CipherSenseAI/africa-cloud-cover-bias

Cloud Cover and Structural Observation Gaps in African Agricultural EO (six-zone dataset)

Supporting data for the paper "Cloud Cover and Structural Observation Gaps in African Agricultural Earth Observation: Evidence from Six Agroecological Zones" (Olaoye Anthony Somide, CropSense AI Research / CipherSense AI; Zenodo, doi:10.5281/zenodo.22642336; also EarthArXiv, doi:10.31223/X5J503).

Weekly usable Sentinel-2 optical observation frequency over cropland for six administrative zones (one per African sub-region, spanning the continental climate gradient) across the 2022–2024 growing seasons, plus the matched CHIRPS rainfall series and the analysis code needed to reproduce every figure and statistic in the paper.

Summary

Property	Value
Zones	6 admin-1 (ADM1) zones, one per African sub-region (two in West Africa)
Period	2022-01-01 to 2024-12-31 (156 ISO weeks per zone)
Primary table	2,808 rows (6 zones × 156 weeks × 3 SCL thresholds)
Satellite source	Sentinel-2 L2A Scene Classification Layer (8,986 scenes)
Rainfall source	CHIRPS v2.0
License	CC BY 4.0

Key findings (from the paper)

• Mean weekly usable-observation frequency over cropland ranges from 55% (humid Central Africa) to 81% (arid Nile Delta), a 26-point spread; "blind-spot weeks" (<50% of cropland observable) range 12%–42%.
• The one firmly supported inferential result: within every zone, usable-observation frequency is negatively correlated with weekly rainfall (Spearman ρ −0.34 to −0.70, p < 0.001 after adjusting for temporal autocorrelation; the comparison is internal to a single zone).
• The cross-regional gradient and the concentration of the deficit in the rain-fed growing season are descriptive patterns, not powered tests: the study has six spatial sampling units. The regional difference is significant at the zone-year level (n=18, Kruskal–Wallis p = 0.01); the ~14-point growing-season deficit is only marginal by circular block permutation (pooled p = 0.03, one of six zones individually). A naive Kruskal–Wallis on 936 zone-weeks would report p ≈ 10⁻⁹; that is pseudoreplication and the paper does not use it.
• Nile-irrigated Egypt inverts the seasonal pattern, as the rainfall mechanism predicts.
• The gradient is stable across the SCL threshold (strict/moderate; the two differ only in whether water/unclassified/snow-ice count as usable, a <1.5-point shift), the choice of year (its ends), the cropland mask (<2.3-point shift), and the scene-level cloud pre-filter (<1.2-point upper bound, cloud_prefilter_sensitivity.csv).

Files

data/

File	Rows	Description
weekly_usable_observation_frequency_2022_2024.parquet	2,808	The study dataset. Zone × ISO-week × SCL threshold: usable / observed cropland pixel counts and percentages, scene and tile counts. Cropland-masked (ESA WorldCover class 40).
weekly_usable_observation_frequency_2022_2024_unmasked.parquet	2,808	Same, over the full ADM1 polygons (no cropland mask); the Robustness §6.3 dataset.
chirps_weekly_rainfall_2022_2024.parquet	936	Zone × ISO-week rainfall total and mean (CHIRPS v2.0), keyed to match the observation table exactly.
chirps_weekly_rainfall_2022_2024_daily.parquet	6,576	Daily zonal rainfall.
zones.csv	6	Zone lookup: region, country, area, cropland fraction, scene count, climate/system label (paper Table 1).
crop_stress_windows.csv	7	Estimated mid-season window per zone: the middle third of the sow→harvest cycle (a calendar approximation of maize flowering/grain-fill), with the FAO calendar mapping and its confidence.
maize_fao_calendar.jsonl	40	The FAO GIEWS maize sowing/harvesting records for the six countries, as ingested.
zero_item_weeks_investigation.csv	15	The 15 zone-weeks that returned no scene under the cloud pre-filter, re-queried without it (paper §5.4).
cloud_prefilter_sensitivity.csv	6	Per-zone bound on the effect of the eo:cloud_cover < 95 scene-level pre-filter: scenes it dropped, and the upper bound on each zone's mean if they were all restored (paper §6.5).
modis_crosscheck.csv	6	Per-zone SCL-vs-MODIS agreement: mean SCL usable %, mean MODIS clear-sky %, the bias, and the SCL/MODIS and MODIS/rainfall rank correlations (paper §6.6).
modis_crosscheck_weekly.csv	936	Zone × ISO-week: SCL usable %, MODIS (MCD06COSP) cloud fraction and clear-sky %, contributing MODIS days, and CHIRPS rainfall; the table behind §6.6.
phase2_results_table.csv	6	Consolidated per-zone results (≈ paper Table 2).
smoke_test_output.json	—	Per-scene usable-pixel percentage for one rainy-season week across seven zones (methods sanity check).

scripts/

The analysis code: pandas / scipy / matplotlib on the tables above. This is not the acquisition pipeline (STAC search + windowed COG reads), which is not part of this release; the paper describes the method in enough detail to reimplement it.

