithm is used to extract buildings and the CVA algorithm is used for change detection to obtain the differences of all objects. Finally, the Bayesian threshold calculation method of the EM algorithm is used to determine the change threshold. The spectral characteristics and building index characteristics are superimposed to form multiple features. This more clearly reflects the differences of the image spots, constructs a feature space, and improves the accuracy of the final building change detection results.

[0180] Strategy 2: After using an improved bilateral filter to improve the spatial structural consistency between pixels within the object, the detection error caused by "pseudo-changes" is reduced through on-off channel small target suppression, traditional building extraction based on rectangular aspect ratio, and Otsu method classification.

[0181] Strategy 3: A change detection method based on DS evidence theory that fuses feature information. Based on image segmentation, the change vector analysis method is used to calculate the differences in spectral and texture features and morphological building index before and after the object. The sigmoid function is used as the membership function to construct the BPAF. Finally, the DS evidence theory is used for feature fusion and rule judgment to obtain the building change area.

[0182] 1) General results and analysis

[0183] The change detection results of the method of the present invention and the comparative method are shown in Figures 2(a)-(i) and 3(a)-(i).

[0184] Since this method categorizes building changes into new construction, demolition, and reconstruction, it uses different grayscale values to represent different types of changes and uses boxes to mark typical areas for further visual analysis. Furthermore, since both comparison methods are binary classification methods, in this comparison experiment, new construction, demolition, and reconstruction are all classified as changes, allowing for visual and quantitative evaluation based on binary images.

[0185] In order to further verify the effectiveness of the fine-grained building change detection model proposed in this paper, two representative areas in Figure 2(a) were selected for local visual analysis, and the local magnification results are shown in Figures 4(a)-(i) and 5(a)-(i).

[0186] Comparative analysis of Figures 4(a)-(i) reveals that the upper left portion of the original image represents a newly added building. Due to the high proportion of shadows and the minimal difference between the building and surrounding shadows, Figures 4(e)-(i) show that strategies 1 and 3 using MBI, as well as the traditional methods of strategy 2, are unable to accurately identify this region as a building in the post-processed image. However, the present invention, by extracting shadow features as evidence confidence indicators, can better leverage the strengths of each piece of evidence within the evidence set under the DS decision output rule, classifying this region as a newly added building. The middle portion on the right side represents a renovated building. Due to its irregular segmentation, the traditional building extraction method used in strategy 2 is unable to correctly identify it. However, strategy 3, by combining spectral and texture features, eliminates this renovated building area. Both the MBI and CVA change vector detection methods used in strategy 1 and the present invention's method are able to correctly extract this region. As shown in Figures 4(c) and 4(d), the output model of the present invention more accurately outputs this region as a renovated building.

[0187] As shown in Figures 5(c)-(i), the proposed method demonstrates superior detection performance compared to Strategies 1 and 2. The designed NBI index, utilizing NDWI, area index, and shape features, accurately screens out non-building areas that have been converted to water bodies, vacant land, and other structures during the two years of urban planning, correctly extracting the demolished building area. However, due to the small and dense buildings in the previous imagery of this area, Strategies 1 and 2 were unclear in detecting the building areas, resulting in high false positive and missed detection rates. In contrast, the proposed method achieves higher detection accuracy and is more consistent with the reference image. In the absence of shadow influence, the detection results of Strategy 3 and the proposed method, which also utilize evidence theory, are relatively consistent with the reference image. However, at the proposed NBI area threshold, the proposed method further eliminates two false positive areas compared to Strategy 3, resulting in a more consistent result with actual results.

[0188] 2) Quantitative evaluation

[0189] This paper quantitatively analyzes two sets of fine-grained building change detection results using overall accuracy, false positive rate, and Kappa coefficient. The accuracy evaluation results are shown in Tables 1 and 2, respectively.

[0190] Table 1: Accuracy of fine-grained building change detection using the proposed method in dataset 1

[0191]

[0192] Table 2 Fine-grained building change detection accuracy of the proposed method in dataset 2

[0193]

[0194] The experimental accuracy shown in Tables 1 and 2 demonstrates that the constructed fusion evidence set, after DS decision fusion, can effectively distinguish between buildings and non-buildings in a region and further classify and detect the resulting altered buildings. The overall accuracy of the experiment reached 80%, and the Kappa coefficient was greater than 0.7, indicating that the classification results were highly consistent with the actual results.

[0195] According to Figures 2(e) to (h) and Figures 3(e) to (h), the accuracy of the method of the present invention and the comparative method was evaluated, and the results are shown in Tables 3 and 4 below.

[0196] Table 3 Accuracy of dataset 1 of the present invention and the comparative method

[0197]

[0198] Table 4 Accuracy of dataset 2 of the present invention and the comparative method

[0199]

[0200] From the quantitative analysis based on Tables 3 and 4, the overall accuracy of the change detection results of the proposed method is the highest among the three comparison groups. Although the missed detection rate is slightly improved in Dataset 2, the overall consistency is the highest, reflecting the good complementarity between the evidence within the constructed evidence set. The effectiveness of the DS decision fusion model after extracting evidence confidence based on shadow detection improves the overall accuracy of change detection.

[0201] 3) Area parameter analysis

[0202] Since the evidence set constructed by the present invention has the standard value of the building area of the non-building index part s a The setting of has a significant impact on the non-building index results. Therefore, the standard value of the building area s is set. a The value range is [500, 5000], with a step size of 500. The standard value s of the building area is analyzed in the experiments of dataset 1 and dataset 2. a The relationship between the setting and the accuracy of the output results of fine-grained building change detection, the experimental results and the change trend are as follows Figure 6 shown.

[0203] Depend on Figure 6 It can be seen that the standard value of the built-up area in dataset 1 is s aWhen the value of the building area standard s increases from 500 to 2000, the overall accuracy of fine-grained building change detection increases; a When the number of building area standard values increases from 2000 to 3000, the overall accuracy of fine-grained building change detection remains relatively stable but slowly decreases; when the standard value of building area s a When the number of building area standard values increases from 3000 to 5000, the overall accuracy of fine-grained building change detection shows a downward trend; in dataset 2, the standard value of building area s a When the value of the building area standard s increases from 500 to 1500, the overall accuracy of fine-grained building change detection increases; a When the value of the building area standard value s increases from 1500 to 2000, the overall accuracy of fine-grained building change detection remains relatively flat but increases slowly; when the building area standard value s a When the number of changes increases from 2000 to 5000, the overall accuracy of fine-grained building change detection shows a downward trend. Based on datasets 1 and 2, this paper recommends setting s in practical applications. a is 2000.

Claims

1. A method for detecting building changes in high-resolution remote sensing images, characterized in that: The steps are as follows: S1, obtain heterogeneous high-resolution remote sensing image datasets of a certain area at different time phases, and perform registration, fusion and segmentation on the heterogeneous high-resolution remote sensing image datasets to obtain a unified object set; S2, constructing an evidence set based on the obtained segmentation results; the evidence set includes a non-building index NBI, a building index MBI of the two temporal images, and differential information between the two temporal images; S3, extracts evidence confidence indicators based on the shadow features of the object in images of different time phases; S4, uses DS evidence theory for decision fusion and outputs object-oriented fine-grained building change detection results; In step S2, the evidence set constructed includes: non-building index of the preceding / following time phase, building index of the preceding / following time phase, and multi-time phase differential features; the specific implementation steps are as follows: S21, extract the non-building index NBI before and after the phase For any object T in the segmentation result i , define the non-building index NBI: Among them, NDVI is the normalized difference vegetation index, NDWI is the normalized difference water index, and Pr is T i The rectangularity; Pwl is the aspect ratio, Pwl m is the maximum aspect ratio obtained after traversing all objects; S is the area index, let Among them, s i T i Area: s i =r 2 ×n i , r represents the resolution of the remote sensing image, n i Represents the total number of pixels in the i-th object; s a Defined as the standard value of building area; According to the size of the NBI value, the non-building index NBI before and after the extraction phase is N 1i and N 2i ; S22, extract the building index MBI before and after the phase The steps for extracting the building index MBI of the before / after phase are as follows: S221, calculate the brightness value: Among them, band k (x) is the brightness value of the kth spectral band at pixel x, W is the maximum band number of the visible light spectrum, and the maximum value of each pixel in the visible light band is taken as the brightness value of the pixel; S222, Morphological White Hat Reconstruction: in, is the morphological opening operation of the brightness image b; d and v represent the direction and scale of the linear structure element respectively; S223, calculation of differential morphometric profile DMP: DMP WTH (d,v)=|WTH(d,(v+Δv))-WTH(d,v)| S224, calculate the building index MBI: Where V=(v max -v min ) / Δv+1, D is the number of directions when calculating the building section; v max , v min are the maximum and minimum scale values of the linear structure element, respectively, and Δv is the scale change step size; Fill the holes in the building index MBI results and extract the building index MBI before and after the phases as λ 1i and λ 2i ; S23, obtain evidence collection Define the differential feature: Where z is the object T i The total number of pixels in , σ 1k , σ 2k For object T i The corresponding pixel values of the k-th pixel in the two time phases, σ max is the maximum value of the pixel value of the k-th pixel in the two images; At this time, the final evidence set R is obtained by combining the two-phase NBI index, the two-phase MBI index and the two-phase differential features. i ={N 1i ,N 2i ,λ 1i ,λ 2i ,C i }.

