4  Initial Analysis - Trigger Proposal

Faryab

4.1 Intro

Based on initial partner feedback it was decided to focus on Faryab province.

This script evaluates various environmental indicators for their utility as trigger metrics in Faryab Province in Afghanistna. The triggers are evaluated:

• Forecasted:
  • SEAS5 – cumulative predicted rainfall f or MAM predicted form Nov-March Johnson et al. (2019)
• Observational:
  • CHIRPS – Monthly rainfall: November - May
  • NDSI – Monthly NDSI: November – May
  • NDVI – Monthly NDVI: March- May
  • Combined Monthly CHIRPS & SWE: March- May

For each indicator we determine the worst years on record (drought-wise) across each monitoring moment/time-step. We then see how well the worst years at each of these indicator-timestep combinations line up with the worst years calculated according to ASI Dekad 3 calculation. We look at ASI Dekad 3 of May because ASI is a measure of scale of drought at a geographical level based on NDVI and crop specific coefficients. May Dekad 3 is after the growing season and just as the harvest begins for spring wheat, so is a good time to measure this cumulative indicator. It corresponded with the EMDAT dataset on historical drought in Faryab.

The heatmap & rainfall time-series produced here are included in this slide deck

box::use(
    .. / R / blob_connect
)

box::use(
    dplyr,
    gg= ggplot2,
    lubridate,
    purrr,
    tidyr,
    stringr,
    readr,
    gghdx,
    janitor,
    scales,
    stats,
    gghdx,
    forcats
)

gghdx$gghdx()

#############################
#### PERFORMANCE METRICS ####
#############################

# We calculate performance in a specific way. It is by passing in a dataset
# that has years when we would want to activate, `y`, and and years when we would
# activate for a certain indicator, `x`. These are then used to calculate precision
# and recall. We also use bootstrapping to test if we are better than random.
#
# Since we have a set return period for the historical activation data, and also
# the same desired activation rate as the return period, in this instance, sum(x)
# and sum(y) are equal, so precision and recall are equal.

# functions to calculate metrics
precision <- function(x, y) {
    sum(x & y) / sum(x)
}

recall <- function(x, y) {
    sum(x & y) / sum(y)
}

#' Compares level of metric (calcualted by `fun`) to "random" values
#' by bootstrapping `y` 1,000 times. Returns the % of times the actual metric is higher than the
#' random. Expect 95% of the time or more we are better than random!
#'
#' If the values are the same, considers it "good" (better than random).
metric_calc <- function(x, y, fun, indicator) {
    threshold <- fun(x = x, y = y)
    pct_better <- purrr$map_lgl(
        .x = 1:1000,
        .f = \(a) fun(x = sample(x, size = length(x), replace = FALSE), y = y) <= threshold
    ) |>
        mean()

    metric_name <- deparse(substitute(fun))

    dplyr$tibble(
        "{metric_name}" := threshold,
        "{metric_name}_beq_random" := pct_better
    )
}

calc_metrics <- function(df, x, y) {
    df |>
        dplyr$summarise(
            precision = metric_calc({{ x }}, {{ y }}, precision),
            recall = metric_calc({{ x }}, {{ y }}, recall),
            indicator = unique(indicator),
            .groups = "drop"
        ) |>
        tidyr$unnest(
            cols = tidyr$everything()
        )
}

##############
#### DATA ####
##############

# blob processed data
df_chirps <- blob_connect$read_blob_file("DF_ADM1_CHIRPS", show_col_types = FALSE) |>
    dplyr$filter(
        ADM1_NAME == "Faryab"
    ) |>
    dplyr$group_by(
        mo = lubridate$month(date)
    ) |>
    dplyr$mutate(
        chirps_z = (value - mean(value)) / stats$sd(value)
    ) |>
    dplyr$ungroup()

df_seas5 <- blob_connect$read_blob_file("DF_FARYAB_SEAS5", show_col_types = FALSE) |>
    dplyr$mutate(
        valid_month = lubridate$month(valid_date),
        valid_year = lubridate$year(valid_date)
    )

df_snow <- blob_connect$read_blob_file("DF_ADM1_MODIS_SNOW", show_col_types = FALSE) |>
    dplyr$filter(
        ADM1_NAME == "Faryab",
        parameter == "NDSI_Snow_Cover_mean"
    ) |>
    dplyr$group_by(
        mo = lubridate$month(date)
    ) |>
    dplyr$mutate(
        swe_z = (value - mean(value)) / stats$sd(value)
    ) |>
    dplyr$ungroup()

