Forest fire patch and occurrence time extraction method based on GEE platform and VCT algorithm

Abstract

The present invention discloses a method for extracting forest fire patches and occurrence times based on a GEE platform and a VCT algorithm. The method comprises utilizing the VCT algorithm results and all available Landsat images, and using the obtained decision tree rules based on the standardized vegetation indexes NBRr, NDMIr, NDVIr and the calculated vegetation index differences dNBRr, dNDMIr, and dNDVIr. Ideally, the method can quickly determine the specific time of occurrence of forest disturbance to between the acquisition dates of two images, and can further shorten the time when Landsat8 and Landsat9 data exist simultaneously. On this basis, the method uses the decision tree rules to further distinguish between fire disturbance and non-fire disturbance based on the vegetation index values ​​of the images closest to the image before and after the disturbance occurs. In these processes, it is not necessary to select a large amount of training data for different regions respectively, and the same set of decision tree rules can be used to judge images acquired in different regions and at different times. Compared with many existing methods, the method has better versatility and maintains higher accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of forest fire information extraction, and in particular to a method for extracting forest fire patches and occurrence times based on a GEE platform and a VCT algorithm. Background Art

[0002] As the backbone of Earth's terrestrial ecosystems, forests are inevitably subject to various types of disturbance, such as fire, deforestation, pests and diseases. Forest fires, one of the most frequent types of disturbance, can cause significant damage to forest ecosystems and the well-being of the people who rely on them. Effective monitoring and management of forest fires significantly impacts the safety and well-being of forests and those who depend on them. Traditional methods for monitoring forest disturbances rely primarily on extensive ground-based surveys. Due to their high cost, high risk, long lead times, and limited spatial accessibility, these methods cannot meet the high-precision and timely requirements for monitoring forest disturbances today. With the rapid development and advancement of modern remote sensing technology and computer image processing and analysis algorithms, analyzing multi-temporal satellite remote sensing imagery has provided a more convenient and timely approach for monitoring forest change. Early methods for identifying disturbances using remote sensing data typically relied on comparing images taken over relatively long time intervals. This comparison approach resulted in some low-intensity disturbances potentially recovering within the time interval, making them unrecognizable or unrecorded. With the increase in the types of remote sensing data and the increase in the types of freely available data, such as Landsat and Sentinel satellites, coupled with the development of automatic analysis algorithms, there has been a significant increase in studies using denser time series of remote sensing images, such as annual data, to analyze forest changes.

[0003] Several studies have been conducted on extracting forest fire disturbance information, including those based on low-spatial-resolution remote sensing imagery, such as MODIS, and medium-spatial-resolution remote sensing data, such as Landsat and Sentinel-2. Extracting forest fire disturbance information based on low-spatial-resolution data often overlooks detailed fire information, limiting its guiding role in developing post-fire restoration measures. Landsat data, with its 30-meter spatial resolution, 16-day temporal resolution, and multispectral information, has been used in many studies related to forest fire disturbance. Based on Landsat imagery, several automated time series analysis algorithms have been developed for forest disturbance monitoring, including Vegetation Change Tracker (VCT), LandTrendr, Continuous Change Detection and Classification (CCDC), and Breaks for Additive Seasonal and Trend (BFAST). Among these methods, the VCT, LandTrendr, and BFAST methods can only determine the duration of forest disturbance within a year. Post-fire forests often recover within a year, with some forest types recovering to over 50% of their pre-fire levels within a year. This affects the accuracy of forest fire information extraction and the results of subsequent studies assessing forest fire losses and carbon emissions. While the CCDC method can accurately identify the onset of forest disturbance, its operation requires a large number of forest disturbance samples within the study area. Furthermore, these methods can only determine whether forest disturbance has occurred, but cannot distinguish whether the disturbance is caused by fire. Other studies, building on these forest disturbance monitoring methods, have used algorithms such as random forests (RF) to distinguish various types of forest disturbance, including fire disturbance. However, in these studies, varying conditions across regions can lead to differences in the spectral signatures of the same ground features, often requiring the selection of a large number of training samples with different forest disturbance types based on the study area to distinguish between disturbance types. None of these studies has found a more convenient and rapid method for extracting more accurate temporal and spatial information about forest fires. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the existing technology, the present invention provides a method for extracting forest fire patches and occurrence time based on the GEE platform and VCT algorithm, which can more conveniently and quickly perform more accurate forest fire time information extraction and spatial information extraction.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting forest fire patches and occurrence time based on the GEE platform and the VCT algorithm, comprising the following steps:

[0006] S1. Run the VCT algorithm in the GEE platform using the VCT code provided by the GEE platform. The VCT algorithm is used to obtain the continuous forest mask and the forest disturbance patch mask of each year in the study area for subsequent analysis.

