Uganda: SEAS5 drought forecast × humanitarian caseload

NoteWhy this document exists

Uganda has no HNRP. Its only response plan is the refugee RRP, which the Forecast × HNRP pipeline excludes as not the country’s own plan, so Uganda does not appear in the app. This standalone analysis borrows from and adapts the Forecast × HNRP tab’s methodology for Uganda, with supplementary data from the country team:

• PiN and severity: Uganda JIAF 2.0 Light - Data Collection (v5 census baseline), sheets 1. Severity and 2. PiN (22 districts × host community / refugee population groups; the raw sheets list 26 “District” values, four of which are subtotal rows the loader drops).
• Targeted and reached: Gender-MAX Targeted / Achieved People (2026-08-19), district-level maxima with gender and disability breakdowns.
• Climate: the same SEAS5 / ERA5 sources the app uses. The region level comes straight from the rasterstats DB. The district level was computed for this analysis (Uganda is capped at max_adm_level=1 in the rasterstats registry), with exactextract over the same COGs; the method reproduces the DB’s region means within about 1%, and the results run through the repo’s unchanged skill and return-period code.

It reads in four steps: where the season stands (observed), where it is heading (the forecast), what that timing means for crops and pasture, and who is exposed, ending in a priority readout. Flooding runs on different timing from the drought, so the flood outlook for the wet Oct-Dec season has its own section at the end.

import calendar
import math
import os
import sys
from pathlib import Path

os.environ["TQDM_DISABLE"] = "1"   # tqdm widgets render as opaque bars in HTML
# DuckDB (used by the EM-DAT loader) pops a Jupyter progress widget that
# renders as an opaque bar in the HTML; disable it on every new connection.
import duckdb
_duckdb_connect = duckdb.connect
def _quiet_connect(*args, **kwargs):
    con = _duckdb_connect(*args, **kwargs)
    try:
        con.execute("SET enable_progress_bar = false")
    except duckdb.Error:
        pass
    return con
duckdb.connect = _quiet_connect

import geopandas as gpd
import matplotlib.pyplot as plt
import numpy as np
import ocha_stratus as stratus
import pandas as pd
from matplotlib.colors import ListedColormap
from matplotlib.patches import Patch

REPO = Path.cwd()
if not (REPO / "src").exists():          # rendered from analysis/
    REPO = REPO.parent
sys.path.insert(0, str(REPO))
sys.path.insert(0, str(REPO / "pipeline"))

from src.constants import PROJECT_PREFIX, TRIMESTERS
from src.skill import trimester_lead
from src.datasources import uga_country_team as uct
from export_static_site import (THRESHOLDS, _actual_issued_year, _tri_valid,
                                _tri_label, compute_rainy_set)
from ocha_stratus import codab

UGA_DIR = f"{PROJECT_PREFIX}/processed/uga"

# The app's own palette, so figures read like the Forecast × HNRP tab.
C_DROUGHT_VSEV = "#7f5619"   # strongly below normal (RP >= vsev_rp)
C_DROUGHT_SEV  = "#dda555"   # below normal        (RP >= sev_rp)
C_MUTED        = "#e9eeee"   # no qualifying drought signal
C_EDGE         = "#c4d0d1"
C_PIN, C_TGT, C_REA = "#009EDB", "#F58220", "#D15353"
JIAF_COLORS = ["#e9f2fb", "#d4e5f7", "#82b5e9", "#418fde", "#1f69b3"]

cod1 = codab.load_codab_from_blob("uga", admin_level=1)
cod2 = codab.load_codab_from_blob("uga", admin_level=2)

def worst_drought_slot(skill: pd.DataFrame, rainy_set: set) -> pd.DataFrame:
    """Port of export_hnrp_drought.py's per-unit selection: at the latest
    issuance, among valid (lead -2..4) rainy-season trimesters with skill
    r >= r_mod and a dry-side forecast (percentile < 50), keep the trimester
    with the highest drought return period."""
    # Latest issuance, the app's way: max actual issued year, then max month in it.
    iy = skill[skill["current_forecast_year"].notna()].copy()
    iy["_iy"] = iy.apply(_actual_issued_year, axis=1)
    by_month = iy.groupby("issued_month")["_iy"].max().astype(int)
    issued_year = int(by_month.max())
    issued_month = int(by_month[by_month == issued_year].index.max())
    # Filter on the actual issued YEAR too, never the month alone: if the skill
    # pipeline ran before the current in-season ERA5 month reached the DB, the
    # in-season trimesters at this issued_month still carry LAST year's forecast.
    sub = iy[(iy["issued_month"] == issued_month) & (iy["_iy"] == issued_year)].copy()
    sub = sub[sub.apply(lambda r: _tri_valid(TRIMESTERS[r["trimester"]], issued_month), axis=1)]

    def _mm(r):
        days = sum(calendar.monthrange(2025, m)[1] for m in TRIMESTERS[r["trimester"]])
        fc = round(math.expm1(float(r["current_forecast_mean"])) * days) if pd.notna(r["current_forecast_mean"]) else None
        nrm = round(math.expm1(float(r["era5_mean"])) * days) if pd.notna(r["era5_mean"]) else None
        return fc, nrm

    rows = []
    for pcode, g in sub.groupby("pcode"):
        best, why = None, set()
        for _, r in g.iterrows():
            pct, pr = r["forecast_percentile"], r["pearson_r"]
            if pd.isna(pct) or pd.isna(pr):
                continue
            rainy = (pcode, r["trimester"]) in rainy_set
            if not rainy:
                why.add("off-season"); continue
            if pr < THRESHOLDS["r_mod"]:
                why.add("low skill"); continue
            if pct >= 50:
                why.add("wet/normal"); continue
            if pd.isna(r["forecast_rp"]):
                continue
            if best is None or r["forecast_rp"] > best["drought_rp"]:
                fc_mm, nrm_mm = _mm(r)
                best = dict(trimester=r["trimester"], tri_label=_tri_label(TRIMESTERS[r["trimester"]]),
                            lead=trimester_lead(issued_month, TRIMESTERS[r["trimester"]]),
                            drought_rp=float(r["forecast_rp"]), pct=float(pct), r=float(pr),
                            fc_mm=fc_mm, nrm_mm=nrm_mm)
        row = dict(pcode=pcode, issued_month=issued_month, issued_year=issued_year)
        if best:
            row |= best
            row["cat"] = "strongly below normal" if best["drought_rp"] >= THRESHOLDS["vsev_rp"] \
                else ("below normal" if best["drought_rp"] >= THRESHOLDS["sev_rp"] else "mild dry signal")
        else:
            # no qualifying slot: say why (dominant reason across trimesters)
            row["cat"] = ("no rainy trimester in window" if why == {"off-season"}
                          else " / ".join(sorted(why)) if why else "no data")
        rows.append(row)
    return pd.DataFrame(rows)


def rainy_from_era5(era5: pd.DataFrame) -> set:
    mc = (era5.assign(month=era5["valid_date"].dt.month)
              .groupby(["pcode", "month"])["mean"].mean()
              .reset_index().rename(columns={"mean": "mean_mm_day"}))
    return compute_rainy_set(mc)


# Region level (adm1): straight from the rasterstats DB + app skill parquet.
skill1 = stratus.load_parquet_from_blob(
    f"{PROJECT_PREFIX}/processed/skill_stats_detrended_adm1.parquet", stage="dev")
skill1 = skill1[skill1["iso3"] == "UGA"]
engine = stratus.get_engine(stage="prod")
era5_1 = pd.read_sql(
    "SELECT pcode, valid_date, mean FROM public.era5 WHERE iso3='UGA' AND adm_level=1",
    engine, parse_dates=["valid_date"])
fc1 = worst_drought_slot(skill1, rainy_from_era5(era5_1))

# District level (adm2): computed for this analysis (see callout above).
skill2 = stratus.load_parquet_from_blob(f"{UGA_DIR}/skill_stats_detrended_adm2.parquet", stage="dev")
era5_2 = stratus.load_parquet_from_blob(f"{UGA_DIR}/era5_adm2.parquet", stage="dev")
era5_2["valid_date"] = pd.to_datetime(era5_2["valid_date"])
fc2 = worst_drought_slot(skill2, rainy_from_era5(era5_2))

issued = int(fc2["issued_month"].iloc[0])
issued_yr = int(fc2["issued_year"].iloc[0])
if (issued, issued_yr) != (int(fc1["issued_month"].iloc[0]), int(fc1["issued_year"].iloc[0])):
    raise ValueError(
        f"adm1 and adm2 skill parquets disagree on the latest issuance: "
        f"adm2 {calendar.month_name[issued]} {issued_yr} vs adm1 "
        f"{calendar.month_name[int(fc1['issued_month'].iloc[0])]} {int(fc1['issued_year'].iloc[0])} "
        f"— re-run the stale pipeline before rendering")
print(f"Latest issuance: {calendar.month_name[issued]} {issued_yr} | "
      f"{len(fc2)} districts, {len(fc1)} regions")

Latest issuance: August 2026 | 135 districts, 4 regions

Where the season stands

The place to start is what has actually fallen. The curves below accumulate rainfall over each zone’s own crop-calendar window (planting start to harvest end, from the JRC ASAP calendars introduced in the seasonal-relevance section; Karamoja’s Apr-Sep wet season from the ERA5 climatology) and compare 2026 against every year since 1981. Karamoja and West Nile sit at opposite ends of the range. Karamoja has received 41-54% of its normal season-to-date rain, the driest or near-driest Apr-Jul on record in its districts, with the remaining season forecast at the 0th percentile. West Nile has received 84-99% of normal: the wet March banked most of the season’s water early, so in accumulation terms its main season is close to normal and the record-dry June-July is a harvest-timing concern rather than a water-quantity failure. (Do not confuse this observed season-to-date figure with the forecast deficit map in the next section, which is the forecast for the worst upcoming slot.) The bimodal zones sit in between at roughly 74-100%. Throughout this section “normal” is the 1981-2025 mean season-to-date.

cal = stratus.load_parquet_from_blob(f"{UGA_DIR}/asap_crop_calendar.parquet", stage="dev")
dz = stratus.load_parquet_from_blob(f"{UGA_DIR}/asap_district_zones.parquet", stage="dev")
zones = ["Karamoja", "West Nile", "Mid Northen", "Eastern", "Lake Albert Crescent",
         "Western Highlands", "Southern Drylands", "Lake Victoria Crescent",
         "South Eastern", "Southern Highlands"]
jiaf_pcodes = set(uct.load_jiaf_pin()["pcode"])

dk2mo = lambda d: int(np.ceil(d / 3))
main_cal = cal[cal["crop_name"].fillna("").str.contains("Maize")
               & ~cal["crop_name"].fillna("").str.contains("Second|North")]
season_win = {r["zone"]: (dk2mo(r["sos_s"]), dk2mo(r["eos_e"])) for _, r in main_cal.iterrows()}
season_win["Karamoja"] = (4, 9)   # wet season from ERA5 climatology; no ASAP calendar

e_all = era5_2.merge(dz[["pcode", "zone", "district"]], on="pcode")
e_all["yr"], e_all["mo"] = e_all["valid_date"].dt.year, e_all["valid_date"].dt.month
e_all["mm"] = e_all["mean"] * e_all["mo"].map(lambda m: calendar.monthrange(2025, m)[1])
LATEST_OBS = int(e_all.loc[e_all['yr'] == 2026, 'mo'].max())

fig, axes = plt.subplots(2, 5, figsize=(9.5, 4.8))
for ax, zone in zip(axes.flat, zones):
    m0, m1 = season_win[zone]
    mos = list(range(m0, m1 + 1))
    zm2 = (e_all[(e_all["zone"] == zone) & (e_all["mo"].isin(mos))]
           .groupby(["yr", "mo"])["mm"].mean().unstack().reindex(columns=mos))
    cum = zm2.cumsum(axis=1)
    hist = cum.loc[cum.index < 2026]
    for _, row in hist.iterrows():
        ax.plot(mos, row.values, color="#c4d0d1", lw=0.6, alpha=0.6)
    ax.plot(mos, hist.mean(), color="#3f4748", lw=1.4, ls="--")
    upto = [m for m in mos if m <= LATEST_OBS]
    if 2026 in cum.index and upto:
        ax.plot(upto, cum.loc[2026, upto], color="#7f5619", lw=2.2)
        # mean basis, matching the district ranking below (was median: the two
        # figures silently used different definitions of "normal")
        pct = 100 * cum.loc[2026, upto[-1]] / hist[upto[-1]].mean()
        ax.set_title(f"{zone}\n{pct:.0f}% of normal to date", fontsize=8)
    ax.set_xticks([mos[0], mos[-1]], [calendar.month_abbr[mos[0]], calendar.month_abbr[mos[-1]]], fontsize=7)
    ax.tick_params(axis='y', labelsize=7)
    ax.spines[["top", "right"]].set_visible(False)
fig.suptitle("Cumulative main-season rainfall (mm): 2026 against 1981-2025", y=1.0)
plt.tight_layout()
plt.show()

Cumulative rainfall over each zone’s main-season window, zone mean of district series. Gray: every year 1981-2025; dashed: the 1981-2025 mean; brown: 2026 through July (the ERA5 record’s edge). The percentage is 2026 season-to-date as a share of the normal (1981-2025 mean) season-to-date — the same basis as the district ranking below.

std_rows = []
for (pc, dist, zone), gg in e_all[e_all["pcode"].isin(jiaf_pcodes)].groupby(["pcode", "district", "zone"]):
    m0, m1 = season_win[zone]
    mos = [m for m in range(m0, min(m1, LATEST_OBS) + 1)]
    std = gg[gg["mo"].isin(mos)].groupby("yr")["mm"].sum()
    hist = std[std.index < 2026]
    std_rows.append({"district": dist, "zone": zone, "pct": 100 * std[2026] / hist.mean(),
                     "driest": bool((hist < std[2026]).sum() == 0)})
sd = pd.DataFrame(std_rows)
sd["group"] = np.select([sd["zone"] == "Karamoja", sd["zone"] == "West Nile"],
                        ["Karamoja", "West Nile corridor"], "Bimodal south / west")
GC = {"Karamoja": "#7f5619", "West Nile corridor": "#e69f00", "Bimodal south / west": "#56b4e9"}
sd = sd.sort_values("pct").reset_index(drop=True)
fig, ax = plt.subplots(figsize=(9, 0.32 * len(sd) + 1.2))
for i, r in sd.iterrows():
    ax.plot(r["pct"], i, marker="o", ms=9, color=GC[r["group"]],
            markerfacecolor=GC[r["group"]] if r["driest"] else "white", markeredgewidth=1.6)
ax.set_yticks(range(len(sd)), sd["district"], fontsize=8.5)
ax.axvline(100, color="#9db1b3", lw=1, ls="--")
ax.text(100, len(sd) - 0.2, " normal", fontsize=8, color="#5e6a6b")
ax.set_xlabel("observed season-to-date rainfall, % of normal")
ax.xaxis.set_major_formatter(lambda v, _: f"{v:.0f}%")
ax.legend(handles=[plt.Line2D([], [], marker="o", ls="", color=c, label=g) for g, c in GC.items()]
                  + [plt.Line2D([], [], marker="o", ls="", color="#5e6a6b", markerfacecolor="white", label="open = not record-driest")],
          loc="lower right", frameon=False, fontsize=8)
ax.grid(axis="x", color="#eef1f1", zorder=0)
ax.spines[["top", "right"]].set_visible(False)
ax.set_title("Season-to-date rainfall, JIAF districts")
plt.tight_layout()
plt.show()

Observed season-to-date rainfall as % of the normal (1981-2025 mean) season-to-date, per JIAF district over its zone’s calendar window. Filled markers are districts having their driest season-to-date since 1981.

Where it is heading: the forecast

The table and map below show, per unit, the worst qualifying drought slot at the latest SEAS5 issuance, using the app’s rule: valid trimesters (lead −2…4), rainy-season only, skill \(r \ge 0.30\) (detrended), dry-side forecasts only (percentile < 50), severity = the drought return period. The RP says how unusual the forecast is, never how much water is missing, so each RP is paired with the forecast and climatological-normal seasonal totals (mm).

t = fc1.merge(cod1[["ADM1_PCODE", "ADM1_EN"]], left_on="pcode", right_on="ADM1_PCODE")
t = t[["ADM1_EN", "cat", "tri_label", "lead", "drought_rp", "pct", "r", "fc_mm", "nrm_mm"]]
t.columns = ["Region", "Category", "Trimester", "Lead", "Drought RP (yr)",
             "Percentile", "Skill r", "Forecast (mm)", "Normal (mm)"]
t.sort_values("Drought RP (yr)", ascending=False).style.hide(axis="index").format(
    {"Drought RP (yr)": "{:.1f}", "Percentile": "{:.1f}", "Skill r": "{:.2f}"}, na_rep="—")

Region	Category	Trimester	Lead	Drought RP (yr)	Percentile	Skill r	Forecast (mm)	Normal (mm)
Central	strongly below normal	Jul–Aug–Sep	-1	46.0	0.0	0.75	201	351
Eastern	strongly below normal	Jun–Jul–Aug	-2	46.0	0.0	0.93	270	456
Northern	strongly below normal	Jun–Jul–Aug	-2	46.0	0.0	0.95	192	422
Western	strongly below normal	Jul–Aug–Sep	-1	46.0	0.0	0.74	200	351

g = cod2.merge(fc2, left_on="ADM2_PCODE", right_on="pcode", how="left")
jiaf_pcodes = set(uct.load_jiaf_pin()["pcode"])

def cat_color(row):
    if row["cat"] == "strongly below normal":
        return C_DROUGHT_VSEV
    if row["cat"] == "below normal":
        return C_DROUGHT_SEV
    if row["cat"] == "mild dry signal":
        return "#f0e3cd"
    return C_MUTED

fig, ax = plt.subplots(figsize=(9, 8))
g["color"] = g.apply(cat_color, axis=1)
g.plot(ax=ax, color=g["color"], edgecolor=C_EDGE, linewidth=0.5)
cod1.boundary.plot(ax=ax, color="#9db1b3", linewidth=1.2)
g[g["ADM2_PCODE"].isin(jiaf_pcodes)].boundary.plot(ax=ax, color="#1d2021", linewidth=1.1)
ax.set_axis_off()
ax.set_title(f"SEAS5 drought signal by district, issued {calendar.month_name[issued]} {issued_yr}")
ax.legend(handles=[
    Patch(facecolor=C_DROUGHT_VSEV, edgecolor=C_EDGE, label=f"strongly below normal (RP ≥ {THRESHOLDS['vsev_rp']} yr)"),
    Patch(facecolor=C_DROUGHT_SEV, edgecolor=C_EDGE, label=f"below normal (RP ≥ {THRESHOLDS['sev_rp']} yr)"),
    Patch(facecolor="#f0e3cd", edgecolor=C_EDGE, label="mild dry signal (RP < 3 yr)"),
    Patch(facecolor=C_MUTED, edgecolor=C_EDGE, label="no qualifying drought signal"),
    Patch(facecolor="none", edgecolor="#1d2021", label="JIAF 2.0 Light district (severity & PiN assessed)"),
], loc="lower left", frameon=False, fontsize=9)
plt.tight_layout()
plt.show()

Worst qualifying drought slot per district, latest SEAS5 issuance, in the app’s category palette. The black outline marks the 22 districts assessed in the JIAF 2.0 Light exercise (severity and PiN); other districts are climate context only.

ImportantDry now, wet later

The drought signal is concentrated in June to September 2026. Monthly percentiles below are vs 1981-2025, computed as the mean of the district series (seas5_adm2 / era5_adm2) over the named unit. June was the driest June in the 45-year record at national, Northern-region, and Karamoja aggregation alike (observed); July sat at the 13th percentile nationally (Karamoja: the 0th). The August-issued forecast puts August at the 0th percentile on all three aggregations, and September at the 0th nationally and over Karamoja (Northern region: the 2nd). From October onward the same issuance forecasts wetter than normal (Northern region: Oct 93rd / Nov 96th / Dec 98th percentile), so a “wetter than normal OND” reading elsewhere in the app is correct and does not contradict this document. The Forecast × HNRP methodology is drought-side only and reports each unit’s worst qualifying slot, which lands on the Jun-Sep window. The table below gives every valid trimester so both sides are visible.

Skill: every reported slot passed the detrended \(r \ge 0.30\) gate. The Jun-Jul-Aug slots show \(r\) of 0.92 to 0.96 because they are in-season at an August issuance: June and July are already observed, so two-thirds of that signal is monitoring, not forecast. The fully-forecast Aug-Sep-Oct slots carry genuine forecast skill (\(r\) 0.34 to 0.62, median 0.57), and each table row reports its own \(r\).

Crop relevance: the Jun-Aug window carries 38-44% of total annual rainfall in every Karamoja district (ERA5 climatology, this analysis). What that means for production differs by agro-ecological zone; see Seasonal relevance below, built from the JRC ASAP crop calendars.

# Same vintage guard as worst_drought_slot: month-only selection would render
# last year's in-season rows whenever the adm1 pipeline ran before the current
# ERA5 month landed (this table once showed Aug 2025's JJA/JAS as "issued 2026").
iy1 = skill1[skill1["current_forecast_year"].notna()].copy()
iy1["_iy"] = iy1.apply(_actual_issued_year, axis=1)
iy1 = iy1[(iy1["issued_month"] == issued) & (iy1["_iy"] == issued_yr)]
iy1 = iy1[iy1.apply(lambda r: _tri_valid(TRIMESTERS[r["trimester"]], issued), axis=1)]
iy1["region"] = iy1["pcode"].map(dict(zip(cod1["ADM1_PCODE"], cod1["ADM1_EN"])))
iy1["lead"] = iy1["trimester"].map(lambda t: trimester_lead(issued, TRIMESTERS[t]))
order = iy1.drop_duplicates("trimester").sort_values("lead")
piv = iy1.pivot_table(index="region", columns="trimester", values="forecast_percentile")
piv = piv[order["trimester"]]
piv.columns = [f"{_tri_label(TRIMESTERS[t])} (lead {l})" for t, l in zip(order["trimester"], order["lead"])]
(piv.style.format("{:.0f}")
    .background_gradient(cmap="BrBG", vmin=0, vmax=100, axis=None)
    .set_caption("Forecast percentile vs 1981-2025 for each valid trimester, "
                 f"issued {calendar.month_name[issued]} {issued_yr} (brown = dry, green = wet). "
                 "In-season trimesters (negative leads) mix observed months with the forecast."))

Table 1: Forecast percentile vs 1981-2025 for each valid trimester, issued August 2026 (brown = dry, green = wet). In-season trimesters (negative leads) mix observed months with the forecast.

	Jun–Jul–Aug (lead -2)	Jul–Aug–Sep (lead -1)	Aug–Sep–Oct (lead 0)	Sep–Oct–Nov (lead 1)	Oct–Nov–Dec (lead 2)	Nov–Dec–Jan (lead 3)	Dec–Jan–Feb (lead 4)
region
Central	2	0	9	71	96	89	96
Eastern	0	0	7	60	93	89	96
Northern	0	0	0	24	89	93	96
Western	4	0	0	13	89	89	96

WarningNearly every district is at the maximum return period

The empirical RP is capped at 46 years (the 45-year record + 1), and at this issuance almost all districts hit that cap: the forecast is drier than anything in the historical record across nearly the whole country, so the RP map cannot differentiate within Uganda. The RP says how unusual the season is. The map below shows how deep the deficit is (the forecast seasonal total as a share of the climatological normal), which is where districts differ.

gd = g[g["fc_mm"].notna() & g["nrm_mm"].notna()].copy()
gd["pct_normal"] = 100 * gd["fc_mm"] / gd["nrm_mm"]

fig, ax = plt.subplots(figsize=(9, 8))
g.plot(ax=ax, color=C_MUTED, edgecolor=C_EDGE, linewidth=0.5)
gd.plot(ax=ax, column="pct_normal", cmap="YlOrBr_r", vmin=0, vmax=100,
        edgecolor=C_EDGE, linewidth=0.5,
        legend=True, legend_kwds={"label": "forecast, % of normal seasonal total",
                                  "shrink": 0.6})
cod1.boundary.plot(ax=ax, color="#9db1b3", linewidth=1.2)
g[g["ADM2_PCODE"].isin(jiaf_pcodes)].boundary.plot(ax=ax, color="#1d2021", linewidth=1.1)
ax.set_axis_off()
ax.set_title(f"Forecast rainfall deficit by district, issued {calendar.month_name[issued]} 2026")
plt.tight_layout()
plt.show()

Forecast seasonal total as % of the climatological normal, for each district’s worst qualifying slot. Darker = deeper deficit.

The same issuance also forecasts a strongly wet Oct-Dec; the flood side of that story has its own section at the end of this document, so the drought reading here stays in one piece.

Seasonal relevance: crops, pasture, and timing

Cropping seasons come from the JRC ASAP sub-national crop calendars (planting, growth and harvest windows per agro-ecological zone, in 10-day dekads), with each CODAB district assigned to its zone (pipeline/build_uga_crop_calendar.py). ASAP calendars only maize, beans and millet: Uganda’s root and perennial crops (cassava, banana) and Karamoja’s sorghum are not calendared, and perennials buffer a rainfall shock differently from cereals. The zones fall into three different situations:

• Karamoja is pastoral, not a cropping zone. ASAP publishes no crop calendar for it: 91% of the zone is rangeland and only 4% cropland. The record-dry Jun-Aug, its rainfall peak, is first a pasture, water and livestock crisis, hitting the sparse sorghum patches on top. An unusually wet Oct-Dec can recharge water points and regrow pasture for herds; for the cropped patches it comes after the zone’s single season has already failed, and nothing replanted grows before the next rains around April.
• West Nile is unimodal in ASAP (no second-season crops): planting March to April, harvest June to July. Stage timing matters here: planting rains were wet (March at the 91st percentile) and growth was mixed (April 22nd, May 47th), so the record-dry June hit maturation and harvest rather than grain-fill, and the season-to-date accumulation is 84-99% of normal (see “Where the season stands” above). That points to a season close to normal in water terms, with the risk concentrated in harvest conditions and late-planted fields. What makes it a priority anyway is what comes after: there is no second planting until March 2027, so whatever the season yielded has to carry the refugee-hosting districts (Yumbe, Adjumani, Madi-Okollo & Terego, Obongi, Koboko, Arua/Terego, Moyo) through the whole dry stretch, and the wet OND forecast offers no cropping recovery there.
• The bimodal zones (Mid Northern incl. Lamwo; Lake Albert Crescent incl. Kiryandongo/Kikuube; Western Highlands incl. Kamwenge/Kyegegwa; Southern Drylands incl. Isingiro; Lake Victoria Crescent incl. Kampala): the dry window hits main-season harvest (Jun-Aug), and second-season planting (Aug-Oct) risks poor establishment in the record-dry Aug-Sep, but the strongly wet Oct-Dec then favors whatever gets established. A delayed second-season start with good follow-on rains is the base case there.

def dk(d):  # dekad (1-36) -> month-axis position (1.0 = Jan 1, 13.0 = Dec 31)
    return 1 + (d - 1) / 3

maize = cal[cal["crop_name"].str.contains("Maize", na=False)].copy()
maize["season"] = np.select(
    [maize["crop_name"].str.contains("Second"), maize["crop_name"].str.contains("North")],
    ["second", "north"], "main")

C_PLANT, C_GROW, C_HARV = "#9ecae1", "#a1d99b", "#e6b34c"
fig, ax = plt.subplots(figsize=(9, 5.5))
for i, z in enumerate(zones):
    y = len(zones) - 1 - i
    rows = maize[maize["zone"] == z]
    if rows.empty:
        ax.text(6.9, y, "pastoral: rangeland 91%, no ASAP crop calendar",
                va="center", fontsize=8.5, color="#5b6b6d", style="italic")
        continue
    for _, r in rows.iterrows():
        off = {"main": 0.22, "north": 0.0, "second": -0.22}[r["season"]]
        segs = [(dk(r["sos_s"]), dk(r["sos_e"]), C_PLANT),
                (dk(r["sos_e"]), dk(r["eos_s"]) if r["eos_s"] > r["sos_e"] else 13, C_GROW)]
        if r["eos_s"] > r["sos_e"]:                     # same calendar year
            segs.append((dk(r["eos_s"]), dk(r["eos_e"]), C_HARV))
        else:                                           # wraps into January
            segs += [(1, dk(r["eos_s"]), C_GROW), (dk(r["eos_s"]), dk(r["eos_e"]), C_HARV)]
        for x0, x1, c in segs:
            ax.barh(y + off, x1 - x0, left=x0, height=0.24, color=c, edgecolor="none")
ax.axvspan(6, 10, color="#7f5619", alpha=0.14, zorder=0)
ax.axvspan(10, 13, color="#2a7e43", alpha=0.10, zorder=0)
ax.text(8.0, len(zones) - 0.35, "record dry\nJun-Sep", ha="center", fontsize=8.5, color="#7f5619")
ax.text(11.5, len(zones) - 0.35, "wet forecast\nOct-Dec", ha="center", fontsize=8.5, color="#2a7e43")
ax.set_yticks(range(len(zones)), zones[::-1])
ax.set_xticks(np.arange(1, 13), [calendar.month_abbr[m] for m in range(1, 13)])
ax.set_xlim(1, 13); ax.set_ylim(-0.6, len(zones))
ax.legend(handles=[Patch(color=C_PLANT, label="planting"), Patch(color=C_GROW, label="growth"),
                   Patch(color=C_HARV, label="harvest")],
          loc="upper center", bbox_to_anchor=(0.5, -0.09), ncol=3, frameon=False)
ax.spines[["top", "right"]].set_visible(False)
ax.set_title("Maize seasons by zone (top: main; middle: northern long-cycle; bottom: second season)")
plt.tight_layout()
plt.show()

Maize seasons per ASAP agro-ecological zone. ASAP calendars Uganda for maize, beans and millet; beans and millet share maize’s windows everywhere except Eastern, where millet’s harvest runs about a month longer. Eastern’s long-cycle northern maize is the middle bar. Brown band: the record-dry Jun-Sep window; green band: the wetter-than-normal Oct-Dec forecast. Karamoja has no ASAP crop calendar; it is 91% rangeland.

Did the drought hit when it matters for crops?

Most zones harvest from June or July, so much of the main season’s growth happened under different rain than the record-dry window this document headlines. The table below puts the two side by side: the observed 2026 rainfall percentile per month, by zone. Planting (March) was wet everywhere; April was below median everywhere; June collapsed to the driest on record everywhere; July recovered in the west but not in the north or east. Read against the calendar above, that means the growth-stage failure is real in Karamoja (June-July is the core of its single wet season) and for the Eastern zone’s long-cycle maize (grain-fill through June-July at the 0th-4th percentile), while West Nile’s crop mostly grew under near-normal rain and met the record dry at harvest, which damages less. Note also the shape of the Karamoja collapse: a very wet February-March (89th-98th percentile) would have encouraged early planting and an early pasture flush before April slid below median and June-July failed outright, a false-start season, which tends to cost farmers and herders more than one that is poor from the beginning. The forward-looking risk (the Aug-Sep forecast) falls on second-season establishment in the bimodal zones and on Karamoja’s season tail.

e2z = era5_2.merge(dz[["pcode", "zone"]], on="pcode")
zm = (e2z.groupby(["zone", e2z["valid_date"].dt.year.rename("yr"),
                   e2z["valid_date"].dt.month.rename("mo")])["mean"]
          .mean().reset_index())
obs_rows = []
for zone, gg in zm.groupby("zone"):
    row = {"zone": zone}
    for mo in range(1, 8):
        cur = gg[(gg["yr"] == 2026) & (gg["mo"] == mo)]["mean"]
        hist = gg[(gg["yr"] < 2026) & (gg["mo"] == mo)]["mean"]
        row[calendar.month_abbr[mo]] = 100 * (hist < cur.iloc[0]).mean()
    obs_rows.append(row)
obs = pd.DataFrame(obs_rows).set_index("zone").sort_index()
(obs.style.format("{:.0f}")
    .background_gradient(cmap="BrBG", vmin=0, vmax=100, axis=None)
    .set_caption("Observed 2026 monthly rainfall percentile vs 1981-2025 "
                 "(ERA5, zone mean of district series; brown = dry, green = wet)"))

Table 2: Observed 2026 monthly rainfall percentile vs 1981-2025 (ERA5, zone mean of district series; brown = dry, green = wet)

	Jan	Feb	Mar	Apr	May	Jun	Jul
zone
Eastern	4	96	98	22	36	0	4
Karamoja	9	89	98	38	24	0	0
Lake Albert Crescent	7	93	76	20	47	0	36
Lake Victoria Crescent	7	96	60	4	9	0	7
Mid Northen	11	91	80	24	29	0	20
South Eastern	7	96	69	18	9	0	9
Southern Drylands	4	91	71	13	20	9	31
Southern Highlands	13	87	78	18	33	9	78
West Nile	9	91	91	22	47	0	49
Western Highlands	24	96	87	9	67	4	73

jiaf_pin = uct.load_jiaf_pin()
zj = (dz[dz["pcode"].isin(set(jiaf_pin["pcode"]))]
      .merge(jiaf_pin.groupby("pcode", as_index=False)["joint_pin"].sum(), on="pcode", how="left"))
ztbl = (zj.groupby("zone")
          .agg(districts=("district", lambda s: ", ".join(sorted(s))),
               joint_pin=("joint_pin", "sum"))
          .sort_values("joint_pin", ascending=False).reset_index())
ztbl.columns = ["ASAP zone", "JIAF districts", "Joint PiN"]
ztbl.style.hide(axis="index").format({"Joint PiN": "{:,.0f}"})

ASAP zone	JIAF districts	Joint PiN
West Nile	Adjumani, Arua, Koboko, Madi Okollo, Obongi, Yumbe	1,049,455
Karamoja	Abim, Amudat, Kaabong, Karenga, Kotido, Moroto, Nabilatuk, Nakapiripirit, Napak	847,143
Lake Albert Crescent	Kikuube, Kiryandongo	206,087
Western Highlands	Kamwenge, Kyegegwa	203,079
Mid Northen	Lamwo	195,234
Southern Drylands	Isingiro	194,028
Lake Victoria Crescent	Kampala	0

Why it matters: people already in need

JIAF classifies areas; PiN counts people. The joint PiN per district × population group is the highest sectoral PiN (JIAF 2.0 mosaic rule); the intersectoral severity is the area’s class.

Two honesty notes on reading the table. The intersectoral severity scale is saturated in this dataset: every district × group is class 2 or 3, 21 of the 22 districts include a class-3 area (Kampala is the exception), and 100% of the joint PiN sits in severity-3 areas — so the color-graded Severity column separates almost nothing within the assessed set. Where the data do differentiate is the sectoral classes (Nutrition reaches 4 in Kaabong, Karenga, Kotido and Moroto — exactly the Karamoja core). And the PiN is a pre-shock baseline from the 2026 planning cycle: it cannot respond to the forecast shock this document is about, only be read alongside it.

pin = uct.load_jiaf_pin()
sev = uct.load_jiaf_severity()
j = pin.merge(sev[["district", "pop_group", "intersectoral_severity"]],
              on=["district", "pop_group"], how="left")
tbl = j[["sub_region", "district", "pop_group", "population",
         "intersectoral_severity", "joint_pin", "pct_population"]].copy()
tbl["pct_population"] = 100 * tbl["pct_population"]
tbl.columns = ["Sub-region", "District", "Group", "Population",
               "Severity", "Joint PiN", "PiN % of pop."]
(tbl.sort_values(["Sub-region", "District", "Group"])
    .style.hide(axis="index")
    .format({"Population": "{:,.0f}", "Joint PiN": "{:,.0f}", "PiN % of pop.": "{:.0f}%"}, na_rep="—")
    .background_gradient(subset=["Severity"], cmap=ListedColormap(JIAF_COLORS), vmin=1, vmax=5))

Sub-region	District	Group	Population	Severity	Joint PiN	PiN % of pop.
Acholi	Lamwo	Host community	213,156	3.000000	114,912	54%
Acholi	Lamwo	Refugees	98,882	3.000000	80,322	81%
Kampala (urban)	Kampala	Refugees	170,474	2.000000	—	—
Karamoja	Abim	Host community	144,084	3.000000	74,261	52%
Karamoja	Amudat	Host community	203,358	3.000000	132,183	65%
Karamoja	Kaabong	Host community	264,631	3.000000	142,009	54%
Karamoja	Karenga	Host community	100,375	3.000000	56,210	56%
Karamoja	Kotido	Host community	219,734	3.000000	142,828	65%
Karamoja	Moroto	Host community	103,639	3.000000	53,995	52%
Karamoja	Nabilatuk	Host community	136,785	3.000000	72,814	53%
Karamoja	Nakapiripirit	Host community	111,681	3.000000	58,132	52%
Karamoja	Napak	Host community	211,830	3.000000	114,711	54%
Mid-West	Kikuube	Host community	379,547	2.000000	—	—
Mid-West	Kikuube	Refugees	157,991	3.000000	104,274	66%
Mid-West	Kiryandongo	Host community	364,872	2.000000	—	—
Mid-West	Kiryandongo	Refugees	166,907	3.000000	101,813	61%
South West	Isingiro	Host community	635,077	2.000000	—	—
South West	Isingiro	Refugees	281,357	3.000000	188,509	67%
South West	Isingiro	Refugees	281,357	3.000000	188,509	67%
South West	Isingiro	Refugees	8,238	3.000000	5,519	67%
South West	Isingiro	Refugees	8,238	3.000000	5,519	67%
South West	Kamwenge	Host community	337,167	2.000000	—	—
South West	Kamwenge	Refugees	108,799	3.000000	93,567	86%
South West	Kyegegwa	Host community	501,120	2.000000	—	—
South West	Kyegegwa	Refugees	138,623	3.000000	109,512	79%
West Nile	Adjumani	Host community	300,590	3.000000	72,120	24%
West Nile	Adjumani	Refugees	235,088	3.000000	157,509	67%
West Nile	Koboko	Host community	271,781	2.000000	—	—
West Nile	Koboko	Refugees	6,528	3.000000	2,350	36%
West Nile	Madi-Okollo & Terego	Host community	501,304	3.000000	107,067	21%
West Nile	Madi-Okollo & Terego	Refugees	204,257	3.000000	120,512	59%
West Nile	Obongi	Host community	142,983	3.000000	41,265	29%
West Nile	Obongi	Refugees	144,330	3.000000	83,423	58%
West Nile	Terego	Refugees	79,832	3.000000	48,777	61%
West Nile	Yumbe	Host community	945,100	3.000000	275,780	29%
West Nile	Yumbe	Refugees	209,928	3.000000	140,652	67%

tot = j.groupby("pop_group")[["population", "joint_pin"]].sum()
tot.loc["TOTAL"] = tot.sum()
tot.columns = ["Population baseline", "Joint PiN"]
tot.style.format("{:,.0f}")

	Population baseline	Joint PiN
pop_group
Host community	6,088,814	1,458,287
Refugees	2,300,829	1,430,767
TOTAL	8,389,643	2,889,054

Response: PiN → targeted → reached

Max targeted / achieved people from the response-monitoring workbooks (2026-08-19). Each admin level in the source is an independent rollup, not a sum of the finer level: district figures here are the district-level series, and levels must never be added together.

One join caveat, resolved conservatively: the workbooks use post-split districts and have an “Arua” row but no “Terego” row, while JIAF assesses Terego and not Arua - and the two share a CODAB pcode (Terego split from Arua in 2020) without being the same territory. Response figures therefore attach to JIAF rows by district name: Terego shows “none reported” rather than inheriting Arua’s figures, and whether Arua’s reporting actually covers Terego is a question to put to the country team. The combined JIAF unit “Madi-Okollo & Terego” takes the workbook’s “Madi Okollo” row, its dominant part.

tgt = uct.load_reach("targeted", level=2)[["pcode", "adm2_name", "total"]].rename(columns={"total": "targeted"})
rea = uct.load_reach("achieved", level=2)[["pcode", "adm2_name", "total"]].rename(columns={"total": "reached"})
reach = tgt.merge(rea, on=["pcode", "adm2_name"], how="outer")
reach["_k"] = reach["adm2_name"].str.strip().str.lower()

# The workbooks use post-split districts, JIAF uses operational units, and a
# pcode is not an identity: workbook "Arua" and JIAF "Terego" both map to
# UG3072 (Terego split from Arua in 2020) while being disjoint territories, so
# a pcode join would credit Arua's response to Terego's row. Reach therefore
# attaches by NAME, with one documented exception: the combined JIAF unit
# "Madi-Okollo & Terego" takes workbook "Madi Okollo" (its dominant part).
REACH_NAME_FOR_JIAF = {"madi-okollo & terego": "madi okollo"}

def attach_reach(df, name_col="district"):
    k = df[name_col].str.strip().str.lower().map(lambda s: REACH_NAME_FOR_JIAF.get(s, s))
    return (df.assign(_k=k)
              .merge(reach[["_k", "targeted", "reached"]], on="_k", how="left")
              .drop(columns="_k"))

dpin = pin.groupby(["pcode"], as_index=False).agg(district=("district", "first"),
                                                  pin=("joint_pin", "sum"))
jf = attach_reach(dpin)
matched = set(jf["district"].str.strip().str.lower().map(lambda s: REACH_NAME_FOR_JIAF.get(s, s)))
extra = reach[~reach["_k"].isin(matched)].rename(columns={"adm2_name": "district"})
fun = pd.concat([jf, extra[["pcode", "district", "targeted", "reached"]]], ignore_index=True)
fun = fun[fun[["targeted", "reached"]].notna().any(axis=1) | fun["pin"].notna()]
fun = fun.sort_values("pin", ascending=True).fillna(0)

fig, ax = plt.subplots(figsize=(9, 0.42 * len(fun) + 1.2))
y = np.arange(len(fun))
h = 0.27
ax.barh(y + h, fun["pin"], height=h, color=C_PIN, label="People in need (JIAF)")
ax.barh(y, fun["targeted"], height=h, color=C_TGT, label="Max targeted")
ax.barh(y - h, fun["reached"], height=h, color=C_REA, label="Max reached")
ax.set_yticks(y, fun["district"])
ax.xaxis.set_major_formatter(lambda v, _: f"{v/1e3:,.0f}k")
ax.legend(frameon=False, loc="lower right")
ax.spines[["top", "right"]].set_visible(False)
ax.set_title("Caseload by district")
plt.tight_layout()
plt.show()

Per district: joint PiN (host + refugee groups summed), max targeted, max reached. Districts with any response data, sorted by PiN.

Forecast × caseload

The readout the app’s tab exists for: where a bad drought forecast meets a large caseload. Districts below are the JIAF 2.0 Light districts joined to their forecast slot; the two operational units that share a CODAB polygon (Terego sits inside CODAB’s Arua; “Madi-Okollo & Terego” maps to Madi Okollo) inherit that polygon’s forecast. The last column adds OND flood-extent context: pixel-mean flood recurrence is low across the whole assessed set (0-6%, highest on the Albert Nile at Obongi) - the flood-extent hotspots lie outside the JIAF footprint, in the Teso wetlands (see the flood section).

comb = (j.groupby(["pcode", "sub_region", "district"], as_index=False)
          .agg(population=("population", "sum"), joint_pin=("joint_pin", "sum"),
               worst_sev=("intersectoral_severity", "max"))
          .merge(fc2, on="pcode", how="left"))
comb = attach_reach(comb)   # by name, not pcode — see the response section
comb["pct_normal"] = 100 * comb["fc_mm"] / comb["nrm_mm"]

# OND flood-extent context for the assessed districts (pixel-mean recurrence;
# low everywhere in the JIAF set - the extent hotspots lie outside it).
flrec = stratus.load_parquet_from_blob(f"{UGA_DIR}/flood_ond_adm2.parquet", stage="dev")
comb = comb.merge(flrec[["pcode", "recurrence_any"]], on="pcode", how="left")

# How much of a normal year's rain falls in this trimester (days-weighted ERA5
# climatology): the deficit matters more where the season carries the year.
mc2 = (era5_2.assign(month=era5_2["valid_date"].dt.month)
             .groupby(["pcode", "month"])["mean"].mean().reset_index())
mc2["mm"] = mc2["mean"] * mc2["month"].map(lambda m: calendar.monthrange(2025, m)[1])
annual_mm = mc2.groupby("pcode")["mm"].sum()

def tri_share(pcode, tri):
    if pd.isna(tri):
        return None
    mm = mc2[(mc2["pcode"] == pcode) & (mc2["month"].isin(TRIMESTERS[tri]))]["mm"].sum()
    return 100 * mm / annual_mm[pcode]

comb["tri_share"] = [tri_share(p, t) for p, t in zip(comb["pcode"], comb["trimester"])]
out = comb[["sub_region", "district", "worst_sev", "joint_pin", "targeted", "reached",
            "cat", "tri_label", "drought_rp", "pct_normal", "tri_share", "r",
            "fc_mm", "nrm_mm", "recurrence_any"]].copy()
out.columns = ["Sub-region", "District", "Severity", "Joint PiN", "Targeted", "Reached",
               "Forecast category", "Trimester", "Drought RP (yr)", "% of normal",
               "Season share of annual rain", "Skill r", "Forecast (mm)", "Normal (mm)",
               "OND flood recurrence"]
(out.sort_values(["% of normal", "Joint PiN"], ascending=[True, False])
    .style.hide(axis="index")
    .format({"Joint PiN": "{:,.0f}", "Targeted": "{:,.0f}", "Reached": "{:,.0f}",
             "Drought RP (yr)": "{:.1f}", "% of normal": "{:.0f}%",
             "Season share of annual rain": "{:.0f}%", "Skill r": "{:.2f}",
             "OND flood recurrence": "{:.0%}"}, na_rep="—")
    .background_gradient(subset=["Severity"], cmap=ListedColormap(JIAF_COLORS), vmin=1, vmax=5))

Sub-region	District	Severity	Joint PiN	Targeted	Reached	Forecast category	Trimester	Drought RP (yr)	% of normal	Season share of annual rain	Skill r	Forecast (mm)	Normal (mm)	OND flood recurrence
Karamoja	Moroto	3.000000	53,995	16,463	13,844	strongly below normal	Jun–Jul–Aug	46.0	3%	43%	0.96	11	319	3%
Karamoja	Kaabong	3.000000	142,009	41,345	19,047	strongly below normal	Jun–Jul–Aug	46.0	7%	44%	0.95	31	443	1%
Karamoja	Amudat	3.000000	132,183	—	—	strongly below normal	Jun–Jul–Aug	46.0	10%	43%	0.94	61	582	3%
Karamoja	Karenga	3.000000	56,210	—	—	strongly below normal	Jun–Jul–Aug	46.0	11%	41%	0.94	42	395	4%
Karamoja	Kotido	3.000000	142,828	34,591	69,284	strongly below normal	Jun–Jul–Aug	46.0	12%	44%	0.94	43	373	6%
Karamoja	Napak	3.000000	114,711	—	—	strongly below normal	Jun–Jul–Aug	46.0	13%	39%	0.95	49	373	6%
Karamoja	Nabilatuk	3.000000	72,814	—	—	strongly below normal	Jun–Jul–Aug	46.0	14%	38%	0.93	57	400	5%
Karamoja	Nakapiripirit	3.000000	58,132	—	—	strongly below normal	Jun–Jul–Aug	46.0	17%	38%	0.92	92	537	4%
Karamoja	Abim	3.000000	74,261	—	—	strongly below normal	Jun–Jul–Aug	46.0	26%	40%	0.95	115	436	5%
South West	Isingiro	3.000000	388,056	223,051	137,112	strongly below normal	Jul–Aug–Sep	46.0	40%	17%	0.59	59	148	3%
South West	Kyegegwa	3.000000	109,512	99,300	70,046	strongly below normal	Jul–Aug–Sep	46.0	43%	23%	0.67	99	231	1%
Mid-West	Kiryandongo	3.000000	101,813	126,451	73,627	strongly below normal	Jul–Aug–Sep	46.0	51%	30%	0.79	220	430	3%
South West	Kamwenge	3.000000	93,567	78,600	44,004	strongly below normal	Jul–Aug–Sep	46.0	52%	20%	0.64	121	233	1%
Acholi	Lamwo	3.000000	195,234	72,749	43,061	strongly below normal	Jun–Jul–Aug	46.0	58%	35%	0.94	305	525	2%
West Nile	Terego	3.000000	48,777	—	—	strongly below normal	Aug–Sep–Oct	46.0	59%	40%	0.58	223	376	0%
West Nile	Madi-Okollo & Terego	3.000000	227,579	210,446	117,815	strongly below normal	Aug–Sep–Oct	46.0	64%	38%	0.58	230	362	4%
West Nile	Koboko	3.000000	2,350	6,314	5,998	strongly below normal	Aug–Sep–Oct	46.0	65%	42%	0.59	385	588	0%
West Nile	Obongi	3.000000	124,688	104,100	58,279	strongly below normal	Jul–Aug–Sep	23.0	67%	33%	0.79	269	402	6%
West Nile	Yumbe	3.000000	416,432	151,800	84,983	strongly below normal	Jun–Jul–Aug	46.0	67%	33%	0.93	247	368	1%
Kampala (urban)	Kampala	2.000000	0	120,000	67,180	strongly below normal	Jul–Aug–Sep	46.0	70%	20%	0.63	255	365	4%
West Nile	Adjumani	3.000000	229,629	172,351	106,177	strongly below normal	Aug–Sep–Oct	11.5	74%	37%	0.58	461	626	2%
Mid-West	Kikuube	3.000000	104,274	115,051	76,064	strongly below normal	Aug–Sep–Oct	23.0	79%	33%	0.57	451	570	2%

s = comb[comb["joint_pin"].notna() & comb["pct_normal"].notna()].copy()
TRI_COLORS = {"Jun–Jul–Aug": "#7f5619", "Jul–Aug–Sep": "#e69f00", "Aug–Sep–Oct": "#56b4e9"}
s["color"] = s["tri_label"].map(TRI_COLORS).fillna("#9db1b3")

fig, ax = plt.subplots(figsize=(9, 6))
ax.scatter(s["pct_normal"], s["joint_pin"], s=90, c=s["color"],
           edgecolor="#5b6b6d", linewidth=0.8, zorder=3)
seen = [t for t in TRI_COLORS if t in set(s["tri_label"])]
ax.legend(handles=[plt.Line2D([], [], marker="o", ls="", markersize=9,
                              markerfacecolor=TRI_COLORS[t], markeredgecolor="#5b6b6d",
                              label=f"{t}{' (in-season)' if t != 'Aug–Sep–Oct' else ''}")
                   for t in seen],
          title="worst qualifying trimester", loc="center",
          bbox_to_anchor=(0.5, 0.45), frameon=False)
for _, r in s.iterrows():
    if r["pct_normal"] <= s["pct_normal"].quantile(0.35) or \
       r["joint_pin"] >= s["joint_pin"].quantile(0.7):
        ax.annotate(r["district"], (r["pct_normal"], r["joint_pin"]),
                    xytext=(6, 4), textcoords="offset points", fontsize=8.5)
ax.set_xlabel("Forecast seasonal total, % of climatological normal (worst qualifying slot)")
ax.set_ylabel("Joint PiN (people)")
ax.xaxis.set_major_formatter(lambda v, _: f"{v:.0f}%")
ax.yaxis.set_major_formatter(lambda v, _: f"{v/1e3:,.0f}k")
ax.invert_xaxis()  # drier to the right, matching "worse →"
ax.grid(color="#eef1f1", zorder=0)
ax.spines[["top", "right"]].set_visible(False)
ax.set_title("Forecast rainfall deficit × people in need, JIAF districts")
plt.tight_layout()
plt.show()

Each point is a JIAF district (host + refugee PiN summed) at its own worst qualifying slot; the trimester differs by district and is shown by color (in-season trimesters mix observed months with forecast). Left = deeper deficit; up = more people in need. Nearly all districts are at the capped 46-yr RP, so the deficit rather than the RP is what separates them.

Priorities: where first, what kind, when

Everything above establishes three facts per district: how deep the forecast deficit is, how many people were already in need before the season failed, and whether a response footprint exists there. This section reads those three dimensions together. It adds no new data or model; the grouping is the agro-ecological grouping from the cropping-seasons section, and every number comes from the tables above.

pri = comb[comb["joint_pin"].notna() & comb["pct_normal"].notna()].copy()
pri = pri.merge(dz[["pcode", "zone"]].drop_duplicates("pcode"), on="pcode", how="left")
pri["group"] = np.select(
    [pri["zone"] == "Karamoja", pri["zone"] == "West Nile"],
    ["Karamoja", "West Nile corridor"], "Bimodal south / west")
pri["coverage"] = pri["targeted"] / pri["joint_pin"]

gs = (pri.groupby("group")
        .agg(districts=("district", "count"), joint_pin=("joint_pin", "sum"),
             targeted=("targeted", "sum"), reached=("reached", "sum"),
             deficit_min=("pct_normal", "min"), deficit_max=("pct_normal", "max"))
        .reindex(["Karamoja", "West Nile corridor", "Bimodal south / west"]))
gs["deficit range"] = gs.apply(lambda r: f"{r['deficit_min']:.0f}-{r['deficit_max']:.0f}% of normal", axis=1)
gs["targeted / PiN"] = gs["targeted"] / gs["joint_pin"]
out_gs = gs[["districts", "joint_pin", "targeted", "reached", "deficit range", "targeted / PiN"]]
out_gs.columns = ["Districts", "Joint PiN", "Targeted", "Reached", "Forecast deficit", "Targeted / PiN"]
(out_gs.style.format({"Joint PiN": "{:,.0f}", "Targeted": "{:,.0f}", "Reached": "{:,.0f}",
                      "Targeted / PiN": "{:.0%}"}, na_rep="none reported"))

	Districts	Joint PiN	Targeted	Reached	Forecast deficit	Targeted / PiN
group
Karamoja	9	847,143	92,399	102,175	3-26% of normal	11%
West Nile corridor	6	1,049,455	645,011	373,252	59-74% of normal	61%
Bimodal south / west	7	992,456	835,202	511,094	40-79% of normal	84%

The reading, group by group:

• Karamoja first. It has received 41-54% of its normal season-to-date rain, the driest or near-driest Apr-Jul on record, in a pastoral zone where the failed Jun-Aug rains are a pasture, water and livestock crisis, and it is the group where the response footprint is thinnest: six of the nine districts have no reported targeted or reached figures at all, and where figures exist (Kotido, Kaabong, Moroto) they target roughly a tenth of the group’s PiN. Whatever explains that gap (reporting scope or actual absence), it is the first thing to resolve, because the dry season from October gives no pasture regrowth until the next rains around April.
• West Nile corridor: a near-normal season that must last the year. The season-to-date accumulation is 84-99% of normal and the record-dry June hit at harvest, after near-normal planting and growth rains; the main season is close to normal in water terms, with risk in harvest conditions and late fields. But this zone has no second season, so this harvest must last until the March 2027 planting. A response exists here but targets well under the JIAF caseload; the question is depth and duration of assistance, not whether to start.
• Bimodal south / west: a decision window, not yet a verdict. Second-season planting (Aug-Oct) is at risk of poor establishment in the record-dry Aug-Sep, but the strongly wet Oct-Dec favors whatever does get established. These districts need a check at establishment time rather than an immediate scale-up.

import matplotlib.cm as cm
from matplotlib.colors import Normalize as MplNorm

order = ["Karamoja", "West Nile corridor", "Bimodal south / west"]
blocks = [pri[pri["group"] == g].sort_values("pct_normal") for g in order]
gap = 1.6
ys, rows_flat = [], []
y = 0.0
for blk in blocks:
    start = y
    for _, r in blk.iterrows():
        rows_flat.append((y, r)); y += 1
    ys.append((start, y)); y += gap
total_h = y

fig, (axL, axR) = plt.subplots(
    1, 2, figsize=(9.5, 0.42 * total_h + 1.4), sharey=True,
    gridspec_kw={"width_ratios": [1, 1.35], "wspace": 0.04})
cmap, nrm = cm.get_cmap("YlOrBr_r"), MplNorm(0, 100)
for yy, r in rows_flat:
    axL.barh(yy, r["pct_normal"], color=cmap(nrm(r["pct_normal"])), height=0.72)
    axL.text(r["pct_normal"] + 1.5, yy, f"{r['pct_normal']:.0f}%",
             va="center", ha="right", fontsize=8, color="#3f4748")
    axR.barh(yy, r["joint_pin"], color=C_PIN, height=0.72, alpha=0.85)
    if pd.notna(r["targeted"]):
        axR.plot([r["targeted"]] * 2, [yy - 0.36, yy + 0.36], color=C_TGT, lw=2.4)
    if pd.notna(r["reached"]):
        axR.plot([r["reached"]] * 2, [yy - 0.36, yy + 0.36], color=C_REA, lw=2.4)
    if pd.isna(r["targeted"]) and pd.isna(r["reached"]):
        axR.text(r["joint_pin"] + 8000, yy, "none reported", va="center",
                 fontsize=7.2, color="#b8272d")
axL.set_yticks([yy for yy, _ in rows_flat], [r["district"] for _, r in rows_flat], fontsize=8.5)
for (start, end), g in zip(ys, order):
    axL.text(99, start - 0.75, g.upper(), ha="left", va="center",
             fontsize=8.5, fontweight="bold", color="#18614c")
axL.set_xlim(0, 100); axL.invert_xaxis(); axL.invert_yaxis()
axL.set_xlabel("forecast, % of normal season", fontsize=9)
axR.set_xlabel("people", fontsize=9)
axR.xaxis.set_major_formatter(lambda v, _: f"{v/1e3:,.0f}k")
for ax in (axL, axR):
    ax.spines[["top", "right"]].set_visible(False)
    ax.grid(axis="x", color="#eef1f1", zorder=0)
axR.legend(handles=[Patch(color=C_PIN, label="people in need"),
                    plt.Line2D([], [], color=C_TGT, lw=2.4, label="targeted"),
                    plt.Line2D([], [], color=C_REA, lw=2.4, label="reached")],
           loc="lower right", frameon=False, fontsize=8.5)
fig.suptitle("Priority board: deficit, caseload, and response footprint", y=0.995)
plt.tight_layout()
plt.show()

One row per JIAF district, grouped by situation and sorted driest-first. Left: the forecast deficit (darker = less of the normal season’s rain). Right: people in need (blue), with targeted (orange tick) and reached (red tick) from the monitoring workbooks; ‘none reported’ marks districts absent from those workbooks.

MONTHS = ["Aug", "Sep", "Oct", "Nov", "Dec", "Jan", "Feb", "Mar", "Apr"]
X = {m: i for i, m in enumerate(MONTHS)}

rows = [
    ("Karamoja", [
        (X["Aug"], X["Oct"] + 1, "#7f5619", 0.30, "failed rains: pasture, water, livestock stress"),
        (X["Nov"], X["Mar"] + 1, "#9db1b3", 0.30, "dry season: no regrowth until the ~Apr rains"),
    ], [(X["Apr"] + 0.5, "next rains\n(climatology)")]),
    ("West Nile corridor", [
        (X["Aug"], X["Mar"] + 1, "#e69f00", 0.30, "this harvest must last until Mar 2027 planting"),
    ], [(X["Mar"] + 0.2, "next planting\n(Mar 2027)")]),
    ("Bimodal south / west", [
        (X["Aug"], X["Oct"] + 1, "#56b4e9", 0.35, "second-season planting: establishment at risk"),
        (X["Oct"], X["Jan"] + 1, "#2a7e43", 0.25, "wet Oct-Dec supports what got established"),
    ], [(X["Oct"] + 0.5, "establishment\ncheck"), (X["Jan"] + 0.5, "second harvest")]),
    ("Flood watch", [
        (X["Oct"], X["Dec"] + 1, "#1f69b3", 0.25, "Teso/Kyoga + Elgon/Kasese slopes"),
    ], []),
    ("Monitoring", [], [
        (X["Sep"] + 0.2, "SEAS5\nissuance"), (X["Oct"] + 0.2, "SEAS5 + ERA5\nconfirm Aug-Sep"),
        (X["Nov"] + 0.2, "SEAS5\nissuance"),
    ]),
]

fig, ax = plt.subplots(figsize=(9.5, 4.2))
for i, (name, bands, marks) in enumerate(rows):
    yy = len(rows) - 1 - i
    for k, (x0, x1, c, a, lab) in enumerate(bands):
        ax.barh(yy, x1 - x0, left=x0, height=0.52, color=c, alpha=a)
        dy = 0.24 if (len(bands) > 1 and k == 0) else (-0.24 if len(bands) > 1 else 0.02)
        ax.text((x0 + x1) / 2, yy + dy, lab, ha="center", va="center", fontsize=8, color="#1f2324")
    for x, lab in marks:
        ax.plot(x, yy, marker="D", ms=7, color="#18614c")
        ax.annotate(lab, (x, yy), xytext=(0, -30), textcoords="offset points",
                    ha="center", fontsize=7.3, color="#18614c")
ax.set_yticks(range(len(rows)), [r[0] for r in rows][::-1], fontsize=9.5)
ax.set_xticks(range(len(MONTHS)), [f"{m}\n'26" if i < 5 else f"{m}\n'27" for i, m in enumerate(MONTHS)], fontsize=8.5)
ax.set_xlim(-0.3, len(MONTHS) + 0.4); ax.set_ylim(-0.8, len(rows) - 0.3)
ax.spines[["top", "right", "left"]].set_visible(False)
ax.set_title("When it bites, and when to look again")
plt.tight_layout()
plt.show()

What happens when, from the crop calendars, the ERA5 climatology and the SEAS5 issuance cycle. Diamonds are decision or verification points.

The monitoring row is the workload plan for this analysis itself: SEAS5 issues around the 5th of each month, so the September and October issuances (and the ERA5 observations that confirm how Aug-Sep actually landed) are the natural points to re-run this document and re-check the bimodal establishment window. The pipeline is parameterized, so re-running just means re-rendering the document.

Flood outlook: the OND wet season

When and where OND flooding recurs

The same issuance that forecasts the drought puts October, November and December at the 91st-98th wet percentile, so the flood question is part of the outlook. Worth knowing before the map: Oct-Dec is not Uganda’s main flood season everywhere. In the FloodScan record (daily SFED, 1998-2026) the national annual flood peak falls in April or May in 19 of 29 years, driven by the Northern (April peak) and Eastern (May peak) regions; November is a genuine secondary peak in the Central and Western regions. The map below shows OND because that is the window the current forecast puts wet, not because it is the only season that floods.

fs1 = pd.read_sql("SELECT pcode, valid_date, mean FROM public.floodscan "
                  "WHERE iso3='UGA' AND adm_level=1 AND band='SFED'",
                  engine, parse_dates=["valid_date"])
cyc = (fs1.assign(mo=fs1["valid_date"].dt.month)
          .groupby(["pcode", "mo"])["mean"].mean().unstack())
cyc = 100 * cyc.div(cyc.max(axis=1), axis=0)
names1 = dict(zip(cod1["ADM1_PCODE"], cod1["ADM1_EN"]))
fig, ax = plt.subplots(figsize=(9, 3.4))
RC = {"UG1": "#1f69b3", "UG2": "#e69f00", "UG3": "#7f5619", "UG4": "#2a7e43"}
for pc, row in cyc.iterrows():
    ax.plot(row.index, row.values, color=RC[pc], lw=2, label=names1[pc])
    pk = row.idxmax()
    ax.plot(pk, row[pk], marker="o", ms=6, color=RC[pc])
ax.axvspan(10, 12, color="#2a7e43", alpha=0.08)
ax.text(11, 8, "OND", ha="center", fontsize=9, color="#2a7e43")
ax.set_xticks(range(1, 13), [calendar.month_abbr[m] for m in range(1, 13)])
ax.set_ylabel("% of region's peak month")
ax.legend(frameon=False, fontsize=8.5, ncol=4, loc="lower center", bbox_to_anchor=(0.5, -0.38))
ax.grid(color="#eef1f1", zorder=0)
ax.spines[["top", "right"]].set_visible(False)
ax.set_title("When Uganda floods: monthly mean SFED by region")
plt.tight_layout()
plt.show()

Seasonal cycle of flooding: monthly mean FloodScan SFED per region, 1998-2026, scaled to each region’s peak month. April-May dominates the north and east; November is the Central/Western secondary peak.

import rasterio as rio
from rasterio.io import MemoryFile
tif = stratus.load_blob_data(f"{UGA_DIR}/flood_ond_recurrence.tif", stage="dev")
from rasterio.features import geometry_mask
with MemoryFile(tif) as mf, mf.open() as ds:
    rec = ds.read(1)
    ext = [ds.bounds.left, ds.bounds.right, ds.bounds.bottom, ds.bounds.top]
    inside = geometry_mask(cod1.geometry, out_shape=rec.shape,
                           transform=ds.transform, invert=True, all_touched=True)
fig, ax = plt.subplots(figsize=(9, 8))
masked = np.ma.masked_where(~inside | (rec < 0.02), rec)
im = ax.imshow(masked, extent=ext, cmap="Blues", vmin=0, vmax=0.6, zorder=2)
cod2.boundary.plot(ax=ax, color=C_EDGE, linewidth=0.4, zorder=3)
cod1.boundary.plot(ax=ax, color="#9db1b3", linewidth=1.1, zorder=3)
cod2[cod2["ADM2_PCODE"].isin(jiaf_pcodes)].boundary.plot(ax=ax, color="#1d2021", linewidth=1.0, zorder=4)
ax.set_xlim(29.4, 35.1); ax.set_ylim(-1.6, 4.4)
cb = plt.colorbar(im, ax=ax, shrink=0.6)
cb.set_label("share of OND seasons with flooding (SFED >= 0.05)")
cb.ax.yaxis.set_major_formatter(lambda v, _: f"{100*v:.0f}%")
ax.set_axis_off()
ax.set_title("OND flood recurrence, 1998-2025 (FloodScan)")
plt.tight_layout()
plt.show()

Share of Oct-Dec seasons 1998-2025 with any meaningful flooding (seasonal max SFED >= 0.05, the team’s noise floor), per ~9 km FloodScan pixel. Black outline: JIAF districts. Persistent wetland systems (Lake Kyoga fringes in Teso) recur most; this is historical recurrence, not a forecast.

fl = stratus.load_parquet_from_blob(f"{UGA_DIR}/flood_ond_adm2.parquet", stage="dev")
fl = fl.merge(cod2[["ADM2_PCODE", "ADM2_EN"]], left_on="pcode", right_on="ADM2_PCODE")
top = fl.nlargest(12, "recurrence_any")[["ADM2_EN", "recurrence_any", "recurrence_substantial"]]
top.columns = ["District", "Any flooding (share of OND seasons)", "Substantial (SFED >= 0.2)"]
(top.style.hide(axis="index")
    .format({"Any flooding (share of OND seasons)": "{:.0%}", "Substantial (SFED >= 0.2)": "{:.0%}"})
    .set_caption("The 12 most flood-recurrent of Uganda's 135 districts, ranked by "
                 "pixel-mean OND flood recurrence, 1998-2025. Selected on the flood metric "
                 "alone, country-wide; none is in the JIAF set (see the note below the map)."))

Table 3: The 12 most flood-recurrent of Uganda's 135 districts, ranked by pixel-mean OND flood recurrence, 1998-2025. Selected on the flood metric alone, country-wide; none is in the JIAF set (see the note below the map).

District	Any flooding (share of OND seasons)	Substantial (SFED >= 0.2)
Ngora	58%	26%
Pallisa	32%	11%
Kumi	30%	22%
Butaleja	24%	3%
Katakwi	23%	10%
Soroti	20%	7%
Serere	20%	5%
Budaka	19%	5%
Bukedea	16%	10%
Ntoroko	14%	5%
Tororo	13%	1%
Bulambuli	12%	2%

The recurrence concentrates in the Teso / Lake Kyoga wetland systems (Ngora 57%, Pallisa, Kumi, Katakwi, Soroti) and along the Butaleja lowlands, none of which are JIAF districts, but Katakwi and Soroti border Karamoja: a wet OND that recharges Karamoja’s water points can simultaneously flood the districts just downstream of it. District-level FloodScan statistics below region level do not exist in the rasterstats DB for Uganda; these were computed from the daily COGs (pipeline/compute_uga_flood_recurrence.py).

Extreme seasons: the El Niño / IOD lens

The 2026 OND season arrives with a strong El Niño and a positive Indian Ocean Dipole crossing its event threshold (weekly DMI +0.41 in early August, forecast to stay positive through December; BOM, WMO). What did past seasons like that look like? The panels below show every El Niño OND in the FloodScan record (OND-mean ONI ≥ +0.5, the standard cut: nine seasons) plus 2025, each labeled with its observed OND-mean ONI (ENSO) and DMI (IOD) from NOAA (pipeline/fetch_climate_indices.py), its flooded area, and that area’s empirical return period.

from scipy.stats import spearmanr
ytif = stratus.load_blob_data(f"{UGA_DIR}/flood_ond_yearly_max.tif", stage="dev")
with MemoryFile(ytif) as mf, mf.open() as yds:
    yyears = [int(d) for d in yds.descriptions]
    ystack = yds.read()
    yext = [yds.bounds.left, yds.bounds.right, yds.bounds.bottom, yds.bounds.top]
    yin = geometry_mask(cod1.geometry, out_shape=ystack.shape[1:], transform=yds.transform, invert=True, all_touched=True)
idxs = stratus.load_parquet_from_blob(f"{PROJECT_PREFIX}/raw/climate_indices/ond_indices.parquet", stage="dev").set_index("year")
area_pct = pd.Series({y: 100 * np.nanmean((ystack[i] >= 0.05)[yin]) for i, y in enumerate(yyears)})
area_rank = area_pct.rank(ascending=False)
area_rp = (len(area_pct) + 1) / area_rank

nino_years = sorted(int(y) for y in idxs.index
                    if 1998 <= y <= 2025 and idxs.loc[y, "oni_ond"] >= 0.5)
FACET_YEARS = nino_years + [2025]   # 2025 is not an El Niño OND; shown for scale
fig, axes = plt.subplots(2, 5, figsize=(13, 7.0), layout="constrained")
if len(FACET_YEARS) > axes.size:   # zip would silently drop panels
    raise ValueError(f"{len(FACET_YEARS)} facet years > {axes.size} panels — resize the grid")
for ax in axes.flat[len(FACET_YEARS):]:
    ax.set_visible(False)
for ax, yr in zip(axes.flat, FACET_YEARS):
    arr = ystack[yyears.index(yr)]
    m = np.ma.masked_where(~yin | (arr < 0.05), arr)
    ax.imshow(np.where(yin, 0, np.nan), extent=yext, cmap="Greys", vmin=-1, vmax=3)
    im = ax.imshow(m, extent=yext, cmap="Blues", vmin=0, vmax=0.6)
    cod1.boundary.plot(ax=ax, color="#9db1b3", linewidth=0.6)
    ax.set_xlim(29.4, 35.1); ax.set_ylim(-1.6, 4.4); ax.set_axis_off()
    ax.set_title(f"{yr} | ONI {idxs.loc[yr,'oni_ond']:+.1f}  DMI {idxs.loc[yr,'dmi_ond']:+.1f}\n"
                 f"flooded {area_pct[yr]:.1f}% (RP {area_rp[yr]:.0f} yr)", fontsize=9)
fig.colorbar(im, ax=axes, shrink=0.5, label="OND seasonal max SFED")
fig.suptitle("El Niño ONDs and 2025, on one scale", fontsize=12)
plt.show()

OND seasonal-max SFED in every El Niño OND of the record (OND-mean ONI >= +0.5) plus 2025, against the same color scale. Panel labels give the observed OND-mean ENSO (ONI) and IOD (DMI) indices, the flooded share of pixels (SFED >= 0.05), and its rank-based return period among the 28 seasons.

NoteWhat the statistics support, and what is our interpretation

With 28 seasons, be explicit about which links are demonstrated and which are a reading of the record.

Statistically supported (Spearman unless noted): the ocean drivers set the season’s rainfall (OND rain vs DMI r = +0.63, p < 0.001; vs ONI r = +0.55, p = 0.002), and the wettest-quintile ONDs have larger flooded areas than the rest (Mann-Whitney one-sided p = 0.018).

Suggestive but not established: the continuous rain-to-flooded-area relationship (r = +0.28, p = 0.14) and season-to-season persistence of flooded area (lag-1 r = +0.38, p = 0.053).

Not supported as direct links: neither index predicts flooded area directly (area vs DMI r = +0.17, p = 0.40; vs ONI r = +0.31, p = 0.11). In plain terms: if the ocean drivers matter for flooding here, the data say they matter only by making the season wetter; there is no sign of an additional index-to-flood pathway beyond rainfall.

Interpretation, not tested: the pattern that El Niño years flood more when a positive IOD co-occurs. Group cuts: El Niño = OND-mean ONI ≥ +0.5, +IOD = OND-mean DMI ≥ +0.4. Descriptively, the four El Niño + IOD seasons (2006, 2018, 2019, 2023) averaged 4.9% flooded (median 3.8%) against 3.2% (median 2.9%) in the five El Niño-only seasons (2002, 2004, 2009, 2014, 2015). The split is sensitive to the DMI cut: 2015 (DMI +0.37, flooded 5.9%) sits a hair below it in the El Niño-only group, and moving that one borderline year across strengthens the contrast to 5.1% vs 2.5% (means) - so the lean survives the cut, but the biggest floods of all (2024, 2025) still sit in the neither-driver group. Read it as a lean, not a rule.

Also interpretation, on why 2025 flooded on little rain: two candidate mechanisms, neither measured here. Antecedent wetness: 2023-24 were the second- and third-largest flood ONDs, so the system entered 2025 wet. Regulation and channel state: Lake Victoria’s only outlet is the managed release at Jinja (Nalubaale/Kiira, then Bujagali-Isimba-Kyoga), release policy has been publicly blamed for downstream flooding in earlier high-stand years, and papyrus “floating islands” have blocked the Kyoga outlet before (Monitor, EACCR). One spatial clue from our own data: only 18% of 2025’s flooded pixels lie in the recurrent Kyoga/wetland core, defined here as pixels that flood in at least 30% of ONDs 1998-2025 (recurrence_any ≥ 0.3 on the recurrence raster); for 2019 the same figure is 40%. The 2023-25 floods spread far beyond the lake system, which fits broad antecedent wetness better than a purely lake-stage or dam-release story, without excluding either as a contributor. Settling this needs lake-level and discharge data we have not used.

ond_rain = (e_all[e_all["mo"].isin([10, 11, 12])]
            .groupby(["yr", "pcode"])["mm"].sum().groupby("yr").mean().loc[1998:2025])
drv = pd.DataFrame({"rain": ond_rain, "area": area_pct, "dmi": idxs["dmi_ond"], "oni": idxs["oni_ond"]}).dropna()
top5 = set(area_pct.nlargest(5).index)
fig, ax = plt.subplots(figsize=(9, 5.6))
sc = ax.scatter(drv["rain"], drv["area"], c=drv["dmi"], cmap="BrBG", vmin=-0.9, vmax=0.9,
                s=110, edgecolor="#5b6b6d", linewidth=0.8, zorder=3)
for yr, r in drv.iterrows():
    if (yr - 1) in top5:
        ax.scatter(r["rain"], r["area"], s=230, facecolor="none", edgecolor="#b8272d", linewidth=1.6, zorder=2)
    ax.annotate(str(yr), (r["rain"], r["area"]), xytext=(5, 4), textcoords="offset points", fontsize=8)
ax.set_yscale("log")
ax.set_xlabel("OND rainfall, national mean of district totals (mm)")
ax.set_ylabel("flooded area, % of pixels (log)")
ax.yaxis.set_major_formatter(lambda v, _: f"{v:g}%")
cb = plt.colorbar(sc, ax=ax); cb.set_label("OND-mean DMI (IOD)")
ax.grid(color="#eef1f1", zorder=0)
ax.spines[["top", "right"]].set_visible(False)
ax.set_title("Wet drivers set the rain; the rain and stored water set the flood")
plt.tight_layout()
plt.show()

Each point is one OND season: national mean OND rainfall (x) against flooded area (y, log scale). Fill color: the observed IOD (DMI). Ringed points followed a top-5 flood season. n=28; the callout above says which of these relationships are statistically supported and which are interpretation.

Where flooding amplifies in wet seasons, and where impacts land

Conditioning on the six wettest ONDs of the record (2002, 2006, 2011, 2015, 2019, 2023), a split with statistical support (the callout above), shows where a wet season changes the map, in return-period terms: in the Teso / Kyoga core the flood RP tightens from about one year in 1.7 to one in 1.3, and a set of pixels that rarely flood otherwise gain more than 30 percentage points of flood probability, expanding the flooded footprint east and centre. The right panel shows the other half of the story from EM-DAT (2001-2024): recorded impacts concentrate not in the wetland core but in flash-flood and landslide terrain. Splitting the two types makes the geography sharper. Flood events (36 of 47) cluster on Kasese under the Rwenzoris (12 of its 13 events) and the towns of the Elgon foothills; landslide events (11 of 47) sit almost entirely on the Mount Elgon slopes (Bududa, Sironko, Mbale, Bulambuli) yet carry 765 of the 1,266 recorded deaths, 60% of the total from a quarter of the events. Neither shows up well in a 9 km flood-extent product. And a further reason extent and recorded impact diverge is exposure, not hazard: high rainfall that floods sparsely populated wetland generates extent but no impact record, which is exactly what 2023 looks like (the largest El Niño-year extent, no EM-DAT event). The two maps disagree because they measure different things, and both belong in the watch.

wet6 = set(ond_rain.nlargest(6).index)
wmask = np.array([y in wet6 for y in yyears])
p_wet = np.nanmean(ystack[wmask] >= 0.05, axis=0)
p_oth = np.nanmean(ystack[~wmask] >= 0.05, axis=0)
amp = 100 * (p_wet - p_oth)

from ocha_stratus import emdat
em = emdat.load_emdat_from_blob(iso3="UGA")
emf = em[em["Disaster Type"].isin(["Flood", "Mass movement (wet)"])].copy()
_names = sorted(cod2["ADM2_EN"].tolist(), key=len, reverse=True)
emf["districts"] = emf["Location"].map(
    lambda L: [n for n in _names if isinstance(L, str) and n.lower() in L.lower()])
_x = emf.explode("districts").dropna(subset=["districts"])
m_flood = _x[_x["Disaster Type"] == "Flood"].groupby("districts").size().rename("floods")
m_slide = _x[_x["Disaster Type"] == "Mass movement (wet)"].groupby("districts").size().rename("slides")
emap = (cod2.merge(m_flood, left_on="ADM2_EN", right_index=True, how="left")
             .merge(m_slide, left_on="ADM2_EN", right_index=True, how="left"))

fig, (axA, axB) = plt.subplots(1, 2, figsize=(13, 6.8))
mamp = np.ma.masked_where(~yin | (np.abs(amp) < 5), amp)
axA.imshow(np.where(yin, 0, np.nan), extent=yext, cmap="Greys", vmin=-1, vmax=3)
imA = axA.imshow(mamp, extent=yext, cmap="PuBu", vmin=0, vmax=60, zorder=2)
cod1.boundary.plot(ax=axA, color="#9db1b3", linewidth=0.8, zorder=3)
cod2[cod2["ADM2_PCODE"].isin(jiaf_pcodes)].boundary.plot(ax=axA, color="#1d2021", linewidth=0.8, zorder=4)
axA.set_xlim(29.4, 35.1); axA.set_ylim(-1.6, 4.4); axA.set_axis_off()
plt.colorbar(imA, ax=axA, shrink=0.55, label="extra flood probability in wet ONDs (pp)")
axA.set_title("Amplification: wettest-quintile ONDs vs the rest", fontsize=11)

emap.plot(ax=axB, column="floods", cmap="OrRd", edgecolor=C_EDGE, linewidth=0.4,
          missing_kwds={"color": "#f5f7f7"}, legend=True,
          legend_kwds={"label": "EM-DAT flood events, 2001-2024", "shrink": 0.55})
cod1.boundary.plot(ax=axB, color="#9db1b3", linewidth=0.8)
cod2[cod2["ADM2_PCODE"].isin(jiaf_pcodes)].boundary.plot(ax=axB, color="#1d2021", linewidth=0.8)
sl = emap[emap["slides"].notna()].copy()
sl["pt"] = sl.representative_point()
axB.scatter([g.x for g in sl["pt"]], [g.y for g in sl["pt"]],
            s=28 * sl["slides"], marker="^", facecolor="#1f2324",
            edgecolor="white", linewidth=0.7, zorder=5,
            label="landslide events (size = count)")
axB.legend(loc="lower left", frameon=False, fontsize=8.5)
axB.set_xlim(29.4, 35.1); axB.set_ylim(-1.6, 4.4); axB.set_axis_off()
axB.set_title("Recorded impacts: floods (fill) vs landslides (triangles)", fontsize=11)
plt.tight_layout()
plt.show()

Left: extra probability of flooding (percentage points) in the six wettest ONDs versus all other seasons; the Teso/Kyoga core intensifies and the footprint spreads. Right: EM-DAT recorded impacts 2001-2024 by district named in the event record (all months; 42 of 47 events located), split by type: flood events as the fill, wet mass-movement (landslide) events as triangles sized by count. JIAF districts outlined in black on both.

# No binary "flood event" definition survives the wetland districts (some part
# of Kumi/Katakwi/Soroti shows water every OND), so the district table reports
# the continuous quantity instead: the share of district area flooded.
from exactextract import exact_extract
from exactextract.raster import NumPyRasterSource
yr_srcs = [NumPyRasterSource((ystack[i] >= 0.05).astype("float64"),
                             xmin=yext[0], ymin=yext[2], xmax=yext[1], ymax=yext[3],
                             name=f"y{y}", srs_wkt=cod2.crs.to_wkt())
           for i, y in enumerate(yyears)]
frac = exact_extract(yr_srcs, cod2, ["mean"], include_cols=["ADM2_PCODE", "ADM2_EN"], output="pandas")
fmat = frac.set_index(["ADM2_PCODE", "ADM2_EN"])
fmat.columns = [int(c.split("_")[0][1:]) for c in fmat.columns]
wet_cols = [y for y in fmat.columns if y in wet6]
oth_cols = [y for y in fmat.columns if y not in wet6]
res = pd.DataFrame({
    "typical": 100 * fmat[oth_cols].median(axis=1),
    "wetq": 100 * fmat[wet_cols].median(axis=1),
    "worst": 100 * fmat.max(axis=1),
}).reset_index()
tt = res.nlargest(12, "typical")[["ADM2_EN", "typical", "wetq", "worst"]]
tt.columns = ["District", "Median area flooded, normal ONDs",
              "Median, wettest-quintile ONDs", "Worst season on record"]
(tt.style.hide(axis="index")
   .format({"Median area flooded, normal ONDs": "{:.0f}%",
            "Median, wettest-quintile ONDs": "{:.0f}%", "Worst season on record": "{:.0f}%"})
   .set_caption("The 12 districts with the largest typical flooded area, of Uganda's 135. "
                "HISTORICAL flooding, not the 2026 outlook: share of each district's area "
                "with OND seasonal-max SFED >= 0.05, 1998-2025. These wetland districts show "
                "some water every season, so a binary flood/no-flood count is meaningless "
                "here; the question is how much of the district floods. With a near-record "
                "wet OND forecast, the wettest-quintile column is the operative climatology. "
                "None of these districts is in the JIAF set."))

Table 4: The 12 districts with the largest typical flooded area, of Uganda's 135. HISTORICAL flooding, not the 2026 outlook: share of each district's area with OND seasonal-max SFED >= 0.05, 1998-2025. These wetland districts show some water every season, so a binary flood/no-flood count is meaningless here; the question is how much of the district floods. With a near-record wet OND forecast, the wettest-quintile column is the operative climatology. None of these districts is in the JIAF set.

District	Median area flooded, normal ONDs	Median, wettest-quintile ONDs	Worst season on record
Ngora	67%	67%	72%
Pallisa	28%	48%	68%
Kumi	26%	37%	45%
Serere	20%	20%	46%
Katakwi	20%	21%	38%
Butaleja	18%	39%	66%
Soroti	18%	18%	42%
Budaka	16%	37%	42%
Bukedea	12%	12%	44%
Amuria	8%	12%	18%
Ntoroko	7%	12%	49%
Nabilatuk	3%	6%	21%

The worst events on record, and what FloodScan saw

The four case studies below are the worst flood and landslide events in Uganda’s EM-DAT record, chosen to span the two regimes. For each: the FloodScan flood extent over the event window (left) with the districts named in the event record outlined, and the daily flood-extent series for the event’s region (right) with that year against every other year, and 2026 so far for context. The two wetland-extent events (2007 Teso, Dec 2019) are exactly the regime a record-wet OND forecast makes more likely; the two landslide events (Bududa 2010, Mt Elgon 2024) barely register in flood extent, which is the strongest visual argument for watching the Elgon and Rwenzori slopes with something other than FloodScan.

One live observation falls out of the green lines: Eastern region’s 2026 flood extent has run near record levels all year even while Jun-Sep rainfall was the driest on record. Flood extent there is following the system’s stored water (the Kyoga wetlands after three record flood seasons), not the current season’s rain, and that is the state the record-wet OND forecast will land on.

ev_tif = stratus.load_blob_data(f"{UGA_DIR}/flood_events_maxsfed.tif", stage="dev")
ev_meta = stratus.load_parquet_from_blob(f"{UGA_DIR}/flood_events.parquet", stage="dev")
with MemoryFile(ev_tif) as mf, mf.open() as eds:
    ev_bands = {d: eds.read(i + 1) for i, d in enumerate(eds.descriptions)}
    ev_ext = [eds.bounds.left, eds.bounds.right, eds.bounds.bottom, eds.bounds.top]
    ev_in = geometry_mask(cod1.geometry, out_shape=eds.read(1).shape, transform=eds.transform, invert=True, all_touched=True)

fs1d = fs1.copy()
fs1d["doy"] = fs1d["valid_date"].dt.dayofyear
fs1d["yr"] = fs1d["valid_date"].dt.year

fig, axes = plt.subplots(len(ev_meta), 2, figsize=(11.5, 4.1 * len(ev_meta)),
                         gridspec_kw={"width_ratios": [1, 1.25]})
for i, ev in ev_meta.iterrows():
    axM, axT = axes[i]
    arr = ev_bands[ev["disno"]]
    m = np.ma.masked_where(~ev_in | (arr < 0.05), arr)
    axM.imshow(np.where(ev_in, 0, np.nan), extent=ev_ext, cmap="Greys", vmin=-1, vmax=3)
    imm = axM.imshow(m, extent=ev_ext, cmap="Blues", vmin=0, vmax=0.8, zorder=2)
    cod1.boundary.plot(ax=axM, color="#9db1b3", linewidth=0.6, zorder=3)
    evd = cod2[cod2["ADM2_EN"].isin(ev["districts"].split(";"))]
    evd.boundary.plot(ax=axM, color="#b8272d", linewidth=1.1, zorder=4)
    axM.set_xlim(29.4, 35.1); axM.set_ylim(-1.6, 4.4); axM.set_axis_off()
    axM.set_title(f"{ev['label']}\nmax SFED over event window", fontsize=9.5)

    reg = fs1d[fs1d["pcode"] == ev["region"]]
    pivd = reg.pivot_table(index="doy", columns="yr", values="mean").rolling(7, center=True, min_periods=1).mean()
    ev_yr = pd.Timestamp(ev["start"]).year
    for yr in pivd.columns:
        if yr not in (ev_yr, 2026):
            axT.plot(pivd.index, pivd[yr], color="#c4d0d1", lw=0.5, alpha=0.5)
    axT.plot(pivd.index, pivd[ev_yr], color="#7f5619", lw=2.0, label=str(ev_yr))
    if 2026 in pivd.columns:
        axT.plot(pivd.index, pivd[2026], color="#1e795f", lw=2.0, label="2026 to date")
    x0 = pd.Timestamp(ev["start"]).dayofyear; x1 = pd.Timestamp(ev["end"]).dayofyear
    axT.axvspan(x0 - 1, x1 + 1, color="#b8272d", alpha=0.18)
    axT.set_xticks([1, 60, 121, 182, 244, 305, 366],
                   ["Jan", "Mar", "May", "Jul", "Sep", "Nov", "Jan"], fontsize=8)
    axT.set_ylabel("region-mean SFED", fontsize=8.5)
    axT.legend(frameon=False, fontsize=8.5, loc="upper left")
    axT.grid(color="#eef1f1", zorder=0)
    axT.spines[["top", "right"]].set_visible(False)
    nm1 = {"UG1": "Central", "UG2": "Eastern", "UG3": "Northern", "UG4": "Western"}
    axT.set_title(f"{nm1[ev['region']]} region daily flood extent | "
                  f"{int(ev['deaths'])} deaths, {int(ev['affected']):,} affected", fontsize=9.5)
plt.tight_layout()
plt.show()

Left column: maximum FloodScan SFED over each event window (padded a week each side); red outlines are the districts named in EM-DAT. Right column: daily region-mean SFED (7-day rolling mean, the team’s convention) for the event’s region; gray = every other year, brown = the event year, green = 2026 to date, red band = the event window.

Totals are not intensity: what a seasonal forecast can and cannot see

A seasonal forecast is a statement about three-month rainfall totals. Floods and landslides are triggered by rain at much shorter scales, so it is fair to ask how much the one says about the other. Two answers from the data:

• Totals carry real but incomplete information about bursts. Across the 112 region-seasons of the IMERG record (daily rain, 1998-2025), the OND total and the season’s largest 1-day and 3-day rains correlate at Spearman r = 0.70 and 0.73: about half the rank information is shared, and half is not. A wet season makes big bursts likelier; it does not schedule them.
• Events split exactly along that line. Taking every located EM-DAT event and asking how the two weeks before it ranked against the same calendar window in all other years: flood events sat under elevated 3-day bursts (median 73rd percentile, Wilcoxon p = 0.001, n = 36) and in elevated-total months (median 70th, p = 0.003). Landslide events sat under elevated bursts (median 79th, p = 0.019, n = 11) but their event-month totals were not significantly elevated (median 68th, p = 0.12).

One reconciliation, because these numbers can look contradictory. Three different links are in play. Rain-to-rain: wet seasons contain bigger bursts (r = 0.73 pooled; 0.59-0.75 within each region), the strongest link here. Rain-to-events: recorded flood events cluster in wet months (median 70th percentile, p = 0.003), so seasonal wetness does relate to the events people experience. Rain-to-extent: the season’s rainfall only weakly predicts total flooded area (r = 0.28), because area is dominated by what the wetland systems already hold, as 2025 showed. Rainfall predicts bursts and relates to impact events; it is specifically the year-to-year extent that runs on stored water.

Would monthly granularity help? Half of it would. Monthly totals are a better burst proxy than seasonal ones (monthly total vs that month’s largest 3-day rain: r = 0.82, against 0.73 for the season), and flooding inside OND is not uniform (November is the Central/Western climatological peak, and the 15 OND events in EM-DAT split Oct 4 / Nov 7 / Dec 4). But the forecast cannot deliver that granularity: SEAS5’s skill for individual OND months from the August issuance is r = 0.17-0.43 by region and month, mostly below the r >= 0.30 gate this document applies everywhere else, against 0.35-0.48 for the OND lump - the trimester average exists precisely to cancel the month-level noise. The one exception is November in the Northern and Eastern regions (r = 0.42-0.43), which happens to be the flood-peak month.

So the practical structure is a ladder of timescales. The trimester outlook says how wet the season is likely to be (extent-type flood risk, which tracks totals). The monthly climatology says when within the season flooding usually happens (November). Bursts and landslide timing belong to short-range rainfall monitoring (IMERG and weather-scale forecasts); no seasonal product reaches them. Watching the Elgon and Rwenzori slopes means the short end of that ladder, not seasonal totals.

imr = pd.read_sql("SELECT pcode, valid_date, mean FROM public.imerg WHERE iso3='UGA' AND adm_level=1",
                  engine, parse_dates=["valid_date"])
im0 = pd.read_sql("SELECT valid_date, mean FROM public.imerg WHERE iso3='UGA' AND adm_level=0",
                  engine, parse_dates=["valid_date"]).set_index("valid_date")["mean"].sort_index()
imr["yr"], imr["mo"] = imr["valid_date"].dt.year, imr["valid_date"].dt.month
iond = imr[imr["mo"].isin([10, 11, 12]) & (imr["yr"] < 2026)].sort_values("valid_date")
g = iond.groupby(["pcode", "yr"])
tot = g["mean"].sum()
mx3 = g["mean"].apply(lambda s: s.rolling(3).sum().max())

ev = emf.dropna(subset=["Start Year", "Start Month"]).copy()
ev["d"] = pd.to_datetime(dict(year=ev["Start Year"], month=ev["Start Month"], day=ev["Start Day"].fillna(15)))
def _burst_pct(d, win=14, agg=3):
    end = d + pd.Timedelta(days=2)
    obs = im0.loc[end - pd.Timedelta(days=win):end].rolling(agg).sum().max()
    hist = []
    for y in range(1998, 2026):
        try:
            e2 = end.replace(year=y)
        except ValueError:
            continue
        v = im0.loc[e2 - pd.Timedelta(days=win):e2].rolling(agg).sum().max()
        if not np.isnan(v):
            hist.append(v)
    return 100 * np.mean(np.array(hist) < obs)
im0m = im0.resample("MS").sum()
def _mo_pct(d):
    m = pd.Timestamp(d.year, d.month, 1)
    same = im0m[im0m.index.month == d.month]
    return 100 * np.mean(same.drop(index=m, errors="ignore") < im0m.get(m, np.nan))
ev["burst"] = ev["d"].map(_burst_pct)
ev["montot"] = ev["d"].map(_mo_pct)

fig, (axL, axR) = plt.subplots(1, 2, figsize=(11, 4.6), gridspec_kw={"width_ratios": [1.1, 1]})
RC = {"UG1": ("#1f69b3", "Central"), "UG2": ("#e69f00", "Eastern"), "UG3": ("#7f5619", "Northern"), "UG4": ("#2a7e43", "Western")}
for pc, (c, nm) in RC.items():
    axL.scatter(tot.loc[pc], mx3.loc[pc], s=26, color=c, alpha=0.8, label=nm)
axL.set_xlabel("OND rainfall total (mm)"); axL.set_ylabel("largest 3-day rain in season (mm)")
axL.legend(frameon=False, fontsize=8); axL.grid(color="#eef1f1", zorder=0)
axL.spines[["top", "right"]].set_visible(False)
axL.set_title("Totals vs bursts: related, far from identical (r = 0.73)", fontsize=10)

rows = [("Flood", "burst"), ("Flood", "montot"), ("Mass movement (wet)", "burst"), ("Mass movement (wet)", "montot")]
labels = ["floods:\n3-day burst", "floods:\nmonth total", "landslides:\n3-day burst", "landslides:\nmonth total"]
for k, ((typ, col), lab) in enumerate(zip(rows, labels)):
    yy = len(rows) - 1 - k
    vals = ev.loc[ev["Disaster Type"] == typ, col].dropna()
    axR.scatter(vals, np.full(len(vals), yy) + np.random.default_rng(7).uniform(-0.12, 0.12, len(vals)),
                s=34, color="#1f69b3" if typ == "Flood" else "#1f2324", alpha=0.75)
    axR.plot([vals.median()] * 2, [yy - 0.24, yy + 0.24], color="#b8272d", lw=2.4)
axR.set_yticks(range(len(rows)), labels[::-1], fontsize=8.5)
axR.axvline(50, color="#9db1b3", lw=1, ls="--")
axR.set_xlabel("pre-event rainfall, percentile vs other years (red = median)")
axR.set_xlim(0, 100); axR.grid(axis="x", color="#eef1f1", zorder=0)
axR.spines[["top", "right"]].set_visible(False)
axR.set_title("What preceded the recorded events", fontsize=10)
plt.tight_layout()
plt.show()

Left: each point is one region-season; the OND total against the season’s largest 3-day rain (IMERG, 1998-2025) - correlated (r = 0.73) but with wide spread at any given total. Right: rainfall conditions before each located EM-DAT event, as percentiles of the same calendar window in other years; flood events follow bursts and wet months, landslide events follow bursts only.

Reading the 2026 flood season

What follows is interpretation built on the pieces above: an outlook, not a statistical result. Both levers point the same way this year. The ocean drivers are aligned: a strong El Niño with a positive IOD crossing its threshold in August and forecast positive through December, the 2015/2023 configuration, which is what the 91st-98th percentile wet OND forecast in the section above is expressing. And the landscape is measurably wet already: 2023, 2024 and 2025 are the three largest flood ONDs of the FloodScan record (1998-2025), and 2026’s January-August standing flood extent ranks among the top three of all 29 years in every region (Central 1st, Eastern and Western 2nd, Northern 3rd). Neither guarantees a flood year: the driver-to-rain step is statistically solid, the rain-to-flood step is only categorically supported, and year-to-year persistence is suggestive at best. But the historically informative conditions are simultaneously present.

The closest analogues sharpen the point. Ranking every August-issued SEAS5 OND forecast in the 1981-2026 hindcast (46 issuances), 2026 is the fourth wettest, the 93rd percentile. The only wetter ones are 1997 (the strongest El Niño + IOD combination on record) and 1994 (a major positive-IOD year), both before FloodScan coverage begins, and 2015. Within the FloodScan era, the four wettest August-issued outlooks all flooded above the median season: 2015 (96th percentile; strong El Niño, and at 5.9% the largest flood extent outside 2023-25), 2006 (91st; Teso floods, with Kumi and Butaleja among eight districts affected in EM-DAT), 2023 (89th; the largest El Niño-year extent, though EM-DAT records no OND event that season), and 2011 (87th; 4.7% flooded, the seventh-largest OND extent of the 28). Just below that set, 2019 (83rd) also flooded above the median and was the deadliest OND in the record, with 104 deaths across the Rwenzori and Elgon events. The counterexamples run the other way. 2018 (33rd) was not forecast wet, and its worst damage came from a rainfall-triggered landslide at Bududa, not from large flood extent. 2025 (30th) was not forecast wet either, yet produced the record extent: the water was already in the system from 2023-24, not delivered by that season’s rain. In this record a wet August forecast has usually been followed by a wet flood season, but plenty of flooding has happened without one. Practically: treat the amplification map’s Teso/Kyoga core and its expansion fringe as the flood-watch zone for Oct-Dec, watch Kasese and the Elgon corridor for the flash-flood and landslide impacts EM-DAT says cause fatalities, and re-check at the September and October SEAS5 issuances alongside FloodScan itself once the rains start.

Method notes & caveats

• Forecast selection is the app’s rule verbatim (code imported from pipeline/export_static_site.py). Skill and return periods come from the repo’s src.skill.run_all_combinations, detrended variant, identical to the Forecast × HNRP tab.
• The district climate stats are new. Uganda sits outside the rasterstats DB’s adm2 coverage, so SEAS5/ERA5 zonal statistics over the 135 CODAB districts were computed with exactextract from the same COGs (pipeline/compute_uga_district_stats.py), then pipeline/compute_skill_uga_adm2.py ran the standard skill code. Region-level exactextract means reproduce the DB’s stored means within about 1%.
• JIAF 2.0 Light covers 22 districts (× host/refugee groups; the raw sheets’ 26 distinct “District” values include four subtotal rows); all other districts appear on the map for climate context only. Joint PiN is the mosaic-rule maximum of sectoral PiN, not their sum. Severity classifies areas, not people.
• Targeted/reached levels are independent rollups in the source workbooks (the region series does not equal the sum of its districts); this document only ever uses one level at a time.
• The empirical return period saturates at 46 years (record length + 1). A district shown at 46 is “drier than anything since 1981”, not “exactly a 46-year event”. Use the deficit (% of normal) to compare districts once the cap binds.
• Flood layers: FloodScan daily SFED (0.083 deg) seasonal maxima per OND; flooded = SFED >= 0.05 (the noise floor used by ds-floodexposure-monitoring); return periods are rank-based empirical ((n+1)/rank), same convention as the drought side. Climate indices are OND means of NOAA PSL’s DMI (HadISST) and ONI series, stored at raw/climate_indices/ for reproducibility. EM-DAT events are located by matching district names in the Location text (42 of 47 matched); flood and wet mass-movement (landslide) events are analyzed as separate types. EM-DAT under-records slow wetland inundation, records no impact where flooding meets no people, and its reporting density grows over time, so event counts are a geography of recorded impacts, not a hazard climatology.
• Intensity analysis: IMERG daily rainfall from public.imerg (adm0/adm1, 1998-2025); pre-event windows are the 14 days ending two days after the EM-DAT start date, compared against the same calendar window in all other years; monthly EM-DAT start dates without a day use the 15th.
• Terego (created 2020) is not in the CODAB vintage: its rows keep their own people figures but inherit UG3072 (Arua)’s forecast; “Madi-Okollo & Terego” inherits UG3084 (Madi Okollo).