CN112052756A - Post-earthquake high-resolution remote sensing image earthquake damage building detection method - Google Patents
Post-earthquake high-resolution remote sensing image earthquake damage building detection method Download PDFInfo
- Publication number
- CN112052756A CN112052756A CN202010855069.8A CN202010855069A CN112052756A CN 112052756 A CN112052756 A CN 112052756A CN 202010855069 A CN202010855069 A CN 202010855069A CN 112052756 A CN112052756 A CN 112052756A
- Authority
- CN
- China
- Prior art keywords
- earthquake
- earthquake damage
- building
- area
- rgb
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/176—Urban or other man-made structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
- G06F40/242—Dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/155—Segmentation; Edge detection involving morphological operators
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Probability & Statistics with Applications (AREA)
- Geometry (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention discloses a method for detecting earthquake damage of a building by using post-earthquake high resolution remote sensing images. Firstly, extracting a potential building set through WJSEG segmentation and a group of non-building screening rules; secondly, a seismic damage visual dictionary model is constructed by utilizing the characteristics of spectrum, texture and geometric morphology, and a semantic gap from pixels to seismic damage characteristics is spanned; on the basis, a visual dictionary optimization strategy based on intra-class and inter-class penalty factors is designed, and information redundancy and evidence conflict are reduced; and finally, further dividing the buildings into intact buildings, partial earthquake damage buildings and ruins through a random forest classifier. The experimental result shows that the overall precision of the method reaches more than 85 percent, so that key decision support information can be provided for emergency response rescue after earthquake and reconstruction after disaster.
Description
Technical Field
The invention belongs to the field of image detection, and particularly relates to a method for detecting earthquake damage buildings.
Background
After an earthquake occurs, the position of the earthquake-damaged building is judged quickly and accurately, and the important function is achieved for carrying out emergency rescue and disaster assessment after the earthquake. Compared with the traditional field investigation relying on ground personnel, the earthquake damage building detection based on the high-resolution remote sensing image has the advantages of strong real-time performance, wide coverage range, low safety risk and the like, and becomes one of the research hotspots in the field of remote sensing application.
At present, methods for extracting earthquake-damaged buildings from high-resolution remote sensing images can be mainly divided into two types:
(1) a classification method based on post-earthquake single-temporal images. The method converts the detection problem of the earthquake damage building into the classification problem of the remote sensing image after the earthquake, thereby breaking through the limitation caused by the dependence on the image before the earthquake. For example, leaf Xin and the like can achieve the total precision of more than 80% in earthquake damage building detection based on Quickbird and aerial remote sensing images after earthquake by analyzing the spatial correlation of the internal gradient of the roof of the building.
(2) A change detection method based on a pre-earthquake time phase image and a post-earthquake time phase image. The method introduces the pre-earthquake image, can extract the change information of the building as the earthquake damage detection basis, and therefore has higher detection precision. For example, Roberta and the like propose a method for detecting the change of earthquake damage buildings by combining multiple parameters and multiple descriptors, and the overall accuracy of the method can reach more than 96%. The limitations of such methods are the difficult acquisition of the pre-earthquake images, and the high-precision registration of the pre-earthquake and post-earthquake images. Therefore, the classification method based on the single-temporal image after the earthquake has more application and popularization in emergency response after the earthquake. However, since such methods lack pre-earthquake reference information and the post-earthquake scene is more complex in structure and spatial layout, how to perform effective abstract representation on the earthquake damage characteristics of the building is a key and difficult problem in determining whether to accurately distinguish the earthquake damage building from other ground objects.
Disclosure of Invention
In order to solve the technical problems mentioned in the background technology, the invention provides a method for detecting earthquake-damaged buildings by high resolution remote sensing images after an earthquake.
In order to achieve the technical purpose, the technical scheme of the invention is as follows:
a method for detecting earthquake damage of buildings by high-resolution remote sensing images after earthquake comprises the following steps:
(1) sequentially carrying out image segmentation and non-building object elimination on the obtained post-earthquake high-resolution remote sensing image to obtain a potential building object set;
(2) constructing a seismic hazard visual dictionary from three angles of spectrum, texture and geometric morphology based on the visual bag-of-words model;
(3) constructing an optimized earthquake damage visual dictionary model based on the intra-class penalty factors and the inter-class penalty factors;
(4) and based on the optimized earthquake damage visual vocabulary model and the potential building object set, a random forest classifier is adopted to obtain a final earthquake damage building detection result.
Further, in step (1), the non-building object culling follows the following rules:
(a) area regulation: the area of each object, namely the number N of pixels contained in each object is countedpixelsIf N is presentpixelsIf the area is smaller than or equal to the preset area threshold value, the object is considered as a weak target and is removed;
(b) morphological constructionBuilding index rule: determining a separation threshold T for a morphological building index using a maximum inter-class variance methodMBICalculating the mean value of the morphological building index of all pixels in each objectRejection satisfiesAll of the objects of (1);
(c) the rectangle degree and the length-width ratio are regular: the degree of rectangularity of an object is defined as Rd ═ Npixels/NrectangleIn which N ispixelsThe number of pixels contained in the minimum bounding rectangle of the object, NrectangleIs the aspect ratio of the smallest circumscribed rectangle; if an object satisfies Rd < 0.8 and NrectangleIf the distance is more than 5, the object is considered as a long and narrow object and is removed.
