What a low-cost multispectral rig can and can't tell you about pasture

The question. Farm advisors keep being sold aerial NDVI as a decision tool. The gear that makes it affordable at farm scale is not the gear the research was validated on. So: how far can you trust a cheap multispectral rig before a decision needs ground truth behind it?

What we did. Six weeks, the same twelve Waikato paddocks, flown weekly at a fixed altitude and time of day, with a low-cost multispectral setup. Every flight paired against calibrated reference readings and physical cuts on a subset of plots. Same pilot, same flight plan, same processing pipeline throughout — the point was to isolate the sensor, not to test our own consistency.

Where it held up

Ranking paddocks is fine. Relative ordering — which paddock is ahead of which — tracked the reference closely and stayed stable across the six weeks. If the question is "where do I put the mob next" or "which block is lagging", the cheap rig answers it.

Spotting change over time is fine, as long as you compare like with like. Week-on-week movement within a single paddock was consistent enough to be useful, because the systematic error mostly cancels when you difference two flights taken under similar conditions.

Where it drifted

Absolute values are not trustworthy. Cloud cover and sun angle moved readings enough that the same paddock could look materially different on two flights a day apart with no biological change. Anything that treats the index as an absolute number inherits that noise.

The top of the range compresses. On the densest paddocks the index flattened out well before growth did, which is a known saturation behaviour but is worth stating plainly: at the top end the sensor stops distinguishing "good" from "better".

Edges lie. Fence lines, tracks, troughs and shade from shelter belts pulled paddock averages around more than we expected. Masking the boundary inward by a few metres changed the answer more than most of the biological signal we were chasing.

Where we drew the line

We would use this class of rig to prioritise attention — to decide where to walk, where to graze next, which blocks to watch. We would not put a variable-rate fertiliser prescription on it without ground truthing, and we would not report absolute values to a third party.

That distinction matters commercially, not just scientifically. A tool that ranks paddocks is a tool that saves a farm manager time. A tool that prescribes inputs is a tool that carries financial and environmental consequence when it's wrong — and this rig, in these conditions, is not there.

What we'd change next time

• Fly a reference panel on every flight, not on a subset — the calibration overhead is smaller than the analysis it saves.
• Fix a tighter time-of-day window, and abandon flights outside it rather than adjusting afterwards.
• Mask paddock boundaries inward by default, and report the masked area so the number is auditable.
• Report a ranking and a confidence band, never a bare index value.

Why this is on our website. We price client work on outcomes. That's only honest if we know where our tools actually break — so we find that out on our own time and our own money first. This one now sits in our internal notes as "ranking yes, prescription no", and any client engagement that touches aerial sensing starts from that line.