CN111008565B

CN111008565B - Fire trace detection method based on short wave infrared and thermal infrared data feature fusion

Info

Publication number: CN111008565B
Application number: CN201911065798.7A
Authority: CN
Inventors: 柳思聪; 郑永杰; 杜谦; 童小华; 谢欢; 冯毅; 金雁敏; 刘世杰; 陈鹏
Original assignee: Tongji University
Current assignee: Tongji University
Priority date: 2019-11-04
Filing date: 2019-11-04
Publication date: 2023-05-02
Anticipated expiration: 2039-11-04
Also published as: CN111008565A

Abstract

The invention relates to a fire trace detection method based on short wave infrared and thermal infrared data feature fusion, which comprises the following steps: 1) Acquiring Landsat-8 image data before and after the occurrence of fire in a region to be detected, and preprocessing front and back time phase images; 2) Respectively calculating a front time phase MNBR index difference characteristic dMNBR and a rear time phase bright temperature difference characteristic dBT according to the preprocessed time phase images; 3) A fusion algorithm GTF based on gradient transfer and total variation minimization is used for fusing the index difference value characteristic and the bright temperature difference value characteristic to generate a gray level map of the fire trace; 4) And carrying out self-adaptive threshold segmentation on the gray level map of the fire land, generating a binary change detection map of the fire land and the non-fire land, and further realizing extraction of the fire land. Compared with the prior art, the invention has the advantages of inhibiting background interference, accurately and rapidly extracting the burning place and the like.

Description

Fire trace detection method based on short wave infrared and thermal infrared data feature fusion

Technical Field

The invention relates to the field of automatic detection of multi-temporal remote sensing images of fire spots, in particular to a fire spot detection method based on Landsat-8 short wave infrared and thermal infrared data feature fusion.

Background

Forest fires are a world-class major natural hazard, whose occurrence is often accompanied by an imbalance in the ecosystem and destruction of the forest structure. Especially for a large-scale forest fire, the traditional fire trace land extraction depends on ground investigation and measurement, is greatly influenced by weather, and is time-consuming and labor-consuming.

By adopting the satellite observation technology, the remote sensing image can be used for directly detecting the burning place under a large scale, so that a plurality of defects of the traditional detection are avoided to a great extent. In recent decades, scholars at home and abroad develop extensive research on forest fire information extraction based on remote sensing images. Many studies have identified fire traces based on spectral variation characteristics. Such as: spectral indexing, principal component analysis algorithms, variation vector analysis, and the like. Among them, the spectral index method is often used to directly extract the fire spot with higher accuracy because of its simple form. The vast majority of spectral indexes for extracting fire traces are based on the red, near infrared, short wave infrared, thermal infrared and other bands. The proposed principle is mainly derived from the spectral reflectance variation of green vegetation before and after fire. The spectrum index based on the short-wave infrared design is most common, and the large difference of the spectrum reflectivity of the fire trace and the green vegetation in the short-wave infrared band is considered. The burning area can be extracted quickly based on the spectral characteristics, but the background pixel is misdetected as the target pixel in a large amount due to the existence of the phenomenon of 'same spectrum foreign matter'. The occurrence of forest fires is accompanied not only by a change in the spectrum but also by an increase in the local temperature. Meanwhile, within one month after the fire disaster occurs, the residual temperature of the fire trace still can be 5-6 ℃ higher than that of the background area, and some scholars can invert the temperature information through the thermal infrared wave band so as to extract the fire trace. Since temperature information is also greatly affected by weather or the like, there is a high possibility that a large number of target pixels will be missed by temperature extraction of fire traces alone.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide a fire trace detection method based on fusion of short-wave infrared and thermal infrared data characteristics.

The aim of the invention can be achieved by the following technical scheme:

a fire trace detection method based on fusion of short wave infrared and thermal infrared data features comprises the following steps:

1) Acquiring Landsat-8 image data before and after the occurrence of fire in a region to be detected, and preprocessing front and back time phase images;

2) Respectively calculating a front time phase MNBR index difference characteristic dMNBR and a rear time phase bright temperature difference characteristic dBT according to the preprocessed time phase images;

3) Fusing the index difference value characteristic and the bright temperature difference value characteristic to generate a gray scale map of the fire trace;

4) And carrying out self-adaptive threshold segmentation on the gray level map of the fire land, generating a binary change detection map of the fire land and the non-fire land, and further realizing extraction of the fire land.

