Dfyjkt

I have a dataframe with a time series that looks like this:

df<-structure(list(date = structure(c(-6905, -6891, -6853, -6588, 
-6588, -6586, -6523, -6515, -5856, -5753), class = "Date"), flow = c(2.22, 
2.56, 3.3, 1.38, 4, 1.4, 1.32, 1.26, 6, 35.69)), .Names = c("date", 
"flow"), row.names = c(NA, 10L), class = "data.frame")

I want to remove all the lines that are not the maximum within 2 days forward or backward of its date. So in the case above, lines 4 and 6 will be removed. I couldn't find similar answered questions.

I wrote this code that doesn't work and it is ugly, long and doesn't take care of the edges of the dataframe:

 idx <- c()
 for (j in 3:(length(df$date)-2))
 as.Date(df$date[j])-as.Date(df$date[j-2])<3)
 if (df$flow[j]!=max(df$flow[(j-2):(j+2)]))
 idx <- c(idx,j)
 
 else if (as.Date(df$date[j+1])-as.Date(df$date[j])<3 

Notice that the dates in the dataframe are not consecutive.

Using the zoo library.

library(zoo)

# convert into a zoo time series
dtf.zoo <- zoo(dt$flow, order.by=dt$date)

# remove duplicate dates by keeping the maximum value
dtf.zoo <- aggregate(dtf.zoo, time(dtf.zoo), max)

# pad with NAs to make the time series regular
dtf.zoo <- merge(
 dtf.zoo, 
 zoo(, seq(min(index(dtf.zoo)), max(index(dtf.zoo)), "day"))
)

# find rows that are less than a value two days prior or hence
rem <- which(dtf.zoo < rollapply(dtf.zoo, 5, max, na.rm=TRUE, partial=TRUE))

# remove those rows
dtf.zoo2 <- dtf.zoo[-rem]

# remove NAs
dt2 <- data.frame(flow=na.omit(dtf.zoo2))
dt2
# flow
# 1951-02-05 2.22
# 1951-02-19 2.56
# 1951-03-29 3.30
# 1951-12-19 4.00
# 1952-02-22 1.32
# 1952-03-01 1.26
# 1953-12-20 6.00
# 1954-04-02 35.69

which(!(dt$flow %in% dt2$flow))
# 4 6

You can also use the tidyverse approch:

require(tidyverse)

df %>% 
 #Arrange by date
 arrange(date) %>%
 #Picking the max for each da
 group_by(date) %>% 
 top_n(1, flow) %>% 
 ungroup() %>%
 #Adding missing dates with NAs
 complete(date = seq.Date(min(date), max(date), by="day")) %>% 
 #Remove Two up/down
 mutate(
 remove = case_when(
 flow < rowMeans(data.frame(lag(flow, 1), 
 lag(flow, 2)), na.rm = TRUE) ~ "remove", 
 flow < rowMeans(data.frame(lead(flow, 1),
 lead(flow, 2)), na.rm = TRUE) ~ "remove", 
 TRUE ~ "keep")) %>% 
 na.omit() %>%
 filter(remove == "keep") %>% 
 select(-remove)


# A tibble: 8 x 2
 date flow
 <date> <dbl>
1 1951-02-05 2.22
2 1951-02-19 2.56
3 1951-03-29 3.30
4 1951-12-19 4.00
5 1952-02-22 1.32
6 1952-03-01 1.26
7 1953-12-20 6.00
8 1954-04-02 35.7 

I use lapply() to check the range : [date - 2 days , date + 2 days] of each date.

rm.list <- lapply(df$date, function(x) 
 ind <- which(abs(df$date - x) <= 2)
 flow <- df$flow[ind]
 if(length(ind) > 1) which(flow < max(flow)) + min(ind) - 1
 else NULL
)

rm <- unique(unlist(rm.list)) # [1] 4 6
df[-rm, ]

# date flow
# 1 1951-02-05 2.22
# 2 1951-02-19 2.56
# 3 1951-03-29 3.30
# 5 1951-12-19 4.00
# 7 1952-02-22 1.32
# 8 1952-03-01 1.26
# 9 1953-12-20 6.00
# 10 1954-04-02 35.69