Further, in the step (2), the spectrum-based earthquake damage visual dictionary is constructed by respectively performing K-means clustering on the three-band images of the RGB model according to gray values to obtain any object Ri∈RpotVisual dictionary WR of seismic injury of K-dimensional spectrum corresponding to three wave bands of R, G, Bi、WGi、WBiWherein R ispotThe set of potential building objects obtained in the step (1);
the construction method of the earthquake damage visual dictionary based on the texture is that local homogeneity indexes J-value are adopted as a class of earthquake damage visual vocabulary, and the definition of the J-value is as follows:
J-value=(ST-SW)/SW
wherein S isTIs the total variance, S, of all pixels within a window of a certain sizeWIs the sum of variances of pixels belonging to the same gray level within a window of a certain size; setting the number of scales, extracting an object R by calculating a multi-scale J-value image set and adopting the same clustering strategy as the built spectrum seismic damage visual dictionary for each scaleiCorresponding J-value earthquake damage visual dictionary WJi;
Seism based on geometric morphologyThe method for constructing the vicious vision dictionary comprises the steps of adopting area attributes, diagonal attributes and normalized moment of inertia attributes in morphological attribute profile MAPs as a class of earthquake vicious vision vocabulary respectively, and designing different morphological attribute operators to enable the characteristics of a target on specific scale parameters and attributes to have maximum response to be distinguished from other ground features, so as to obtain an object RiEarthquake damage visual dictionary WAREA of extracted area, diagonal and normalized moment of inertia attributesi、WDIAGi、WNMIi;
Defining an object R by integrating a visual dictionary of earthquake damage based on spectrum, texture and geometric morphologyiInitial earthquake damage visual dictionary Wi=[WRi,WGi,WBi,WJi,WAREAi,WDIAGi,WNMIi]。
Further, the specific process of step (3) is as follows:
(301) respectively calculating the structural similarity SSIM between any two wave bands or scale images under the same characteristic, and calculating the in-class penalty factor alpha of the spectral characteristic according to the structural similarity SSIMRGBJ-value, within class penalty factor alphaJWithin class penalty factor alpha for area attributesAREAWithin class penalty factor alpha for diagonal attributesDIAGAnd an intra-class penalty factor alpha for the normalized moment of inertia attributeNMI;
(302) Respectively adopting a mean value fusion strategy to R, G, B three-band images, multi-scale J-values and multi-scale sections with three attributes of area, diagonal and normalized moment of inertia to obtain five types of fusion images IMGRGB、IMGJ-value、IMGAREA、IMGDIAGAnd IMGNMI;
(303) Respectively calculating the structural similarity SSIM between each type of fusion image and other types of fusion images, and calculating the inter-class penalty factor beta of the five types of characteristics according to the structural similarity SSIMRGB、βJ、βAREA、βDIAGAnd betaNMI;
(304) Constructing an optimized earthquake damage visual dictionary W based on the punishment factors in the classes and the punishment factors between the classesi=[αRGBβRGBWRi,αRGBβRGB WGi,αRGBβRGB WBi,αJβJ WJi,αAREAβAREA WAREAi,αDIAGβDIAGWDIAGi,αNMIβNMI WNMIi]。
Further, the specific process of step (4) is as follows:
(401) dividing the objects in the potential building object set into four categories of 'intact buildings', 'partial earthquake damage buildings', 'ruins' and 'other ground objects';
(402) counting an optimized earthquake damage visual dictionary histogram of each object to serve as an input feature space of a random forest classifier;
(403) selecting a training sample set, and randomly and retractably extracting a plurality of sample subsets by using a Bagging method to ensure that the number of samples in each sample subset is equal to that of the samples in the training sample set;
(404) determining the number of the random features corresponding to the nodes, and respectively constructing a decision tree model for each sample subset;
(405) classifying each sample to be classified, voting according to the classification labels, and determining the final classification labels according to the number of votes.
Adopt the beneficial effect that above-mentioned technical scheme brought:
the method is based on the inherent boundary of the building, comprehensively utilizes the spectrum, the texture and the geometric morphological characteristics, and carries out multi-angle fine portrayal on the earthquake damage building by constructing and optimizing the earthquake damage visual dictionary model, thereby spanning the semantic gap from pixels to the building earthquake damage characteristics. On the basis, the earthquake damage semantic histogram is used for image expression, so that the subsequent training and recognition of the random forest classifier are facilitated. In experiments conducted on post-earthquake high-resolution remote sensing images in the Wenchuan and Yushu areas, the method provided by the invention shows excellent performance.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is an original image, a reference image and a sub-image of data sets 1 and 2;
FIG. 3 is a graph of the seismic hazard building extraction results of the data set 1;
FIG. 4 is a graph of seismic hazard building extraction results for data set 2;
FIG. 5 is a graph of the detection of buildings by seismic injury for sub-image 1 and sub-image 2 of data set 1;
FIG. 6 is a graph of the detection of buildings of seismic origin for sub-images 3 and 4 of data set 2;
fig. 7 is a graph of the overall accuracy versus K.