In the step 1), preprocessing is performed on the front-back time phase images, specifically:

and eliminating radiation errors generated by the atmosphere through radiation calibration, atmosphere correction and image clipping in sequence.

In the step 2), for two short wave infrared bands of the Landsat-8OLI image, an exponential difference characteristic dMNBR reflecting the spectrum information of the fire trace is obtained by adopting band operation, specifically:

dMNBR＝MNBR _post -MNBR _pre

wherein MNBR is an improved normalized combustion ratio, ρ ₆ 、ρ ₇ MNBR in the 6 th and 7 th bands of Landsat-8OLI, respectively _post MNBR index, MNBR, which is post-disaster image _pre Is MNBR index of the pre-disaster image.

In the step 2), for the thermal infrared band of the Landsat-8TIRS image, the bright temperature difference value characteristic dBT reflecting the temperature information of the fire trace is obtained through a bright Wen Fanyan formula, which is specifically as follows:

dBT＝BT _post -BT _pre

wherein BT is the bright temperature, ρ _TIR1 Band 1, K of Landsat-8TIRS image ₁ 、K ₂ Are all constant, BT _post BT is the bright temperature of the post-disaster image _pre Is the bright temperature of the pre-disaster image.

In the step 3), a GTF algorithm is adopted to perform feature level fusion on the index difference feature dMNBR and the bright temperature difference feature dBT.

The feature level fusion specifically comprises the following steps:

Y＝X-dMNBR

X*＝Y*+dMNBR

where ε (X) is an empirical error, smaller empirical error indicates better fusion. m, n are the rows and columns of the image, respectively, λ is the input parameter, Y represents the variance map of X and dMNBR, Y is the optimal solution obtained with minimum total variance, J (Y) is the gradient at pixel i of the Y image,

the linear operators corresponding to the horizontal first-order difference and the vertical first-order difference are respectively adopted, and X is the gray level diagram of the fire trace after feature fusion.

In the step 4), an OTSU algorithm is adopted to perform adaptive threshold segmentation on the gray level graph X of the fire land after feature fusion to obtain a binary change detection graph of the fire land and the non-fire land.

In the step 4), the extraction of the fire trace specifically comprises the following steps:

in the binary change detection map for the fire land and the non-fire land, the value 1 is white, which represents the fire land, and the value 0 is black, which represents the non-fire land, thereby realizing the extraction of the fire land.

Compared with the prior art, the invention has the following advantages:

compared with other fire indexes, the dMNBR characteristics extracted by the MNBR index can accurately identify the fire land, but water and other complex backgrounds are easily divided into the fire land by mistake, the fire land is usually distinguished from other places by using the DBT characteristics extracted by BT, but the resolution is lower, so that the detail information of the fire land cannot be well represented, and as the GTF algorithm can effectively transfer the detail information of the dMNBR to dBT through total variation minimization, the dMNBR characteristic is fused by using the GTF algorithm, the fire land can be accurately identified, the change of water and other complex backgrounds can be inhibited, and the advantage complementation of the two characteristics is realized.

The gray level map of the fire trace after feature fusion contains the detail information of dMNBR, so that the fire trace is clearly visible, meanwhile, the gray level map of the fire trace also contains the temperature information of dBT, the change of water and other complex backgrounds is well restrained, and the accurate extraction of the fire trace is realized by carrying out self-adaptive threshold segmentation on the gray level map of the fire trace after feature fusion.

Drawings

FIG. 1 shows the degree of separation of the fire and non-fire for different methods in the GNPF study area.

FIG. 2 is a reference graph of the change in burn area for the GNPF study area.

