Curve-shape
irrigation logic.

Almost every commercial smart-irrigation product on the market today decides whether to water by comparing a single moisture reading against a fixed threshold. If the reading sits below a setpoint, the valve opens; if it sits above, the valve closes. The setpoint is usually entered by the operator at install time and changed manually.

This approach has a class of failure modes that are well known to anyone who has actually deployed it. A sensor placed two centimetres off-target reads ten percent low forever. A sensor in a row that drains faster than its neighbours triggers irrigation that the neighbouring rows then receive whether they need it or not. Seasonal evapotranspiration changes invalidate the setpoint every six to eight weeks. A fertigation event spikes EC and pulls the moisture reading off-calibration for a day. None of these failures show up in a single-reading comparison.

The Sankhya Intelligence irrigation recommendation engine does not use thresholds. It uses the shape of the moisture curve over the preceding twenty-four hours, compared against the same zone's recent and historical curve behaviour. This page explains how.

Each instrumented zone publishes a moisture reading approximately every hour through its node. Over twenty-four hours this produces a curve with a characteristic shape for a healthy, well-irrigated zone: a sharp rise on the irrigation pulse, a wet-band plateau, an exponential decay during root uptake, and a low-band that ideally sits just above the wilt point at the moment the next irrigation event begins.

Three numbers describe the shape compactly:

• Range — the difference between the day's maximum and minimum moisture, in sensor units. A wide range means a deep irrigation pulse followed by full uptake; a narrow range means either over-watering (the soil never dries down) or under-watering (the soil never wets up).
• Position-in-day — where the current reading sits within today's min-to-max envelope, normalised to 0–1. A value of 0.1 means we are near today's low-band; a value of 0.9 means we are near today's wet-band.
• Delta-from-baseline — the difference between today's current reading and the trailing 7-day average for the same zone. This catches drift slowly, in either direction.

The recommendation engine reads these three shape numbers, compares them against the zone's own history and the cross-zone median, and emits one of five recommendation bands.

The historical reference is built from the same zone's daily curves over the prior several weeks. For every prior day with enough samples to characterise the curve, the system stores the day's min, max, range, and average. From those it computes a per-zone median daily range, median daily minimum, and median daily maximum.

The reason the comparison is per-zone rather than per-orchard is that zones have genuinely different soil character. A heavier-clay zone holds water longer than a sandier zone in the same orchard and produces a flatter curve at the same operational health. A threshold tuned to one is wrong for the other. The historical curve baseline absorbs this difference automatically — each zone is judged against itself.

The system also maintains a cross-zone median daily range as a fallback. If a zone is new and lacks its own history (the first three to seven days after node assignment), the cross-zone median is used as the reference until enough days accumulate to characterise the zone individually.

Confidence is reported alongside every recommendation. A zone with at least seven days of history and at least eight readings in the last twenty-four hours is reported at medium confidence; a zone with three to seven days is reported at low confidence; below three days the system reports experimental and the operator is encouraged to ignore the volume guidance until more data accumulates.

The output is one of five bands, each tied to an explicit volume change. The bands are deliberately coarse — finer granularity would imply a precision the underlying signal does not justify. Every band carries a one-line reason the operator can read at a glance.

The Emergency band uses absolute moisture limits — a sensor reading below about 12 % or above about 55 % triggers the emergency branch regardless of curve shape. These are deliberately broad guardrails that exist to catch supply failures (blocked drip line, valve stuck closed, burst pipe) where the curve has effectively gone flat.

All of the more nuanced bands — Strong, Mild, Hold — are curve-relative. They make their decision by comparing today's curve to the zone's own recent history. This matters for three reasons.

It absorbs sensor calibration drift. A capacitive moisture sensor that reads two percent low for three months is still useful, because the curve shape against its own history is unchanged. A threshold-based system would trigger spurious irrigation increases throughout that period.

It absorbs seasonal evapotranspiration change. The zone needs more water in late summer than in spring. The 7-day rolling baseline tracks the seasonal shift automatically without operator intervention.

It surfaces real anomalies. A zone whose curve range suddenly halves, while neighbouring zones continue to show their normal range, is a strong signal that something has changed at the zone level — emitter blockage, root encroachment in the drip line, a leaking lateral. The curve-relative engine flags this for inspection where a threshold-based system would simply continue irrigating at the operator-set value and miss the failure entirely.

Of the three shape numbers, position-in-day has been the most agronomically informative in practice. A healthy zone with adequate irrigation tends to sit near position 0.4–0.6 across most of the day; the curve is mid-band more often than it is at the wet edge or the dry edge.

A zone whose position-in-day clusters near 0.1 throughout the afternoon is dropping too far between cycles — either irrigation duration is too short, or the interval is too long, or both. A zone whose position-in-day clusters near 0.9 for hours after the irrigation pulse is staying saturated longer than the soil's drainage rate justifies — usually either over-volume or a heavier soil texture than the schedule assumes.

When the strong-increase branch fires, it is almost always because position-in-day has been below 0.35 at the same time that the day's range is much larger than the zone's reference range. That combination — a wide swing that bottoms out low — is the specific signature of a zone whose irrigation event is doing real work but is not delivering enough volume to hold the wet-band long enough.

Every recommendation the engine produces requires explicit operator approval before any change is made to the irrigation schedule. The system does not autonomously open valves, change pump run-times, or modify fertigation programs. This is by design: the recommendation engine is decision support, not autonomous control.

The operator sees the recommendation, the band, the reason in plain English, and the confidence. They can accept, modify, or override. Every override is logged and used as feedback — if the same zone is consistently overridden in one direction, the system can flag the historical reference for re-calibration.

The deliberate constraint on autonomy is what makes the system trustworthy in production. An irrigation mistake in a high-value orchard can damage trees that take years to recover. The cost of an extra approval click is negligible compared to the cost of an unsupervised valve opening at the wrong moment.

The curve-shape recommendation engine operates at zone resolution. The companion Uptake Index operates at per-tree resolution and answers a different question: when is each individual tree actually pulling water and nutrient from the root zone, and how strongly. The two layers are complementary — Uptake Index identifies the active window; the curve-shape engine sizes the volume.

The sensor data that feeds both layers comes from RS-485 Modbus in-soil sensors on the open-hardware v4 carrier board. For the sensor wiring details see RS-485 Modbus soil sensor wiring; for the board's power and MCU architecture see ESP32-S3 solar-powered sensor node design.

For the broader system architecture — how the curve engine fits with visual pest detection, fertigation recommendations, and the longitudinal data model — see /technical.