Count­ing craters

Ma­chine learn­ing could help re­searchers date plan­e­tary sur­faces by tal­ly­ing up their craters

Sky at Night Magazine - - BULLETIN -

Count­ing the num­ber of craters in a re­gion of a plan­e­tary sur­face gives sci­en­tists im­por­tant in­for­ma­tion on how old the planet is. Like specks of rain­drops ac­cu­mu­lat­ing on a pave­ment, the longer a sur­face has been ex­posed, the more craters it will have. Crater den­si­ties tell you the rel­a­tive age of fea­tures like lava flows or wind de­posits, and if we’ve also been able to an­a­lyse a sam­ple from the sur­face (as we have with lu­nar rocks re­turned by the Apollo pro­gramme and then dated pre­cisely by mea­sur­ing ra­dioac­tive iso­topes, for ex­am­ple), we can then also cal­cu­late the ab­so­lute age of whole re­gions.

The prob­lem is that crater count­ing has tra­di­tion­ally been an ex­ceed­ingly slow and la­bo­ri­ous ex­er­cise; up un­til now it’s been mostly done by hu­man eye. The data we’ve been able to gather this way ei­ther cov­ers large ar­eas of a sur­face tak­ing note of only the largest craters, or it in­cludes the smaller craters but cov­ers only a very spe­cific, lim­ited geo­graphic re­gion. Au­to­mated meth­ods us­ing com­puter al­go­rithms have been de­vel­oped, but they can of­ten be con­fused by over­lap­ping or de­graded craters, vari­a­tions in il­lu­mi­na­tion or other land­scape fea­tures, such as ridges.

Ari Sil­burt at the Univer­sity of Toronto and his col­leagues have been try­ing to change all of this. They have ap­plied a new com­puter tech­nique to the chal­lenge based on Deep Learn­ing, which uses ar­ti­fi­cial neu­ral net­works, where the com­puter em­u­lates the func­tion­ing of a sim­ple brain. As Sil­burt him­self ex­plains, “Sim­i­lar to how a hu­man learns to recog­nise a cat by see­ing many dif­fer­ent ex­am­ples of cats, a com­puter can learn to recog­nise a cat via ma­chine learn­ing by re­ceiv­ing many ex­am­ples of what is and isn’t a cat.”

Sil­burt and his team ap­plied their neu­ral net­work to land­scape maps of the Moon’s sur­face cre­ated by the Lu­nar Re­con­nais­sance Or­biter and Kaguya probes. The im­me­di­ate ad­van­tage of start­ing with these ‘dig­i­tal el­e­va­tion maps’, rather than sim­ply pho­to­graphs, is that they are not af­fected by vary­ing shad­ows from dif­fer­ent an­gles of sun­light.

They first tested their neu­ral net­work on im­ages that had al­ready been counted by sci­en­tists to check that it worked re­li­ably. Their tech­nique suc­cess­fully lo­cated 92 per cent of the craters iden­ti­fied by the hu­man ex­perts. Cru­cially, it also found a large num­ber of new craters – al­most twice as many, in fact, and in par­tic­u­lar scores of those very small craters that are of­ten ne­glected in crater counts car­ried out with the hu­man eye. Sil­burt then ap­plied his neu­ral net­work to the planet Mer­cury and found that his tech­nique also worked very well on this com­pletely dif­fer­ent ter­rain.

While this is a very promis­ing new ap­proach, Sil­burt’s crater-count­ing com­puter code still had an er­ror rate of about 11 per cent. This isn’t too bad, but prob­a­bly not quite good enough just yet to be used for com­pletely au­tomat­ing the process of build­ing ac­cu­rate crater cat­a­logues. Sil­burt and his team are now work­ing on tweak­ing and per­fect­ing their sys­tem. Nev­er­the­less, this rep­re­sents a very promis­ing ap­proach for au­tomat­ing the la­bo­ri­ous process of iden­ti­fy­ing the num­ber and sizes of craters, and so ul­ti­mately im­prov­ing our un­der­stand­ing of how dif­fer­ent land­scapes on plan­ets in the So­lar Sys­tem formed.

LEWIS DARTNELL was read­ing… Lu­nar Crater Iden­ti­fi­ca­tion via Deep Learn­ing by Ari Sil­burt et al Read it on­line at

“Their tech­nique lo­cated 92 per cent of the craters iden­ti­fied by ex­perts. It also found a large num­ber of new crater"

Count­ing craters by eye is easy for the big ones, but we may need to use AI to in­clude smaller ones

LEWIS DARTNELL is an as­tro­bi­ol­ogy re­searcher at the Univer­sity of West­min­ster and the au­thor of The Knowl­edge: How to Re­build our World from Scratch (www.the-knowl­

Newspapers in English

Newspapers from UK

© PressReader. All rights reserved.