1 00:00:01,590 --> 00:00:06,180 My name is Compton James Tucker, and I am a scientist at the 2 00:00:06,180 --> 00:00:10,560 Goddard Space Flight Center. We're very interested to improve 3 00:00:10,590 --> 00:00:15,270 our knowledge of the carbon cycle globally. Where is carbon 4 00:00:15,270 --> 00:00:19,410 going in vegetation? And how long does it persist? In the 5 00:00:19,410 --> 00:00:23,220 study, we use a large volume of commercial satellite data, 6 00:00:24,840 --> 00:00:27,990 hundreds of thousands of commercial satellite images at 7 00:00:27,990 --> 00:00:32,880 the 50 centimeter scale, to map trees to identify trees in a 8 00:00:32,880 --> 00:00:36,510 semi-arid region, from the Atlantic Ocean to the Red Sea in 9 00:00:36,510 --> 00:00:40,260 Africa, what we actually mapped were tree crowns. 10 00:01:13,210 --> 00:01:17,800 We then use our tree crown data to make predictions from the 11 00:01:17,830 --> 00:01:21,100 allometry, which was also collected on the same region. 12 00:01:21,460 --> 00:01:26,260 And the data are very important. The the processing code is 13 00:01:26,260 --> 00:01:29,710 important, the training data is important. The allometry is 14 00:01:29,710 --> 00:01:32,500 important. And then understanding the results that 15 00:01:32,500 --> 00:01:38,410 come out of those four components. In the study, the 16 00:01:38,410 --> 00:01:43,990 study has been in the works since 2015, or 2016. I started 17 00:01:43,990 --> 00:01:48,550 five or six years ago, draining the archive of all of the data 18 00:01:48,550 --> 00:01:53,050 from Africa. This has taken me three or four years to get all 19 00:01:53,050 --> 00:02:00,250 the data. Secondly, Ankit, who's one of our team members, as a 20 00:02:00,250 --> 00:02:05,080 graduate student in computer science, he wrote our processing 21 00:02:05,080 --> 00:02:08,560 code. And it's a highly optimized neural net code, it 22 00:02:08,560 --> 00:02:12,670 works very well. He worked on that for two or three years, 23 00:02:12,730 --> 00:02:16,420 then you need the training data to go with the processing code. 24 00:02:16,450 --> 00:02:20,290 When you use machine learning or artificial intelligence, you 25 00:02:20,290 --> 00:02:24,880 need to train on something so you have confidence that that's 26 00:02:24,880 --> 00:02:27,460 what you're measuring. Training data is where you go out and you 27 00:02:27,460 --> 00:02:31,900 select all different types of trees. And they have to have a 28 00:02:31,900 --> 00:02:35,170 green tree crown and an associated shadow to be a tree. 29 00:02:35,800 --> 00:02:39,490 And Martin Brandt did this over three or four months, and 30 00:02:39,490 --> 00:02:45,490 selected 89,000 or 90,000 individual trees, it's a heroic 31 00:02:45,490 --> 00:02:53,560 effort. Now there are people like Pierre Hiernaux, one of our 32 00:02:53,560 --> 00:02:56,590 co-authors who go out and they sample trees and they measure 33 00:02:56,590 --> 00:03:00,430 the tree crown, they then cut the tree down, they then measure 34 00:03:00,430 --> 00:03:05,560 the volume of leaves in the tree crown. The same for the wood and 35 00:03:05,560 --> 00:03:10,480 the same for the roots. And so we then convert the tree crown 36 00:03:10,480 --> 00:03:16,060 data which we measure into the predicted leaf mass or carbon, 37 00:03:16,270 --> 00:03:21,250 the root carbon and the wood carbon of every individual tree. 38 00:03:22,020 --> 00:03:26,790 You know, individual tree crown is probably the highest 39 00:03:27,810 --> 00:03:36,570 resolution you're gonna get. And like knowing the exact number of 40 00:03:36,570 --> 00:03:41,460 trees, and also when they have leaves throughout the year is 41 00:03:41,460 --> 00:03:45,000 going to be really, really important for improving our 42 00:03:45,000 --> 00:03:45,900 climate models. 43 00:03:47,130 --> 00:03:50,430 Then you put all this together, and you run out on a 44 00:03:50,430 --> 00:03:54,270 supercomputer. So we would run the data of this way, run it 45 00:03:54,270 --> 00:03:58,290 that way, then you take the results. That's really the fun 46 00:03:58,290 --> 00:04:02,190 part of seeing what you did, how well you did it, and what it can 47 00:04:02,190 --> 00:04:02,790 be used for. 48 00:04:03,950 --> 00:04:08,720 So the viewer is an important tool for NGOs that are 49 00:04:08,720 --> 00:04:13,700 interested in understanding if the tree restoration programs 50 00:04:13,730 --> 00:04:17,990 have paid off, but it can also be used for the local farmer who 51 00:04:17,990 --> 00:04:22,070 would be interested in knowing how many trees are standing on 52 00:04:22,250 --> 00:04:25,220 the fields, and are they alive, are they dead, 53 00:04:25,400 --> 00:04:29,300 etc. With a viewer you can zoom into individual trees and see 54 00:04:29,300 --> 00:04:33,050 how much carbon is there and the leaves and the wood and the 55 00:04:33,050 --> 00:04:38,240 roots and the specific location of that tree. Or you can 56 00:04:38,240 --> 00:04:42,350 aggregate the data up to an area of 100 meters by 100 meters or 57 00:04:42,350 --> 00:04:48,170 one hectare. We plan to expand our work next to Australia and 58 00:04:48,170 --> 00:04:52,160 then maybe to Eastern Africa, Southern Africa, Central Asia, 59 00:04:52,610 --> 00:04:56,840 and possibly other arid and semi arid areas.