Deployment Readiness of TinyML and Edge AI for Precision Apiculture: A Systematic Review and Evidence Map

Abstract

This systematic review and evidence map will assess how far TinyML and edge-AI systems for managed honey-bee (Apis mellifera) hive monitoring have progressed from offline algorithm development to physical on-device inference and validation under realistic apiary conditions. Deployment readiness is operationalised along two ordinal axes. The first is a computing-deployment taxonomy (C0-C4), ranging from offline, cloud, workstation, or otherwise non-target-device evaluation (C0) to physical inference on MCU-class hardware (C4). Intermediate categories distinguish claims of edge suitability from demonstrated local inference, including gateways, conventional local computers, single-board computers, and edge accelerators. The second is a validation taxonomy (V0-V4), ranging from benchmark or laboratory evaluation to sustained validation in operating apiaries. Full definitions and decision rules are specified in the registered protocol. Studies will be characterised by sensing modality, application, machine-learning or computational method, hardware platform, dataset, deployment characteristics, validation setting, and reported performance. Relevant acoustic, visual, environmental, chemical, and multimodal hive-monitoring studies will be considered according to the predefined eligibility criteria. Six primary bibliographic sources will be searched from database inception to the production-search date: Scopus, Web of Science Core Collection, IEEE Xplore, the ACM Guide to Computing Literature, PubMed, and CAB Abstracts. Supplementary identification methods include backward and forward citation searching and predefined targeted discovery approaches described in the protocol. English-language full texts are required for inclusion; potentially eligible non-English records encountered during screening will be logged explicitly rather than silently removed. Database-specific search strategies were developed and validated before registration using a predefined 15-study sentinel set spanning relevant sensing, modelling, and deployment characteristics. Validation was assessed against the subset of sentinel studies indexed by each database. Pilot searches were conducted only for syntax development, query refinement, and sentinel validation; pilot records and pilot result counts will not enter screening, synthesis, or PRISMA reporting. Following registration, the frozen database-specific strategies will be rerun as production searches. Only records retrieved through the production searches and predefined supplementary identification methods will proceed to deduplication, screening, extraction, synthesis, and PRISMA accounting. Title/abstract screening will be conducted by the lead reviewer, with a predefined independent sample assessed by the second reviewer and inter-reviewer agreement evaluated according to the protocol. Potentially eligible full texts will undergo independent second-reviewer assessment, with disagreements resolved using the predefined reconciliation and escalation procedure. Data extraction and quality assessment will follow the verification procedures specified in the protocol. Synthesis will be descriptive and will include a C0-C4 × V0-V4 evidence map, cross-tabulations by sensing modality, application, and hardware class, deployment-evidence summaries, and reporting-completeness measures. No statistical meta-analysis is planned because substantial methodological, task, dataset, hardware, and validation heterogeneity is expected. Reporting will follow PRISMA 2020 and PRISMA-S principles. The review aims to quantify the extent of operational edge and TinyML deployment in precision apiculture, distinguish deployment claims from demonstrated physical inference and field validation, and identify the reporting practices required for deployment claims to be independently assessed. The observed density and distribution of evidence across deployment and validation levels will be reported irrespective of the pattern identified.