The FY-3D Global Active Fire product: Principle, 1 Methodology and Validation

. Wild fires have a strong negative effect on environment, ecology and public health. However, 17 the potential degradation of mainstream global fire products leads to large uncertainty on the effective 18 monitoring of wild fires and its influence. To fill this gap, we produced FY-3D global fire products with 19 a similar spatial and temporal resolution, aiming to serve as the alternative and continuity and 20 replacement for MODIS global fire products. Firstly, the sensor parameters and major algorithms for 21 noise detection and fire identification in FY-3D products were introduced. For visual-check-based 22 accuracy assessment, five typical regions, Africa, South America, Indo-China Peninsula ， Siberia and 23 Australia, across the globe were selected and the overall accuracy exceeded 94%.For accuracy 24 assessment, five typical regions, Africa, South America, Indo-China Peninsula ， Siberia and Australia, 25 across the globe were selected. The overall consistence between FY-3D fire products and reference data 26 exceeded 94%, with a more than 90% consistence

is comprehensively compared with the other mainstream fire products, especially MODIS global fire 124 products at the global and regional scale. Thanks to its good global consistence and regional suitability, 125 The new FY-3D global fire products has the potential toaim to serve as a continuity of the globalexisting, 126 yet degrading MODIS fire products and better support regional (especially Asia) and global ecological 127 and environment research in China.  the fire-spot significantly higher than surrounding pixels. For rapid monitoring of global wildfires, it is 163 necessary to develop an algorithm for the automatic identification of fire spots. 164 MERSI-Ⅱ fire monitoring products from FY-3D satellite can provide fire spot location, sub-pixel fire 166 spot area, temperature, and fire spot intensity, in inland areas around the world and generate global fire-167 spot pixel information (including day and night) in an HDF format. FY-3D fire products are produced 168 following a projection with the equal latitude and longitude (0.01°). Fire spot intensity is classified 169 according to sub-pixel fire spot area and temperature, with an overall accuracy above 85%. Based on 170 daily monitoring products, SMART (Satellite Monitoring Analyzing and Remote sensing Tools) system 171 can generate the images of global monthly fire spot distribution, with a resolution of 0.25°. 172

173
The algorithm for fire spot identification depends on the sensitivity of mid-infrared channels to high-174 temperature heat sources. The radiance and brightness temperature of the pixels in the mid-infrared 175 channels with sub-pixel fire spots are higher than those of the surrounding non-fire pixels and those of 176 the pixels in the far-infrared channels. Therefore, the pixels with fire spots can be identified by setting 177 an appropriate threshold, and the estimation of background temperature is the key to high detection 178 accuracy and sensitivity. 179 Sub-pixel fire spot estimation relies on the brightness temperature in mid-infrared channels, and the far-181 infrared channels are employed when the mid-infrared channels have saturated brightness temperature. 182 In the single-channel estimation formula, the temperature of the open flame spot is set to 750 K. covers all global fire spot pixels in this month. Concerning the multi-time monitoring information of the 206 same pixel, the maximum fire spot area is taken as the current-month fire spot information for the pixel. 207 This section mainly introduces the specific algorithm and steps for generating FY-3D global fire products 219 based on the original data obtained from MERSI-Ⅱ. The input data include MERSI-Ⅱ global orbital 220 Earth observations, MERSI-Ⅱ global orbital geographical locations, MERSI-Ⅱ global orbital cloud 221 detection data, and global land and sea template data, as shown in Table 2. 222 Automatic identification of fire spots is the major step for generating fire products. Firstly, the 5-minute 225 L1 data segments of MERSI-Ⅱ and various auxiliary data are read in, and the noise lines are identified 226 to generate the noise line mark. Next, the 5-minute data segments are projected according to rule of the 227 equal latitude and longitude, and cut as 5° × 5° grids to generate a local map. 228

