Sensor-Fusion Irrigation Scheduling for Drip-Irrigated Tomato: FAO-56 Water Balance vs Soil-Moisture Thresholds

SFIS (Sensor-Fusion Irrigation Scheduling study) compares three families of irrigation-scheduling rules for a drip-irrigated tomato field in Dharwad district, Karnataka, using real weather data and a declared, reproducible evaluation protocol — the kind of disciplined comparison 22RMI16 (Research Methodology & IPR) expects of an M.Tech dissertation.

Indian drip farms increasingly have both information streams available: free daily weather (NASA POWER archives cover any Indian district) and cheap capacitive soil-moisture sensors. Yet the common scheduling rules still conflict in practice — irrigate by the FAO-56 soil-water balance, irrigate when the sensor crosses a threshold, or fuse the two (the balance proposes an event, the sensor vetoes or confirms it). Vendors sell all three; almost nobody says which loses less water under which sensor quality.

The study builds a daily FAO-56 dual-crop-coefficient water balance for a 1 ha sandy-loam field driven by NASA POWER weather for Dharwad, implements four policies (fixed-interval farmer practice, ET-balance, sensor-threshold, fused), and evaluates them with a Monte Carlo protocol over declared sensor-error levels (±2 %, ±4 % volumetric water content) and dropout scenarios. Metrics are pre-declared: seasonal irrigation depth, number of events, deep percolation, stress days via the FAO-56 Ks coefficient and a CROPWAT cross-check of reference ET.

The deliverable is a paper-style study with an auditable Python pipeline — honest about what is real (weather) and what is simulated (sensor error), which is exactly the distinction examiners probe.

Drip-irrigated vegetable growers in Karnataka's dry zone receive contradictory scheduling advice: agronomy extension pushes FAO-56 evapotranspiration-based water balance, while sensor vendors push soil-moisture thresholds, and fused rules are marketed without evidence. Field experiments comparing methods are expensive, site-specific and rarely quantify how sensor error changes the ranking; simulation studies usually assume perfect sensors, so their conclusions do not transfer to the ±2–4 % error reality of low-cost capacitive probes. There is no reproducible, declared-protocol study on real Indian weather that (a) implements the ET-balance, sensor-threshold and fused scheduling policies in one auditable framework, (b) quantifies water applied, deep percolation and crop stress for each under declared sensor-error models, and (c) identifies the sensor-quality level at which the best policy changes. This dissertation fills that gap with a fully scripted Monte Carlo evaluation anchored in NASA POWER weather for Dharwad and FAO-56 methodology, and states its limitations — simulated sensor stream, single soil texture — explicitly rather than hiding them.

Abstract

This dissertation compares irrigation-scheduling policies for a drip-irrigated tomato field in Dharwad, Karnataka, under a pre-declared evaluation protocol. A daily FAO-56 dual-crop-coefficient soil-water balance is driven by three seasons of real NASA POWER weather. Four policies are evaluated: a fixed-interval baseline representing farmer practice, a pure ET-balance rule, a soil-moisture sensor-threshold rule, and a fused rule in which the balance schedules an event and the sensor confirms or vetoes it. Sensor behaviour is simulated with declared error models (Gaussian noise at two magnitudes plus dropout), because no audited open Indian dataset pairs collocated weather and sensor streams. Results are reported across 500 Monte Carlo realisations per configuration on four metrics: irrigation depth, event counts, deep percolation and Ks-weighted stress days. The study quantifies the sensor-error level where the fused rule overtakes the pure threshold rule.

Introduction

Karnataka's northern dry zone grows tomato and chilli under drip on sandy loams where over-irrigation wastes pumped groundwater and under-irrigation costs yield. Micro-irrigation schemes have made controllers and soil-moisture probes affordable, but scheduling advice remains folkloric: "water every second day", "irrigate when the meter shows 20 %". The agronomy literature offers rigorous methods — FAO-56 evapotranspiration water balance, sensor-threshold scheduling — but rarely compares them head-to-head under realistic sensor imperfection for an Indian drip system.

Existing system and literature gap

Field studies that compare scheduling methods usually run on a single experimental plot, so their conclusions carry the weather and soil of one season and one farm, and sensor-error robustness is rarely quantified. Simulation studies, conversely, often assume perfect sensors. What is missing is a declared-protocol comparison anchored in real Indian weather, with every modelling choice stated in advance and every number regenerable from source data — the 22RMI16 discipline applied to irrigation engineering.

Proposed work

• Download and quality-check daily weather for Dharwad from NASA POWER for three complete tomato seasons.
• Compute reference ET0 with the FAO-56 Penman–Monteith equation and build the dual-Kc (Kcb + Ke) daily water balance for sandy loam; cross-check ET0 against FAO CROPWAT 8.0 with CLIMWAT climate normals.
• Implement four scheduling policies behind one interface; simulate a capacitive soil-moisture sensor with declared Gaussian error (σ = 2 % and 4 % VWC), quantisation and 5 % dropout days.
• Evaluate all configurations over 500 seeded Monte Carlo realisations, reporting mean ± standard deviation per metric and locating the error level where the policy ranking changes.

Feasibility

• Technical: the study runs in Python (pandas, NumPy, SciPy) on a 16 GB laptop; each season is ~180 daily rows, so Monte Carlo evaluation completes in minutes.
• Data: weather is free and official (NASA POWER); crop and soil parameters come from FAO-56 tables and standard agricultural-engineering texts — no paid data.
• Economic: zero licence cost; CROPWAT 8.0 is a free FAO cross-check.
• Operational (research): the protocol, error models and seeds are frozen before evaluation; every table and figure regenerates from two commands — the audit trail an M.Tech viva demands.

In scope

• One field archetype: 1 ha drip-irrigated tomato on sandy loam in Dharwad district, Karnataka; three complete seasons of NASA POWER daily weather (approx. 540 days).
• FAO-56 methodology: Penman–Monteith ET0, dual crop coefficient, Ks stress accounting, TAW/RAW depletion logic with root growth.
• Four scheduling policies; one simulated capacitive soil-moisture sensor at 15 cm depth with two declared error levels plus dropout.
• Monte Carlo evaluation (500 realisations, fixed seeds), sensitivity of the ranking to sensor error, and a CROPWAT ET0 cross-check.

Out of scope

• Physical field trials or laboratory sensor calibration — the sensor stream is simulated by declared models and the report says so.
• Yield modelling beyond the FAO-56 Ks stress proxy; economic analysis beyond water-volume comparisons.
• Other soils, crops and districts (listed as replication work in future scope); fertigation and salinity dynamics.
• Automatic controller hardware: the study ends at validated scheduling rules, not firmware.

The work follows a quantitative, comparative-simulation research design in the phased pattern of an M.Tech dissertation, with the evaluation protocol (metrics, seeds, error models) frozen at the end of Phase-1 — before any policy is scored.

Validation inside the study: unit tests for the water-balance closure (inflow − outflow = Δstorage within tolerance), hand-computed ET0 on sample days, policy-decision unit tests on synthetic days, and a full re-run reproducing results bit-for-bit from fixed seeds.

The study is a linear, scripted pipeline in which the real-data path and the simulated-sensor path stay visibly separate, so no simulated quantity can masquerade as measured.

flowchart TD
  A["NASA POWER daily weather\nDharwad, 3 seasons"] --> B["Ingestion and QC\nmissing days, spike checks"]
  B --> C["FAO-56 Penman-Monteith ET0"]
  C --> D["Dual-Kc daily water balance\nKcb + Ke, root growth, TAW/RAW"]
  E["Soil and crop parameters\nFAO-56 tables, sandy loam"] --> D
  D --> F["Policy library\nfixed-interval, ET-balance, threshold, fused"]
  D --> G["Sensor simulator\nGaussian error, quantisation, dropout"]
  G --> F
  F --> H["Monte Carlo harness\n500 realisations, fixed seeds"]
  H --> I["Metrics: irrigation depth,\npercolation, Ks stress days"]
  H --> J["Sensitivity analysis:\nranking vs sensor error"]
  I --> K["Results tables and figures"]
  J --> K
  C -.ET0 cross-check.-> L["FAO CROPWAT 8.0\nCLIMWAT normals"]

Three design decisions matter at the viva. First, the weather ingest writes one versioned, QC'd data artefact; every later stage reads only that file, so a download or cleaning change forces an explicit rebuild rather than silently entering old tables. Second, the sensor simulator wraps the balance state and is the only source of sensor readings — policies never see true water content when configured for sensor mode, which makes information leakage structurally impossible and lets the error model be quoted verbatim in the report. Third, the Monte Carlo harness owns all scoring: policies are scored only inside it, with fixed seeds recorded in the configuration file, so the reported distributions are exactly reproducible and the sensitivity analysis (ranking versus σ) is a re-parameterisation, not a new experiment. The CROPWAT cross-check sits outside the Python pipeline deliberately: it is an independent FAO implementation, and agreement between two independent ET0 implementations is stronger evidence than any internal test.

• Replicate on a second soil (the vertic black soils of north Karnataka) and a second crop (chilli) to test whether the policy ranking transfers.
• Replace the simulated sensor with data from a low-cost probe logging campaign on a college farm plot, and re-run the comparison on measured readings.
• Extend the balance to salinity and fertigation scheduling for drip systems, coupling Ks with osmotic stress.
• Add an uncertainty-aware fused rule (Bayesian update of depletion from the sensor reading) and compare against the veto/confirm heuristic.
• Package the winning policy as a simple advisory calculator (spreadsheet or phone form) for KVK extension use.