News & views fully convolutional neural networks. This Ecology deep­learning method is designed to recog­ nize objects (in this case, tree canopies) on the basis of their characteristic shapes and colours within a larger image. Convolutional networks Satellites could soon rely on the availability of training data, which in this case consisted of satellite images in map every tree on Earth which the visible outlines of tree and shrub canopies were manually traced. Through train­ Niall P. Hanan & Julius Y. Anchang ing using these samples, the computer learnt how to identify individual tree canopies with An analysis of satellite images has pinpointed individual tree high precision in other images. The result is a canopies over a large area of West Africa. The data suggest that wall­to­wall mapping of all trees larger than it will soon be possible, with certain limitations, to map the 2 m in diameter across the whole of southern Mauritania, Senegal and southwestern Mali. location and size of every tree worldwide. See p.78 A previous estimate2 of the total number of trees on a global scale was obtained using field data from approximately 430,000 forest Terrestrial ecosystems are defined in large Brandt et al. provide a striking demon­ plots around the world. The authors of that part by their woody plants. Grasslands, shrub­ stration of this transformation in terrestrial study used statistical regression models to lands, savannahs, woodlands and forests rep­ remote sensing. The authors analysed more estimate tree density between the field sites, resent a series of gradations in tree and shrub than 11,000 images, at a spatial resolution of on the basis of vegetation type and climate. density, from ecosystems with low­density, 0.5 m, to identify individual trees and shrubs Their analysis suggested that there are approx­ low­stature woody plants to those with taller with canopy diameters of 2 m or more. The imately three trillion trees globally. However, trees and overlapping canopies. Accurate authors completed this giant task using arti­ this approach to tree­density estimation has information on the woody­vegetation struc­ ficial intelligence — exploiting a computa­ inherent errors and uncertainties, particularly ture of ecosystems is, therefore, fundamental tional approach that involves what are called for drylands, for which relatively few field to our understanding of global­scale ecology, biogeography and the bio geochemical cycles a of carbon, water and other nutrients. On 150 Trees per page 78, Brandt et al.1 report their analysis of hectare a massive database of high­resolution satellite MAURITANIA 0 1–25 images covering more than 1.3 million square 26–50 kilometres of the western Sahara and Sahel 51–100 300 regions of West Africa. The authors mapped > 100 the location and size of more than 1.8 billion individual tree canopies; never before have trees been mapped at this level of detail across such a large area. SENEGAL 600 The spatial resolution of most satellite data MALI is relatively coarse, with individual image pixels generally corresponding to areas on the GAMBIA ground that are larger than 100 square metres, and often larger than one square kilometre. b c d This limitation has forced researchers in the field of Earth observation to focus on meas­ uring bulk properties, such as the proportion of a landscape covered by tree canopies when viewed from above (a measurement known as canopy cover). However, during the past two decades, a variety of commercial satellites have begun to collect data at a higher spatial resolution, 5 km 5 km 50 m capable of capturing ground objects measur­ JULIUS ANCHANG ing one square metre or less. This resolution Figure 1 | Large-scale tree mapping. Accurate information about tree distribution provides useful ecological improvement places the field of terrestrial insights, but such data are difficult to obtain for large areas of land. a, A previous study2 estimating global remote sensing on the threshold of a funda­ tree density per hectare relied on data from field plots — samples of these data are shown for western Africa. mental leap forward: from focusing on aggre­ The inset box in a is in a dry region (with an average annual rainfall of less than 600 millimetres per year), and corresponds to b. Dotted lines indicate the boundaries of average rainfall in millimetres per year. c, d, Brandt gate landscape­scale measurements to having et al.1 report the detection of individual tree canopies across western Africa, obtained using an artificial­ the potential to map the location and canopy intelligence approach to analyse high­resolution satellite images. The authors found a higher tree density size of every tree over large regional or global in dry regions of Africa than did the earlier study. For example, Brandt and colleagues’ analysis of the area scales. This revolution in observational capa­ corresponding to the inset box in a produced the tree density per hectare shown in c. They identified the size bilities will undoubtedly drive fundamental