[0007] S2. Forest fire disturbance analysis was performed using three vegetation spectral indices: Normalized Burning Ratio (NBR), Normalized Moisture Difference Index (NDMI), and Normalized Vegetation Index (NDVI);

[0008] S3, normalize the three vegetation indices using the persistent forest mask obtained in S1. The three normalized indices are expressed as NBRr, NDMIr, and NDVIr;

[0009] S4. Based on the index values ​​of all available images in the interference year, the three index differences between two adjacent images are further calculated, which are expressed as dNBRr, dNDMIr, and dNDVIr respectively;

[0010] S5. Construct a decision tree based on the changes of the standardized vegetation index in the remote sensing image time series during the disturbance year;

[0011] S6. Using a decision tree rule to compare the standardized index difference value images, determine between which two images the forest disturbance pixel interference occurs, thereby determining the specific time when the forest disturbance occurs, and further use the decision tree to determine the type of disturbance based on the vegetation index difference value of the two images before and after the disturbance time;

[0012] S7. After determining the specific time range of forest disturbance occurrence within the disturbance year, the disturbance occurrence time range obtained by the decision tree rule was compared with the actual disturbance occurrence time to evaluate the time accuracy of the forest disturbance occurrence time extraction;

[0013] S8. Compare the fire or non-fire forest disturbance types obtained by the decision tree rules based on random points and the actual fire or non-fire disturbance types, and calculate the overall accuracy and Kappa coefficient by constructing a confusion matrix to evaluate the spatial accuracy of forest disturbance type extraction.

[0014] Preferably, the NBR is applied to forest fire detection and fire severity analysis, the NDMI reflects the change in surface humidity before and after forest disturbance occurs, and the NDVI represents the change in surface vegetation cover.

[0015] Preferably, the values ​​of the three vegetation indices are expanded to between 0 and 2000 without affecting the results, and the calculation formulas for the three vegetation indices using Landsat OLI images are as follows:

[0016]

[0017]

[0018]

[0019] For Landsat OLI image data, ρ5 is the reflectance of the near-infrared band, with a central wavelength of 0.86 μm; ρ7 is the reflectance of the shortwave infrared 2 band, with a central wavelength of 2.20 μm; ρ6 is the reflectance of the shortwave infrared 1 band, with a central wavelength of 1.61 μm; and ρ4 is the reflectance of the red light band, with a central wavelength of 0.66 μm.

[0020] Preferably, the formula used for index normalization is as follows:

[0021] I r =I i -I f

[0022] Among them I r is the standardized pixel index value; I f is the mean index value of all persistent forest pixels in an image; I i is the pixel index value before normalization.

[0023] Preferably, the calculation formula of the index difference value is as follows:

[0024] dI=I latter -I former

[0025] Where dI is the index difference value calculated from two adjacent images; I former is the pixel index value in the earlier image of the two adjacent images; I latter It is the pixel index value in the later of the two images. If a pixel in an image is masked by a cloud pixel, the index difference value of the pixel is calculated from the index value of the nearest cloud-free pixel.

[0026] Preferably, the decision tree rules include two sets of decision tree rules obtained by performing decision tree analysis on sample pixels from the two images, and the accuracy of the decision tree rules in forest disturbance time extraction and fire patch extraction in the two areas is verified respectively. The decision tree rules include rule one and rule two. Rule one is used to determine the specific time when forest disturbance occurs in the pixel, and rule two is used to determine whether the disturbance is fire disturbance or non-fire disturbance.

[0027] Preferably, the calculation formula of time accuracy TA is as follows:

[0028]

[0029] Where n is the number of random points whose actual interference occurrence time matches the interference occurrence time obtained by decision tree rule 1, and N is the number of all random points.

[0030] Preferably, the overall accuracy P A The calculation formula of Kappa coefficient is as follows:

[0031]

[0032]

[0033]

[0034] Among them, n is the number of random points whose judgment results of the decision tree rule are consistent with the actual interference type; m is the number of random points whose judgment results are inconsistent with the actual interference type, n fm is the number of random points judged as fire disturbance using the decision tree rule; n fr is the number of random points whose real disturbance type is fire disturbance; n nm is the number of random points judged as non-fire disturbance using the decision tree rule; n nr is the number of random points whose true disturbance type is non-fire disturbance.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention is based on the results of the VCT algorithm and all available Landsat images obtained by the GEE platform. According to the standardized vegetation index NBRr, NDMIr, NDVIr and the calculated vegetation index difference dNBRr, dNDMIr, dNDVIr, the obtained decision tree rules can, under ideal conditions, quickly determine the specific time of forest disturbance to between the dates of acquisition of the two images. When Landsat 8 and Landsat 9 data exist at the same time, the time can be further shortened. On this basis, the decision tree rules are used to further distinguish between fire disturbance and non-fire disturbance based on the vegetation index value of the image closest to the time before and after the disturbance occurs. In these processes, there is no need to select a large amount of training data for different areas. The same set of decision tree rules can be used to judge images acquired in different areas and at different times. Compared with many existing methods, it has better versatility and maintains a higher accuracy.