Detailed Description
The technical scheme of the invention is explained in detail in the following with the accompanying drawings.
The invention provides a method for detecting earthquake damage of a building by using post-earthquake high resolution remote sensing images, which mainly comprises the following four steps: (1) extracting a set of potential buildings; (2) constructing a multi-feature earthquake damage visual dictionary; (3) optimizing and expressing the image of the earthquake damage visual vocabulary; (4) earthquake damage building detection based on random forest classification. The specific implementation flow is shown in fig. 1.
(1) Potential building set extraction
The invention firstly adopts a high-resolution remote sensing image segmentation algorithm WJSEG to divide adjacent isolated pixels into object sets with semantic information, and specific implementation steps can be seen in documents WANG C, et al.A novel multi-scale segmentation algorithm for high resolution displaced images based on wavelet transform and improved JSEG algorithm, Optik-International Journal for Light and Electron Optics,2014,125(19):5588-iniThereby providing a basic analysis unit for subsequent processing.
Considering the ROI (region of interest) of the present invention as a seismic structure, in the initial object set RiniOn the basis, the non-building objects are subjected to primary screening, so that the number of subsequent analysis units is reduced, and the reduction of classification false detection rate is facilitated. By analysing dictionaries of non-building objectsType characteristics, the design screening rules are as follows:
area regulation: the area of each object, namely the number N of pixels contained in each object is countedpixelsIf N is presentpixelsIf the target is less than or equal to 80, the target is regarded as a small target such as a vehicle, noise and the like, and is removed. In addition, it should be specifically noted that the area threshold may be modified accordingly according to the actual situation.
Morphological Building Index (MBI, Morphological Building Index) rule: the MBI value of the pixel is obtained by calculating a multi-scale difference sequence by utilizing the characteristic that the building pixel shows high-brightness characteristics in the gray-scale image after the top hat transformation. If the MBI is larger, the probability that the pixel belongs to the building is higher. On the basis, the maximum inter-class variance method is used for determining the separation threshold T of the MBIMBI. Computing the MBI mean of all pixels in each objectRejection satisfiesAll of the objects of (1);
the rectangle degree and the length-width ratio are regular: the degree of rectangularity of an object is defined as Rd ═ Npixels/Nrectangle(wherein N ispixelsThe number of pixels contained in the minimum bounding rectangle of the object, NrectangleThe aspect ratio of the smallest circumscribed rectangle). At this time, if an object satisfies Rd < 0.8 and NrectangleIf the number is more than 5, the object is considered as a long and narrow target such as a road, a river channel and the like, and is removed.
Using the above 3 rules to discriminate RiniAll objects in (1) and all objects which are not eliminated are combined into a potential building object set Rpot。
(2) Construction of multi-feature earthquake damage visual dictionary
In the post-earthquake high-resolution remote sensing image, earthquake damage semantic knowledge implied in fine underlying visual features brought by improvement of spatial resolution is very rich. Therefore, the earthquake damage Visual dictionary is constructed from three angles of spectrum, texture and geometric morphology based on a Visual Bag of Words model (BoVW, Bag of Visual Words), and spans from the bottom layer Visual features to the semantic gap of the earthquake damage features of the building. The Visual Bag of Words (BoVW) model is an image processing algorithm developed from the field of natural language processing, and can provide an effective semantic feature description means for detection and identification of a specific target. According to the BoVW theory, the building earthquake damage visual dictionary is specifically as follows.
In post-earthquake scenes, areas of low grey value are often present because earthquake damage to buildings is rougher than damage sections of intact buildings. Therefore, the invention firstly carries out K-means clustering on the three-band images of the RGB model according to the gray values. On the basis, any object R can be obtainedi∈RpotThe corresponding K-dimensional spectral earthquake damage visual dictionary in the R, G, B three bands is: WR (pulse Width modulation)i=[redi1,redi2,…,rediK]、WGi=[greeni1,greeni2,…,greeniK]、WBi=[bluei1,bluei2,…,blueiK]。
In the aspect of texture characteristics, a perfect building generally has the characteristics of small internal gray level difference and strong texture structure consistency; the inner texture of the earthquake-damaged building has disordered structure and large gray level difference. Therefore, the local homogeneity index J-value is used as a class of earthquake damage visual vocabulary. J-value can effectively describe the texture complexity inside an object, and has the advantages of multi-scale and rotation invariance, which is defined as follows:
J-value=(ST-SW)/SW
wherein S isTIs the total variance, S, of all pixels within a window of a certain sizeWIs the sum of the variances of pixels belonging to the same gray scale within a window of a certain size. The smaller the value of J-value, the lower the degree of homogeneity inside the object, the more complex the texture, and thus the higher the possibility of belonging to a seismic structure. Setting the number of scales as 4, calculating a multi-scale J-value image set, and adopting the same clustering strategy as an RGB dictionary for each scale to extractTo take object RiThe corresponding J-value earthquake damage visual dictionary is WJi=[Ji11,…,Ji1K,Ji21,…,Ji2K,…,Ji41,…,Ji4K]。
In the aspect of geometric morphology, the composition of a perfect building in an earthquake scene is single, the shape is regular, and the structure is compact; earthquake damage to buildings presents a top surface with a mixture of intact and damaged areas, and debris consisting of collapsed building debris.
