CN109697425A - A kind of highway threat information recognition methods - Google Patents

Abstract

The invention relates to a method for identifying road threat information. The steps are: 1) constructing a ground object sample set according to photos or aerial photographs, selecting a typical area to establish a sample subset and a label set; 2) standardizing the sample data, and then synthesizing red, Green and blue band standardized pixel value sets, divide the sample set into training set and test set, build a threat identification model, and calculate the error rate; 3) Digitize the pictures or aerial photos of the information to be extracted, and carry out standardization processing, and then standardize The result is brought into the threat identification model to realize the extraction of threat information. The present invention takes into account the problem of identifying the threat information of color photos or aerial photographs by applying the threat identification model, and the calculation example shows that the provided threat information can provide guidance for assessing the degree of disaster, determining the preparation for rescue work, and formulating emergency rescue plans, etc. It is a beneficial supplement to the shortage of data such as remote sensing images, which can improve the efficiency of threat identification and play an important role in the rapid implementation of emergency rescue decision-making.

Description

A kind of highway threat information recognition methods

Technical field

The invention belongs to threaten source identification and contingency management field, it is related to using high-resolution remote sensing image, aviation phase The information such as piece or high-resolution picture identify that a kind of highway threat information identification in place occur in landslide scene or barrier lake Method.

Background technique

Up to the present, it is carried out both at home and abroad in the remote Sensing Interpretation of the threat identifications such as landslide mainly by visual interpretation, this needs Interpretation mark is established according to contents such as the geometrical characteristic of image, textural characteristics and context relations, and by veteran profession Personnel carry out interpretation, and still without being formed completely and the knowledge base of the threat identification of system, and the foundation for interpreting mark has area Domain characteristic, generalization are limited；How independent of professional interpretation personnel, the migration of interpretation knowledge is realized, without each area Interpretation re-establish interpretation mark, and realize the threat based on the amateur image data such as universal color photo quickly identify It is a problem to be solved, no matter the solving of this problem to the method for threat identification or instructs the progress of emergency management and rescue all to have Extremely important application value.

Summary of the invention

The object of the present invention is to provide it is a kind of can overcome drawbacks described above, using sampling camera photochrome or aerophotograph can know Not Hua Po, the highway threat information recognition methods that threatens of the highways such as barrier lake.Its technical solution is using following steps:

1) training, test set are established:

(1) it takes pictures to highway, atural object sample point is acquired on the photo or aerophotograph taken, construct ground object sample collection:

Wherein, in formula (1) n be each subset in air route sum, S indicates total sample set, S₁、S₂、S_i、S_nRespectively For the 1st, 2, i and n sample set；

(2) it selects representative region to establish sample set: selecting typical region for each sample set, and in representative region Sample point acquisition is carried out, mathematical description is as follows:

S_i={ x_j|x_j∈(z₁, z₂... z_m), x_j={ x_jr, x_jg, x_jb} (2)

x_jIndicate sample set S_iJ-th of sampled point, m is the typical sampling area number selected, z₁、z₂、z_mRespectively 1st, 2 and m typical sampling area, x_jr、x_jg、x_jbRespectively indicate the picture of the corresponding red, green, blue wave band acquisition of j-th of sampled point Element value；

(3) the tally set LB assignment of sample set S, mathematical description are as follows:

Wherein, LB is the corresponding tally set of sample set, LB₁、LB₂、LB_nRespectively sample set S₁, sample set S₂, sample Subset S_nThe 1st, 2 and n sub-set of tags；

2) typical feature information extraction, comprising the following steps:

(1) standardization of sample set S data:

Zx in formula (4), (5) and (6)_jr、zx_jg、zx_jbThe picture of red, green, blue wave band acquisition after respectively indicating standardization Minimum value and max function are sought in element value, min () and max () expression；

(2) synthesis of red, green, blue wave band standardization pixel value collection and its tally set:

S_j=[zx_jr, zx_jg, zx_jb] (7)

LB in formula (8)_jr, LB_jgAnd LB_jbRespectively indicate zx_jr、zx_jgAnd zx_jbCorresponding label value, bracket [] table Show matrix, the comma in matrix is for segmentation between the column and the column；

(3) sample set S is divided for training set S_trWith test set S_te, tally set LB is divided for training set LB_trAnd test set LB_te:

NS_tr≥NS_Te,NLB_tr≥NLB_te, NS_tr+NS_te=n, NLB_tr+NLB_te=n (11)

NS in formula (9)~(11)_tr、NS_teRespectively indicate the sample number of training set, the sample number of test set, NLB_tr、 NLB_teRespectively indicate the tally set sample number of training set and the tally set sample number of test set；

(4) threat identification model is constructed:

LModel={ Weight, error, ThrV } (12)

Weight=log₁₀(1-error)/error) (13)

ThrV=min (sum (LB_tr-LB_trf)) (14)

In formula (12), (13) and (14), Weight be forecast sample weight, error be forecast sample error and, ThrV is the threshold value of label assignment, and sum () is summing function, LB_trfFor the training sample label value of prediction；

Classifier number range 1~20, according to the threat identification model of building, to step since first classifier (1)~(3) training set, test set and its tally set obtained is trained, and the classification of training set is calculated according to formula (15) Device error rate, until classifier error rate is not more than ThrV, deconditioning:

E_tri=(NS_tr-R_tr)/NS_tr (15)

E in formula (15)_triFor the error in classification rate that i-th of classifier of training set obtains, R_trFor the correct number of classification；

(5) test data set and its tally set are substituted into threat identification model, and computation model error rate:

E_tei=(NS_te-R_te)/NS_te (16)

E in formula (16)_teiFor the error in classification rate that i-th of classifier of test set obtains, R_teFor the correct number of classification；

3) highway threat information identifies:

(1) the picture aerophotograph of information to be extracted is digitized；

(2) the picture aerophotograph after digitlization is standardized；

(3) standardization result of photo or aerophotograph is substituted into threat identification model formation (12)~(14), realizes and threatens letter Breath identification.

Compared with prior art, the present invention advantage are as follows:

A. this method has the ability of the highways threat information identification such as processing landslide and barrier lake, for similar scene occurs Emergency relief provides accident treatment template and guidance；Meanwhile by different photos or the pixel data of the red, green, blue color of aerophotograph The threat identification of standardization building can be improved the recognition efficiency of threat, provide one kind effectively to solve the highway threat identification Method；

B. position and the range of threat information can be quickly identified by aerophotograph or photochrome, this method is to remote sensing shadow As the advantageous supplement in the case of data deficiencies, the efficiency of threat identification can be improved, have in emergency management and rescue decision Rapid Implementation It plays an important role.

Specific embodiment

Technical solution of the present invention is described further below with reference to embodiment.Specific steps are as follows:

1) training, test set are established:

(1) photochrome comprising road is selected, constructs ground object sample collection S, sample set S includes road sample set S₁With non-rice habitats sample set S₂, and

It (2) is sample set S₁, sample set S₂Respectively 5 typical region z of selection₁~z₅, and in sample set S₁, sample Subset S₂Representative region in respectively acquire 100 sample points, x_jr、x_jg、x_jbRespectively indicate j-th of sampled point it is corresponding it is red, green, The pixel value of blue wave band acquisition:

x_1r=[134 114 147 111 112 108 104 122 112 134 132 114 117 115 115 129 112 138 113 120 142 151 124 130 130 130 121 123 121 125 122 119 161 147 126 123 131 140 153 113 127 121 129 122 133 125 130 131 126 119 108 152 136 118 129 127 129 130 138 152 120 151 147 150 150 141 145 148 147 145 154 154 165 158 162 163 162 159 165 153 147 148 149 136 149 147 138 106 126 144 88 93 100 102 94 96 97 81 98 104 99 89 100 92 122 95 98 95 84 87 104 95 91 99 88 103 96 80 89 97 94 104 85 88 104 95 91 99 88 103 95 81 88 98 94 104 86 88]

x_1g=[96 76 77 69 72 76 69 81 77 90 85 70 71 74 81 88 73 88 68 87 80 94 84 89 89 82 81 89 77 94 81 79 100 84 86 85 86 85 86 69 75 82 80 78 84 80 77 81 93 77 69 83 92 77 88 75 88 82 75 8681 81 126 131 129 119 125 126 130 127 134 133 148 143 141 142 137 137 146 131 125 131 129 117 126 126 117 92 108 116 78 81 90 86 82 86 87 70 84 90 89 78 86 82 103 83 82 84 84 79 92 89 81 85 77 91 86 74 77 82 84 92 79 76 92 89 81 85 77 91 87 74 78 82 84 93 79 76]