The general principle of fire detection based on MERSI-Ⅱ 239
Channel 20 of FY-3D MERSI-Ⅱ is mid-infrared, with a wavelength of 3.55-3.95 m, while Channels 240 24 and 25 are far-infrared, with a wavelength of 10.3-11.3 m and 11.5-12.5 m, respectively. According 241 to Wien's displacement law, 243 where is the peaks at the wavelength, T is the absolute temperature, b is a constant of proportionality 244 called Wien's displacement constant, equal to about 2898 μm ⋅ K. Bblackbody temperature T is inversely 245 proportional to peak radiation wavelength λmax, as the higher temperature can lead to the smaller peak 246 radiation wavelength. The peak radiation wavelength of the surface at normal temperature (about 300 K) 247 is close to that of Channels 24 and 25; the combustion temperature of forest fires is generally 500 K-248 1200 K, and the peak wavelength of thermal radiation is close to that of Channel 20. When a fire spot 249 appears in the observed pixel, the radiance increment in Channel 20 caused by the high temperature in 250 the small sub-region of the pixel, where the fire spot is located (Since the pixel resolution of the scanning 251 radiometer is 1.1 km, it is usually not be all open flame areas at the same time in such a large range), is 252 much higher than surrounding pixels without an open flame and also greater than that in Channels 24 253 and 25. In this case, the weighted average of radiance increase and brightness temperature increase of 254 each channel differ notably in this pixel, based on which the fire information can be extracted and 255 analyzed. 256

257
As indicated by Fig. 4(a), when the fire spot temperature grows, the brightness temperature of CH20 258 pixels increases rapidly. Even if the fire spot only accounts for 0.1% the pixel area, the brightness 259 temperature increment can reach 10 K (44K) when the fire spot is 500 K (900 K). Although the brightness 260 temperature increase of CH24 also rises with the higher fire spot temperature, it is far lower than that of 261 CH20. Fig. 4(b) illustrates that as the fire spot area gets larger, the brightness temperature of CH20-262 mixed-pixels grows rapidly. It reaches 12K when the fire spot is 900 K, even if the fire spot only accounts 263 for 0.01% of the pixel area. Similarly, the brightness temperature increment of CH24 grows at a much 264 lower rate than CH20.

Detection of cloud pixels 273
Effective cloud detection is required for generating reliable fire products for the following reasons. Firstly, 274 the existence of cloud in the atmospheric layers may block the emitted information of fire spots, leading 13 to missed identification. Secondly, specular reflection of cloud can lead to wrong identification of fire 276 spots. Therefore, cloud identification was conducted before fire identification. Similar to MODIS, FY-277 3D also included radiation information from multiple bands and the principle of cloud identification for 278 FY-3D fire products was similar to that of MODIS. Based on the reflectance difference between cloud 279 and land pixels, we classified cloud pixels following the rules listed in Table 3.  Note: These eight rules are set to exclude a diversity of cloud bias. And a pixel that meets any rule 285 any rule in Table  286 287 288

Calculation of background temperature 289
According to the principle of fire spot identification, when a fire spot appears in a pixel (i.e., open flame), 290 the brightness temperature of the pixel in Channel 20 is significantly higher than the background 291 brightness temperature (the brightness temperature of surrounding non-fire pixels); the brightness 292 temperatures of Channels 24 and 25 are also higher than the background, but the temperature difference 293 is much smaller than Channel 20. In this case, the difference of brightness temperature between fire-spot 294 pixels and background in both the mid-infrared channel and far-infrared channels can be employed as 295 important factors for automatic identification of fire spots. Therefore, the background temperature of the 296 detected pixel is required for identifying fire spots. Since the background temperature cannot be obtained 297 from the fire-spot pixels, it should be calculated according to the average of their surrounding pixels. 298 However, the reflection of solar radiation during the daytime also causes a higher brightness temperature 299 in the mid-infrared channel, which mainly occurs in the zone bare of vegetation, cloud surface, and water 300 bodies (specular reflection). In particular, the difference of brightness temperature between mid-infrared 301 and far-infrared channels caused by specular reflection of solar radiation can reach tens of K on the cloud 302 surface and water bodies. Since the reflection of solar radiation on the bare surface is relatively weak in 14 due to the high sensitivity requirement for fire identification. When the background brightness 305 temperature is calculated, pixels that already contain fire spots should also be excluded. Therefore, 306 suspected high-temperature pixels, which may already contain fire spot pixels, cloudy pixels, water 307 pixels and those pixels affected by solar flare should be removed for background temperature calculation. 308 After above-mentioned disturbing pixels were removed, the average and standard deviation of 318 background temperature in the mid-infrared channel, and the background average and standard deviation 319 of brightness temperature difference between the mid-infrared and far-infrared channels were calculated 320 with peripheral pixels as background pixels. 321

