The Econmist - USA (2021-10-30)

The Economist October 30th 2021 Science&technology 91

Renewableenergy

A census of solar cells

R

ebuildingan entireplanet’senergy

system  is  a  big  job.  Just  ask  the  dele­

gates  at  the  cop­26  climate  conference

scheduled to kick off in Britain on October

31st.  The  most  basic  problem  is  knowing

what, exactly, you are trying to rebuild. Ac­

ademic­research  groups,  think­tanks,

charities  and  other  concerned  organisa­

tions try to keep track of the world’s wind

turbines,  solar­power  plants,  fossil­fu­

elled power stations, cement factories and

so on. To this end, they rely heavily on data

from  national  governments  and  big  com­

panies, but these are often incomplete. The

most  comprehensive  database  covering

American solar­power installations, for in­

stance, is thought to miss around a fifth of

the  photovoltaic  panels  actually  installed

on the ground.

In  a  paper  just  published  in  Nature,  a

team of researchers led by Lucas Kruitwa­

gen, a climate scientist and ai researcher at

Oxford  University,  demonstrate  another

way to keep tabs on the green­energy revo­

lution.  Dr  Kruitwagen  and  his  colleagues

have  put  together  an  inventory  of  almost

69,000  big  solar­power  stations  (defined

as  those  with  a  rated  capacity  of  10kwof

electricity  or  more)  all  over  the  world—

more than four times as many as were pre­

viously  listed  in  public  databases.  This

new  inventory  includes  their  locations,

the  date  they  entered  service  and  a  rough

estimate of their generating capacity. 

Conceptually,  the  team’s  method  is

simple. Instead of relying on top­down re­

ports,  they  worked  from  the  bottom  up,

looking at the entire planet from space and

counting  how  many  solar  panels  they

could see. This is not the first time people

have  hunted  from  orbit  for  solar­power

stations.  But  previous  analyses  have  been

limited  to  a  few  countries.  As  far  as  Dr

Kruitwagen knows, his is the first attempt

to survey the entire planet for a particular

type of infrastructure. Earth is a big place,

of course, which means practice is a great

deal harder than theory. His approach has

been made possible by two big technologi­

cal trends. 

One  is  a  growing  abundance  of  cheap,

easily  available  satellite  imagery.  In  the

20th  century,  reconnaissance  satellites

were  the  jealously  guarded  property  of  a

handful of governments. These days, a cot­

tage  industry  of  Earth­observation  firms

and agencies sells images on the open mar­

ket.  Dr  Kruitwagen’s  pictures  came  from

two sets of satellites, Sentinel­2 and spot,

run by the European Space Agency and Air­

bus  respectively.  These  peer  down  on  the

world, recording visible light and also the

infrared  and  ultraviolet  parts  of  the  spec­

trum.  The  images  Dr  Kruitwagen  used

amounted to around 550 terabytes of data,

spanning  the  period  between  2016  and

2018.  That  is  enough  to  fill  more  than  a
      hundred desktop hard drives. 
      Sifting  through  this  many  pictures  by
      eye  would  have  been  impractical.  That  is
      where  the  second  technological  trend
      comes  in.  Dr  Kruitwagen  and  his  col­
      leagues  trained  a  machine­learning  sys­
      tem to spot the solar panels for them. 
      Computer  vision  is  a  hot  field.  But  the
      specifics  of  orbital  reconnaissance  meant
      that off­the­shelf software was not suitable

forthetasktheresearchershadinmind.

Machine­learningsystemsaretaughtwhat

todobyexamininga “trainingset”,which

contains examples of what is being

searchedfor.Forcommontaskssuchasfa­

cialrecognition,pre­builttrainingsetsare

oftenavailable.ButDrKruitwagen’steam

hadtobuildtheirown.

Forthis,theyturnedtoOpenStreetMap,

anopen­sourcerival to GoogleMapsin

whichvolunteershadalreadytaggedlarge

numbersofsolarplants.Buttherewaslit­

tle consistency. “Some people had just

drawn rough outlines aroundan entire

field,” Dr Kruitwagen says. “Others had

goneinandtracedtheoutlineofeachrow

ofpanelsseparately.”Fixingthatinvolveda

greatdealofmanuallabour.

Oncethetrainingdatahadbeencleaned

up, the learning algorithms had to be

tweakedaswell.Fromspace,evenbigsolar

installationslooksmall.Eachpixelinthe

Sentinelimagesrepresenteda ten­by­ten­

metresquare.Evenforthehigher­resolu­

tionspotsatellites,thesquares’sidesare

one and a half metres long. Existing classi­

fiers, trained for things like facial recogni­

tion  or  self­driving  cars,  are  used  to  spot­

ting objects that loom large in their field of

vision.  Hunting  for  smaller  ones  meant

tinkering  with  the  software  to  boost  its

ability  to  detect  tiny  features.  False  posi­

tives—things  like  tennis  courts  and  agri­

cultural  greenhouses  that  resemble  solar

panels from space—had to be removed.

Panel games

Though extraordinary, Dr Kruitwagen’s re­

sults are already out of date. The data­gath­

ering  phase  of  the  project  ended  in  2018,

meaning that the thousands of new plants

built  since  then  are  not  included.  But  the

project,  he  says,  proves  that  the  method

works. He intends to make his results, in­

cluding  the  labour­intensive  training  set,

available for others to use. One logical ex­

tension of his project, he says, would be to

expand the analysis to include solar panels

installed  on  domestic  rooftops.  Such  “be­

hind­the­meter”  installations  are  particu­

larly tricky to track in other ways. 

More  generally,  Dr  Kruitwagen  hopes

that  his  eye­in­the­sky  approach—which,

despite  the  planetary  scale  of  the  project,

cost  only  around  $15,000  in  cloud­com­

puting time—could presage more accurate

estimates  of  other  bits  of  climate­related

infrastructure,  such  as  fossil­fuel  power

stations,  cement  plants  and  terminals  for

ships  carrying  liquefied  natural  gas.  The

eventual result could be the assembly of a

publicly available, computer­generated in­

ventory  of  every  significant  bit  of  energy

infrastructure  on  Earth.  Quite  apart  from

such  a  model's  commercial  and  academic

value,  hesays,an  informed  public  would

be one betterable to hold politicians’ feet

to the fire.n

An accurate count of Earth’s solar-powerstationshasnowbeenmade

Bring me sunshine

PV solar-energy facilities*

Source: Nature, 2021

*Over ten kilowatts capacity.

Study from Jun 1st 2016 to Sep 30th 2018

An inland sea of solar panels