and location of individual tree canopies (green), as shown in d for an area corresponding to the inset box in changes in how we think about, monitor, model c. Tree information at this level of detail was not available in the earlier study. (Images made using data from and manage global terrestrial ecosystems. refs 1 and 2.) (Springer Nature remains neutral with regard to jurisdictional claims in published maps.) 42 | Nature | Vol 587 | 5 November 2020 ©2020 Spri nger Nature Li mited. All rights reserved. ©2020 Spri nger Nature Li mited. All rights reserved. measurements are available to calibrate the be particularly tricky at regional and global wood biomass. Continuing research will models. scales and across biodiverse ecosystems. be needed to develop more­efficient cano­ A comparison (Fig. 1) of that earlier result The mapping of individual tree canopies by py­classification algorithms. In the meantime, with Brandt and colleagues’ findings in the species will probably remain at the top of the Brandt and colleagues have clearly demon­ western Sahel, for example, shows that the Earth­observation research community’s wish strated the potential for future global mapping previous study tended to underestimate the list for some time6. of tree canopies at submetre scales. number of trees in the drier regions (areas In the years ahead, remote sensing will with annual rainfall of less than 600 milli­ undoubtedly provide unprecedented detail Niall P. Hanan is with the Jornada Basin metres). Moreover, the previous estimates about vegetation structure as data from a Long-Term Ecological Research Program, provided no information on the location and range of sources — including light detection Department of Plant and Environmental size of individual trees within each square kilo­ and ranging (lidar), radar and high­resolution Sciences, and Julius Y. Anchang is in the metre, whereas Brandt and colleagues provide visible and near­infrared sensors — become Department of Plant and Environmental detailed information on the location and size more readily available7. Satellite­derived Sciences, New Mexico State University, of every individual canopy. The improvement Las Cruces, New Mexico 88003, USA. provided in the latest study can also be seen “Never before have trees e-mails: [email protected]; in the much higher level of detail it gives for [email protected] the wetter regions (those with annual rainfall been mapped at this level greater than 600 mm), and shows local spatial of detail across such variability in trees that is presumably associ­ a large area.” ated with contrasting soil types, water availa­ bility, land use and land­use history. There are, of course, caveats and limita­ high­resolution data on tree canopy size and 1. Brandt, M. et al. Nature 587, 78–82 (2020). tions to Brandt and colleagues’ work and the density could contribute to the inventory 2. Crowther, T. W. et al. Nature 525, 201–205 (2015). potential for scaling up their approach to a and management of forests and woodland, 3. Axelsson, C. R. & Hanan, N. P. Biogeosciences 14, 3239–3252 (2017). global analysis. Successful canopy detection deforestation monitoring, and assessment 4. Reichstein, M. et al. Nature 566, 195–204 (2019). declined drastically below a canopy diameter of the carbon sequestered in biomass, tim­ 5. Engler, R. et al. Forest Ecol. Mgmt 310, 64–73 (2013). of 2 m, owing to constraints imposed by the ber, fuel wood and tree crops. The ability to 6. Draper, F. C. et al. Trends Ecol. Evol. https://doi. org/10.1016/j.tree.2020.08.003 (2020). spatial resolution of the imagery, and consist­ map the size and location of individual tree 7. National Academies of Sciences, Engineering, and 3 ent with earlier work . Although we can expect canopies using such satellite data will com­ Medicine. Thriving on Our Changing Planet: A Decadal further improvements in the spatial resolu­ plement information available from other Strategy for Earth Observation from Space (Natl Academies Press, 2018). tion of satellite images, it becomes pertinent instruments that provide data for tree height, to ask what minimum canopy size is needed vertical canopy profiles and above­ground This article was published online on 14 October 2020. to characterize woody­plant communities in various regions. For global tree­canopy map­ Astronomy ping, if we assume that the computational and storage challenges associated with large data volumes can be overcome, the biggest roadblock would lie in developing efficient A fast radio burst approaches for automated classification and delineation of canopies. Brandt and col­ in our own Galaxy leagues’ deep­learning method required an input of approximately 90,000 manually digi­ Amanda Weltman & Anthony Walters tized training points. This approach becomes untenable for work on a global scale, and The origins of millisecond­long bursts of radio emissions, more­automated (unsuper vised) methods for known as fast radio bursts, from beyond our Galaxy have been extracting information from satellite imagery 4 enigmatic.