Intel methodology

News (C-INTEL)

The bottom line: Each week C-INTEL reads the African conservation press across 46 in-scope countries and surfaces only the few items that should change what a grant officer does next.

C-INTEL screens the African conservation press and surfaces the handful of items each week that should move a grant officer's hand. It is screening, not due diligence: items are neutral, attributed, and carry no editorial synthesis. Its in-scope set spans 46 African countries, and items surface as qualifying press appears.

How an item reaches the digest

1. Ingest articles from a registered set of regional and national outlets (English and Portuguese).
2. Classify each article with a language model against a fixed taxonomy: in-scope countries, themes, severity, and which protected areas are named.
3. Score relevance and confirm protected-area matches.
4. Translate non-English items to English.
5. Summarise with a draft-then-critic pass, so each summary is checked against the source before it ships.
6. Assemble the weekly digest, ordered by severity.

Screening philosophy

The filter is exclusion-based with a positive bar. It deliberately drops whole categories of noise (listed under Exclusions), and keeps only items that clear the bar of being materially useful to a funding decision. The model is told to be conservative: when unsure, it under-flags rather than over-flags, and prefers "informational" to "material".

Tiers drive priority

Every topic in the taxonomy carries a tier. The tier sets the default severity and where the item lands in the digest. This is the key point for proposing themes: a topic is not just a label - its tier decides how loudly it surfaces.

Theme taxonomy

Twelve themes, each with tiered topics. This is a working draft, open to revision.

Entities tracked

Alongside the themes, the system actively watches named entities. These are also open to additions.

Exclusions

The following categories are deliberately filtered out:

• Zoo and captive-breeding news
• Aquarium and dolphinarium news
• Domestic dog and cat welfare
• Wildlife documentary releases (unless tied to operator news)
• Generic Africa wildlife-wonder tourism puff
• Generic climate policy (only surface when keystone-country-specific)
• Stock market news (unless directly affecting listed conservation-relevant entity)
• Capital-city-only mentions without conservation context

A source blocklist additionally hard-excludes specific outlets and authors judged unreliable for evidentiary purposes, so their material is never treated as a signal.

Fire

The trigger, in plain terms: a park is flagged when it burns hotter than its own normal range for the season and stays elevated for several days running. Because the bar is each park's own history rather than a fixed number, parks that burn every year are not flagged for burning - only for breaking their own pattern.

The fire watch surfaces anomalous fire activity at keystone protected areas, drawn from NASA satellite detections. It is a screening signal: it points to a protected area that may warrant a closer look, not a confirmation of cause or damage.

Where the data comes from

Detections come from NASA FIRMS (Fire Information for Resource Management System), the near-real-time active-fire product derived from the VIIRS and MODIS satellite instruments. The feed ingests these detections daily.

How an event is built

1. Ingest the day's active-fire detections for every keystone with a known boundary.
2. Aggregate and cluster detections to each protected area.
3. Score the day against that park's own recent-history baseline (a rolling window of roughly the past 60 days), not against an absolute number.
4. Track persistence: how many consecutive days a park stays elevated, and whether the streak is still running or has recently ended.
5. Select the events worth surfacing and classify each against the press record.

What "anomalous" means

A park is flagged when its fire activity is high relative to its own normal range for the season, not when it simply has fire. Many African savanna parks burn every year, and a fixed threshold would flag them constantly. Scoring against each park's own baseline means a normally fire-prone park is only flagged when it departs from its own pattern.

This-time-of-year context

The daily flag above measures each park against its own recent weeks. That is the right screen for "is something happening now", but it cannot say whether the fire is normal for the time of year. A separate layer answers that: roughly 24 years of MODIS fire history (2003 on) folded into a week-of-year climatology for each park, a typical range for every week of the calendar built from all the prior years. The current year is read against it and given a plain verdict, normal, above, or unusual for the season. It appears as the Fire seasonality chart on each park page and the vs the seasonal normal line in the fire alert. Because the norm is learned per park, hemisphere and biome are automatic: a southern savanna's August peak, a northern park's January one, and a rainforest's year-round near-zero all fall out with no hand-coded rule. An absolute floor keeps a normally near-quiet park from flagging on a few stray pixels. This layer is context only: like every satellite signal it never moves a PACE score, and it is a coarser 1km MODIS lens shown alongside, not in place of, the recent-history flag above.

