Simple Data Science Of Global Warming In KDnuggets

Would love to get a post from you for KDnuggets (Gregory Piatetsky, KDnuggets President)

logoSome days ago, Gregory Piatetsky invited me to write a post for KDnuggets. I couldn’t say no. He suggested to me some topics and I decided to experiment around climate change to demonstrate how easy is to see some evidences of this alarming menace. You can read the post here.

This is the code I wrote to do this experiment:

require(sqldf)
require(googleVis)
require(ggplot2)
require(ggthemes)
setwd("YOUR WORKING DIRECTORY HERE") #This line doen's work until you type a valid path
#Data Avaliable in http://eca.knmi.nl/utils/downloadfile.php?file=download/ECA_blend_tg.zip
files=list.files(pattern = "TG_STAID")
results=data.frame(staid=character(0), trnd=numeric(0), nobs=numeric(0))
#Loop to calculate linear trends
for (i in 1:length(files))
{
  table=read.table(files[i], header=TRUE, sep = ',', skip=20)
  table=table[table$Q_TG==0,]
  table$date=as.Date(as.character(table$DATE), "%Y%m%d")
  results=rbind(data.frame(staid=files[i], trnd=lm.fit (x = matrix(table$date), y = table$TG)$coefficients, nobs=nrow(table)), results)
}
write.csv(results, file="results.csv", row.names = FALSE)#Save your work
results=read.csv(file="results.csv")#Read your work. You can start here if you stop your work further this line
#Remove outliers
results=results[!results$trnd %in% boxplot.stats(results$trnd, coef = 4)$out,]
#Histogram
hist(x=results$trnd, breaks = 50,
     col = "orange",
     border = "pink",
     freq=TRUE,
     xlim=c(-0.03, 0.03),
     ylim=c(0, 300),
     xlab="Historical trend of daily mean temperatures",
     ylab="Number of stations",
     main="Evolution of temperatures in Europe and the Mediterranean",
     sub="Source: European Climate Assessment & Dataset project")
results$staid2=as.numeric(gsub("[^0-9]","",results$staid)) #To join results with geographical coordinates
#Read table of geographical coordinates
staids=read.table("http://www.ecad.eu/download/ECA_all_stations.txt", header=TRUE, sep = ',', skip=17)
#Right tail of the distribution
hotpoints=sqldf("SELECT a.staid, a.staid2, a.trnd, a.nobs, b.staname, b.lat, b.lon
      FROM results a INNER JOIN staids b ON (a.staid2=b.staid) WHERE TRND >= 0.02")
#Convert degrees:minutes:seconds to decimal coordinates
hotpoints=within(hotpoints, {
  dms=do.call(rbind, strsplit(as.character(LAT), ":"))
  lat=sign(as.numeric(dms[,1]))*(abs(as.numeric(dms[,1]))+(as.numeric(dms[,2]) + as.numeric(dms[,3])/60)/60);rm(dms)
})
hotpoints=within(hotpoints, {
  dms=do.call(rbind, strsplit(as.character(LON), ":"))
  lon=sign(as.numeric(dms[,1]))*(abs(as.numeric(dms[,1]))+(as.numeric(dms[,2]) + as.numeric(dms[,3])/60)/60);rm(dms)
})
#To make readable tha name of the station in the map
hotpoints$staname=gsub("^\\s+|\\s+$", "", hotpoints$STANAME)
#To prepare coordinates to gvis function
hotpoints$LatLong=with(hotpoints, paste(lat, lon, sep=":"))
#The amazing gvisMap function of googleVis package
hotpointsMap=gvisMap(hotpoints, "LatLong" , "STANAME",
        options=list(showTip=TRUE,
                     showLine=TRUE,
                     enableScrollWheel=TRUE,
                     mapType='terrain',
                     useMapTypeControl=TRUE))
plot(hotpointsMap)
#The plot one of this hot stations
InAmenas=read.table("TG_STAID000312.txt", header=TRUE, sep = ',', skip=20)
InAmenas=InAmenas[InAmenas$Q_TG==0,]
InAmenas$date=as.Date(as.character(InAmenas$DATE), "%Y%m%d")
ggplot(InAmenas, aes(date, TG)) +
  geom_line(color="red", alpha=0.8) +
  xlab("") +
  ylab("Mean temperature in 0.1C")+
  ggtitle("Mean temperature in IN-AMENAS (ALGERIA) 1958- 1998")+
  geom_smooth(method = "lm", se=FALSE, color="red", lwd=2)+
  theme_economist(base_size = 20, base_family = "sans", horizontal = TRUE,
                  dkpanel = FALSE, stata = FALSE)

Leave a Reply

Your email address will not be published. Required fields are marked *