Protocol: From benchmark to field: a systematic review of external-validation and deployment evidence for deep learning in precision agriculture

Registered protocol of a systematic review reported according to PRISMA 2020. Question: how much performance do deep-learning models for precision agriculture lose when evaluated outside their training distribution (another farm, season, region, sensor or cultivar) or deployed in real field conditions, and how is deployment evidence (latency, energy, memory, field-trial duration) reported? Status at registration: searches completed on 15-16 October 2026 (Scopus, Web of Science Core Collection, OpenAlex, SpringerLink, plus citation searching); 14,174 records after deduplication; title/abstract screening not started. The protocol defines the research questions, search strategy, eligibility criteria (IN.1-IN.6, EX.1-EX.8), dual-reviewer active-learning screening with ASReview LAB and a pre-registered stopping rule (SAFE-based), the data-extraction codebook, the quality-appraisal rubric and the synthesis plan (internal-to-external relative performance loss). Version 1.1 adds Amendment 1 (18 October 2026, before any screening decision): AI-assisted title/abstract screening with human verification.