Script	Reproduces
analyze_independence.py	The paper's inferential numbers. Regional gradient at the zone (n=6) and zone-year (n=18) level + a GEE with zone clusters; within-zone rainfall correlation with autocorrelation-adjusted p and moving-block bootstrap; crop-stress by circular block permutation (§4.3, §5.1–5.3)
analyze_phase1.py	Descriptive per-zone / per-region baseline (§5.1). Its pooled Kruskal–Wallis is pseudoreplication for a regional claim and is not used as evidence; see the file header and analyze_independence.py
analyze_phase2.py	Rainfall correlation, same-week and lagged, per zone (§5.2)
crop_stress_overlay.py	The descriptive mid-season-window gap and the FAO-calendar → window mapping (§5.3). Its Mann–Whitney overstates significance; the block-permutation test is in analyze_independence.py
analyze_phase3.py	SCL-threshold, season-year and cropland-mask sensitivity (§6)
make_figures.py	The six figures + phase2_results_table.csv

figures/

The paper's figures as PNG (also embedded in the paper PDF).

Load

import pandas as pd

obs  = pd.read_parquet("hf://datasets/CipherSenseAI/africa-cloud-cover-bias/data/weekly_usable_observation_frequency_2022_2024.parquet")
rain = pd.read_parquet("hf://datasets/CipherSenseAI/africa-cloud-cover-bias/data/chirps_weekly_rainfall_2022_2024.parquet")

merged = obs[obs.tier == "moderate"].merge(rain, on=["zone", "iso_year", "iso_week"])

Reproduce the paper's numbers

pip install -r scripts/requirements.txt
python scripts/analyze_independence.py  # the inferential numbers (unit-appropriate)
python scripts/analyze_phase1.py        # descriptive per-zone / per-region baseline
python scripts/analyze_phase2.py        # rainfall correlation, per zone
python scripts/crop_stress_overlay.py   # descriptive crop-stress gap + calendar mapping
python scripts/analyze_phase3.py        # SCL-threshold / year / cropland-mask robustness
python scripts/make_figures.py          # figures + results table

Column reference

weekly_usable_observation_frequency_2022_2024[_unmasked].parquet

Column	Type	Description
zone	str	<ISO3>/<ADM1 name>, e.g. COD/Kasai-Oriental
region	str	African sub-region
iso_year, iso_week	int	ISO 8601 week
week_start	str	Monday of the ISO week (YYYY-MM-DD)
tier	str	SCL usable-pixel threshold: strict, moderate (headline), or lenient
n_items	int	Sentinel-2 scenes composited for the zone-week
n_tiles	int	Distinct MGRS tiles touched
total_in_zone_pixels	int	Cropland pixels in the zone (full-polygon pixels in the _unmasked file)
observed_pixels, observed_pct	int, float	Pixels with any valid data that week
usable_pixels, usable_pct	int, float	Pixels classified usable by at least one scene that week; usable_pct is the paper's primary variable

chirps_weekly_rainfall_2022_2024.parquet

Column	Type	Description
zone, region, iso_year, iso_week, week_start		join key, as above
n_days	int	CHIRPS days in the week (7 except at the series ends)
rainfall_mm_sum, rainfall_mm_mean	float	Zonal-mean rainfall, weekly total and daily mean

See DATASHEET.md for the remaining files, the pre-specified hypotheses, and caveats.

Citation

@misc{somide2026cloudcover,
  title     = {Cloud Cover and Structural Observation Gaps in African Agricultural
               Earth Observation: Evidence from Six Agroecological Zones},
  author    = {Somide, Olaoye Anthony},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22642336},
  url       = {https://doi.org/10.5281/zenodo.22642336}
}

License

CC BY 4.0. Derived from Sentinel-2 (Copernicus), CHIRPS (Climate Hazards Center), ESA WorldCover, geoBoundaries, and FAO GIEWS; see LICENSE.