Unsupervised learning of satellite images enhances discovery of late Miocene fossil sites in the Urema Rift, Gorongosa, Mozambique

Abstract

Background Paleoanthropological research focus still devotes most resources to areas generally known to be fossil rich instead of a strategy that first maps and identifies possible fossil sites in a given region. This leads to the paradoxical task of planning paleontological campaigns without knowing the true extent and likely potential of each fossil site and, hence, how to optimize the investment of time and resources. Yet to answer key questions in hominin evolution, paleoanthropologists must engage in fieldwork that targets substantial temporal and geographical gaps in the fossil record. How can the risk of potentially unsuccessful surveys be minimized, while maximizing the potential for successful surveys? Methods Here we present a simple and effective solution for finding fossil sites based on clustering by unsupervised learning of satellite images with the k-means algorithm and pioneer its testing in the Urema Rift, the southern termination of the East African Rift System (EARS). We focus on a relatively unknown time period critical for understanding African apes and early hominin evolution, the early part of the late Miocene, in an overlooked area of southeastern Africa, in Gorongosa National Park, Mozambique. This clustering approach highlighted priority targets for prospecting that represented only 4.49% of the total area analysed. Results Applying this method, four new fossil sites were discovered in the area, and results show an 85% accuracy in a binary classification. This indicates the high potential of a remote sensing tool for exploratory paleontological surveys by enhancing the discovery of productive fossiliferous deposits. The relative importance of spectral bands for clustering was also determined using the random forest algorithm, and near-infrared was the most important variable for fossil site detection, followed by other infrared variables. Bands in the visible spectrum performed the worst and are not likely indicators of fossil sites. Discussion We show that unsupervised learning is a useful tool for locating new fossil sites in relatively unexplored regions. Additionally, it can be used to target specific gaps in the fossil record and to increase the sample of fossil sites. In Gorongosa, the discovery of the first estuarine coastal forests of the EARS fills an important paleobiogeographic gap of Africa. These new sites will be key for testing hypotheses of primate evolution in such environmental settings.

The first Miocene fossils from coastal woodlands in the southern East African Rift

“…After four field seasons (2016-2019), extensive surveys, and new approaches in the search of paleontological sites (d’Oliveira Coelho et al, 2021), the Paleo-Primate Project Gorongosa has 1) documented ten new paleontological localities, 2) established a preliminary stratigraphic and sedimentological framework for the fossil sites (Habermann et al, 2019), 3) provided the first radiometric age determinations for the Mazamba Formation, 4) provided the first reconstructions of past vegetation in the region based on pedogenic carbonates and fossil wood, and 5) described the first fossil teeth from the southern East African Rift System. The Gorongosa sample includes new species of fossil mammals, and a unique combination of fossil specimens straddling the terrestrial/marine biomes.…”

“…Furthermore, none of the well-known Miocene fossil sites in the EARS provides evidence of eastern African coastal forests, a major ecosystem that may have played a key role in hominin origins and the evolution of several mammalian lineages (Joordens, Feibel, Vonhof, Schulp, & Kroon, 2019; Kingdon, 2003). Although the necessity of documenting new fossil sites in previously unknown areas is widely appreciated and advocated (Almécija et al, 2021; Cote, 2018), discovering entirely new paleontological areas is a rare event (d’Oliveira Coelho, Anemone, & Carvalho, 2021). Here we describe the first dentognathic specimens of fossil vertebrates discovered in the East African Rift of central Mozambique.…”

Remote Sensing Reveals Lasting Legacies of Land-Use by Small-Scale Foraging Communities in the Southwestern Indian Ocean

“…Our paper aims to build on this previous work by combining archaeological satellite-based remote sensing and innovative computational automation approaches to reveal legacies of forager land-use in coastal Madagascar. Prior remote sensing studies of archaeological foraging societies (and recently paleoanthropological sites) have taken place around the world, and many have relied on unsupervised land-cover classifications and environmental proxies to narrow down survey areas to where ephemeral cultural deposits are likely to exist (Keeney and Hickey, 2015;Davis et al, 2020a;d'Oliveira Coelho et al, 2021;Lim et al, 2021). Here, we attempt to use multispectral satellite data to directly pinpoint ecological changes associated with archaeological activity.…”

Multispectral satellite imaging improves detection of large individual fossils

“…They ran advanced statistics to objectively consider the results, stressing again the importance of creating spatial predictive models before planning palaeontological surveys. Finally, a recent study by d’Oliveira Coelho et al (2021) applied unsupervised ( k -means) classification on a portion of the East African Rift System (Urema Rift, Mozambique) reaching 84.6 % accuracy in fossil site detection. In recent years, researchers have begun pointing their attention to extraterrestrial ichnofossil signs on other planets, such as Mars (McKeown et al 2009, 2011; Noe Dobrea et al 2010, 2017; Baucon et al 2021), giving a new motivation to using satellite images and spatial analyses to detect life.…”