Three Seasons in the Sahel

Last November, I stood in a millet field thirty kilometres south of Mopti with a handheld spectrometer and a notebook already half-full of contradictions. The satellite imagery for this exact parcel — downloaded the night before from a hotel room in Bamako with spotty WiFi — classified it as "severely degraded, below-replacement productivity." The farmer beside me, a man named Ousmane Cissé who has worked this soil since 1998, had just finished showing me the root systems of his intercropped cowpea and sorghum. The soil was dark, crumbly, alive. We dug down thirty centimetres and found earthworms. The satellite had never turned a single shovel.

This is not an indictment of remote sensing. The NDVI time series that flagged Cissé's field was correct about surface reflectance: the sparse canopy in late dry season does look like degradation from 700 kilometres up. But the conclusion it enabled — the one that appears in dashboards, policy briefs, and funding allocation spreadsheets — was wrong in ways that matter for the 23 million people farming the Sahelian belt. The field was not degraded. It was following a fallow-and-rotation rhythm that no pixel-based classifier is trained to recognise.

What the indices cannot index

Between January and March this year, our team walked 147 farm plots across the Ségou and Mopti regions, conducting structured interviews and taking soil cores at each site. We cross-referenced every plot against three widely used land-cover products: the ESA WorldCover 2021, the Copernicus Global Land Service dry-matter productivity layer, and a custom Google Earth Engine composite built by our partners at the Institut d'Économie Rurale in Bamako. On 41 percent of the plots, at least two of the three products misclassified the land-use status. The errors were not random noise; they clustered systematically around smallholder polyculture systems and fields in active fallow recovery.

The underlying problem is not sensor resolution — it is category design. Global land-cover taxonomies are optimised for large, contiguous, monoculture landscapes. The patchwork of intercropped millet, cowpea, and baobab that defines Sahelian agroecology reads as "mixed" or "degraded" to classifiers that were trained on European and North American training datasets. The categories are clean. The ground is not.

On 41 percent of the plots, at least two of the three products misclassified the land-use status.

What surprised us most was not the misclassification rate itself — colleagues in dryland ecology have been documenting this gap for years — but the downstream consequences in funding pipelines. We traced fourteen separate grant decisions made between 2019 and 2023 by international donors operating in central Mali. In eleven of them, the initial site-selection shortlist was generated from remotely sensed degradation indices. Plots like Cissé's were either excluded from intervention zones or flagged as "priority restoration" sites that needed external inputs. The farmers who had been building soil carbon for decades were told, by the data, that their land was broken.