Therefore, the area, the diagonal line and the normalized moment of inertia NMI (normalized Mutual information) in the morphological Attribute profiles MAPs (morphological Attribute profiles) are respectively used as a class of earthquake damage visual vocabulary, and the implementation steps of the three attributes can be referred to in the literature. MAPs theory is developed from set theory, and the method distinguishes the characteristic appearance of a target on a specific scale parameter and attribute from the maximum response of other ground objects by designing different morphological attribute operators. Wherein, the area attribute is used for describing the scale of the internal structure of the object; the diagonal attribute reflects the diagonal length of the minimum circumscribed rectangle of the object and is used for describing the rule degree of the building; the NMI attribute reflects the mass distribution of the object, which is used to describe the compactness of the building. Taking the NMI attribute as an example, attribute filtering will progressively suppress objects smaller than a given NMI scale parameter. As the scale parameter increases, pixels in a sound building are progressively suppressed because of the higher degree of compactness. Therefore, the change trends of the area earthquake damage visual words of the intact buildings and the earthquake damage buildings in the multi-scale attribute section are obviously different, and the area and diagonal attributes have similar principles. The number of scales and the clustering strategy are the same as the RGB dictionary, and an object R is definediThe extracted earthquake damage visual dictionary of area, diagonal and NMI attributes are respectively: WAREAi=[AREAi11,…,AREAi1K,AREAi21,…,AREAi2K,…,AREAi41,…,AREAi4K]、WDIAGi=[DIAGi11,…,DIAGi1K,DIAGi21,…,DIAGi2K,…,DIAGi41,…,DIAGi4K]、WNMIi=[NMIi11,…,NMIi1K,NMIi21,…,NMIi2K,…,NMIi41,…,NMIi4K]。
On the basis, a 3-class earthquake damage visual dictionary is integrated and defined as an object RiThe initial earthquake damage visual dictionary is Wi=[WRi,WGi,WBi,WJi,WAREAi,WDIAGi,WNMIi]。
(3) Earthquake damage visual vocabulary image optimized expression
According to BoVW theory, on the basis of the constructed initial earthquake damage visual dictionary, for RpotRespectively counting the occurrence frequency of each visual vocabulary and constructing a visual vocabulary histogram. After all objects are traversed, the initial earthquake damage visual dictionary expression of the original image can be realized.
Nevertheless, different earthquake damage visual vocabularies have different expression effects on earthquake damage characteristics of the building, and redundant information possibly existing among the earthquake damage visual vocabularies not only increases the calculation complexity, but also reduces the detection precision due to mutual conflict among the different vocabularies as earthquake damage evidences. Therefore, the invention provides an optimized earthquake damage visual dictionary model based on intra-class penalty factors and inter-class penalty factors, so as to distinguish the contribution of different visual vocabularies to image expression, and the specific steps are as follows:
step 1: and calculating the intra-class penalty factor. Since there is usually a strong correlation between images of the same kind, different scales or bands, a lot of redundant information is contained. For this purpose, firstly, the structural similarity ssim (structural similarity) is used to calculate the correlation between any two bands or scale images under the same feature. SSIM combines the use of vector means, variances, and covariances with the unique advantages of bounded, symmetric, and maximum values over traditional "distance" similarity measures. SSIM between arbitrary vectors x and y is defined as follows:
wherein the content of the first and second substances,μx,μy,σxyx and y means, standard deviation, variance and covariance, respectively. The value interval of SSIM is [0,1 ]]The larger the value, the stronger the correlation between the two vectors, and the more redundant information. Taking the area attribute as an example, if p and q are serial numbers of any two different scales, the SSIM between the two different scales can be represented as SSIMpqAnd obtaining the intra-class penalty factor of the area attribute as follows:
similarly, the in-class penalty factors for which J-value, diagonal attribute and NMI attribute are available are divided into alphaJ、αDIAGAnd alphaNMI. The same strategy is adopted to calculate R, G, B the correlation between the three-band images, and the in-class penalty factor alpha of the spectral feature can be obtainedRGB。
Step 2: an inter-class penalty factor is calculated. Adopting mean value fusion strategy for R, G, B three-band image, multi-scale J-value and three-attribute multi-scale section respectively to obtain IMGRGB、IMGJ-value、IMGAREA、IMGDIAGAnd IMGNMIFive types of fused images. Next, inter-class penalty factors for RGB features are obtained using the following formula:
wherein, SSIMRGB&J,SSIMRGB&AREA,SSIMRGB&DIAGAnd SSIMRGB&NMIRespectively, the SSIM values between the RGB features and the four other feature fusion images. By analogy, the inter-class penalty factor beta of other similar characteristics can be obtainedRGB、βJ、βAREA、βDIAG、βNMI。
Step 3: updating visual dictionary and vision of earthquake damageA perceptual vocabulary histogram. Based on the punishment factors in and among classes, the constructed optimized earthquake damage visual dictionary is as follows: wi=[αRGBβRGB WRi,αRGBβRGB WGi,αRGBβRGB WBi,αJβJ WJi,αAREAβAREAWAREAi,αDIAGβDIAGWDIAGi,αNMIβNMI WNMIi]And correspondingly updating the visual vocabulary histogram as the input of a subsequent random forest classifier.