The press tags

Every fire event is checked against the C-INTEL press corpus and carries one of three honest labels: press corroborated (a matching report exists, with the lead or lag in days recorded), scoop (inside the in-scope footprint but no matching press found), and no press baseline (outside the in-scope footprint, so there is no press to compare against - stated plainly because it is not a claim the event went unreported, and not a scoop).

Dates

A fire event is stamped with the day it was last flagged by the satellite (shown as flagged [date]). This is a detection date, not a publication time.

Coverage and limits

• Coverage. 161 of the 162 declared keystones have a usable boundary and are scored. The remaining one (Boumba-Djombi) is a genuine keystone with no publicly available boundary; it is documented as such and held out of geometry scoring rather than dropped or given a fabricated extent.
• Bounding boxes, not exact polygons. Scoring uses each park's bounding box, not its precise mapped boundary. A fire inside the park and a fire just at its edge are not yet distinguished.
• Cloud and canopy. Optical fire detection under-sees fire beneath heavy cloud or dense forest canopy. A quiet reading over a cloud-forest park is a limit of the sensor, not proof of no fire.
• What it does not claim. The watch does not identify a fire's cause and does not confirm damage. It flags that something is anomalous and worth examining.

Provenance

The pipeline runs daily. Every event carries the version of the scoring logic that produced it, the time it was scored, and a short statement of the basis for the score.

Forest

The trigger, in plain terms: a park is flagged when it loses canopy faster than its own recent pattern. The comparison is always against the park's own baseline, so a normally stable forest stands out the moment its loss accelerates.

The forest watch tracks forest-cover loss at keystone protected areas - clearing, conversion, and degradation visible from space. Like the fire watch, it is a screening signal: it points to a protected area whose canopy is changing faster than its own recent pattern, not a confirmation of cause.

Status

Live. The daily job has accrued one baseline reading per park per day since the seed on 20 May 2026, and anomaly scoring now surfaces flagged parks into the unified feed as standalone satellite items. A flagged park carries a Satellite signal badge, or Provisional while its baseline window is still short.

Where the data comes from

The live signal is DIST-ALERT, the near-real-time land-disturbance alert layer derived from harmonised Sentinel-2 and Landsat optical imagery. It sits on the established global products - JRC Tropical Moist Forest and the Hansen Global Forest Change series - which provide the longer-run context named in the press taxonomy under landscape-scale forest-cover change.

Roadmap

The baseline, anomaly scoring, and feed band are live. The pipeline scores each park's loss signal against its own accrued baseline, tracks multi-day persistence rather than a fixed threshold (mirroring the fire watch), and renders flagged parks into the feed as standalone satellite items. Forest carries no press pairing - every item is a pure satellite signal, labelled as such rather than posed as a press-matched event.

1. Radar complement (later build). Add a cloud-penetrating radar alerting layer to close the optical blind spot described below.

Vegetation

The trigger, in plain terms: a park is flagged when its greenness falls below what is normal for that park at that time of year. Scoring against each park's own seasonal curve means ordinary dry-season browning is ignored while a real drought or dieback still surfaces.

The vegetation watch is intended to track greenness and vegetation-condition anomalies - the drought stress, browning, and seasonal shifts that bear on forage, water, and habitat at keystone protected areas. It is designed as a screening signal of the same family as fire and forest: a departure from a park's own seasonal norm, not an absolute reading.

Status

Live. The watch scores each park against its own seasonal greenness baseline and surfaces flagged parks into the unified feed as standalone satellite items. A flagged park carries a Satellite signal badge, or Provisional while its baseline is still shallow.