Fig. 3 is a gray scale map of a fire trace obtained by different methods in GNPF research area, wherein fig. 3a is a CVA gray scale map, fig. 3b is a dNBR gray scale map, fig. 3c is a dMNBR gray scale map, fig. 3d is a dBT gray scale map, and fig. 3e is a gray scale map after feature fusion of the present invention.

Fig. 4 is a diagram of detecting a change in a fire trace binary image (abbreviated as a binary image) extracted by different methods in GNPF research areas, wherein fig. 4a is a CVA binary image, fig. 4b is a dNBR binary image, fig. 4c is a dMNBR binary image, fig. 4d is a dBT binary image, and fig. 4e is a binary image obtained by the present invention.

Fig. 5 is a partial binary contrast diagram of a GNPF study area water body (W1), wherein fig. 5a is a CVA water body region binary diagram, fig. 5b is a dNBR water body region binary diagram, fig. 5c is a dMNBR water body region binary diagram, fig. 5d is a dBT water body region binary diagram, and fig. 5e is a water body region binary diagram obtained by the present invention.

Fig. 6 is a partial binary contrast diagram of the GNPF study area fire trace (F1), wherein fig. 6a is a partial binary diagram of the CVA fire trace, fig. 6b is a partial binary diagram of the dNBR fire trace, fig. 6c is a partial binary diagram of the dMNBR fire trace, fig. 6d is a partial binary diagram of the dBT fire trace, and fig. 6e is a partial binary diagram of the fire trace obtained by the present invention.

Fig. 7 is a flow chart of the method of the present invention.

Detailed Description

The invention will now be described in detail with reference to the drawings and specific examples.

As shown in FIG. 7, the invention provides a fire trace detection method based on Landsat-8 short wave infrared and thermal infrared data feature fusion, which mainly comprises the following four steps:

(1) Image preprocessing

The pretreatment of the image data of the GNPF research area sequentially adopts radiometric calibration, atmospheric correction and image cutting, so that the radiation error generated by the atmosphere is eliminated, and the true reflectivity of the ground object is inverted.

(2) Calculating an index difference feature dMNBR and a bright temperature difference feature dBT

And (3) utilizing a formula 1 to carry out band operation on two short wave infrared bands of Landsat-8OLI to obtain an MNBR index map of a time phase before and after a fire disaster.

/>

Wherein MNBR is a novel fire trace index based on short wave infrared, ρ ₆ 、ρ ₇ The 6 th and 7 th bands of Landsat-8OLI, respectively.

And (3) utilizing a formula 2 to perform difference on the calculated front and back time phases MNBR indexes to obtain a difference characteristic of the MNBR indexes of the research area: dMNBR.

dMNBR＝MNBR _post -MNBR _pre (2)

Wherein, MNBR _post MNBR index, MNBR, which is post-disaster image _pre As MNBR index of the pre-disaster image, dMNBR is a difference characteristic of MNBR index of front and back phases.

The thermal infrared wave Duan Fanyan bright temperature of the Landsat-8TIRS front-rear time phase image is utilized in the formula 3:

K ₂ ＝1321.0789Kelvin

wherein BT represents the bright temperature, ρ _TIR1 Band 1, K for Landsat-8TIRS ₁ 、K ₂ Are constant.

The difference characteristic of the bright temperature of the research area is obtained by using the equation 4 to obtain the bright Wen Zuocha of the front and back phases obtained by inversion: dBT:

dBT＝BT _post -BT _pre (4)

wherein BT _post BT is the bright temperature of the post-disaster image _pre The dBT is the difference characteristic of the bright temperature of the front and back time phases.

(3) Gray scale map of fire trace obtained by feature fusion

And (3) performing feature fusion on the obtained index difference feature dMNBR and the bright temperature difference value feature dBT by using formulas (5) - (9) to obtain a gray scale image X reflecting the burning trace.

Y＝X-dMNBR (6)

X*＝Y*+dMNBR (9)

Where ε (X) is the empirical error, smaller empirical error represents better fusion. m, n are the rows and columns of the image, respectively, λ is the input parameter, Y represents the variation map of X and dMNBR, and Y is the optimal solution obtained by minimizing the empirical error. J (Y) represents the gradient at pixel i of the Y image, where

Linear operators corresponding to horizontal and vertical first order differences, respectively. X represents the gray scale of the fire trace after feature fusion.