(4) Earthquake damage building detection based on random forest classification
Optimizing a set R of seismic damage visual vocabulary models and potential building objects based on the extractionpotIn the invention, an integrated classifier random forest RF (random forest) is adopted to obtain a final earthquake damage building detection result. The method comprises the following specific steps:
step 1: r is to bepotThe objects in (1) are divided into four categories of 'intact buildings', 'partially earthquake damaged buildings', 'ruins' and 'other ground objects'. The reason for further identifying the ruins in earthquake-damaged buildings is that such areas are usually the primary targets for emergency rescue after earthquake, and the 'partial earthquake-damaged buildings' can be used as reference for evaluation and reconstruction of earthquake damage after the earthquake;
step 2: statistics of each object RiThe optimized earthquake damage visual dictionary histogram is used as an input feature space of the random forest classifier;
step 3: selecting a training sample set, and randomly and repeatedly extracting C sample subsets by using a Bagging method to ensure that the number of samples in each subset is equal to that of the training sample set;
step 4: determining the number of the random features corresponding to the nodes, and respectively constructing a decision tree model for each sample subset;
step 5: classifying each sample to be classified, voting according to the classification labels, and determining the final classification labels according to the number of votes.
Analysis of experiments
Two groups of high-resolution remote sensing images acquired in a short time period after the earthquake are adopted in the experiment, and compared with the experiment results of two different methods, the high-resolution remote sensing images are analyzed.
The data set 1 is a post-earthquake QuickBild panchromatic-multispectral fusion remote sensing image in Wenchun area of Sichuan province in China, and comprises R, G, B three wave bands. The earthquake occurrence time is 12 days at 5 months in 2008, the highest earthquake magnitude is 8.0 levels, the image acquisition time is 3 days at 6 months in 2008, and the size is 1024 × 1024 pixels, as shown in (a) in fig. 2. The data set 2 is a post-earthquake GE01 panchromatic-multispectral fusion remote sensing image in Yushu area of Qinghai province of China, and comprises R, G, B three wave bands. The earthquake occurrence time is 4/2010, the highest earthquake magnitude is 7.1, the image acquisition time is 5/6/2010, and the size is 1024 × 1024 pixels, as shown in fig. 2 (b). In addition, the reference image is drawn manually using field investigation and visual interpretation. Some representative regions in the original image are extracted and represented by sub-images with different line-shaped frames in fig. 2, where (a) the solid line frame is sub-image 1 and the dashed line frame is sub-image 2, and (b) the solid line frame is sub-image 3 and the dashed line frame is sub-image 4, for further detailed discussion and analysis.
In both sets of experiments, K in the proposed method was set to 14 and 18, respectively, according to the trial and error method. Meanwhile, in order to analyze the effectiveness of the constructed spectrum, texture and geometric morphology visual earthquake damage dictionary, experiments are firstly carried out based on the proposed method and only three different earthquake damage dictionaries are respectively utilized. On the basis, experiments are carried out by utilizing all three earthquake damage visual dictionaries so as to analyze the complementarity of the earthquake damage visual dictionaries in the detection of earthquake damage buildings.
In addition, two advanced methods were chosen for comparative experiments: the method 1 comprises the steps of firstly optimizing an initial feature set by using an improved SEATH algorithm, and further obtaining a detection result by using a classification method based on membership. The initial feature set adopts all the spectral, texture and geometric morphological features extracted by the method, and other implementation steps are consistent with those of the original text. The method 2 adopts color and gradient characteristics to construct a visual bag-of-words model, and obtains the earthquake damage building detection result through classification of a Support Vector Machine (SVM). It should be noted that, in the following description,in order to ensure the consistency of precision evaluation, all comparison methods adopt an initial object set RiniAs a basic analysis unit.
The experimental results include intact buildings, partially earthquake damaged buildings, ruins and other four categories, and are represented by different colors, as shown in fig. 3 and 4.
