Skip to content
Shyenova

Remote Sensing

What satellites can and cannot tell you about a crop

Sentinel data is free, covers the country and revisits every few days. So why do most agricultural earth observation projects still end in a folder of attractive maps?
Published
2 June 2026
Reading time
8 minutes
Topic
Remote Sensing
Author
Shyenova

Earth observation stopped being expensive some years ago. Sentinel-1 and Sentinel-2 are free, cover the whole country and revisit every five to twelve days. Commercial constellations offer sub-metre imagery at daily cadence for anyone willing to pay. Whatever is limiting agricultural monitoring in India, it is not access to pixels.

What limits it is the distance between an index map and a decision. Somebody has to establish what a particular value means for this crop, at this growth stage, in this district, and then demonstrate that the interpretation holds against what is actually in the field. That step is where most projects stall, and it is not a technical problem so much as a discipline problem.

What it genuinely does well

  1. 01Acreage. With a full-season time series and adequate ground truth, major crop classification lands in the high eighties to low nineties for overall accuracy, and, crucially, several weeks ahead of conventional reporting.
  2. 02Sowing progress. The transition from bare soil to emerging canopy is one of the clearest signals in the whole discipline, visible within a fortnight of the event.
  3. 03Relative crop condition. Comparing this field to its neighbours, or this week to the same week last season, is far more robust than any absolute statement about condition.
  4. 04Change. Anything that alters the surface (irrigation, harvest, flooding, abandonment) shows up reliably, which is why change detection is the most dependable product in the toolkit.

Where it is weaker than the brochures suggest

Cloud, which is not a minor caveat in India

Optical imagery cannot see through monsoon cloud, and the kharif season is precisely when a great many people want to know what is happening. The answer is Sentinel-1 radar, which sees through cloud and is harder to work with, less intuitive to interpret, and genuinely less accurate for some crop discriminations. Any proposal that promises consistent optical monitoring through an Indian monsoon is either not serious or not being read carefully.

Small plots

Sentinel-2's ten-metre pixel handles plots above roughly half a hectare reliably. Below that, a single pixel contains several fields and their signals mix, and accuracy degrades in a way that averages conceal. Where holdings are genuinely small, either you pay for commercial high-resolution imagery or you aggregate the output to village level, which is frequently adequate for the decision being made, and should be established at the outset rather than discovered at the end.

Absolute yield

This is the most over-promised output in the field. Satellites observe canopy condition, from which yield is estimated by a model calibrated on ground measurements. Without crop-cutting experiment data or an equivalent, the model is uncalibrated and its output is a plausible-looking guess. With calibration across a genuine range of yields (including the poor plots, not only the good ones) it becomes useful. The difference is entirely in the ground data, not in the satellite.

ProductRealistic accuracyBinding constraint
Crop classification85–93%Ground truth volume and plot size
Acreage estimation±4–8%Mixed pixels at field boundaries
Sowing date±7–10 daysRevisit interval and cloud
Relative conditionStrongComparability across dates
Absolute yield±10–20%Crop-cutting calibration data
Roughly what to expect, given adequate ground truth and a full-season time series.

NDVI is an input, not an answer

A map of NDVI is compelling, easy to produce, and supports no decision on its own. The index saturates in dense canopy, varies by crop and growth stage, and is not comparable across dates without careful handling. Presented as a deliverable it looks like insight; used as a deliverable it produces nothing.

The useful output is never the index. It is a calibrated estimate with an error bar and a stated set of conditions under which it should not be trusted.

None of this is an argument against earth observation. Used within its limits it delivers information weeks earlier and at a fraction of the cost of ground survey, and the gap is widening as the archives deepen. But the limits are real, they are knowable in advance, and stating them clearly at the start is what separates a monitoring programme that gets used from a folder of maps that gets admired once.

Start a conversation

Tell us what is not working.

Most useful first conversations are thirty minutes and involve no proposal. Describe the problem you are looking at and we will tell you honestly whether it is one we can help with.

Or write directly to hello@shyenova.com