Burned area identification method based on unsupervised remote sensing change detection

Abstract

The invention discloses a burned area identification method based on unsupervised remote sensing change detection, and the method specifically comprises the steps: calculating N spectral indexes of images before and after a fire, and generating N difference images; carrying out classification judgment on pixels in all the difference images, and fusing classification judgment results to obtain an initial sample; adopting a logic rule to determine burned area type pixels and non-burned area type pixels in the initial sample as final training labels; calculating local G indexes of the images before and after the fire, and cascading the local G indexes with the original spectral features to generate a training feature set; training a plurality of learners by taking the training label and the feature set as input; and performing decision-level fusion on the output of the trained multiple learners so as to identify the burned area. According to the method, large-area, high-precision and high-efficiency burned area extraction can be realized, and the method can better serve post-field-fire reconstruction and pre-disaster prevention.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a method for identifying burnt areas based on unsupervised remote sensing change detection. Background Art

[0002] In recent years, the incidence and severity of wildfires have been increasing due to climate change and human intervention, and have become one of the most serious disasters in the world. Wildfires not only pose a threat to human health, but also cause huge economic losses. Therefore, it is crucial to quickly and accurately extract the spatial distribution of burned areas after a wildfire occurs, which will help understand fire patterns and their driving forces, analyze the impact of fires on humans and the environment, and become the cornerstone for effective prevention and mitigation of wildfire disasters.

[0003] The traditional method of extracting burnt areas is mainly manual field survey. Although it is highly accurate, it is time-consuming, labor-intensive, and dangerous. Remote sensing technology has the advantages of wide coverage, low cost, and high safety. It has become the mainstream method for extracting burnt areas. In general, remote sensing methods for extracting burnt areas can be divided into two categories: spectral index methods using single-phase images and change detection methods using multi-phase images. The spectral index method constructs an index for identification based on the reflectivity differences of burnt areas in different bands. Common spectral indices include the normalized burn index and the normalized burn index 2. Although these spectral index methods are simple and effective, they have two shortcomings. First, they are easily affected by "same object, different spectrum, same spectrum, different objects". Second, it is impossible to determine whether the identified burnt areas are caused by recent wildfires or historical ones. In contrast, the change detection method that compares the images before and after the fire can detect a more accurate spatial distribution of burnt areas, so it is more competitive in practical applications.

[0004] According to whether manual labeling of training samples is required, burnt scar extraction methods based on change detection can be divided into two categories: supervised methods and unsupervised methods. Compared with supervised methods, unsupervised methods do not require manual labeling of training samples and can automatically identify burnt scars, thus better meeting the urgent need to quickly extract burnt scars after wildfires. However, existing burnt scar extraction methods based on unsupervised change detection still face problems such as difficulty in determining the optimal segmentation threshold and insufficient utilization of spatial features, resulting in poor performance in complex situations. Therefore, it is urgent to develop effective and efficient burnt scar extraction methods. Summary of the invention

[0005] Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the present invention provides a method for identifying burnt areas based on unsupervised remote sensing change detection.

[0006] Technical solution: The present invention provides a method for identifying burnt areas based on unsupervised remote sensing change detection, which is specifically:

[0007] Calculate N spectral indices of the images before and after the fire, and generate N difference images; N is an integer greater than or equal to 1;

[0008] Classify and judge the pixels in all difference images, fuse the classification results to get the initial sample, use logical rules to determine the burned area pixels and non-burned area pixels in the initial sample, and use them as the final training labels;

[0009] The local G index of the images before and after the fire is calculated and concatenated with the original spectral features to generate a training feature set.

[0010] Train multiple learners using training labels and training feature sets as input;

[0011] The outputs of multiple trained learners are fused at the decision level to identify burnt areas.

[0012] Furthermore, the spectral index includes: normalized combustion index NBR, normalized combustion index 2NBR2 and normalized combustion shortwave infrared index NBRSWIR, and the expressions of NBR, NBR2 and NBRSWIR are as follows:

[0013]

[0014] Among them, NIR, SWIR1 and SWIR2 are the reflectances of the near-infrared band, short-wave infrared band 1 and short-wave infrared band 2, respectively.