The quantitative accuracy evaluation results are shown in tables 1 and 2. The identification effect of the method is obviously superior to that of other comparison methods through visual observation and quantitative precision evaluation. The overall precision of the method reaches more than 85%, and the fluctuation is only 0.2%, so that the optimized earthquake damage visual dictionary constructed by the method is feasible, effective and reliable in earthquake damage building detection. In addition, after the three types of earthquake damage visual dictionaries are comprehensively utilized, the overall precision is higher than that of a detection result obtained by singly adopting a certain type of earthquake damage dictionary, so that the complementarity of the spectrum, the texture and the geometric morphological characteristics in the detection of the earthquake damage building is proved. Nevertheless, the overall accuracy is still significantly lower than that of the method of the present invention, that is, the intra-class and inter-class penalty factors introduced by the present invention help to reduce redundant information and evidence conflicts, thereby improving the accuracy of subsequent classification. Compared with the method of the invention, the method 1 adopts the same initial feature set for optimization, and the precision difference of more than 10% in two groups of experiments further proves the excellent performance of the optimized earthquake damage visual dictionary constructed by the invention. The method 2 ignores the role of geometric morphological characteristics in the detection of earthquake-damaged buildings, and the detection precision is obviously lower than that of the method and the fluctuation is obvious.
Table 1: data set 1 quantitative accuracy evaluation results
Method/index | Overall accuracy (%) | False detection Rate (%) | Missing rate (%) | Kappa coefficient |
The method of the invention | 85.60% | 5.29% | 14.30% | 0.741 |
Spectral features | 72.70% | 11.10% | 27.30% | 0.553 |
Morphological characteristics | 72.30% | 11.30% | 27.70% | 0.547 |
Texture features | 69.50% | 12.80% | 30.50% | 0.502 |
Not optimized | 73.50% | 10.90% | 25.50% | 0.556 |
Method 1 | 74.90% | 10.00% | 25.10% | 0.596 |
Method 2 | 82.80% | 6.49% | 17.20% | 0.697 |
Table 2: data set 2 quantitative accuracy evaluation results
The recognition results of the representative sub-images are shown in fig. 5 and 6. As shown in fig. 5 and 6, the selected sub-images are all representative post-earthquake scenes, and buildings with different earthquake damage degrees are mixed with ground objects such as vegetation, roads, wastelands and the like, so that the development analysis of the sub-images is favorable for further verifying the effect of the method. Wherein, only the method of the invention makes correct judgment on non-building objects which have similar spectral characteristics and regular shapes with buildings, such as vegetation and bare land; roads and ruins are easy to be confused due to similar highlight features which are usually shown in images, and only the method 2 of the invention makes correct judgment; for the identification of the intact building, only the method of the invention makes correct judgment, and the other two methods generate missing detection or error detection; for the identification of part of earthquake-damaged buildings, only the method 2 of the invention makes correct judgment. Therefore, further analysis of the typical region after the earthquake shows that the method of the invention achieves more ideal effect and is obviously superior to the two comparison methods, and the result is consistent with the results of visual observation and quantitative evaluation.
In the earthquake damage visual word bag model constructed by the invention, the setting of the dictionary length K has a remarkable influence on the overall precision. Although the local optimal value can be determined by adopting a trial-and-error method, the degree of automation is low. For this reason, the overall accuracy was further analyzed as a function of K.
As shown in fig. 7, the overall accuracy in both experiments showed a tendency of gradually increasing first and gradually decreasing after reaching the peak as K increased. Meanwhile, dataset 1 and dataset 2 experiments achieved a maximum overall at K14 (overall 85.6%) and K18 (overall 85.8%), respectively. In addition, when K is taken as a value in the value taking intervals [12 and 20], the overall precision can reach more than 80%. Therefore, it is suggested that in practical applications, the optimal value can be determined manually or by trial and error in this interval.
Conclusion
On the premise of lacking pre-earthquake reference information, the invention provides a high-resolution remote sensing image earthquake damage building detection method based on an optimized visual dictionary. The method spans the semantic gap from the pixels to the building earthquake damage features, and constructs an efficient optimized visual earthquake damage dictionary model, thereby realizing multi-angle fine portrayal of the earthquake damage building. By carrying out experiments on a plurality of groups of high-resolution remote sensing images after the earthquake of different regions and different sensors, the overall accuracy of the method disclosed by the invention is up to more than 85%. Meanwhile, through further division of intact buildings, partial earthquake-damaged buildings and ruins in detection results, key decision support can be provided for emergency response rescue after earthquake and reconstruction after disaster.
The embodiments are only for illustrating the technical idea of the present invention, and the technical idea of the present invention is not limited thereto, and any modifications made on the basis of the technical scheme according to the technical idea of the present invention fall within the scope of the present invention.
Claims (5)
1. A method for detecting earthquake damage of buildings by high-resolution remote sensing images after earthquake is characterized by comprising the following steps:
(1) sequentially carrying out image segmentation and non-building object elimination on the obtained post-earthquake high-resolution remote sensing image to obtain a potential building object set;
(2) constructing a seismic hazard visual dictionary from three angles of spectrum, texture and geometric morphology based on the visual bag-of-words model;
(3) constructing an optimized earthquake damage visual dictionary model based on the intra-class penalty factors and the inter-class penalty factors;
(4) and based on the optimized earthquake damage visual vocabulary model and the potential building object set, a random forest classifier is adopted to obtain a final earthquake damage building detection result.