[0037] 2. The results of the VCT, LandTrendr, and BFAST algorithms commonly used in existing research can only determine the time of forest disturbance within a single disturbance year and cannot further determine the type of disturbance that occurred. Further research based on these algorithms to distinguish forest disturbance types often requires a large number of forest disturbance type samples within the study area to train the classification model. Compared with existing research, the present invention does not require the selection of a large number of training samples based on different regions. It can quickly determine the specific time of forest disturbance and whether the forest disturbance is fire disturbance or non-fire disturbance. It also shows good results in images from different regions and has better versatility.

[0038] 3. The decision tree rule used in the present invention is constructed based on the standardized vegetation index. During the standardization process, the differences between images in different regions and at different times are reduced, thereby obtaining a more general decision tree rule for distinguishing between fire disturbance and non-fire disturbance. At the same time, based on the present invention, if the spectral characteristics of forest disturbance patches, such as surface temperature, brightness, greenness, spatial characteristics, morphological characteristics and texture characteristics, can be further combined with other types of data, such as high spatial resolution satellite image data, LiDAR data, etc., then it may be possible to further distinguish different types of disturbance in non-fire disturbance, such as logging, pests and diseases, etc., laying the foundation for further research. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0040] In the attached figure:

[0041] Figure 1 is a flow chart of the method of the present invention;

[0042] Figure 2 This is the VCT forest disturbance analysis result map from 2013 to 2021. The left map is Xichang City, strip number 130 / 41; the right map is Muli County, strip number 131 / 41;

[0043] Figure 3 This is a comparison of vegetation index before and after standardization in Xichang and Muli areas;

[0044] Figure 4 It is the change of dNBRr, dNDMIr, and dNDVIr values ​​of disturbed / non-disturbed pixels in the disturbance year and the extracted threshold map of forest disturbance occurrence;

[0045] Figure 5 This is a map of the specific time range of forest disturbance occurrence extracted in Xichang area;

[0046] Figure 6 This is a map of the specific time range of forest disturbance occurrence extracted in Muli area;

[0047] Figure 7 It is a three-dimensional decision tree rule boundary used to distinguish forest fire disturbance from non-fire disturbance and the distinction effect map of some disturbance pixels in Xichang and Muli areas;

[0048] Figure 8 This is the fire disturbance and non-fire disturbance patch map extracted in Xichang area;

[0049] Figure 9 This is the fire disturbance and non-fire disturbance patch map extracted from Muli area. DETAILED DESCRIPTION

[0050] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0051] The method for extracting forest fire patches and occurrence time based on the GEE platform and VCT algorithm includes the following steps:

[0052] 1. Specific methods and steps to achieve functions or results:

[0053] The raster data used in this paper comes from the Google Earth Engine (GEE) platform and includes Landsat 8 surface reflectance image data, 30m resolution digital elevation DEM image data, and 30m resolution land use classification data (GlobeLand 30). The Landsat 8 surface reflectance image data has track numbers 130 / 41 and 131 / 41, and includes one-year interval imagery used to run the VCT algorithm, as well as all available imagery from the intervening years used to extract forest fire information. The Landsat 8 data information is shown in the following table: Landsat 8 OLI image data information used in the VCT process acquired by the GEE platform, and the table of all available Landsat OLI images from the intervening years within the GEE platform:

[0054]

[0055]

[0056] Landsat 8 surface reflectance data acquired through the GEE platform undergoes radiometric calibration and atmospheric correction, and cloud pixels are removed using the cloud mask algorithm CFMASK. For orbits 131 / 41, no high-quality vegetation growth-season imagery was available in 2017 for use in the VCT algorithm. Therefore, only changes between 2016 and 2017 were analyzed.

[0057] Based on the results of the VCT algorithm, the present invention standardized the normalized burn rate (NBR), normalized moisture difference index (NDMI), and normalized vegetation index (NDVI). Through decision tree analysis, a method was obtained that can quickly extract the specific time of forest fire occurrence and distinguish forest fire interference from non-fire interference, and the accuracy of the method was evaluated. The specific process of the method is as follows: Figure 1 The real forest disturbance patch information used in the decision tree analysis and accuracy verification process was obtained from local institutions. The disturbance patch information includes the time when the forest disturbance occurred and whether the disturbance type is fire disturbance or non-fire disturbance, as shown in the following real forest disturbance patch information table:

[0058]

[0059]

[0060] 1.1. Extracting forest disturbance events using the VCT algorithm

[0061] The VCT algorithm, proposed in 2010, has been applied in various regions around the world with good performance. Therefore, the results of the VCT algorithm were directly used in the subsequent analysis. The VCT code provided by the GEE platform was used to run the VCT algorithm within the GEE platform. The VCT algorithm was run using one-year intervals of Landsat OLI imagery, 30-meter resolution digital elevation map (DEM) imagery, and 30-meter resolution land use classification data (GlobeLand 30). The VCT algorithm generated a persistent forest mask for the study area and a forest disturbance patch mask for each year for subsequent analysis.

[0062] 1.2 Calculation and standardization of spectral indices

[0063] 1.2.1 Spectral index calculation

[0064] After a forest fire occurs, the spectral characteristics of the surface will change, based on which the occurrence of forest fires can be identified and analyzed. In this invention, three vegetation spectral indices are used to analyze forest fire disturbances, namely the normalized burn rate (NBR), the normalized moisture difference index (NDMI), and the normalized vegetation index (NDVI). Among them, NBR has been widely used in forest fire detection and fire severity analysis; NDMI can reflect the changes in surface moisture before and after forest disturbance occurs; NDVI is one of the most widely used vegetation indices and can well represent changes in surface vegetation cover. In the process of calculating the three vegetation indices, in order to facilitate calculation and mapping, the values ​​of the three vegetation indices are expanded to between 0 and 2000 without affecting the results. The calculation formulas for the three vegetation indices using Landsat OLI images are as follows:

[0065]

[0066]

[0067]

[0068] For Landsat OLI image data, ρ5 is the reflectance of the near-infrared band, with a central wavelength of 0.86 μm; ρ7 is the reflectance of the shortwave infrared 2 band, with a central wavelength of 2.20 μm; ρ6 is the reflectance of the shortwave infrared 1 band, with a central wavelength of 1.61 μm; and ρ4 is the reflectance of the red light band, with a central wavelength of 0.66 μm.

[0069] 1.2.2 Spectral index standardization

[0070] In the present invention, in order to better determine the specific time when forest disturbance occurs, it is necessary to use all available Landsat image data in the year of disturbance. However, due to the differences in atmospheric conditions and light-object geometric characteristics at different times and in different regions, the surface reflectance characteristics at different spatial locations and at different times will also show differences, and the spectral consistency between different images cannot be guaranteed. Therefore, in the present invention, the three vegetation indices (NBR, NDMI, NDVI) are standardized using the continuous forest mask obtained in 1.1. The three standardized indices are expressed as (NBRr, NDMIr, NDVIr); the formula used for index standardization is as follows:

[0071] I r =I i -I f

[0072] Among them I r is the standardized pixel index value; I f is the mean index value of all persistent forest pixels in an image; I i is the pixel index value before normalization.

[0073] 1.3. Determining the time and type of interference

[0074] 1.3.1 Calculation of Index Difference Value

[0075] After normalizing the three vegetation indices to obtain NBRr, NDMIr, and NDVIr, the differences between the three indices between two adjacent images were further calculated based on the index values ​​of all available images in the interference year, and expressed as dNBRr, dNDMIr, and dNDVIr, respectively. The calculation formulas are as follows:

[0076] dI=I latte -I former

[0077] Where dI is the index difference value calculated from two adjacent images; I former is the pixel index value in the earlier image of the two adjacent images; I latter is the pixel index value in the later of the two images. If a pixel in an image is masked by a cloud pixel, the index difference value of that pixel is calculated from the index value of the nearest cloud-free pixel.

[0078] 1.3.2. Building a Decision Tree

[0079] The present invention uses the C4.5 algorithm to perform decision tree analysis to construct a decision tree. The samples used to train the decision tree come from the band number 130 / 41 and 131 / 41 regions. The sample information used is shown in the following table.

[0080]

[0081] Using the three standardized index differences (dNBRr, dNDMIr, and dNDVIr) of sample pixels, decision tree rule 1 was constructed to determine the specific time of forest disturbance in a pixel, and decision tree rule 2 was constructed to determine whether the disturbance was fire or non-fire. The method used sample pixels from two images to perform decision tree analysis, generating two sets of decision tree rules. The accuracy of these rules for extracting forest disturbance time and fire patches in two regions was then verified. Ultimately, the set with the higher accuracy was selected as the final result.

[0082] 1.3.3. Determine when and what type of interference occurs

[0083] Decision tree rule 1 is used to iteratively compare the dI value images generated in 1.3.1 within the disturbance year to determine between which two images the forest disturbance pixel's disturbance occurred. If the three dI index values ​​of the disturbance pixel in an image are classified as disturbance in decision tree rule 1, the disturbance is considered to have occurred between the date of that image and the date of the next image. In the present invention, the use of Landsat 8 image data can only ideally determine the specific time of forest disturbance to within two adjacent images, that is, within 16 days. However, with the use of Landsat 9 imagery, the present invention can further determine the forest disturbance time to within an 8-day interval when Landsat 8 and Landsat 9 images are present simultaneously. After determining the specific time of forest disturbance, decision tree rule 2 is further used to determine whether the forest disturbance is fire disturbance or non-fire disturbance based on the vegetation index dI values ​​of the two images before and after the disturbance time.

[0084] 1.4 Accuracy Verification

[0085] 1.4.1. Verification of the accuracy of forest disturbance occurrence time

[0086] After determining the specific time range of forest disturbance occurrence within the disturbance year, the present invention compared the disturbance occurrence time range obtained by decision tree rule 1 in 1.3.3 with the actual disturbance occurrence time to evaluate the temporal accuracy (TA) of the forest disturbance occurrence time extraction. The present invention randomly generated 300 random points within the actual disturbance patches in two study areas (strip number 130 / 41 and strip number 131 / 41). The number of random points whose actual disturbance occurrence time matched the disturbance occurrence time obtained by decision tree rule 1 was n, and the total number of random points was N. The temporal accuracy (TA) was calculated as follows:

[0087]

[0088] 1.4.2. Verification of Fire Interference Type Extraction Accuracy

[0089] The present invention compares the fire / non-fire forest disturbance type obtained by decision tree rule 2 in 1.3.3 and the real fire / non-fire disturbance type based on random points. The overall accuracy P is calculated by constructing the confusion matrix A The spatial accuracy of forest disturbance type extraction was evaluated by Kappa coefficient. The overall accuracy P was calculated. A The calculation formula of Kappa coefficient is as follows:

[0090]

[0091]

[0092]

[0093] Among them, n is the number of random points whose judgment results of the decision tree rule are consistent with the actual interference type; m is the number of random points whose judgment results are inconsistent with the actual interference type, n fm is the number of random points judged as fire disturbance using the decision tree rule; n fr is the number of random points whose real disturbance type is fire disturbance; n nm is the number of random points judged as non-fire disturbance using the decision tree rule; n nr is the number of random points whose true disturbance type is non-fire disturbance.

[0094] 2. Obtain experimental data on the performance of the product or result:

[0095] The present invention uses Xichang City (strip number 130 / 41) and Muli County (strip number 131 / 41) in Liangshan Prefecture, Sichuan Province, China as the study area, and uses the method of the present invention to extract the time of forest fire occurrence and the spatial information of fire patches. The data used include Landsat 8 surface reflectance image data, 30m resolution digital elevation (DEM) image data, and 30m resolution land use type data (GlobeLand 30) obtained from the GEE platform. And real forest disturbance patch information data obtained from local institutions. The track numbers of Landsat8 surface reflectance image data are 130 / 41 and 131 / 41, including 1-year interval images for running the VCT algorithm and all images available in the disturbance year for extracting forest fire information. Landsat 8 data information is shown in the Landsat 8 OLI image data information table used in the VCT process obtained by the GEE platform and the table of the number of all available Landsat OLI images in the disturbance year in the GEE platform. Landsat 8 surface reflectance data acquired through the GEE platform undergoes radiometric calibration and atmospheric correction, and cloud pixels are removed using the cloud mask algorithm (CFMASK). For track 131 / 41, no high-quality vegetation growth-season imagery was available in 2017 for use in the VCT algorithm. Therefore, only changes between 2016 and 2017 were analyzed. The actual forest disturbance patch information used includes the time of occurrence and whether the disturbance is fire or non-fire, as shown in the table above.

[0096] Based on the results of the VCT algorithm, the present invention standardized the normalized burn rate (NBR), normalized moisture difference index (NDMI), and normalized vegetation index (NDVI). Through decision tree analysis, a method was obtained that can quickly extract the specific time of forest fire occurrence and distinguish forest fire interference from non-fire interference, and the accuracy of the method was evaluated. The specific process of the method is as follows: Figure 1 .

[0097] 2.1. Extracting forest disturbance results year by year using the VCT algorithm

[0098] In this paper, the VCT algorithm was run on the GEE platform using Landsat 8 image data with a one-year interval, 30m resolution DEM data, and GlobeLand 30, a 30m resolution land use classification data. The VCT results obtained the persistent forest pixel masks and forest disturbance pixel masks for the two study areas from 2013 to 2021. The results are as follows: Figure 2The green portion represents persistent forest pixels in the two regions between 2013 and 2021. Except for gray (persistent non-forest pixels) and blue (persistent water pixels), the other colors correspond to disturbed pixels in each year. Given the low reliability of forest disturbance information in the initial years of the VCT algorithm results, the forest disturbance pixels in 2013 were not considered in this study.

[0099] 2.2 Vegetation Index Standardization

[0100] After the three vegetation indices, NBR, NDMI, and NDVI, were standardized using persistent forest pixels, the three vegetation indices were less affected by seasonal changes within a year and showed better stability ( Figure 3 ).from Figure 3 It can be seen that the NBR, NDMI, and NDVI index values ​​in Xichang and Muli areas fluctuated greatly within a year before standardization, while the standardized index values ​​tended to fluctuate within the range of -100 to 100, with the average value close to 0.

[0101] 2.3. Forest disturbance occurrence time extraction and accuracy verification

[0102] Figure 4 The figure shows the changes in the dNBRr, dNDMIr, and dNDVIr values ​​of pixels that experienced interference and pixels that did not experience interference during the interference year. As can be seen from the figure, the dNBRr, dNDMIr, and dNDVIr values ​​of pixels that did not experience interference remained at a low level throughout the interference year, fluctuating only within a range close to 0. If, on the other hand, a pixel experienced interference at a certain time during the interference year, its index value would rise sharply at that time and then return to its original state. Based on this characteristic, the present invention believes that the changes in dNBRr, dNDMIr, and dNDVIr can effectively observe the occurrence of forest interference. The present invention uses the C4.5 decision tree algorithm to analyze the thresholds for extracting forest interference using several indices. The results show that dNBRr is the most sensitive for identifying the occurrence of forest interference. Therefore, dNBRr>190 is used as the basis for judging the occurrence of forest interference in the present invention. In the figure, 190 is the threshold used to distinguish whether forest interference has occurred. If the dNBRr value of a certain image is greater than 190, it is judged that the forest interference occurred between the date of the image and the date of the next image.

[0103] Figure 5 and Figure 6The extracted timeframes for forest disturbance events in Xichang and Muli are presented, using a dNBRr > 190 as the criterion for forest disturbance. However, due to the limited number of Landsat images available on the GEE platform and the presence of cloud pixels, the timeframes for many forest disturbance events cannot be determined to be within 16 days, potentially extending to 32 or even 48 days. Comparing the extracted timeframes with the actual disturbance occurrences and calculating the temporal accuracy (TA) using 300 random points in each region, the TA for the extracted disturbance timeframes reached 94.33% and 90.56%, respectively. The following table provides statistical validation data for the accuracy of forest disturbance time extraction in Xichang and Muli.

[0104]

[0105] 2.4. Forest fire / non-fire disturbance type extraction results and accuracy verification

[0106] Figure 7 The figure shows the decision tree rule 2 obtained by training with the training sample data in Xichang area and the effect of the rule on distinguishing some forest interference pixels. The rule forms a decision boundary in three-dimensional space to separate forest fire interference and non-fire interference. The decision tree rule believes that when an interference pixel is considered to interfere between a certain image and the next image, if the dNBRr of the interference pixel in the image is ≥580 or dNDMIr ≥400 or dNDVIr ≥350, then the interference is considered to be caused by fire factors, otherwise it is considered to be caused by non-fire factors. Figure 7 In the left image, we can see that among the interference pixels in the Xichang region, three real fire pixels and three real non-fire pixels are misclassified. In the right image, we can see that among the interference pixels in the Muli region, three real fire pixels and one real non-fire pixel are misclassified. This indicates that although misclassification still occurs, the decision tree rule 2 obtained by this invention generally has a good effect on distinguishing fire interference pixels from non-fire interference pixels in both Xichang and Muli.

[0107] Figure 8 and Figure 9 The following table shows the fire-disturbed and non-fire-disturbed patches extracted using decision tree rule 2 in Xichang and Muli. The present invention constructed confusion matrices for the extraction results in the two regions using random points and calculated the overall accuracy and Kappa coefficient. The results show that the overall accuracy of forest fire / non-fire disturbance differentiation in Xichang and Muli reached 85.33% and 89.67%, respectively, with Kappa coefficients of 0.71 and 0.74. The following two tables show the details of the confusion matrices for the two regions.

[0108]

[0109]

[0110] Based on the extracted fire / non-fire disturbance patches, the present invention calculated the disturbance area in each disturbance year in Xichang and Muli regions;

[0111]

[0112]

[0113] As shown in the table above, the data indicates that the area of ​​fire disturbance in Xichang between 2020 and 2021 was relatively small, less than 60 hectares, while in 2019-2020, it exceeded 600 hectares. In Muli, the area of ​​fire disturbance between 2019 and 2020 was also significantly larger than in other years, exceeding 3,800 hectares. The years with high fire disturbance areas in these two regions correspond to two relatively severe fires reported in actual forest disturbance data.

[0114] The method of the present invention is based on the results of the VCT algorithm and all available Landsat images obtained by the GEE platform. The decision tree rules in the method can quickly determine the time when forest disturbance occurs between two adjacent images, ideally with an interval of 16 days, and on this basis, determine whether the forest disturbance is a fire disturbance or a non-fire disturbance based on the vegetation index values ​​in the images before and after the disturbance occurs. The VCT algorithm results, LandTrendr algorithm results, and BFAST algorithm results commonly used in existing studies can only determine the time when forest disturbance occurs within one disturbance year, and cannot further determine the type of disturbance that occurs. Further research based on these algorithms to distinguish the types of forest disturbance often requires a large number of forest disturbance type samples in the study area to train the classification model. Compared with existing research, the method of the present invention does not need to select a large number of training samples according to different regions, can quickly determine the specific time when forest disturbance occurs and whether the forest disturbance is a fire disturbance or a non-fire disturbance, and shows good results in images in different regions, with better versatility.

[0115] In the process of extracting the specific time of forest disturbance, the present invention uses all available Landsat images provided in the GEE platform within the disturbance year. This increases the number of images used compared to some methods that only use 1-year or 2-year interval images, providing a more sufficient data basis for shortening the time range of forest disturbance occurrence. Compared to the forest disturbance time determination method using low spatial resolution image data such as MODIS and AVHRR, the present invention uses Landsat images with a spatial resolution of 30m, which is not easily affected by mixed pixels, can better extract the spatial information of forest disturbance patches, and also has a better extraction effect for some smaller-scale fire events. The method in the present invention determines the forest disturbance occurrence time to the date between two adjacent images as much as possible (ideally 16 days), and with the commissioning of Landsat 9 data, in future research, due to the simultaneous existence of available Landsat 8 and Landsat 9 data within the disturbance year, the forest disturbance occurrence time can be further determined to between 8 days. The shortening of the time range for the occurrence of disturbances means that the two images closest to each other before and after the occurrence of forest disturbances can be quickly determined based on this range, and subsequent research can be conducted based on this range, such as further distinguishing the types of forest disturbances, evaluating the losses caused by forest disturbances, estimating the carbon emissions caused by forest fire disturbances, etc. In these studies, if the time interval of the remote sensing images used is long, the forest may have recovered after the forest disturbance occurs. Some studies have shown that different types of forests recover at different speeds after disturbances. Broadleaf forests and mixed coniferous and broadleaf forests can recover to 50% of their pre-fire levels within 250 days after the fire, while coniferous forests may take more than 400 days to recover to 50% of their pre-fire levels. Therefore, the time range for the occurrence of forest disturbances will also have a great impact on the calculation and evaluation of these indicators. The present invention will shorten the time range for the occurrence of forest disturbances as much as possible, which is more conducive to subsequent research. Moreover, compared with the time range of one-year intervals, the present invention shortens the occurrence time of forest disturbance to a range of 16 days or even 8 days, which can better express the seasonal characteristics of forest disturbance in a region. This is more beneficial for the formulation and implementation of various policies after the occurrence of forest disturbance. For example, forest fires in some areas may have seasonal characteristics. For example, if the peak season for fires coincides with the rainy season in a certain area, when formulating post-fire vegetation restoration strategies, it may be possible to reduce the difficulty of post-disaster reconstruction to a certain extent. In addition, forest disturbances in different study areas are often caused by different factors, such as climate, topography, vegetation type, economic development status, etc. More accurate forest disturbance occurrence time can better reveal the relationship between disturbance occurrence and these factors, and can also help to better formulate relevant policies.

[0116] In the present invention, the three vegetation indices NBR, NDMI, and NDVI are standardized (NBRr, NDMIr, NDVIr) using the continuous forest mask in the VCT algorithm results, and the changes in the standardized indices in the remote sensing image time series during the disturbance year are analyzed to obtain a decision tree rule to distinguish forest fire disturbances and non-fire disturbances. The present invention can quickly distinguish between fire disturbances and non-fire disturbances based on the two remote sensing images closest to each other before and after the forest disturbance occurs, and has high accuracy in remote sensing images of different regions. Most of the existing research on distinguishing different types of forest disturbances is based on multispectral impact data and uses machine learning algorithms to classify forest disturbance types. However, due to the differences between images in different regions and at different times, the methods in these studies often require a large number of training samples of different forest disturbance types from different image areas to train the classifier. The decision tree rule used in the present invention is constructed based on the standardized vegetation index. During the standardization process, the differences between images in different regions and at different times are reduced, thereby obtaining a more general decision tree rule for distinguishing between fire disturbances and non-fire disturbances. At the same time, based on the present invention, if we can further combine the spectral characteristics (such as surface temperature, brightness, greenness), spatial characteristics (morphological characteristics) and texture characteristics of forest disturbance patches, and combine other types of data (such as high spatial resolution satellite image data, LiDAR data, etc.), then we may be able to further distinguish different types of disturbance in non-fire disturbance (such as logging, pests and diseases, etc.), laying the foundation for further research.

