A Trust-Boundary Reference Architecture for Secure Cyber-Physical Agriculture: Design and Software-Based...

Smart irrigation is an operational technology (OT) problem as much as an Internet of Things (IoT) one: software decisions act directly on valves, pumps, pressure and water delivery, so a spoofed reading or a replayed command has physical consequences. This paper presents a technology-neutral reference architecture that makes those consequences controllable. It is organised into seven functional layers and three cross-cutting planes, six trust-boundary classes that fix which properties must be re-established at each crossing, and an auditable crown-asset inventory with an explicit trust-level rubric. Two governing execution principles enforce this structure: local safety authority, which keeps final actuation permission at the field gateway, remote terminal unit (RTU) or programmable logic controller (PLC), and feedback-aware execution, which treats a command as complete only when independent physical feedback confirms it. Building on process-aware and safety-supervision ideas from industrial control, the architecture makes physical confirmation a mandatory completion condition at the local enforcement point, which catches a fully authorised command that is nonetheless physically unsafe, a failure mode that authorisation checks alone cannot detect. These properties are examined through a deterministic, openly available software proof-of-concept that verifies the design against its own acceptance criteria; it is not a field validation. The software executes six attack and failure scenarios plus one nominal-operation control against an intentionally unsecured negative control and a secure controller, under identical, controlled conditions: the negative control is compromised in all 6 attack scenarios and the secure controller in 0/6, with 0/1 false positives on legitimate operation. Analytically, the design costs 21 bytes and 50-60% additional Long Range Wide Area Network (LoRaWAN) time-on-air per record, and 78% of single-step perturbations of the trust-scoring inputs leave asset classification unchanged: the actuator class (A2) is completely stable, and the gateway/RTU/PLC class (A3) is stable under 6 of 7 admissible perturbations (86%), changing only when its credential/update authority reaches the trust-anchor threshold. A sensor-level mapping from sampling rate, measurement drift and physical fault modes to the edge-verification controls they dictate, provisions for secure time, counter recovery and the choice between fail-closed and fail-open safe states, an illustrative hypothetical worked deployment example, and a transfer of the model to greenhouse, fertigation, livestock-watering and cold-chain settings show how the architecture could be applied in practice.