# not used later on because the time series is much more limited than the rest
df_smi <- blob_connect$read_blob_file("DF_ADM1_SMI", skip = 1, show_col_types = FALSE) |>
    dplyr$slice(-1) |> #  first line is just an extra line of headers
    dplyr$transmute(
        date = as.Date(paste0(Province, "-01")),
        year = lubridate$year(date),
        month = lubridate$month(date),
        smi = Faryab
    )

# FAO GIEWS data
df_ndvi <- readr$read_csv(
    "https://eo.apps.fao.org/asis/data/country/AFG/GRAPH_NDVI_AGRI/ndvi_adm1_data.csv",
    show_col_types = FALSE
) |>
    dplyr$filter(Province == "Faryab")

df_vhi <- readr$read_csv("https://www.fao.org/giews/earthobservation/asis/data/country/AFG/MAP_NDVI_ANOMALY/DATA/vhi_adm1_dekad_data.csv", show_col_types = FALSE) |>
    dplyr$filter(Province == "Faryab")
df_asi <- readr$read_csv("https://www.fao.org/giews/earthobservation/asis/data/country/AFG/MAP_ASI/DATA/ASI_Dekad_Season1_data.csv", show_col_types = FALSE) |>
    dplyr$filter(Province == "Faryab")

# EMDAT
df_emdat <- blob_connect$read_blob_file("DF_EMDAT")

# Production data
df_prod_init <- blob_connect$read_blob_file("DF_PRODUCTION_DATA", skip = 3, .name_repair = "unique_quiet")

names_2 <- as.character(df_prod_init[1, 4:ncol(df_prod_init)])

df_prod <- df_prod_init |>
    dplyr$slice(
        2
    ) |>
    dplyr$select(
        -c(1:3)
    ) |>
    tidyr$pivot_longer(
        cols = tidyr$everything(),
        values_transform = as.numeric
    ) |>
    dplyr$mutate(
        name = ifelse(
            stringr$str_detect(name, "\\.\\.\\."),
            NA_character_,
            name
        )
    ) |>
    tidyr$fill(
        name
    ) |>
    dplyr$mutate(
        metric = names_2
    )

##########################
#### FIND WORST YEARS ####
##########################

# worst end of season vegetation health index
df_vhi_wy <- df_vhi |>
    dplyr$filter(
        Month == "05",
        Dekad == 3
    ) |>
    dplyr$arrange(
        Data
    ) |>
    dplyr$slice(
        1:8 # 1 in 5 year RP
    ) |>
    dplyr$transmute(
        indicator = "VHI",
        year = Year
    )

# look from 2000 onward
df_vhi_wy_2000 <- df_vhi |>
    dplyr$filter(
        Month == "05",
        Dekad == 3,
        Year >= 2000
    ) |>
    dplyr$arrange(
        Data
    ) |>
    dplyr$slice(
        1:5
    ) |>
    dplyr$transmute(
        indicator = "VHI",
        year = Year
    )

df_seas5_wy <- df_seas5 |>
    dplyr$filter(
        valid_month %in% 3:5
    ) |>
    dplyr$group_by(
        pub_date,
        valid_year
    ) |>
    dplyr$filter(
        dplyr$n() == 3
    ) |>
    dplyr$summarize(
        precipitation = mean(precip_mm_day),
        leadtime = min(leadtime),
        pub_month = lubridate$month(min(pub_date)),
        .groups = "drop"
    ) |>
    dplyr$filter(
        valid_year >= 1984
    ) |>
    dplyr$group_by(
        leadtime
    ) |>
    dplyr$arrange(
        precipitation,
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:8 # 1 in 5 year RP
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "SEAS5",
        year = valid_year,
        leadtime,
        pub_month
    )

df_ndvi_wy <- df_ndvi |>
    dplyr$filter(
        Month %in% c("03", "04", "05")
    ) |>
    dplyr$group_by(
        Year
    ) |>
    dplyr$mutate(
        ndvi_cum = cumsum(Data),
        ndvi_cum_avg = cumsum(Data_long_term_Average),
        ndvi_cum_anom = ndvi_cum / ndvi_cum_avg
    ) |>
    dplyr$group_by(
        Month,
        Dekad
    ) |>
    dplyr$arrange(
        ndvi_cum,
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:8 # 1 in 5 year RP
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "NDVI",
        month = as.numeric(Month),
        dekad = as.numeric(Dekad),
        year = Year
    )

