A low-cost sensor can help schedule irrigation, but only when its reading is calibrated to the soil, placed in the active root zone, combined with weather and hydraulic evidence, and connected to a controller that remains safe when the cloud or network disappears.
Soil is spatially variable. Texture, compaction, slope, root depth, shade, irrigation hardware, and crop stage change water behavior. A single probe cannot represent an entire farm. Define management zones and place sensors at representative locations and depths.
For each zone, store crop, growth stage, soil profile, root depth, field capacity, refill threshold, sensor type, installation depth, calibration version, irrigation method, nominal flow, and operator.
Low-cost capacitive probes often expose a raw electrical value rather than universal volumetric water content. ADC variation, supply voltage, cable length, salinity, temperature, sensor aging, and soil contact affect readings. Calibrate dry and wet reference points for the sensor, soil, depth, and electronics path.
ESP-IDF provides ADC calibration support because raw ADC output can vary. Record raw value and calibrated engineering value; keep the calibration ID so history remains interpretable after a sensor or board replacement.
The cloud can calculate recommendations, compare zones, and schedule work. It should not be the only component able to stop a stuck valve, dry pump, burst pipe, empty reservoir, or overlong irrigation. Local hardware or a nearby controller enforces maximum runtime, maximum volume, pressure/flow bounds, cooldown, and emergency stop.
Commands need ID, zone, requested start, expiry, target volume/runtime, preconditions, policy version, issuer, and signature/authentication. The ESP32 rechecks current local state before acting and reports accepted, started, flowing, completed, stopped, rejected, or faulted.
The chart should also display sample freshness, calibration, sensor depth, irrigation interval, rain, and delivered volume. A percentage without that context can create false confidence.
Power the probe only when sampling where hardware permits, allow stabilization, read several values, use median or robust filtering, and retain min/max or variance. Label disconnected, saturated, out-of-range, unstable, stale, uncalibrated, and estimated states.
Do not smooth away a real irrigation response or sudden leak signature. Keep raw samples for a short diagnostic window and store operational aggregates longer.
Wi-Fi may work in a greenhouse or near buildings. LoRaWAN can serve low-rate battery nodes over wider rural areas through a gateway. Cellular fits remote gateways with coverage. Satellite or delayed physical synchronization may be needed in isolated sites. Select per payload, range, power, terrain, ownership, cost, and downlink needs.
Keep control local during outages. Buffer telemetry with sequence IDs and event time. Commands expire; old cloud commands must not execute after hours of disconnection.
MQTT moves telemetry and commands efficiently, but QoS does not make a valve operation exactly once. Give every reading and command a stable ID. Deduplicate at ingest and device. Separate desired command from observed state and final result.
Use retained messages carefully: a retained stale command is dangerous. Prefer short-lived command topics or explicit expiry and current-state checks. Device status should include firmware, boot reason, queue depth, last calibration, battery, signal, and fault.
Soil sensors indicate current local water status. Evapotranspiration estimates atmospheric demand and crop use. FAO's crop-coefficient approach combines reference ET with crop coefficient. Use it as a model input, then calibrate against local sensors, rain, irrigation, and agronomic observation.
Forecast rain can delay a recommendation, but uncertain forecasts should not silently override critical stress. Keep forecast source, run time, horizon, probability, and realized error.
Before adding machine learning, calculate rate of moisture decline, expected time to refill threshold, water-balance residual, flow per valve, pressure deviation, battery discharge, packet loss, and sensor disagreement. These features are explainable and often sufficient.
A model can later predict irrigation need, sensor failure, clogging, leakage, or pump maintenance. It should return horizon, confidence, contributing evidence, and recommended inspection, not directly command water without bounded policy.
| Condition | Evidence | Automated response | Safety gate | Operator alert |
|---|---|---|---|---|
| Zone below refill threshold | Multiple valid samples, depth profile, ET/rain, crop stage and freshness. | Recommend or schedule bounded irrigation. | Allowed window, reservoir, pressure, recent runtime and manual mode. | Alert if stress forecast is near and no eligible window exists. |
| Valve commanded, no flow | Command accepted, valve feedback, flow unchanged, pressure context. | Stop after short verification timeout and mark fault. | Never extend runtime to “see if it starts.” | Valve, wiring, pump or blocked-line inspection. |
| Unexpected flow while valves off | Flow above baseline, all commands off, pressure/reservoir changes. | Close master valve or stop pump under approved local policy. | Physical emergency path and manual override remain available. | High-severity leak or unauthorized-flow alert. |
| Moisture does not rise after delivery | Verified volume, sensor quality, time lag, depth and neighboring sensors. | Do not repeat automatically; open anomaly. | Maximum daily water and cooldown. | Check sensor placement, leak, runoff, clogging or soil channeling. |
| Forecast rain with adequate soil water | Current profile, ET, forecast probability/amount and irrigation deadline. | Delay recommendation within policy. | Minimum moisture and maximum delay. | Notify only when forecast uncertainty could cause stress. |
| Sensor stale or drifting | Last seen, variance, peer disagreement, calibration age and battery. | Exclude from automatic control and use safe fallback. | No consequential action from one invalid probe. | Maintenance work order with zone and replacement details. |
Use deep sleep, duty cycling, low-quiescent-current regulators, measured battery voltage, and a realistic energy budget for sensing, radio retries, cold weather, and aging. Solar sizing must consider cloudy periods and panel contamination, not only average sunshine.
Brownouts can corrupt state or chatter relays. Use proper drivers, flyback protection, decoupling, watchdogs, boot-safe output states, and durable storage designed for bounded writes.
Enclosure rating, UV exposure, condensation, insects, cable glands, corrosion, lightning, grounding, sensor replaceability, and connector labeling affect uptime. Record enclosure, sensor serial, installation photo/reference, depth, calibration, battery, firmware, last visit, and next maintenance.
A dashboard should make a technician's route actionable: exact zone, likely part, fault history, last readings, safe isolation step, and verification after repair.
Use signed firmware, secure boot and flash encryption where supported and justified, protected credentials, per-device identity, and authenticated transport. OTA should be staged by hardware revision and zone, with rollback and post-update health checks.
Never update all irrigation controllers simultaneously. Preserve a local safe schedule and manual operation if a release fails. Track firmware, configuration, calibration compatibility, updater result, and rollback cause.
Dashboards need soil profile, threshold, ET, rain, recommendation, irrigation events, flow, pressure, water volume, pump energy, device freshness, battery, signal, firmware, queue, and faults. Plot sensor response after each irrigation to verify that water reached the intended root zone.
Business outcomes include water per hectare, irrigation efficiency, stress-hours avoided, leaks found, truck rolls, pump energy, yield/quality context, and time to repair. Do not claim yield improvement without controlled agronomic evidence.
Bench-test sensor calibration, relay/valve drivers, no-flow and overcurrent shutdown, watchdog, brownout, local manual mode, command expiry, offline queue, duplicate messages, clock drift, full storage, low battery, packet loss, gateway outage, and OTA rollback.
Then run a staged field trial across different soil and irrigation zones. Compare sensor readings with manual reference measurements and inspect roots/soil response with an agronomist before enabling automatic control.
I would build ESP32 sensor nodes with calibrated probes and deep sleep, a separate locally safe valve/pump controller, optional LoRaWAN or Wi-Fi gateway, MQTT ingestion, raw time-series storage, zone digital twin, ET/water-balance service, recommendation engine, anomaly detector, command service, and field operations dashboard.
Automation would begin in advisory mode. After enough verified cycles, selected low-risk zones could use bounded automatic irrigation with flow/pressure confirmation, daily limits, command expiry, local interlocks, audit, and immediate manual override.
Smart irrigation is not “moisture below 30%, turn relay on.” It is a measured control system: calibrated root-zone evidence, explicit water policy, local safety, verified delivery, resilient telemetry, explainable prediction, and human ownership of exceptions.
Article about agricultural IoT with ESP32, including calibrated soil-moisture sensors, root-zone management, local irrigation safety, valves, pumps, flow and pressure, MQTT, LoRaWAN, offline buffering, evapotranspiration, predictive alerts, dashboards, OTA, security, and field validation.