2. The method for detecting earthquake-damaged buildings according to the post-earthquake high-resolution remote sensing images as claimed in claim 1, wherein in the step (1), the non-building object elimination follows the following rules:
(a) area regulation: the area of each object, namely the number N of pixels contained in each object is countedpixelsIf N is presentpixelsIf the area is smaller than or equal to the preset area threshold value, the object is considered as a weak target and is removed;
(b) morphological building index rules: determining a separation threshold T for a morphological building index using a maximum inter-class variance methodMBICalculating the mean value of the morphological building index of all pixels in each objectRejection satisfiesAll of the objects of (1);
(c) the rectangle degree and the length-width ratio are regular: the degree of rectangularity of an object is defined as Rd ═ Npixels/NrectangleIn which N ispixelsThe number of pixels contained in the minimum bounding rectangle of the object, NrectangleIs at a minimum outsideConnecting the length-width ratio of the rectangle; if an object satisfies Rd < 0.8 and NrectangleIf the distance is more than 5, the object is considered as a long and narrow object and is removed.
3. The method for detecting earthquake damage buildings according to the post-earthquake high-resolution remote sensing images as claimed in claim 1, wherein in the step (2), the spectrum-based earthquake damage visual dictionary is constructed by respectively performing K-means clustering on the three-band images of the RGB model according to gray values to obtain any object Ri∈RpotVisual dictionary WR of seismic injury of K-dimensional spectrum corresponding to three wave bands of R, G, Bi、WGi、WBiWherein R ispotThe set of potential building objects obtained in the step (1);
the construction method of the earthquake damage visual dictionary based on the texture is that local homogeneity indexes J-value are adopted as a class of earthquake damage visual vocabulary, and the definition of the J-value is as follows:
J-value=(ST-SW)/SW
wherein S isTIs the total variance, S, of all pixels within a window of a certain sizeWIs the sum of variances of pixels belonging to the same gray level within a window of a certain size; setting the number of scales, extracting an object R by calculating a multi-scale J-value image set and adopting the same clustering strategy as the built spectrum seismic damage visual dictionary for each scaleiCorresponding J-value earthquake damage visual dictionary WJi;
The method for constructing the earthquake damage visual dictionary based on the geometric morphology comprises the steps of adopting area attributes, diagonal attributes and normalized moment of inertia attributes in morphological attribute profile MAPs as a class of earthquake damage visual vocabularies respectively, and designing different morphological attribute operators to enable the characteristics of a target on specific scale parameters and attributes to appear and distinguish the maximum response of the target with other ground objects so as to obtain an object RiEarthquake damage visual dictionary WAREA of extracted area, diagonal and normalized moment of inertia attributesi、WDIAGi、WNMIi;
Defining an object R by integrating a visual dictionary of earthquake damage based on spectrum, texture and geometric morphologyiInitial earthquake damage visual dictionary Wi=[WRi,WGi,WBi,WJi,WAREAi,WDIAGi,WNMIi]。
4. The method for detecting earthquake damage buildings by using post-earthquake high-resolution remote sensing images as claimed in claim 3, wherein the concrete process of the step (3) is as follows:
(301) respectively calculating the structural similarity SSIM between any two wave bands or scale images under the same characteristic, and calculating the in-class penalty factor alpha of the spectral characteristic according to the structural similarity SSIMRGBJ-value, within class penalty factor alphaJWithin class penalty factor alpha for area attributesAREAWithin class penalty factor alpha for diagonal attributesDIAGAnd an intra-class penalty factor alpha for the normalized moment of inertia attributeNMI;
(302) Respectively adopting a mean value fusion strategy to R, G, B three-band images, multi-scale J-values and multi-scale sections with three attributes of area, diagonal and normalized moment of inertia to obtain five types of fusion images IMGRGB、IMGJ-value、IMGAREA、IMGDIAGAnd IMGNMI;
(303) Respectively calculating the structural similarity SSIM between each type of fusion image and other types of fusion images, and calculating the inter-class penalty factor beta of the five types of characteristics according to the structural similarity SSIMRGB、βJ、βAREA、βDIAGAnd betaNMI;
(304) Constructing an optimized earthquake damage visual dictionary W based on the punishment factors in the classes and the punishment factors between the classesi=[αRGBβRGB WRi,αRGBβRGB WGi,αRGBβRGB WBi,αJβJ WJi,αAREAβAREA WAREAi,αDIAGβDIAGWDIAGi,αNMIβNMI WNMIi]。
5. The method for detecting earthquake damage buildings by using post-earthquake high-resolution remote sensing images as claimed in claim 1, wherein the specific process of the step (4) is as follows:
(401) dividing the objects in the potential building object set into four categories of 'intact buildings', 'partial earthquake damage buildings', 'ruins' and 'other ground objects';
(402) counting an optimized earthquake damage visual dictionary histogram of each object to serve as an input feature space of a random forest classifier;
(403) selecting a training sample set, and randomly and retractably extracting a plurality of sample subsets by using a Bagging method to ensure that the number of samples in each sample subset is equal to that of the samples in the training sample set;