(4) Binary change detection diagram of fire trace obtained by threshold segmentation

And (3) performing self-adaptive threshold segmentation according to the gray level map of the fire land by using an Otsu method (OTSU) to obtain a binary change detection result map of the fire land and the non-fire land.

To comprehensively compare the performance of the fire land extracted by the algorithm before and after fusion and verify the effectiveness and the advantages of the method, the qualitative and quantitative test analysis is carried out by selecting the change vector analysis (Change Vector Analysis, CVA for short), the normalized combustion ratio difference characteristic (The difference feature of Normalized Burn Ratio, dNBR for short), dMNBR and dBT together with the method result of the method. Of these, NBR and CVA are typical fire trace extraction indexes and change detection methods, respectively.

Examples:

the experimental data adopts medium resolution Landsat-8 satellite remote sensing 1 grade (L1) product data, and the map projection is UTM-WGS84 antarctic polar projection, which is derived from the United States Geological Survey (USGS) official network. The study was aimed at a granny screen national park fire (Grampians National Park Fire, GNPF for short) in victoria, southeast australia, and the image data of 2014, 1 month, 12 days (before disaster) and 2014, 1 month, 28 days (after disaster) were used in this experiment.

Experimental results:

1. different methods for separating fire and non-fire capability analysis

The separation degree of the gray level map of the fire land extracted in the research area 5 methods (CVA, dNBR, dMNBR, dBT and the invention) is calculated based on 0.1% of random samples, wherein the separation degree of more than 1 indicates better separation, and the higher the separation degree is, the stronger the capability of separating the fire land and the non-fire land is represented. The results are shown in FIG. 1. As can be seen from FIG. 1, the gray scale map of the fire trace obtained by fusion of the present invention has the highest degree of separation (3.26) which is far higher than dMNBR (2.49) and dBT (1.25) before fusion.

2. Qualitative analysis of gray scale map of fire trace obtained by different methods

The ability of different methods to distinguish between fire and non-fire spots is qualitatively analyzed, see in particular fig. 3. The fire disaster in the research area is mainly caused by thunder and lightning, and strong wind causes more and scattered small fire disaster areas. Qualitative comparative analysis shows that the CVA method (fig. 3 a) introduces many background changes, causing a large number of background areas to appear as fire-like highlight, and further detection is likely to occur for a large number of misplaced pixels. dNBR (FIG. 3 b)) and dMNBR (FIG. 3 c)) methods, while better able to suppress other clutter, appear in higher gray levels at some water edges. While dBT (fig. 3 d)) can well distinguish the fire spot from the background, it is difficult to obtain accurate fire spot boundaries and small fire spots due to its low spatial resolution, and further detection is highly likely to occur with large leakages. The feature fusion method (figure 3 e) provided by the invention realizes advantage complementation by fusing the two features of dMNBR and dBT, not only improves the overall contrast, ensures that the fire trace is clearly visible, but also has a strong inhibition effect on the water body.

3. Quantitative analysis of two-value change detection diagram of burning place obtained by different methods

The accuracy evaluation was performed based on the 5-method binary change detection results and verification data (change reference diagram (fig. 2)). The method comprises the following specific steps: and performing self-adaptive threshold segmentation on the gray level map of the fire trace by using an Otsu method (OTSU) to obtain binary change detection maps of the fire trace and the non-fire trace, and calculating a confusion matrix of the binary change detection maps and the change reference map. The specific evaluation index selects five types, namely Overall Accuracy (OA), kappa coefficient (Kappa Coefficient, K), error (CE), error (OE), total Error (TE). The evaluation results are shown in Table 1, and then the binary change detection graphs of 5 methods are compared and analyzed, and two local areas (water body and fire land) are selected to be deeply compared and analyzed.