322
The calculation of background temperature was acquired in the following steps. For each 3×3 window, 323 the background temperature is calculated as the mean temperature of all background pixels. Suspicious 324 high-temperature pixels can be identified according to the following conditions: 325 Where TMir is the bright temperature in the middle-infrared channel. Tth is the threshold for high-327 temperature pixels in the middle-infrared channel, usually set as sum of the mean bright temperature of 328 all pixels in the window and 2 × its corresponding standard deviation. T'Mir_bg is the mean bright 329 temperature of background pixels. 330 331 △TMir_bg is the allowed difference between the mean background bright temperature and the suspicious 332 high-temperature pixel, usually set as 2.5 × standard deviation of background pixels. If there were less 333 than 20% of pixels were cloudless pixels, then the 3 ×3 window was extended to 5 ×5, 7×7, 9 × 9…51 × 334 51. If still not applicable, then this pixel was marked as a non-fire pixel. 335

Identification of fire pixels 336
With obtained background temperature, the difference between brightness temperature and background 337 temperature in the mid-infrared channel, as well as the difference of brightness temperature and 338 background temperature between mid-infrared and far-infrared channels, at the candidate pixels could 339 be calculated, based on which we could decide whether the threshold of fire spot identification was 340 reached. If the threshold was reached, the pixel will be preliminarily marked as a fire pixel. Next, for 341 daytime observation data, it is necessary to further check whether the increase of brightness temperature check, fire pixels could be effectively extracted. 344 345 When the following two conditions are met, a pixel can be identified as fire pixel: 346 (1) T3.9 > T3.9bg + n1 ×δT3.9bg 347 Where T39 is the bright temperature of the pixel at 3.9 um. T3.9bg is the background bright temperature. 349 δT3.9bg is the standard deviation of bright temperature of background pixels. △T3.9_11 is the difference of 350 bright temperature between 3.9 um and 11 um. △T3.9bg_11bg is the difference of background bright 351 temperature between 3.9 um and 11 um. The setting of this condition aimed to identify the difference of 352 land cover types in the window. When the land cover types in the window were generally consistent, 353 δT3.9bg_11bg is relatively small. For the identification of fire pixels, when δT3.9bg_11bg was smaller than 2k, 354 this value was replaced using 2K. When δT3.9bg_11bg was larger than 4k, this value was replaced using 4K. 355 n1 and n2 are background coefficients, which varies across regions, observation time and observation 356 angles. For instance, for Northern grasslands, n1 and n2 was set as 3 and 3.5, respectively. 357

Identification of noise line 358
Satellite data received by the ground system contain noise. For instance, some scanning lines may contain 359 many noisy pixels that affect fire spot identification. In this case, noise lines, referred to multiple 360 consecutive noisy pixels in one scanning line, should be checked firstly. Since the identification of fire 361 was carried out on the areal map projected with an equal latitude and on the same circle of longitude, the 362 identified latitude and longitude of fire spots failed to reflect the original positions of scanning lines. 363 Therefore, the noise line was identified on the 5-minute data segments before projection. Firstly, the 5-364 minute data segments were employed to identify fire spots, and the line number of identified fire spot 365 pixels was recorded. Following this, the number of fire spot pixels in each line was counted. When the 366 number of fire spot pixels in a line exceeded the empirical threshold, it was identified as a noise line, and 367 all pixels in the line are marked as noisy ones. In the following process, all pixels in this line were no 368 longer considered for fire-spot identification.

Estimation of sub-pixel fire spot area and temperature 373
MERSI-Ⅱ data is 12 bits, with a quantization level of 0-4095 and high radiation resolution. The spatial 374 resolution is 1.1 km, and the radiance of a pixel observed by the satellite is the weighted average of the 375 radiance of all the ground objects within the pixel range, as 376 where Nt is the radiance of the pixel observed by the satellite; t is the brightness temperature 378 corresponding to Nt; △Si is the area of the i th sub-pixel; NTi is the radiance of the sub-pixel; Ti is the Due to different FRP and temperature, underlying surfaces containing fire spots can be divided into fire 382 zones and non-fire zones (background). When fire spots appear, the radiance of pixels containing fire 383 spots (i.e. mixed pixels) can be expressed by the following formula: When a single channel was adopted to estimate the sub-pixel fire spot area, the fire spot temperature was 400 set to an appropriate value, which was 750 K in this product.   Table 4 shows accuracy of GFR fire spots in the five typical regions. The accuracy of automatically 464 identified fire spot in all regions was generally consistent and all exceeded 90%. Since these selected 465 regions represented distinct vegetation types and located in different hemispheres, the verification of FY-466 3D fire products based on 0.24 SMART proved its stability and reliable high-accuracy at the global scale.
It is worth mentioning that the visual-check based accuracy assessment mainly considered the 469 commission error, while omission error cannot be effectively revealed for the following reason. The 470 omitted fires were mainly caused by the requirement of minimum burning area. Since the spatial 471 resolution of FY-3D and MODIS active fire products is 1km, small fires (less than 100m 2 ) could not be 472

Cross-verification with between FY-3D and other global fire productsMODIS global fire 480 products 481
The cross-verification between FY-3D fire products and the mainstream MODIS fire products, 482 MYD14A1 V6 (https://firms.modaps.eosdis.nasa.gov/map/) with a daily temporal resolution and 1km 483 spatial resolution was conducted using the entire 2019 datasets. The data sets with observation time less 484 than 1 h were selected; the underlying surfaces were visually checked to remove areas covered by non-485 vegetation such as water, ice and snow, and bare land. According to the criterion that the distance 486 matching between the two fire spot pixels was less than 0.03°, cross-verification was conducted with 487 different months, underlying surfaces, regions, and fire intensities. In 2019, there were 2,237,714 fire 488 spot pixels in MODIS fire products, 1,866,920 of which were matched with FY-3D fire products, with 489 an overall consistence of 84.4% (as shown in Fig. 6). As shown in Figure 6, global fire spots were mainly 490 distributed in America, south-central Africa, East, and Southeast Asia, Australia, and parts of Europe, 491 and there were notable spatiotemporal variations of identified fire spots. Specifically, given the overall 492 data volume and spatial distribution, the total number of fire spot pixels from MODIS fire products was 493 larger than FY-3D products. For individual regions, the more fire spots, the higher consistence between 494 FY and MODIS fire products. Africa is the region with the most fire spots across the globe. From May 495 to October, a majority of fire spots was located in southern Africa whilst a majority of fire spots from 496 MODIS and FY-3D products was higher than other regions. The distribution of fire spots in South 498 America also presented seasonal characteristics. From July to October, fire spots mainly concentrated in 499 middle parts of South America. For other seasons, fire spots in South America mainly concentrated in 500 the North and other parts. The consistence between MODIS and FY-3D fire products also demonstrated 501 seasonal differences, with a high consistence from August to November and a relatively low consistence 502 in other seasons. For Eurasia, there were notable seasonal variations of spatial patterns of fire spots. 503 During March to August, there were relatively many fire spots and the consistence between MODIS and 504 FY-3D fire products was relatively high in this region. 505 In addition to the overall consistence between MODIS and FY-3D fire products, we also conducted cross-510 verification of between the two global fire products in terms of different months, underlying surfaces, 511 regions and fire intensities as follows. 512 Fig. 7(a) illustrates the monthly precision test of consistence between FY-3D and MODIS fire products 514 in 2019. The consistenceprecision in the remaining months is over 80% except that in April, October, 515 and November. The highest appears in July, exceeding 90%, while the lowest is in April, 71%. Detailed 516 parameters can be found in Table 5. From the global perspective, the number of fire spots was larger in 517

Cross-verification of between MODIS and FY-3D in terms of different months 513
July, August and September and the mean consistence between MODIS and FY-3D fire products was 518 larger than 85%. For July when the fire products were the most, the consistence achieved 90%. From 519 January to May, the number of fire spots was relatively small, and the mean consistence was around 80%. 520 The consistence for April was 71%, lowest among all months. The notable monthly variations of the 521 consistence between MODIS and FY-3D fire products was mainly attributed to the uneven spatial 522 distribution of fire spots across the globe. As shown in Fig 6, in June and July, a large number of fire 523 spots mainly concentrated in Africa, South America and Eurasia, leading to a high consistence of fire 524 identification. In April, there were limited and sparsely distributed fire spots in Africa and South America, 525 leading to a low consistence. According to the statistics, the number of fire spots was positively correlated 526 with the consistence between different fire products. Meanwhile, in seasons when fire could last longer, 527 the consistence was relatively higher. 528