# since 2000
NDVI <- df_ndvi |>
    dplyr$filter(
        Month %in% c("03", "04", "05"),
        Year >= 2000
    ) |>
    dplyr$group_by(
        Year
    ) |>
    dplyr$mutate(
        ndvi_cum = cumsum(Data),
        ndvi_cum_avg = cumsum(Data_long_term_Average),
        ndvi_cum_anom = ndvi_cum / ndvi_cum_avg
    ) |>
    dplyr$group_by(
        Month,
        Dekad
    ) |>
    dplyr$arrange(
        ndvi_cum,
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:5
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "NDVI",
        month = as.numeric(Month),
        dekad = as.numeric(Dekad),
        year = Year
    ) |>
    dplyr$filter(
        month == 5,
        dekad == 3
    )

df_chirps_wy <- df_chirps |>
    dplyr$mutate(
        month = lubridate$month(date),
        year = lubridate$year(date)
    ) |>
    dplyr$filter(
        month %in% 1:5,
        year >= 1984
    ) |>
    dplyr$group_by(
        year
    ) |>
    dplyr$mutate(
        precipitation_cum = cumsum(value)
    ) |>
    dplyr$group_by(
        month
    ) |>
    dplyr$arrange(
        precipitation_cum,
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:8 # 1 in 5 year RP
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "CHIRPS",
        month,
        year
    )

df_chirps_wy_2000 <- df_chirps |>
    dplyr$mutate(
        month = lubridate$month(date),
        year = lubridate$year(date)
    ) |>
    dplyr$filter(
        month %in% 1:5,
        year >= 2000
    ) |>
    dplyr$group_by(
        year
    ) |>
    dplyr$mutate(
        precipitation_cum = cumsum(value)
    ) |>
    dplyr$group_by(
        month
    ) |>
    dplyr$arrange(
        precipitation_cum,
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:5
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "CHIRPS",
        month,
        year
    )

# look at CHIRPS just across MAM

df_chirps_mam_wy <- df_chirps |>
    dplyr$mutate(
        month = lubridate$month(date),
        year = lubridate$year(date)
    ) |>
    dplyr$filter(
        month %in% 3:5,
        year >= 1984
    ) |>
    dplyr$group_by(
        year
    ) |>
    dplyr$mutate(
        precipitation_cum = cumsum(value)
    ) |>
    dplyr$group_by(
        month
    ) |>
    dplyr$arrange(
        precipitation_cum,
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:8 # 1 in 5 year RP
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "CHIRPS",
        month,
        year
    )

df_emdat_wy <- df_emdat |>
    dplyr$transmute(
        indicator = "EMDAT",
        year = c(2001, 2007, 2008, 2011, 2018, 2023),
        faryab = stringr$str_detect(`Location (Affected provinces)`, "Faryab")
    )

# used for potential monitoring
df_asi_dek_wy <- df_asi |>
    dplyr$mutate(
        month = as.numeric(Month),
        dekad = Dekad,
        year = Year
    ) |>
    dplyr$filter(
        month <= 5
    ) |>
    dplyr$group_by(
        month,
        dekad
    ) |>
    dplyr$arrange(
        dplyr$desc(Data),
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:8 # 1 in 5 year RP
    ) |>
    dplyr$transmute(
        indicator = "ASI",
        year = Year
    ) |>
    dplyr$ungroup()


df_asi_wy <- df_asi |>
    dplyr$filter(
        Month == "05",
        Dekad == 3
    ) |>
    dplyr$arrange(
        dplyr$desc(Data)
    ) |>
    dplyr$slice(
        1:8 # 1 in 5 year RP
    ) |>
    dplyr$transmute(
        indicator = "ASI",
        year = Year
    )

df_asi_wy_2000 <- df_asi |>
    dplyr$filter(
        Month == "05",
        Dekad == 3,
        Year >= 2000
    ) |>
    dplyr$arrange(
        dplyr$desc(Data)
    ) |>
    dplyr$slice(
        1:5
    ) |>
    dplyr$transmute(
        indicator = "ASI",
        year = Year
    )

df_snow_wy <- df_snow |>
    dplyr$mutate(
        month = lubridate$month(date)
    ) |>
    dplyr$filter(
        month %in% c(1:5, 11:12)
    ) |>
    dplyr$group_by(month) |>
    dplyr$slice_min(
        order_by = value,
        n = 5,
        with_ties = FALSE
    ) |>
    dplyr$mutate(
        year = ifelse(month > 5, lubridate$year(date) + 1, lubridate$year(date))
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "NDSI",
        month,
        year
    )

df_smi_wy <- df_smi |>
    dplyr$group_by(
        month
    ) |>
    dplyr$arrange(
        smi,
        .by_group = TRUE
    ) |>
    dplyr$slice(
        1:2
    ) |>
    dplyr$ungroup() |>
    dplyr$transmute(
        indicator = "SMI",
        month,
        year
    )

# CHIRPS AND SWE COMBINATION

df_combo_wy <- df_chirps |>
    dplyr$select(
        date,
        chirps_z
    ) |>
    dplyr$left_join(
        dplyr$select(
            df_snow,
            date,
            swe_z
        )
    ) |>
    dplyr$mutate(
        year = lubridate$year(date),
        month = lubridate$month(date),
        combo = chirps_z + swe_z
    ) |>
    dplyr$filter(
        year >= 2000,
        month %in% 1:5
    ) |>
    dplyr$group_by(
        month
    ) |>
    dplyr$arrange(
        combo
    ) |>
    dplyr$slice(
        1:5
    )

# look at worst years just for production on rainfed agri.
df_prod_wy <- df_prod |>
    dplyr$filter(
        stringr$str_detect(metric, "Prod"),
        stringr$str_detect(name, "Rainfed")
    ) |>
    dplyr$arrange(
        value
    ) |>
    dplyr$slice(1:3)

#####################################
#### CREATE SOME VALIDATION SETS ####
#####################################

# seasonal forecasts
df_val <- dplyr$tibble(
    year = 1984:2024,
    vhi = year %in% df_vhi_wy$year,
    asi = year %in% df_asi_wy$year,
    emdat = year %in% df_emdat_wy$year,
    asi_2000 = year %in% df_asi_wy_2000$year
)

#################################
#### EXPLORATION OF TRIGGERS ####
#################################

# Seasonal forecasts
df_seas5_perf <- df_seas5_wy |>
    dplyr$mutate(
        worst_year = TRUE
    ) |>
    tidyr$complete(
        year = 1984:2024,
        pub_month = c(11, 12, 1:3),
        fill = list(
            indicator = "SEAS5",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(
        pub_month
    ) |>
    calc_metrics(
        x = worst_year,
        y = asi
    )

# df_smi$year |> range()
# df_smi_perf <- df_smi_wy |>
#   dplyr$mutate(
#     worst_year = TRUE
#   ) |>
#   tidyr$complete(
#     year = 1984:2024,
#     month = 1:5,
#     fill = list(
#       indicator = "SMI",
#       worst_year = FALSE
#     )
#   ) |>
#   dplyr$left_join(
#     df_val,
#     by = "year"
#   ) |>
#   dplyr$group_by(
#     month
#   ) |>
#   calc_metrics(
#     x = worst_year,
#     y = asi
#   )


# CHIRPS

df_chirps_perf <- df_chirps_wy |>
    dplyr$mutate(
        worst_year = TRUE
    ) |>
    tidyr$complete(
        year = 1984:2024,
        month = 1:5,
        fill = list(
            indicator = "CHIRPS",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(
        month
    ) |>
    calc_metrics(
        x = worst_year,
        y = asi
    )

# performance just looking at MAM
df_chirps_mam_perf <- df_chirps_mam_wy |>
    dplyr$mutate(
        worst_year = TRUE
    ) |>
    tidyr$complete(
        year = 1984:2024,
        month = 1:5,
        fill = list(
            indicator = "CHIRPS",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(
        month
    ) |>
    calc_metrics(
        x = worst_year,
        y = asi
    )

# CHIRPS (from 2000)
df_chirps_from_2000_mam_perf <- df_chirps_wy_2000 |>
    dplyr$mutate(
        worst_year = TRUE
    ) |>
    tidyr$complete(
        year = 2000:2024,
        month = 1:5,
        fill = list(
            indicator = "CHIRPS",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(
        month
    ) |>
    calc_metrics(
        x = worst_year,
        y = asi_2000
    )


# SNOW

df_snow_perf <- df_snow_wy |>
    dplyr$mutate(
        worst_year = TRUE
    ) |>
    tidyr$complete(
        year = 2000:2024,
        month = c(11:12, 1:3),
        fill = list(
            indicator = "NDSI",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(
        month
    ) |>
    calc_metrics(
        x = worst_year,
        y = asi_2000
    )

# ASI

df_asi_perf <- df_asi_dek_wy |>
    dplyr$mutate(
        worst_year = TRUE
    ) |>
    tidyr$complete(
        year = 1984:2024,
        month = 1:5,
        dekad = 1:3,
        fill = list(
            indicator = "ASI",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(
        month,
        dekad
    ) |>
    calc_metrics(
        x = worst_year,
        y = asi
    )

# Combo
df_combo_perf <- df_combo_wy |>
    dplyr$ungroup() |>
    dplyr$mutate(
        worst_year = TRUE,
        indicator = "CHIRPS\n& NDSI"
    ) |>
    tidyr$complete(
        year = 2000:2024,
        month = 1:5,
        fill = list(
            indicator = "CHIRPS\n& NDSI",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(month) |>
    calc_metrics(
        x = worst_year,
        y = asi_2000
    )

# NDVI

df_ndvi_perf <- df_ndvi_wy |>
    dplyr$mutate(
        worst_year = TRUE
    ) |>
    tidyr$complete(
        year = 1984:2024,
        month = 3:5,
        dekad = 1:3,
        fill = list(
            indicator = "NDVI",
            worst_year = FALSE
        )
    ) |>
    dplyr$left_join(
        df_val,
        by = "year"
    ) |>
    dplyr$group_by(
        month,
        dekad
    ) |>
    calc_metrics(
        x = worst_year,
        y = asi
    )

4.2 Rainfall Timeline

df_chirps |>
    dplyr$mutate(
        month = factor(months(date, abbreviate = TRUE), levels = month.abb)
    ) |>
    gg$ggplot(
        mapping = gg$aes(x = month, y = value)
    ) +
    gg$geom_boxplot() +
    gg$labs(
        x = "",
        y = "Precipitation (mm)",
        title = "Distribution of monthly rainfall, 1984 to 2024, CHIRPS"
    )

4.3 Joint Performance

###########################
#### JOINT PERFORMANCE ####
###########################


dplyr$bind_rows(
    df_chirps_perf,
    df_seas5_perf |> dplyr$rename(month = pub_month),
    df_ndvi_perf |> dplyr$filter(dekad == 3) |> dplyr$select(-dekad), # last dekad best performance
    df_snow_perf,
    df_combo_perf,
    df_asi_perf |> dplyr$filter(dekad == 3) |> dplyr$select(-dekad), # last dekad
) |>
    dplyr$mutate(
        month_label = factor(month.name[month], levels = month.name[c(11:12, 1:5)])
    ) |>
    gg$ggplot(
        mapping = gg$aes(
            x = indicator,
            y = forcats$fct_rev(month_label),
            fill = precision
        )
    ) +
    gg$geom_tile(
        linewidth = 2,
        color = "white"
    ) +
    gg$geom_text(
        mapping = gg$aes(
            label = scales$label_percent(accuracy = 1)(precision)
        ),
        check_overlap = TRUE,
        fontface = "bold",
        color = "white"
    ) +
    gg$labs(
        y = "",
        x = "",
        fill = "Precision/Recall",
        title = "How well can we predict worst years",
        subtitle = "Measured against ASI end-of-season (May)",
        caption = "Methodology compares how well a 1 in 5 year RP level threshold correlates\nto the the worst to 1 in 5 year level drought at the end of the spring wheat season as measured by ASIS"
    ) +
    gg$scale_fill_gradient(
        low = gghdx$hdx_hex("sapphire-ultra-light"),
        high = gghdx$hdx_hex("sapphire-dark"),
        labels = scales$label_percent(accuracy = 1),
        breaks = scales$trans_breaks(identity, identity, n = 3)
    ) +
    gg$theme(
        axis.text.x = gg$element_text(vjust = 1),
        axis.line.x = gg$element_blank(),
        panel.grid = gg$element_blank()
    )

4.4 Next Steps

This is only an initial exploration of when we could potential activate, but needs to be guided by the crisis timeline and proposed operations.

1. Fill out crisis timeline, proposed activities, and clear framework of when triggers would need to be available

2. Compare this to trigger performance and proposal to decide when to trigger (e.g. look at overall activation rates, activation points, thresholds)

3. Design monitoring framework, ideally performed by OCHA Afghanistan with support from the Centre for Humanitarian Data

Johnson, Stephanie J, Timothy N Stockdale, Laura Ferranti, Magdalena A Balmaseda, Franco Molteni, Linus Magnusson, Steffen Tietsche, et al. 2019. “SEAS5: The New ECMWF Seasonal Forecast System.” Geoscientific Model Development 12 (3): 1087–117.