[0117] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for extracting forest fire patches and occurrence time based on the GEE platform and VCT algorithm, characterized in that: The following steps are involved: S1. Run the VCT algorithm in the GEE platform using the VCT code provided by the GEE platform. The VCT algorithm is used to obtain the continuous forest mask and the forest disturbance patch mask of each year in the study area for subsequent analysis. S2. Forest fire disturbance analysis was performed using three vegetation spectral indices: Normalized Burning Ratio (NBR), Normalized Moisture Difference Index (NDMI), and Normalized Vegetation Index (NDVI); S3, normalize the three vegetation indices using the persistent forest mask obtained in S1. The three normalized indices are expressed as NBRr, NDMIr, and NDVIr; S4. Based on the index values ​​of all available images in the interference year, the three index differences between two adjacent images are further calculated, which are expressed as dNBRr, dNDMIr, and dNDVIr respectively; S5. Construct a decision tree based on the changes of the standardized vegetation index in the remote sensing image time series during the disturbance year; S6. Using a decision tree rule to compare the standardized index difference value images, determine between which two images the forest disturbance pixel interference occurs, thereby determining the specific time when the forest disturbance occurs, and further use the decision tree to determine the type of disturbance based on the vegetation index difference value of the two images before and after the disturbance time; S7. After determining the specific time range of forest disturbance occurrence within the disturbance year, the disturbance occurrence time range obtained by the decision tree rule was compared with the actual disturbance occurrence time to evaluate the time accuracy of the forest disturbance occurrence time extraction; S8. Compare the fire or non-fire forest disturbance types obtained by the decision tree rules based on random points and the actual fire or non-fire disturbance types, and calculate the overall accuracy and Kappa coefficient by constructing a confusion matrix to evaluate the spatial accuracy of forest disturbance type extraction.

2. The method for extracting forest fire patches and occurrence times based on the GEE platform and the VCT algorithm according to claim 1, wherein: The NBR is used for forest fire detection and fire severity analysis, the NDMI reflects the change in surface moisture before and after forest disturbance occurs, and the NDVI indicates the change in surface vegetation cover.

3. The method for extracting forest fire patches and occurrence times based on the GEE platform and the VCT algorithm according to claim 1 is characterized in that: Without affecting the results, the values ​​of the three vegetation indices are expanded to between 0 and 2000, and the calculation formulas for the three vegetation indices using Landsat OLI images are as follows: For Landsat OLI image data, ρ5 is the reflectance of the near-infrared band, with a central wavelength of 0.86 μm; ρ7 is the reflectance of the shortwave infrared 2 band, with a central wavelength of 2.20 μm; ρ6 is the reflectance of the shortwave infrared 1 band, with a central wavelength of 1.61 μm; and ρ4 is the reflectance of the red light band, with a central wavelength of 0.66 μm.

4. The method for extracting forest fire patches and occurrence times based on the GEE platform and the VCT algorithm according to claim 1, characterized in that: The formula used for index normalization is as follows: I r =I i -I f Among them I r is the standardized pixel index value; I f is the mean index value of all persistent forest pixels in an image; I i is the pixel index value before normalization.

5. The method for extracting forest fire patches and occurrence times based on the GEE platform and the VCT algorithm according to claim 1 is characterized in that: The formula for calculating the index difference value is as follows: dI=The atter -THE former Where dI is the index difference value calculated from two adjacent images; I former is the pixel index value in the earlier image of the two adjacent images; I latter It is the pixel index value in the later of the two images. If a pixel in an image is masked by a cloud pixel, the index difference value of the pixel is calculated from the index value of the nearest cloud-free pixel.

6. The method for extracting forest fire patches and occurrence times based on the GEE platform and the VCT algorithm according to claim 1, characterized in that: The decision tree rules include two sets of decision tree rules obtained by performing decision tree analysis on sample pixels from the two images, and the accuracy of the decision tree rules in extracting forest disturbance time and fire patches in the two areas was verified respectively. The decision tree rules include rule one and rule two. Rule one is used to determine the specific time when forest disturbance occurs in the pixel, and rule two is used to determine whether the disturbance is fire disturbance or non-fire disturbance.

7. The method for extracting forest fire patches and occurrence times based on the GEE platform and the VCT algorithm according to claim 6 is characterized in that: The calculation formula of time accuracy TA is as follows: Where n is the number of random points whose actual interference occurrence time matches the interference occurrence time obtained by decision tree rule 1, and N is the number of all random points.

8. The method for extracting forest fire patches and occurrence times based on the GEE platform and the VCT algorithm according to claim 6 is characterized in that: Overall accuracy P A The calculation formula of Kappa coefficient is as follows: Among them, n is the number of random points whose judgment results of the decision tree rule are consistent with the actual interference type; m is the number of random points whose judgment results are inconsistent with the actual interference type, n fm is the number of random points judged as fire disturbance using the decision tree rule; n fr is the number of random points whose real disturbance type is fire disturbance; n nm is the number of random points judged as non-fire disturbance using the decision tree rule; n nr is the number of random points whose true disturbance type is non-fire disturbance.

Images