Thread by @Gangryong

Just published the first chapter of my PhD on remote fossil site detection using satellite images & #unsupervisedlearning in @GorongosaPark with my supervisors @paleobob @carvalhoprimate

It is #openaccess at peerj.com/articles/11573/

@primobevolab @oxford_anthro @PeerJLife 1/n

The paper starts by exploring the spatial and temporal gaps of the #primate & #hominid fossil record of Africa in the late #Miocene. @GorongosaPark in #Mozambique is shown to be a strategic location, with great potential to fill some major gaps in #paleoanthropology 2/n

We also discuss the difficulties of surveying for new paleontological localities, specially within modern forests/woodlands as in #Gorongosa, since dense vegetation cover reduces visibility (i.e. finding clues in topography and landscape). 3/n

To increase the chances of a successful discovery of fossil sites, we introduced an algorithmic pipeline. 1) Download #Landsat 8 satellite image ➜ 2) crop image to study area ➜ 3) clustering algorithm ➜ 4) binarize clusters and calculate variable importance #randomForest 4/n

4 new fossil sites were discovered in @GorongosaPark using this approach. Overall accuracy of the binarized #kmeans clusters was ~ 85%. This indicates the high potential of our remote sensing pipeline for exploratory paleontological surveys. 5/n

Relative importance of spectral bands for #clustering was determined using the randomForest algorithm, and #nearinfrared was the most important variable for fossil site detection, followed by other infrared bands. The visible spectrum are not good indicators of fossil sites. 6/n

This tool can be used for locating new fossil sites. In Gorongosa, the discovery of the first estuarine coastal forests of the #EARS fills an important paleobiogeographic gap of Africa. The new sites will be key for testing hypotheses of primate evolution in such settings. 7/n