Responsible AI for Climate-Driven Invasive Species Management: A Governance-Embedded Early Warning System for Fall Armyworm in Kakamega County and Western Kenya

Climate change is accelerating the geographic expansion and outbreak frequency of Fall Armyworm (Spodoptera frugiperda) in East Africa, causing 20-50% annual maize yield losses and threatening smallholder livelihoods in Kenya. Current pest early warning systems remain predominantly reactive—responding after economic damage occurs—due to weak integration between predictive analytics, institutional thresholds, gender-responsive mechanisms, and accountable response systems.

This project develops and evaluates a governance-embedded AI early warning system for climate-driven Fall Armyworm management in Kakamega County and Western Kenya. The system integrates ERA5 climate reanalysis, MODIS vegetation indicators, CHIRPS rainfall data, publicly available pest occurrence records (GBIF, icipe), land use, elevation, and institutional vulnerability metrics within an R-based analytical pipeline.

The study implements a three-arm experimental comparison: (1) baseline machine learning (XGBoost classification) without governance; (2) governance-enhanced system with adaptive thresholds, HITL/HOTL oversight, MEAL feedback loops, and audit logging; and (3) AI-assisted workflow support using LLMs for preprocessing and documentation (LLMs excluded from prediction). The framework integrates gender and intersectional equity as core design principles, operationalizing women's differentiated exposure, sensitivity, and adaptive capacity.

Expected outputs include: an operational R-based early warning prototype; comparative evidence on governance effects; gender-equitable pest intelligence; policy recommendations for Kenya's Ministry of Agriculture and Kakamega County; and a transferable governance architecture applicable to drought, locusts, floods, and crop diseases.