[0015] Furthermore, the expressions of the difference images D1, D2 and D3 are as follows:

[0016] D1=|NBR 2 -NBR 1 |

[0017] D2=|NBR2 2 -NBR2 1 |

[0018] D3=|NBRSWIR 2 -NBRSWIR 1 |

[0019] Among them, NBR 1 and NBR 2 They are NBR and NIR2 of the images before and after the fire. 1 and NIR2 2 NBR2 and NBRSWIR before and after the fire.1 and NBRSWIR 2 These are NBRSWIR images before and after the fire.

[0020] Furthermore, the C-means clustering method is used to classify and judge the pixels in all difference images. The expression is as follows:

[0021]

[0022] Where x(i,j) represents the pixel in the i-th row and j-th column of any difference image, D(i,j) represents the pixel value in the i-th row and j-th column of any difference image, and K 1 and K 2 are the strength values ​​of the two cluster centers respectively; ω BA ,ω u and ω Non-BA They are classified into burned land, uncertain land and non-burned land.

[0023] Furthermore, the logic rule is used to determine the burned land pixels and the non-burned land pixels in the initial sample, specifically:

[0024] Step 1: If the number of times that the pixel in the i-th row and j-th column is judged as the burned land class is greater than or equal to the preset threshold, the pixel in the i-th row and j-th column is taken as a candidate burned land class pixel; if the number of times that the pixel in the i-th row and j-th column is judged as the non-burned land class is greater than or equal to the preset threshold, the pixel in the i-th row and j-th column is taken as a candidate non-burned land class pixel; in other cases, the pixel in the i-th row and j-th column is considered to be an uncertain class;

[0025] Step 2: For any candidate burned land pixel, if all pixels in the neighborhood of the pixel are candidate burned land pixels, the pixel is identified as a burned land pixel; for any candidate non-burned land pixel, if all pixels in the neighborhood of the pixel are candidate non-burned land pixels, the pixel is identified as a non-burned land pixel.

[0026] Furthermore, in step 2, a 3×3 range around the pixel is used as a neighborhood range.

[0027] Furthermore, the expression of the local G index is as follows:

[0028]

[0029] Among them, x′ j represents the pixel value of the jth pixel in the image before or after the fire, n represents the total number of pixels, and w ijIt represents the spatial weight value between the i-th pixel and the j-th pixel in the image before or after the fire, and Gi represents the local G index of the i-th pixel in the image before or after the fire.

[0030] Furthermore, the learner includes: random forest, support vector machine and K nearest neighbor classifier.

[0031] Furthermore, the decision-level fusion is specifically as follows: if the number of learners that output the i-th row and j-th column pixels as burnt-land pixels is greater than or equal to a preset threshold, then the i-th row and j-th column pixels are identified as burnt-land pixels.

[0032] Beneficial effects: The present invention can achieve large-area, high-precision, and high-efficiency extraction of burned areas, and better serve post-wildfire reconstruction and pre-disaster prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the true value map of the spatial distribution of two phases of Sentinel-2 images and burnt areas in the study area of ​​this embodiment.

[0034] Figure 2 It is a flow chart of the present invention.

[0035] Figure 3 is the initial training sample.

[0036] Figure 4 is the final training sample.

[0037] Figure 5 This is the result of burnt area extraction based on random forest classifier.

[0038] Figure 6 It is the result of burnt area extraction based on support vector machine classifier.

[0039] Figure 7 This is the result of burnt area extraction based on the K nearest neighbor classifier.

[0040] Figure 8 It is a diagram of the application results of the present invention in the study area of ​​this embodiment.