Where the data comes from

The live signal is MODIS NDVI - the Normalised Difference Vegetation Index, a satellite-derived measure of how green and healthy vegetation appears, with a long, stable record well suited to seasonal baselining. It is calculated from how plants reflect light: healthy, actively growing plants absorb red light for photosynthesis and reflect a lot of near-infrared, while sparse, stressed, dead, or absent vegetation reflects less near-infrared and so scores lower.

where NIR is the near-infrared light reflected by the surface and Red is the red light reflected. The index runs from -1 to +1. Typical ranges: around 0.6 to 0.9 for dense, healthy vegetation; around 0.2 to 0.5 for grassland, shrubs, and sparse cover; and around 0 or below for bare soil, water, and built-up surfaces. An anomaly is a drop in NDVI against the park's own normal range for the time of year, so ordinary seasonal green-up and senescence are normalised out and a genuine browning or drought-driven decline still surfaces.

Roadmap

Source selection, seasonal baseline, anomaly scoring, and the feed band are live. The pipeline flags parks departing from their own seasonal range, tracks persistence, and renders them into the feed as standalone satellite items. As with forest, vegetation carries no press pairing - every item is a pure satellite signal, labelled as such.

1. Higher-resolution complement (later build). Add a Sentinel-2 vegetation-index layer to sharpen the coarse MODIS footprint over smaller parks.

Water

The trigger, in plain terms: a park is flagged when its open water shrinks below the normal level for that calendar month. Measuring each month against its own history means the yearly flood-and-dry cycle is expected, and only a genuine shortfall is surfaced.

The water watch tracks surface-water extent at keystone protected areas - the shrinking and filling of rivers, floodplains, pans, and reservoirs that drive wildlife movement and dry-season survival. As with the other satellite signals, it is designed to flag a departure from a park's own pattern, not a raw water level.

Status

Live. The watch scores each park against its own month-of-year surface-water baseline, built from a three-year backfill, and surfaces flagged parks into the unified feed as standalone satellite items. A flagged park carries a Satellite signal badge, or Provisional while that calendar month still holds a thin sample.

Where the data comes from

The live signal is Sentinel-2 surface reflectance. Open water is measured as the share of the park footprint flagged as water by the normalised difference water index, after a scene-classification cloud and shadow mask drops scenes too obscured to trust. An anomaly is a drop in that open-water fraction against the same calendar month across the backfill, so the ordinary wet-dry seasonal swing is normalised out and a drought-driven drying still surfaces.

Roadmap

The baseline, anomaly scoring, and feed band are live. The pipeline scores each park's open-water fraction against its own month-of-year median, surfaces departures below that seasonal normal, and renders flagged parks into the feed as standalone satellite items. Water carries no press pairing - every item is a pure satellite signal, labelled as such.

1. Bidirectional anomalies (later build). The current watch surfaces unusual drying only. Unusual flooding and expansion is a real signal but seasonal flood pulses are normal, so a bidirectional version waits until the drops distribution is trusted.

Night-time lights

The trigger, in plain terms: a park is flagged when new light appears at night, brighter than its own normal for that calendar month. Stable existing lights - a lodge, a gate post, a nearby town's glow - are part of the baseline and are expected; only a genuine rise above that pattern is surfaced. New light where there was none is a proxy for human encroachment: settlement, camps, mining, or road lighting.

Status

Live, in beta. The watch scores the full keystone cohort monthly and surfaces flagged parks into the unified feed as standalone satellite items. Because the source is a monthly composite that matures around five weeks after the month ends, each item states its composite month (the activity) alongside the date it surfaced (the detection) - a July item about April light says so.

Where the data comes from

The live signal is the NOAA VIIRS Day/Night Band monthly composite - calibrated low-light imaging designed to see faint artificial light at night. The watch reads the park's average radiance from each new composite and compares it against the park's own month-of-year baseline, so moonlight-corrected seasonal patterns and stable existing lighting are normalised out.