x_1b=[87 73 67 70 64 77 67 75 71 81 77 67 73 72 79 84 68 81 65 78 69 87 82 85 85 82 79 80 74 92 75 77 81 79 84 84 83 78 78 69 77 75 83 75 80 77 69 80 86 78 64 76 89 73 82 77 84 82 66 70 76 71 107 114 108 106 100 103 104 103 109 112 122 122 114 115 115 114 131 108 102 105 105 103 110 109 100 92 104 105 77 83 89 86 84 87 86 76 84 89 90 86 86 81 99 83 79 82 85 77 92 89 82 85 83 95 87 74 79 79 83 92 79 78 92 89 82 85 83 95 88 75 79 79 84 92 77 78]

x_2r=[85 83 80 78 79 56 49 76 56 51 73 65 67 67 61 48 65 57 66 54 65 23 67 66 59 45 60 49 37 57 57 63 69 58 53 65 50 62 43 62 75 73 66 64 64 64 75 52 59 55 60 56 41 49 72 45 70 65 58 53 43 30]

x_2g=[72 68 68 66 68 52 44 68 57 50 70 61 59 57 58 43 57 51 58 47 62 23 57 59 56 39 59 43 36 51 47 64 63 53 49 64 49 56 41 59 64 63 58 64 64 66 61 48 55 46 55 50 40 45 61 42 71 62 51 52 41 30]

x_2b=[66 63 68 66 66 51 38 66 49 45 63 58 56 56 53 39 54 51 56 39 57 23 55 53 49 39 54 45 32 51 46 59 63 49 48 60 45 56 44 54 60 61 56 62 62 65 60 45 52 39 52 50 35 42 59 37 65 57 52 48 44 28]

(3) to sample set S₁Tally set LB₁, sample set S₂Tally set LB₂Assignment, mathematical description are as follows:

LB₁=1, LB₂=2 (3)

2) typical feature information identifies, the specific steps are as follows:

(1) standardization of sample set S data:

Zx in formula (4), (5) and (6)_jr、zx_jg、zx_jbThe picture of red, green, blue wave band acquisition after respectively indicating standardization Minimum value and max function are sought in element value, min () and max () expression；

(2) synthesis of red, green, blue wave band standardization pixel value collection and its tally set:

S_j=[zx_jr, zx_jg, zx_jb] (7)

LB in formula (8)_jr、LB_jgAnd LB_jbRespectively indicate zx_jr、zx_jgAnd zx_jbCorresponding label value, bracket [] table Show matrix, the comma in matrix is for segmentation between the column and the column；

(3) sample set S is divided for training set S_trWith test set S_te, tally set LB is divided for training set LB_trAnd test set LB_te, n value is 200, NS_tr、NS_teRespectively indicate the sample number of training set, the sample number of test set, NLB_tr、NLB_teIt respectively indicates The tally set sample number of training set and the tally set sample number of test set, are 100:

(4) threat identification model is constructed, classifier number 20 is set, training set, the test set that step (1)~(3) are obtained And its tally set substitutes into threat identification model, and calculates 20 training set classifier error rates:

LModel={ Weight, error, ThrV } (12)

Weight=log₁₀(1-error)/error) (13)

ThrV=min (sum (LB_tr-LB_trf)) (14)

E_tri=(100-R_tr)/100 (15)

(5) experiment scene of embodiment is 4: first experiment scene is the high-resolution remote sensing image of residential quarters, Second environment is the landslide image picture downloaded on network, and third environment is similarly the high-resolution weir downloaded on network Lake picture is filled in, the scene of the 4th rescue is the mud-rock flow photo downloaded on network.By the test of 4 experiment scene photochromes Data set and its tally set substitute into threat identification model respectively, calculate its error rate:

E_tei=(100-R_te)/100 (16)

3) highway threat information identifies:

(1) digitlization of the picture aerophotograph of information to be extracted is realized；

(2) digitized picture aerophotograph is standardized；

(3) standardization result of photo is substituted into threat identification model formation (12)~(14), realizes threat information identification.

