Bio-credit scoring

A conventional lender scores a borrower on repayment history. A smallholder farmer usually has none, so the alternative is to score the field instead: infer productive capacity from how the crop reflects sunlight, and lend against that. The idea works because the measurement is real. Its limits are real too, and they are set by optics and by weather correlation rather than by modelling effort.

What the index actually measures

Green leaves absorb strongly in the red — chlorophyll’s absorption band — and scatter strongly in the near-infrared, because leaf mesophyll is structurally transparent at those wavelengths. The contrast between the two is large and gets larger as canopy develops, which is what vegetation indices such as NDVI, the normalised difference of the near-infrared and red bands, exploit. The signal is not “yield”. It is roughly the amount of photosynthetically active canopy the sensor can see.

That distinction sets the first limit. The index saturates: once leaf area index passes roughly three, the upper leaves already absorb nearly all incident red light and additional layers change the reflectance very little. A dense, well-managed crop and an exceptional one look similar from orbit at peak greenness. Second, yield is not biomass. It is biomass multiplied by a harvest index, and grain filling happens after peak canopy, so a heat event during flowering can destroy yield while leaving the season’s greenness curve looking normal.

Pixels, clouds and field size

Sentinel-2 carries its red and near-infrared bands at 10 m ground sampling with a five-day revisit from the two-satellite constellation, and the data is open. Ten metres is adequate for a several-hectare field and marginal for a quarter-hectare one, where most pixels straddle a boundary and mix the target crop with a neighbour’s, a path or a tree line. Mixed pixels do not average out; they bias the estimate toward whatever surrounds the field.

Cloud is the harder constraint, because optical sensing simply stops. In humid tropical growing seasons the usable clear-sky observations can fall to a handful across the whole cycle — and they cluster in dry spells, which is a non-random sample of the season. Radar, as on Sentinel-1, sees through cloud, but it responds to canopy structure and surface moisture, not chlorophyll, so it is a different observable rather than a substitute.

The constraint imagery cannot lift

Even a perfect yield estimate leaves the structural problem of agricultural lending. A credit decision is taken before sowing, about a season no sensor has yet observed; the imagery describes history, and history predicts the future only while climate and management are stationary.

Worse, the risk is correlated. Consumer defaults are largely idiosyncratic, so a large portfolio diversifies them away. A drought hits every borrower within the same rainfall system in the same season, so a book of satellite-scored farm loans concentrated in one district behaves like a single large exposure regardless of how many names are in it. This is why such lending is usually paired with index insurance or a guarantee rather than priced purely off the score — and why geographic spread, not better imagery, is the binding requirement.