2. The method for detecting building changes in high-resolution remote sensing images according to claim 1, characterized in that: In step S3, based on the evidence set R i , using DS evidence theory, taking shadow as the "pseudo-change" in change detection, the specific steps of extracting evidence confidence indicators are as follows: S31, performing shadow detection using a shadow detection method; S32, convert the color RGB image into an HSV image, and determine the binarization threshold M by the inter-class variance g value: g=ω0(μ0-μ) 2 +ω1(μ1-μ) 2 Where ω0 is the ratio of the number of foreground pixels with grayscale values less than M to the entire image, and μ0 is its average grayscale; ω1 is the ratio of the number of background pixels with grayscale values greater than M to the entire image, and μ1 is its average grayscale; μ is the total average grayscale of the image; When the traversal is completed, g = g max When , the segmentation threshold M is the optimal binary segmentation threshold; then the image is subjected to morphological opening and closing operations to repair holes and filter out isolated points to obtain shadow detection results; At this point, the object T is calculated i The shadow area ratios in the dual-phase image are p and q respectively, so T i Confidence of evidence I i : I i =(1-p)×(1-q)。 3. The method for detecting building changes in high-resolution remote sensing images according to claim 1, wherein: In step S4, the specific implementation steps of outputting the building change detection result are as follows: S41, define the recognition framework F = [N, D, R, U, O], divide the objects into new construction class N, demolition class D, reconstruction class R, unchanged class U and other class O, then the focal elements include [N], [D], [R], [U], [O], [N, D, R, U, O]; S42, constructing a total of five mass functions m1, m2, m3, m4, and m5 based on the evidence set extracted in step S2; If we assume that the newly created class N is a focal element, The Dempster composition rule for the five mass functions on F is: Among them, the normalization coefficient N k is the focal element N judged in the k-th mass function; S43, for any object T i Establish a probability assignment function BPAF and set the mass function m1 of the NBI index N1 of the previous phase. The expression of the probability assignment function BPAF is as follows: m 1i ({N})=0.5×N 1i ×I i m 1i ({D})=0.35×(1-N 1i )×I i m 1i ({R})=0.3×(1-N 1i )×I i m 1i ({U})=0.35×(1-N 1i )×I i m 1i ({O})=0.5×N 1i ×I i m 1i ({N,D,R,U,O})=1-I i Among them, I i is the confidence level of the evidence extracted in step S31; R i ={N 1i ,N 2i ,λ 1i ,λ 2i ,C i } The full set of indices provided for each scale evidence, in the rest of the mass functions R i Take the remaining evidence corresponding items in the set; N 1i For object T i The probability of it being a non-building is obtained from the NBI index in the previous phase, N 2i For object T i In the latter phase, the probability of it being a non-building is obtained from the NBI index; 1i For object T i The probability of it being a building is obtained from MBI in the previous phase, λ 2i The probability that it is a building is obtained from MBI in the next phase; C i Used to describe the object T i The degree of change between the two phases; S44, determine the object T according to the probability assignment function value i The category to which it belongs.

4. The method for detecting building changes in high-resolution remote sensing images according to claim 3, wherein: Determine the object T i The category to which the is assigned is determined as follows: P1, T belonging to N category i satisfy: m i ([N])>0.25, or m i ([N])>0.15∧m i ([R])>0.1; P2, T belonging to Class D i satisfy: m i ([D])>0.25, or m i ([D])>0.15∧m i ([R])>0.1; P3, T belonging to R category i satisfy: m i ([R])>0.25, or {m i ([N])>0.1∨m i ([D])>0.1}∧m i ([R])>0.1∧m i ([O])<0.3; P4, T belonging to U category i satisfy: m i ([U])>0.3, and m i ([O])<0.3; P5, otherwise, T i Belongs to category O; Among them, "∧" represents logical "and", "∨" represents logical "or"; According to the above judgment rules, the fine-grained building change detection results are output, and all objects in the image are finally divided into new construction class, demolition class, reconstruction class, unchanged class and non-building class.

Images