Table 1 OTSU method accuracy evaluation table

The fire trace detection method based on Landsat-8 short wave infrared and thermal infrared data feature fusion provided by the invention has the highest precision, wherein the OA is as high as 99.70%. Next, dMNBR and OA was 99.38%. The lowest is the CVA method, and the total error pixels are about 60 ten thousand pixels higher than the method of the present invention. Fig. 4 shows a global binary image of 5 methods, the white box F1 circles the main fire spot of the investigation region, and the white box W1 circles a body of water. Fig. 5 corresponds to a partial enlarged view of W1, and fig. 6 corresponds to an enlarged view of F1. Wherein the graph (4 a) is a binary graph of CVA, and a great amount of background mistakes exist in the binary change detection result and are consistent with the qualitative analysis result. The CVA method is not a change detection method specifically proposed for the fire trace, but also for all the changed features, so that the misclassification error is too high. FIG. 4b shows the result of detecting the binary change of dNBR, and it can be seen from the partial water chart enlarged in FIG. 5b that dNBR divides the water into fire spots. In fig. 5c, the dMNBR method also has a small amount of misclassification at the boundary of the body of water. As can be seen from fig. 5d, the dBT method can well inhibit the water from being divided into fire spots. In the binary change detection result (figure 5 e) obtained by fusing dMNBR and dBT, the problem that the water body is wrongly divided into fire spots is solved. Fig. 4d shows the binary change detection result of dBT, and although there is no background misclassification, many burned spots are missed, and the spectrum index difference result: dNBR and dMNBR can avoid excessive omission of fire traces (see the partial contrast plot of fire traces given in FIG. 6 for details). By fusing dMNBR with the best precision in the spectrum index difference characteristic and dBT with the temperature difference characteristic, the problem of large amount of missed separation of the fire trace is solved.

In summary, a series of qualitative and quantitative test analysis proves that the provided fire spot detection method based on Landsat-8 short wave infrared and thermal infrared data feature fusion has higher fire spot detection performance. Particularly has obvious advantages in highlighting fire spots, inhibiting water bodies and other complex background changes.

Claims

1. A fire trace detection method based on the fusion of short-wave infrared and thermal infrared data features is characterized by comprising the following steps:

4) Performing self-adaptive threshold segmentation on the gray level map of the fire land to generate a binary change detection map of the fire land and the non-fire land, thereby realizing extraction of the fire land;

dMNBR＝MNBR _post -MNBR _pre

wherein MNBR is an improved normalized combustion ratio, ρ ₆ 、ρ ₇ MNBR in the 6 th and 7 th bands of Landsat-8OLI, respectively _post MNBR index, MNBR, which is post-disaster image _pre MNBR index as pre-disaster image;

dBT＝BT _post -BT _pre

2. The method for detecting fire marks based on fusion of short-wave infrared and thermal infrared data features according to claim 1, wherein in the step 1), preprocessing is performed on front-back time phase images, specifically:

3. The method for detecting fire trace based on the fusion of the features of the short-wave infrared data and the thermal infrared data according to claim 1, wherein in the step 3), the feature level fusion is carried out on the index difference feature dMNBR and the bright temperature difference feature dBT by adopting a GTF algorithm.

4. The method for detecting fire marks based on short-wave infrared and thermal infrared data feature fusion according to claim 3, wherein the feature level fusion specifically comprises the following steps:

Y＝X-dMNBR

/>

X*＝Y*+dMNBR

wherein epsilon (X) is an empirical error, and the smaller the empirical error is, the better the fusion effect is; m, n are the rows and columns of the image, respectively, λ is the input parameter, Y represents the variance map of X and dMNBR, Y is the optimal solution obtained with minimum total variance, J (Y) is the gradient at pixel i of the Y image,

5. The method for detecting fire trace based on feature fusion of short-wave infrared and thermal infrared data according to claim 1, wherein in the step 4), an OTSU algorithm is adopted to perform adaptive threshold segmentation on the gray level graph X of the fire trace after feature fusion, so as to obtain a binary change detection graph about fire trace and non-fire trace.

6. The method for detecting fire marks based on fusion of short-wave infrared and thermal infrared data features according to claim 5, wherein in the step 4), the extraction of the fire marks is specifically as follows: