首页 > , Vol. , Issue () : 1993-2002
Abstract: Objective : Forest fires not only damage the stability of forest ecology seriously, but also threaten the safety of human life and property. It is difficult for traditional methods to detect fires in time due to its limited observation region and time window. Remote sensing can accurately detect fires in a larger area continuously. Besides, it improves the detection efficiency. At present, meteorological satellites and polar satellites are mainly used to detect forest fires. However, it is difficult for these satellites to detect small fire scars and deal with the influence of cloud and rain. In this study, multi-sensor remote sensing data were combined to monitor the forest fire spots and progress in Muli Tibetan Autonomous County of the Sichuan Province since March 30, 2019. Method : Firstly, we located the fire spots and time using the GF-4, a geostationary satellite of China. It has a multispectral camera (PMS), which has a high spatial (50 m) resolution. This satellite also carries a mid-infrared sensor (IRS,400m) which is sensitive to abnormal high temperatures. Secondly, we found the spectral difference between burning and unaffected forest stands using Sentinel-2. Thirdly, based on Sentinel-2, we classified the fire scars using OTSU algorithm to set the threshold of dNBR. In the end, we used SAR data from Sentinel-1A to relate NDVI to polarization ratio (PR). Result : Results show that: (1) The location of the fire spot can be determined accurately using IRS and PMS data from GF-4. (2) Combing the fire spot location and an NDVI sharp reduction (from 0.7 to 0.25), we confirmed the burning time to be March 30 using the PMS data from GF-4. (3) There is a significant difference in spectral curves among the burning forest stands, the unaffected areas and different type of burned areas in 490 nm and 2200 nm from Sentinel-2 data. (4) The total area of damaged fire scars was classified (41.56hm2), with an accuracy of 94.67% using the dNBR of 0.35 as the threshold from Sentinel-2 derived dNBR map. The lightly damaged fire scars were also classified (66.56 hm2, 90.94%). (5) The polarization ratio from Sentinel-1A data increased from 6.6dB to 10.8dB after the burning. NDVI linearly relates to the polarization ratio (R2=0.58 for fitting and 0.50 for verification). Conclusion : The above results agree well with the fire spot location and area from local field reports, and the burning time error is less than 12 hours. This research provides an efficient method which can ignore the influence of cloud and rain and other complex environment to monitor forest fires. If GF-4 could focus on monitoring this area, more details would be found to capture the progress of the fire. This study provides a reference for small fire identification after the fire occurrence. The research methods and results can provide technical supports for forest fire emergency.