Cross-verification between of MODIS and FY-3D onin terms of different underlying surfaces 530
Statistical analysis of consistenceprecision is carried out with different types of underlying surfaces. The 531 data of underlying surfaces is the global land use are detailed in Table 6. 532

533
The 15 types of underlying surfaces were selected for verification. Table 6 and Fig. 7(c)  Generally, these surfaces were all covered by sparse or unstable vegetation, the fire on which can last for 551 a relatively short period. Meanwhile, the observation time lag between FY-3D and MODIS was larger 552 than 30 minutes. Therefore, the consistence of FY-3D and MODIS fire products on these surface types 553 was lower than other surface types. 554

Cross-verification of between MODIS and FY-3D in terms of different regions 559
The global monitoring area is divided into Africa, America, Asia, Europe, and Oceania. The verification 560 demonstrates the results with the highest consistenceprecision (over 80%) are found in Africa and Asia, 561 and those in America, Europe, and Oceania show the consistenceprecision over 70%. The FY-562 3D/MERSI-Ⅱ fire identification algorithm draws lessons from the MODIS algorithm and has been 563 improved on that basis, and targeted development has been made for the underlying surface and climatic 564 conditions in China, so it is necessary to test the matching results in China separately. It shows that 565 China's regional consistency of results in China is lower than other continents, only 65%. Compared with 566 other continents, the low consistence between FY-3D and MODIS fire products in China may be 567 attributed to the following reason. Thanks to the field-collected data, the algorithm for fire detection 568 using FY-3D specifically included the underlying surfaces and surrounding geographical conditions in 569 China. Therefore, FY-3D has the potential to provide more reliable fire products for China. 570

571
According to the feedback on practical application in China, especially during the period from July to 572 September, when there were much precipitation, cloud cover, there should be limited fire spots identified. 573 However, based on MODIS fire products, there were many fire spots during this period, which were 574 much more than FY-3D detected fire spots. The consistence between MODIS and FY-3D fire products 575 in China was only 65%. To further examined the suitability of FY-3D fire products in China, the accuracy 576 assessment of FY-3D and MODIS fire products was conducted based on ground truth data and explained 577 in the following sections. Specifically, the fire spot precision of FY-3D/MERI-Ⅱ was higher than 85%, 578 which indicated that the precision of the MODIS algorithm is inferior to FY-3D/MERI-Ⅱ in China with 579 the decline in instrument performance (see Fig. 7(b) for details).

Cross-verification of MODIS and FY-3D in terms of fire intensities 588
The confidence of fire spots and the fire intensity represented by FRP are analyzed respectively, and the 589 data comes from the MODIS fire spot list. Fig. 8(a) and Fig. 8(b) are statistical diagrams of confidence 590 and FRP, respectively. From Fig. 8(a), the confidence of the matched pixels of the two satellites is above 591 66%, while that of the mismatched ones is less than 60% and even lower than 50% in some months. In 592 other words, the higher confidence indicates the higher matching degree. As indicated by Fig. 8(b), the 593 FRP of the matched pixels of two satellites is mostly above 40 MW, while that of the unmatched pixels 594 leads to the greater probability of simultaneous observation by the two satellites and the higher matching 596 degree between their results. 597 598 Two major findings were identified based on the comparison between FY-3D and MODIS fire products 599 in terms of fire intensity: Firstly, the higher the credential of the identified fire, the higher consistence 600 between FY-3D and MODIS fire products. When the credential was larger than 65%, both FY-3D and 601 MODIS could effectively identify the candidate pixel as fire pixel. In other words, the parameter of 602 credential in MODIS fire product provides important reference for fire detection. Secondly, FRP is an 603 index for the heat radiation of the fire. The larger FRP, the larger consistence between FY-3D and MODIS 604 was, indicating a higher accuracy of fire detection. Therefore, the difficulty for fire detection mainly lies 605 in the detection of weak fires. misidentified fires could be effectively recognized (As shown in Figure 9). Based on the field collected 624 reference of fires, we evaluated the suitability of FY-3D fire products in China (Table 8).  As shown in Figure 9 and Table 8, FY-3D products achieved a good accuracy of 79.43% in China. 631 Meanwhile, MODIS also achieved a good accuracy of 74.23%. As introduced above, the omission error 632 of FY-3D and MODIS fire products was mainly attributed to small fire area, which failed to meet the 633 minimum fire area recognizable by sensors. When simply considering the commission error, FY-3D fire 634 products achieved an accuracy of 88.50%, notably higher than that of MODIS (79.69%). This result 635 proved that with the consideration of local underlying surfaces, FY-3D fire products are more suitable 636 for fire monitoring in China. monitoring, which is mainly attributed to its short service periods. On one hand, due to its long time 676 series and general reliability, MODIS fire products remained a major choice for monitoring long-term 677 variations of fire spots across the world. However, the long-term running continuous degradation of 678 MODIS sensors led to the growinglarge uncertainties to the quality of recent and future MODIS fire 679 products. In this case, thanks to its similar spatio-temporal resolution and, high consistence and less-than-680 1h difference of visiting timeprecision, FY-3D fire products haves the potential to be widely employed 681 as the potential alternativereplacement and continuity of global MODIS fire products. Meanwhile, FY-682 3D fire products have a higher reliability in China and its surrounding regions than other fire products. 683 Therefore, FY-3D fire products are an ideal selection for fire monitoring in across China. crop-residue burning and wildfires (e.g. forest fires and grassland fires) emit airborne pollutants (e.g. 694 PM2.5, PM10, CO). In this regard, FY-3D fire products can be used as the emission sources for estimating 695 its environmental effects. 696