[0041] Fig. 9 It is a local example diagram of the application results of the present invention in the study area of ​​this embodiment. DETAILED DESCRIPTION

[0042] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0043] The study area of ​​this example is located in Xichang City, Liangshan Yi Autonomous Prefecture, Sichuan Province. The true values ​​of the spatial distribution of burnt areas from two phases of Sentinel-2 images and manual visual interpretation are as follows: Figure 1 The images before and after the fire were taken on March 25, 2020 and April 9, 2020, respectively. The size is 1200 pixels × 1200 pixels. The spatial resolution after resampling is 10 meters, and the bands included are B2, B3, B4, B5, B6, B7, B8, B8a, B11 and B12.

[0044] like Figure 2 As shown, this embodiment specifically includes the following steps:

[0045] Step 1: Calculate the three spectral indexes (normalized combustion index, normalized combustion index 2 and normalized combustion short-wave infrared index) of the images before and after the fire, generate three difference images based on the simple difference method, and use fuzzy C-means clustering to segment the difference images to produce three segmentation results. Specifically include:

[0046] (a) The calculation methods of normalized burn index (NBR), normalized burn index 2 (NBR2) and normalized burn shortwave infrared index (NBRSWIR) are as follows:

[0047]

[0048] Where NIR, SWIR1 and SWIR2 are the reflectances of near infrared band, shortwave infrared band 1 and shortwave infrared band 2 respectively.

[0049] (b) The calculation method of the difference image is as follows:

[0050] D1=|NBR 2 -NBR 1 |

[0051] D2=|NBR2 2 -NBR2 1 |

[0052] D3=|NBRSWIR 2 -NBRSWIR 1 |

[0053] Where, NBR 1 and NBR 2 They are NBR and NIR2 of the images before and after the fire. 1 and NIR2 2 NBR2 and NBRSWIR before and after the fire. 1 and NBRSWIR 2These are NBRSWIR images before and after the fire.

[0054] (c) The method of using fuzzy C-means clustering to segment the three difference images and generate three segmentation results is as follows:

[0055]

[0056] Where x(i,j) represents the pixel in the i-th row and j-th column of any difference image, D(i,j) is the pixel value in the i-th row and j-th column of any difference image, and ω BA ,ω u and ω Non-BA The K 1 and K 2 are the intensity values ​​of the two cluster centers respectively.

[0057] Step 2: Use logical rules to fuse the three segmentation results to generate initial samples, such as Figure 3 As shown, the initial samples are then refined through neighborhood rules to generate the final training samples, as shown in Figure 4 Specifically including:

[0058] (a) The logical rules are described as follows:

[0059]

[0060] Where x(i,j) represents the pixel in the i-th row and j-th column, M(i,j) is the number of times the pixel in the i-th row and j-th column is judged as the burned area in the three segmentation results, and N(i,j) is the number of times the pixel in the i-th row and j-th column is judged as the non-burned area in the three segmentation results.

[0061] (b) The neighborhood rule is described as follows: in the initial sample, if all pixels in the 3×3 neighborhood around a burned area pixel are of the burned area class, then the pixel is selected as a burned area sample; if all pixels in the 3×3 neighborhood around a non-burned area pixel are of the non-burned area class, then the pixel is selected as a non-burned area sample; the remaining pixels are considered to be uncertain classes and are not used for subsequent classifier training.

[0062] Step 3: Calculate the local G index of the images before and after the fire using the following formula to extract spatial features:

[0063]

[0064] Among them, x′ j represents the pixel value of the jth pixel in the image before or after the fire, n represents the total number of pixels, and w ijIt represents the spatial weight value between the i-th pixel and the j-th pixel in the image before or after the fire, and Gi represents the local G index of the i-th pixel in the image before or after the fire.

[0065] Step 4: train the random forest, support vector machine and K nearest neighbor classifier respectively, and the burnt scar extraction results are as follows: Figure 5 , Figure 6 and Figure 7 As shown in the figure, the following formula is used to integrate the learning algorithm to generate the final burnt area extraction result:

[0066]

[0067] Where x(i,j) represents the pixel in the i-th row and j-th column, and T(i,j) is the number of times the pixel in the i-th row and j-th column is judged as a burned area in the three classification results.

[0068] The fire scar extraction results in this embodiment are as follows: Figure 8 As shown, a local example is Fig. 9 shown.

