Skip to content
Shyenova

Remote Sensing

Crop Intelligence

A yield model that has never been checked against a crop-cutting experiment is a hypothesis with a colour ramp.

Agricultural decisions are made on information that arrives too late. Procurement teams commit to volumes before they know the acreage. Insurers price before they know the condition. Input companies plan production against a sowing estimate assembled from dealer conversations. Everyone is working from a picture that will be accurate in about six weeks, which is roughly when it stops mattering.

Satellite observation collapses that lag. Sowing progress is visible within a fortnight of the event, crop stress within days of onset, and acreage to within a few per cent well before harvest. The difficulty is not producing the map. It is producing a number the commercial team will actually bet on, which requires validation against ground observation and an honest statement of where the model is weak.

How the work runs

Four phases, and what comes out of each.

Durations are indicative and assume reasonable data availability. Where data is poor we say so in the first fortnight rather than modelling on sand.
01

1–2 weeks

Fix the question and the geography

We establish precisely what decision the output feeds (a procurement commitment, a premium, a production plan) and therefore what accuracy at what spatial unit is actually required. A district-level acreage number good to five per cent and a plot-level condition flag are entirely different engineering problems, and conflating them is how these projects become expensive.

Outputs

  • Decision statement with the accuracy and cadence it genuinely needs
  • Area of interest, crop calendar and growth-stage windows
  • Inventory of existing ground data usable for training and validation
02

3–4 weeks

Build the observation stack

Sensor selection follows the crop calendar and the cloud regime, not preference. Optical imagery is richer and useless under monsoon cloud, so kharif work in most of India leans on Sentinel-1 SAR with optical where it clears. We assemble the time series, mask cloud and shadow, harmonise across sensors, and build the phenology curves the classifier will read.

Outputs

  • Analysis-ready cloud-masked time series over the full season
  • Sensor and index selection with the reasoning recorded
  • Phenology profiles by crop and growth stage
03

4–6 weeks

Classify, model and validate against ground truth

Crop classification and yield modelling are trained on ground observation and tested on held-out ground observation, never on themselves. We report a confusion matrix, producer and user accuracy per class, and where the model fails: smallholdings below the pixel, intercropping, and the crop pairs that are spectrally near-identical at the stage you need them separated.

Outputs

  • Classified crop maps with per-class accuracy
  • Yield model calibrated to crop-cutting or equivalent ground data
  • Explicit statement of failure modes and confidence by district
04

3–4 weeks

Put it in the operating rhythm

The pipeline is automated to run on each satellite pass and to deliver into whatever the commercial team already uses, a dashboard, a scheduled file, an API into the planning system. We run it alongside the existing estimation method for a full season so the organisation can watch the two diverge and decide for itself which to trust.

Outputs

  • Automated pipeline running on the satellite revisit cycle
  • Delivery into existing systems, with alerting on material change
  • A parallel-run season comparing satellite estimates to the incumbent method

What good looks like

The ranges we design toward.

These are the improvement ranges these levers are known to produce where the underlying practice has not been examined recently. They are targets to test against your baseline, not promises.

85–93%

Classification accuracy

Typical overall accuracy for major crops with adequate ground truth and a full-season time series.

4–8 weeks

Earlier acreage signal

How far ahead of conventional reporting a reliable acreage estimate typically becomes available.

60–80%

Survey cost

Reduction in ground survey spend once field effort is redirected from measurement to model validation.

Deliverables

What you are left holding.

All of it handed over, documented, and with a session for whoever inherits it.
  1. 01Running pipeline producing acreage, condition and yield estimates on the revisit cycle, handed over with documentation
  2. 02Accuracy assessment by class and district, including where the model should not be trusted
  3. 03Crop calendar and phenology reference for your geographies
  4. 04Ground-truth collection protocol your field teams can follow to keep the model calibrated
  5. 05A season of parallel running against your existing estimation method

What an engagement looks like

A representative shape.

Illustrative. This is the shape of work in remote sensing, not an account of a specific client. We do not publish named results, and we would be wary of a firm that publishes them without the client in the room.

16 weeks

  1. 01Situation

    An agri-processor was committing to procurement volumes months before it had any reliable view of planted acreage, relying on dealer and field-officer reporting whose historical error nobody had measured.

  2. 02What we did

    Built a Sentinel-1 and Sentinel-2 time series across the sourcing districts, trained crop classification on collected ground truth, and ran the satellite estimate in parallel with the incumbent reporting for a full season.

  3. 03Where it landed

    The satellite estimate was available several weeks earlier and diverged from field reporting by a consistent margin in two districts, which ground checks confirmed. The parallel season is what earned the commercial team's trust; the technology alone would not have.

Questions

The ones we get asked.

Does this work during the monsoon, when it is always cloudy?

Optical imagery does not, and anyone claiming otherwise is not being straight with you. Kharif monitoring in most of India runs primarily on Sentinel-1 radar, which sees through cloud, supplemented by optical on the days it clears. Radar is harder to work with and less intuitive, and the accuracy is genuinely lower for some crop discriminations. We state which of your seasons are strong and which are compromised rather than averaging the two together.

Our farms are small. Is the resolution good enough?

This is the right question and the answer depends on your plot size distribution, which we measure before promising anything. Sentinel-2 at ten metres handles plots above roughly half a hectare reliably; below that, mixed pixels dominate and accuracy falls away. Where holdings are genuinely small, either commercial high-resolution imagery is required (at real cost) or the output has to be aggregated to village level, which is often perfectly adequate for the decision being made.

How much ground data do we need?

For classification, a few hundred well-distributed labelled plots per crop per agro-climatic zone, split into training and validation. For yield modelling, crop-cutting experiment data or equivalent across a range of yields, models trained only on good plots predict badly on poor ones. If you have none, the first season is a collection season and we say so up front rather than shipping an uncalibrated model.

Can you just give us the NDVI maps?

We can, and we would rather not stop there. NDVI is an input, not an answer. It saturates in dense canopy, varies with crop and growth stage, and is not comparable across dates without care. A map of it looks compelling and supports no decision on its own. The useful deliverable is a calibrated estimate with an error bar.

Start a conversation

Talk to us about crop intelligence.

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