The recognition result of embodiment is as shown in table 1:

1 threat information of table extracts result and evaluation

Table 1 shows that the scene of 4 experiments, first experiment scene are affected since pavement small in cell is more The precision 93.75 of identification, second experiment scene are impended identification using the threat identification model of first experiment, when planning Between also reduce nearly half, recognition correct rate 100%, it is also complete that by the model barrier lake region occurs for third experiment scene It is extracted；The mud-rock flow occurrence scope of 4th experiment scene equally extracts, by being tested above it can be seen that logical The identification threatened can be completed by crossing migration computation model, the threat identification time also much smaller than testing an experiment scene, it can thus be concluded that This method provides safety guarantee for emergency episode rescue.

Claims (1)

1. a kind of highway threat information recognition methods, feature include following implemented step:

1) training, test set are established:

(1) it takes pictures to highway, atural object sample point is acquired on the photo or aerophotograph taken, construct ground object sample collection:

Wherein, in formula (1) n be each subset in air route sum, S indicates total sample set, S₁、S₂、S_i、S_nRespectively the 1st, 2, i and n sample set；

(2) it selects representative region to establish sample set: selecting typical region for each sample set, and carried out in representative region Sample point acquisition, mathematical description are as follows:

S_i={ x_j|x_j∈(z₁, z₂... z_m), x_j={ x_jr, x_jg, x_jb} (2)

x_jIndicate sample set S_iJ-th of sampled point, m is the typical sampling area number selected, z₁、z₂、z_mRespectively the 1st, 2 With m typical sampling area, x_jr、x_jg、x_jbRespectively indicate the pixel value of the corresponding red, green, blue wave band acquisition of j-th of sampled point；

(3) the tally set LB assignment of sample set S, mathematical description are as follows:

Wherein, LB is the corresponding tally set of sample set, LB₁、LB₂、LB_nRespectively sample set S₁, sample set S₂, sample set S_nThe 1st, 2 and n sub-set of tags；

2) typical feature information extraction, comprising the following steps:

(1) standardization of sample set S data:

Zx in formula (4), (5) and (6)_jr、zx_jg、zx_jbThe pixel value of red, green, blue wave band acquisition after respectively indicating standardization, Minimum value and max function are sought in min () and max () expression；

(2) synthesis of red, green, blue wave band standardization pixel value collection and its tally set:

S_j=[zx_jr, zx_jg, zx_jb] (7)

LB in formula (8)_jr, LB_jgAnd LB_jbRespectively indicate zx_jr、zx_jgAnd zx_jbCorresponding label value, bracket [] indicate square Gust, the comma in matrix is for segmentation between the column and the column；

(3) sample set S is divided for training set S_trWith test set S_te, tally set LB is divided for training set LB_trWith test set LB_te:

NS_tr≥NS_te, NLB_tr≥NLB_te, NS_tr+NS_te=n, NLB_tr+NLB_te=n (11)

NS in formula (9)~(11)_tr、NS_teRespectively indicate the sample number of training set, the sample number of test set, NLB_tr、NLB_tePoint It Biao Shi not the tally set sample number of training set and the tally set sample number of test set；

(4) threat identification model is constructed:

LModel={ Weight, error, ThrV } (12)

Weight=log₁₀(1-error)/error) (13)

ThrV=min (sum (LR_tr-LB_trf) (14)

In formula (12), (13) and (14), Weight is the weight of forecast sample, and error is the error and ThrV of forecast sample For the threshold value of label assignment, sum () is summing function, LB_trfFor the training sample label value of prediction；

Classifier number range 1~20, according to the threat identification model of building, since first classifier to step (1)~ (3) training set, test set and its tally set obtained is trained, and the classifier error of training set is calculated according to formula (15) Rate, until classifier error rate is not more than ThrV, deconditioning:

E_tri=(NS_tr-R_tr)/NS_tr (15)

E in formula (15)_triFor the error in classification rate that i-th of classifier of training set obtains, R_trFor the correct number of classification；

(5) test data set and its tally set are substituted into threat identification model, and computation model error rate:

E_tei=(NS_te-R_te)/NS_te (16)

E in formula (16)_teiFor the error in classification rate that i-th of classifier of test set obtains, R_teFor the correct number of classification；

3) highway threat information identifies:

(1) the picture aerophotograph of information to be extracted is digitized；

(2) the picture aerophotograph after digitlization is standardized；

(3) standardization result of photo or aerophotograph is substituted into threat identification model formation (12)~(14), realizes that threat information is known Not.