(404) determining the number of the random features corresponding to the nodes, and respectively constructing a decision tree model for each sample subset;
(405) classifying each sample to be classified, voting according to the classification labels, and determining the final classification labels according to the number of votes.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010855069.8A CN112052756B (en) | 2020-08-24 | 2020-08-24 | Post-earthquake high-resolution remote sensing image earthquake damage building detection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010855069.8A CN112052756B (en) | 2020-08-24 | 2020-08-24 | Post-earthquake high-resolution remote sensing image earthquake damage building detection method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112052756A true CN112052756A (en) | 2020-12-08 |
CN112052756B CN112052756B (en) | 2023-08-01 |
Family
ID=73599826
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010855069.8A Active CN112052756B (en) | 2020-08-24 | 2020-08-24 | Post-earthquake high-resolution remote sensing image earthquake damage building detection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112052756B (en) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110569751A (en) * | 2019-08-23 | 2019-12-13 | 南京信息工程大学 | High-resolution remote sensing image building extraction method |
CN111191628A (en) * | 2020-01-06 | 2020-05-22 | 河海大学 | Remote sensing image earthquake damage building identification method based on decision tree and feature optimization |
-
2020
- 2020-08-24 CN CN202010855069.8A patent/CN112052756B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110569751A (en) * | 2019-08-23 | 2019-12-13 | 南京信息工程大学 | High-resolution remote sensing image building extraction method |
CN111191628A (en) * | 2020-01-06 | 2020-05-22 | 河海大学 | Remote sensing image earthquake damage building identification method based on decision tree and feature optimization |
Non-Patent Citations (5)
Title |
---|
TU JIHUI等: "Detection of Damaged Areas Based on Visual Bag-of-words Model from Aerial Remote Sensing Images", 《GEOMATICS AND INFORMATION SCIENCE OF WUHAN UNIVERSITY》, pages 691 - 696 * |
WADHAH AYADI等: "A hybrid feature extraction approach for brain MRI classification based on Bag-of-words", 《BIOMEDICAL SIGNAL PROCESSING AND CONTROL》, pages 144 - 152 * |
王超等: "基于优化视觉词典的震后高分遥感影像震害建筑物检测", 《电子测量与仪器学报》, pages 40 - 50 * |
石梵;王超;申?;张艳;仇星;: "基于稀疏词典的高分遥感影像震后建筑物检测", 仪器仪表学报, no. 07, pages 207 - 215 * |
马美娟;景元书;高益波;李谦;赵梦凡;王帅;: "不同灌溉处理稻田小气候特征及生理特性研究", 节水灌溉, no. 09, pages 21 - 27 * |
Also Published As
Publication number | Publication date |
---|---|
CN112052756B (en) | 2023-08-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110399909B (en) | Hyperspectral image classification method based on label constraint elastic network graph model | |
CN108491849B (en) | Hyperspectral image classification method based on three-dimensional dense connection convolution neural network | |
CN103971123B (en) | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) | |
CN111476170A (en) | Remote sensing image semantic segmentation method combining deep learning and random forest | |
CN112464942B (en) | Computer vision-based overlapped tobacco leaf intelligent grading method | |
CN109615008B (en) | Hyperspectral image classification method and system based on stack width learning | |
CN107480620B (en) | Remote sensing image automatic target identification method based on heterogeneous feature fusion | |
CN106503739A (en) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics | |
CN107832797B (en) | Multispectral image classification method based on depth fusion residual error network | |
CN106126585B (en) | The unmanned plane image search method combined based on quality grading with perceived hash characteristics | |
CN105321176A (en) | Image segmentation method based on hierarchical higher order conditional random field | |
CN111191628B (en) | Remote sensing image earthquake damage building identification method based on decision tree and feature optimization | |
CN109409384A (en) | Image-recognizing method, device, medium and equipment based on fine granularity image | |
CN105427309A (en) | Multiscale hierarchical processing method for extracting object-oriented high-spatial resolution remote sensing information | |
CN107895139A (en) | A kind of SAR image target recognition method based on multi-feature fusion | |
CN104850822B (en) | Leaf identification method under simple background based on multi-feature fusion | |
CN104240256A (en) | Image salient detecting method based on layering sparse modeling | |
CN110503613A (en) | Based on the empty convolutional neural networks of cascade towards removing rain based on single image method | |
CN109409389B (en) | Multi-feature-fused object-oriented change detection method | |
CN112488050A (en) | Color and texture combined aerial image scene classification method and system | |
CN103208011A (en) | Hyperspectral image space-spectral domain classification method based on mean value drifting and group sparse coding | |
CN106203522A (en) | Hyperspectral image classification method based on three-dimensional non-local mean filtering | |
CN108960276B (en) | Sample expansion and consistency discrimination method for improving spectral image supervision classification performance | |
Varish | A modified similarity measurement for image retrieval scheme using fusion of color, texture and shape moments | |
CN107358249A (en) | The hyperspectral image classification method of dictionary learning is differentiated based on tag compliance and Fisher |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |