Integrating Nigerian Indigenous Climate Indicators with Regional Climate Forecasts: Framework Development and Retrospective Evaluation

Regional climate models (RCMs) for West Africa exhibit systematic biases in boundary-layer processes and convective initiation, particularly during monsoon onset and cessation. This study develops and retrospectively evaluates a computational framework for encoding indigenous climate taxonomies from Nigeria's three major ethnolinguistic groups—Hausa, Yorùbá and Igbo—as structured metadata interoperable with numerical boundary-layer datasets, aiming to refine RCM performance and develop accessible climate-smart agricultural interfaces. The study employed a multiphase computational and mixed-methods design: a PRISMA 2020 systematic review to identify and document indigenous climate indicators; linguistic and computational encoding of these indicators as structured metadata; development of a four-layer hybrid architecture integrating boundary-layer data, taxonomy encoding, a fuzzy-logic and Random Forest fusion engine, and a multilingual agricultural interface; retrospective computational validation against ERA5 and CHIRPS data for three agro-ecological zones (Kano, Ibadan, Enugu) over 2010–2020; and stakeholder evaluation with 48 farmers. The review identified 23 eligible studies yielding 47 unique indicators (18 Hausa, 15 Yorùbá, 14 Igbo). Retrospective validation indicated that the fusion architecture reduced precipitation RMSE from 1.45 to 1.12 mm/day, a 22.8% reduction (block-bootstrap 95% CI: 18.1–27.4%), and temperature RMSE from 1.82°C to 1.54°C, a 15.4% reduction (95% CI: 11.2–19.6%). Rainfall onset accuracy improved from 68.2% to 86.6% overall (+18.4 percentage points; 95% CI: 14.2–22.6 pp), with region-specific gains of 18.4 pp (Kano), 15.2 pp (Ibadan) and 14.1 pp (Enugu). Random Forest permutation importance indicated that Hausa cloud morphology indicators contributed a relative predictive contribution of 22.3% under mean decrease in impurity, second only to RCM convective variables (38.4%); these values represent model-specific relative contributions and do not establish statistical significance or independent causal effects. Stakeholder evaluation indicated strong receptivity, with ease of understanding in native language rated highest (M = 4.41, SD = 0.62). The study provides evidence of computational feasibility and retrospective predictive improvement.