Is a drought quietly building across my region before the harvest fails?

In the Sahel and the Horn of Africa, early drought detection is the difference between a managed shortfall and a famine that displaces millions. The point of this page is to spot the problem before the crop visibly fails — when soil water has already dropped below normal but the fields still look green.

What you can answer

Whether soil water is below normal for this time of year (FLDAS soil-moisture anomaly vs the 1982–present monthly climatology — this is the famine-early-warning signal).
How dry the root zone and surface are right now, day by day (SMAP enhanced 9km soil moisture, updated within ~2–3 days of acquisition).
Whether crops are already under water stress (ECOSTRESS Evaporative Stress Index — ESI near 0 means plants have closed stomata and stopped transpiring, a direct early sign of crop failure).
Whether the dryness is widespread or patchy by comparing the coarse model field (FLDAS, 0.1°) with the fine ECOSTRESS field (~70m) over the same AOI.

What you can NOT answer with these datasets alone

Final yield in tonnes — these are biophysical stress indicators, not a crop-yield model. Join with FEWS NET / WRSI or a crop model for that.
Cause of the dryness (failed rains vs irrigation withdrawal vs heatwave) — needs IMERG precipitation and land-use/irrigation data.
Sub-field detail every day — ECOSTRESS is high-resolution but revisits irregularly (it flies on the ISS), so you get sharp snapshots, not a daily 70m movie.
Future rainfall — none of these forecast; they tell you the current and recent state. Pair with a seasonal forecast for outlook.

Code template (Python, cloud-direct, ~35 lines)

import earthaccess
import xarray as xr
import numpy as np

earthaccess.login(strategy="netrc")

# 1. Define AOI + the months that lead up to harvest
aoi = (35.0, 2.0, 42.0, 8.0)  # W, S, E, N — Horn of Africa (S. Ethiopia / N. Kenya)
season = ("2024-03-01", "2024-06-30")   # long-rains growing season

# 2. SMAP enhanced daily soil moisture (9km) — the "right now" signal
smap = earthaccess.search_data(
    short_name="SPL3SMP_E",
    bounding_box=aoi,
    temporal=season,
)
# ...open the HDF5 'Soil_Moisture_Retrieval_Data_AM/soil_moisture' group,
#    subset to AOI, average over space -> daily AOI mean soil moisture

# 3. FLDAS famine-warning land model (monthly 0.1°) — current month + baseline
fldas_now = earthaccess.search_data(
    short_name="FLDAS_NOAH01_C_GL_MA",
    bounding_box=aoi,
    temporal=season,
)
fldas_clim = earthaccess.search_data(
    short_name="FLDAS_NOAH01_C_GL_MA",
    bounding_box=aoi,
    temporal=("1991-01-01", "2020-12-31"),  # 30-yr baseline, same calendar months
)
# ...open 'SoilMoi00_10cm_tavg', build the per-month climatology mean & std,
#    then anomaly = (this_month - clim_mean) / clim_std   (a drought z-score)

# 4. ECOSTRESS ESI (~70m) — confirm crops are actually stressed
esi = earthaccess.search_data(
    short_name="ECO_L4T_ESI",
    bounding_box=aoi,
    temporal=season,
)
# ...open the tiled GeoTIFF, ESI -> 0 = severe stress, ~1 = unstressed

# 5. Verdict: drought building if FLDAS anomaly < -1 sigma AND SMAP trending down
#    AND ESI dropping toward 0 across the cropped pixels.
ds = xr.open_dataset(earthaccess.open(fldas_now[:1])[0])   # auth'd cloud-direct read
print("FLDAS top-layer soil moisture:", ds["SoilMoi00_10cm_tavg"].dims)

Expected output

Map: FLDAS soil-moisture anomaly (z-score) across the AOI for the current month — red where soil water sits more than 1 standard deviation below the 30-year normal.
Time-series: AOI-mean SMAP daily soil moisture for the growing season overlaid on the prior-year track, so a downward divergence is obvious early.
Map: ECOSTRESS ESI snapshot over the cropped area — low ESI patches mark fields that have already started shutting down.
One-line verdict combining the three: “Soil water is N sigma below normal, declining, and crop stress is rising — drought building before harvest.”

Caveats

SMAP top ~5cm only: it senses the surface, not the deep root zone. Use FLDAS layered soil-moisture for root-zone context; treat SMAP as the fast daily pulse.
FLDAS is a model, not a measurement: it assimilates rainfall (CHIRPS) and met forcing, so its skill follows the quality of rainfall inputs — sparse-gauge regions are weaker.
ECOSTRESS revisit is irregular and clouds block the thermal sensor; expect gaps, especially during rains. It confirms stress where you have a clear scene rather than providing continuous coverage.
Anomaly baseline matters: a drought “below normal” depends entirely on the climatology window you pick — state it explicitly (here 1991–2020).

Cross-DAAC composition

This is a 3-DAAC join: GES DISC (FLDAS) + NSIDC DAAC (SMAP) + LP DAAC (ECOSTRESS). Auth is uniform (Earthdata Login via earthaccess), but the formats differ — FLDAS and SMAP are HDF5/NetCDF on regular grids, while ECOSTRESS ships as tiled GeoTIFFs. Regrid the fine ECOSTRESS field onto the FLDAS grid (or vice versa) before comparing. See recipes/r01-three-daac-composition.mdx for the general pattern.