Skip to main content

Cropland and conflict in Cabo Delgado

Did armed conflict make agriculture move?

In progress

What it aims to achieve

The project asks whether armed conflict in Cabo Delgado, in northern Mozambique, caused agriculture to move, whether by shrinking, relocating, or expanding.

The original question was narrower, asking whether existing farmland had deteriorated. The broader framing matters for two reasons. First, fields abandoned because of conflict usually revert to scrub, so vegetation tends to become greener rather than browner. Second, displaced people often clear new land where they settle. A fixed cropland mask would miss both of these effects.

The study compares six conflict-core districts with two low-conflict control districts, Chiure and Namuno. The aim is an open and reproducible method that can then be applied in the Democratic Republic of Congo, South Sudan, and Ethiopia. Making it transferable is a core design rule, and it sets this work apart from the one comparable local study (Bacar & Faque, 2025).

How it works

Satellite data

  • Sentinel-2 (10 m) is the main optical source from 2017 on.
  • Sentinel-1 radar (VV/VH backscatter) sees through wet-season cloud and picks up field structure.
  • Landsat 2015–2017, pooled into one composite, gives the pre-conflict baseline, because 2016 alone was too cloudy.
  • All data come free from the Microsoft Planetary Computer, and the analysis runs on a laptop or in Google Colab.

Method

  • Each pixel gets a smoothed time series for each agricultural year. Phenology features are then extracted from it, including amplitude, seasonal integral, timing, and dry-season baseline.
  • Each year, every pixel across the whole landscape is classified as cultivated, natural, water, or settlement, rather than relying on a fixed cropland mask.
  • Movement is measured as new cultivation, abandonment, net change, centroid shift, and distance to the cultivation frontier.
  • Because shifting cultivation moves fields in any case, “normal” movement is measured from the baseline years and the control districts, and the conflict effect is estimated as the departure from that norm.

Conflict and confounders

  • Conflict: UCDP georeferenced event data give a conflict-intensity gradient, used in a difference-in-differences or panel design.
  • Drought: CHIRPS rainfall is the drought control.
  • Cross-checks: IOM Displacement Tracking Matrix data and FEWS NET/IPC food-security phases are used to check the results.

Accuracy

  • A photo-interpreted reference sample.
  • A random forest classifier checked by spatial cross-validation and an area-of-applicability test.
  • Error-adjusted area estimates with confidence intervals (Olofsson et al.), so every area figure comes with its uncertainty.

Back to all work