Future extension of FY-3D fire products 697
China has just launched FY-3E and FY-4B satellites in June and July, 2021. Amid the launch and 698 operation of a new generation of Fengyun meteorological satellites, the accuracy and timeliness of fire 699 monitoring by meteorological satellites have been largely enhanced. Thanks to the improved 700 meteorological data, which provides useful reference to understand the current status of combustibles 701 and potential fire risk, FY-3D satellite will be taken as a better data source to produce various secondary 702 products for fire monitoring and prediction. Based on traditional fire spot identification, further research 703 should concentrate on the assessment of fire area, estimation of biomass carbon emission, prediction of 704 smoke impact, and early warning of forest and grassland fire using the series of Fengyun meteorological 705 satellites. For instance, the water content of combustibles is closely related to temperature, light, and 32 was rarely considered in previous fire products. Based on the a series of products of from Fengyun 708 meteorological satellites, such as surface temperature, vegetation index, surface evapotranspiration, solar 709 radiance, and cloud cover, FY-3D fire products can be improved by establishing an estimation model for 710 the water content of combustibles. Meanwhile, with the fire products such as fire spot and smoke, and 711 the meteorological products such as wind field data from Fengyun series satellites, we can predict the 712 impact of smoke caused by forest and grassland fires on the atmospheric environment in the surrounding 713 and even remote areas. In the future implementations, Fengyun meteorological satellites will play a 714 greater role in monitoring, early warning, and forecast ofing global fires and their ecological impacts. With a similar spatial and temporal resolution, we produced FY-3D global fire products, aiming to serve 726 as the potential alternative and continuity and replacement for MODIS fire products, which has been 727 degrading after long-term service. The sensor parameters and major algorithms for noise detection and 728 fire identification in FY-3D products were introduced. For visual-check-based accuracy assessment, five 729 typical regions, Africa, South America, Indo-China Peninsula，Siberia and Australia, across the globe 730 were selected and the overall consistence between FY-3D fire products and reference dataaccuracy 731 exceeded 94%, with a more than 90% consistence in all regions. We also compared the FY-3D and 732 MODIS fire products for their consistence. The result suggested that the overall consistence was 84.4%, 733 with a fluctuation across seasons, surface types and regions. The high accuracy and consistence with 734 MODIS products proved that FY-3D fire product was an ideal tool for global fire monitoring. Based on 735 field-collected reference data, we further evaluated the suitability of FY-3D fire products in China. The 736 overall accuracy and accuracy (without considering omission errors) was 79.43% and 88.50% 737 respectively, higher than that of MODIS fire products. SSpecially, since detailed geographical conditions 738 in China were considered, FY-3D products should be preferably employed for monitoring fires and 739 estimating its environment effects in China.