[0069] The precision rate, recall rate, F1 score, overall accuracy and Kappa coefficient of the burnt area extraction results in this embodiment are 90.74%, 93.87%, 92.27%, 96.92% and 0.90 respectively, indicating that the present invention has a high degree of automation and strong robustness when used for large-scale burnt area identification.

[0070] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A method for identifying burnt areas based on unsupervised remote sensing change detection, characterized in that: The method is as follows: Calculate N spectral indices of the images before and after the fire, and generate N difference images; N is an integer greater than or equal to 1; Classify and judge the pixels in all difference images, fuse the classification results to get the initial sample, use logical rules to determine the burned area pixels and non-burned area pixels in the initial sample, and use them as the final training labels; The local G index of the images before and after the fire is calculated and concatenated with the original spectral features to generate a training feature set. Train multiple learners using training labels and training feature sets as input; The outputs of multiple trained learners are fused at the decision level to identify burnt areas.

2. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 1, characterized in that: The spectral index includes: normalized combustion index NBR, normalized combustion index 2NBR2 and normalized combustion shortwave infrared index NBRSWIR. The expressions of NBR, NBR2 and NBRSWIR are as follows: Among them, NIR, SWIR1 and SWIR2 are the reflectances of the near infrared band, shortwave infrared band 1 and shortwave infrared band 2 respectively.

3. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 2, characterized in that: The expressions of the difference images D1, D2 and D3 are as follows: D1=|NBR2-NBR1| D2=|NBR22-NBR21| D3=|NBRSWIR2-NBRSWIR1| Among them, NBR1 and NBR2 are the NBRs of the images before and after the fire, respectively; NIR21 and NIR22 are the NBR2s of the images before and after the fire, respectively; NBRSWIR1 and NBRSWIR2 are the NBRSWIRs of the images before and after the fire, respectively.

4. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 1, characterized in that: The C-means clustering method is used to classify and judge the pixels in all difference images. The expression is as follows: Where x(i,j) represents the pixel in the i-th row and j-th column of any difference image, D(i,j) represents the pixel value in the i-th row and j-th column of any difference image, K1 and K2 are the intensity values ​​of the two cluster centers respectively; ω BA ,ω u and ω Non-BA They are classified into burned land, uncertain land and non-burned land.

5. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 1, characterized in that: The logic rule used to determine the burned land pixels and non-burned land pixels in the initial sample is specifically: Step 1: If the number of times that the pixel in the i-th row and j-th column is judged as the burned land class is greater than or equal to the preset threshold, the pixel in the i-th row and j-th column is taken as a candidate burned land class pixel; if the number of times that the pixel in the i-th row and j-th column is judged as the non-burned land class is greater than or equal to the preset threshold, the pixel in the i-th row and j-th column is taken as a candidate non-burned land class pixel; in other cases, the pixel in the i-th row and j-th column is considered to be an uncertain class; Step 2: For any candidate burned land pixel, if all pixels in the neighborhood of the pixel are candidate burned land pixels, the pixel is identified as a burned land pixel; for any candidate non-burned land pixel, if all pixels in the neighborhood of the pixel are candidate non-burned land pixels, the pixel is identified as a non-burned land pixel.

6. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 5, characterized in that: In step 2, a 3×3 range around the pixel is used as a neighborhood range.

7. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 1, characterized in that: The expression for the local G index is as follows: Among them, x′ j represents the pixel value of the jth pixel in the image before or after the fire, n represents the total number of pixels, and w ij G represents the spatial weight value between the i-th pixel and the j-th pixel in the image before or after the fire. i Represents the local G index of the i-th pixel in the image before or after the fire.

8. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 1, characterized in that: The learners include: random forest, support vector machine and K nearest neighbor classifier.

9. The method for identifying burnt areas based on unsupervised remote sensing change detection according to claim 1, characterized in that: The decision-level fusion is specifically as follows: if the number of learners that output the i-th row and j-th column pixels as burnt-land pixels is greater than or equal to a preset threshold, then the i-th row and j-th column pixels are identified as burnt-land pixels.

Images