Open-Source Artificial Intelligence Tools for Irrigation Management: A Critical Narrative Review of Capabilities, Evidence and Implementation Gaps

Artificial intelligence (AI) is increasingly used to estimate crop water demand, forecast soil moisture, classify irrigation need, optimise irrigation schedules and support automated control. In parallel, open-source software has become central to scientific computing, geospatial analysis, crop modelling and Internet of Things (IoT) development. Yet the practical value of combining these two trajectories is often obscured by studies that emphasise predictive accuracy without examining whether the underlying toolchain is open, reproducible, interoperable, maintainable or validated under operational irrigation conditions. This critical narrative review evaluates open-source AI tools for irrigation management across the full decision chain from sensing and data engineering to prediction, optimisation, decision support and actuation. Literature published principally from 2012 to 30 June 2026 was considered, with earlier foundational software publications retained where necessary. The evidence indicates that the enabling software ecosystem is technically mature: open-source machine-learning libraries, crop-water models, geospatial toolkits and IoT components can be combined into transparent irrigation decision-support workflows. Evidence is strongest for reference evapotranspiration estimation, soil-moisture forecasting, crop-water simulation and proof-of-concept scheduling. Evidence is weaker for sustained multiseason water savings, cross-site transfer, autonomous closed-loop control, smallholder usability and independent comparison with well-tuned conventional scheduling. A recurring methodological problem is that high model accuracy is frequently demonstrated against calculated or simulated targets, whereas the operational endpoints of irrigation management are water applied, crop response, energy use, economic return and environmental impact. Open source can improve auditability, adaptation and avoidance of vendor lock-in, but it does not by itself ensure low total cost, robust sensors, trustworthy models or equitable adoption. The most defensible near-term pathway is therefore a modular, hybrid architecture in which open crop-water physics, geospatial data and machine learning are combined with explicit uncertainty handling, local calibration and human oversight. Future progress depends less on another proliferation of algorithms than on shared benchmarks, multienvironment field trials, reproducible software releases and implementation research.