AgroAlert Ghana: A Hyperlocal Drought Prediction and Multilingual Voice Alert System for Smallholder Farmers Using a Hybrid Random Forest–LSTM Ensemble

Climate change is intensifying drought risks in Ghana, threatening the livelihoods of smallholder farmers who often lack access to timely and localized early warning information. This paper presents AgroAlert Ghana, an AI-powered hyperlocal drought prediction and multilingual alert system designed for rural farming communities. The system integrates satellite imagery from Sentinel-2 NDVI, ERA5 weather data and MODIS land surface temperature data to generate weekly drought risk predictions across 15 farming communities covering all 16 regions of Ghana. Using a five-year dataset (2019–2023) with 3,915 weekly observations and 57 drought events, the study develops Random Forest and Long Short-Term Memory (LSTM) models using independent environmental features while eliminating target leakage. A weighted hybrid ensemble model achieved the best performance with a PR-AUC of 0.170, slightly outperforming Random Forest (0.165) and LSTM (0.156). Due to the rare nature of drought events, precision-recall metrics and confidence intervals were used for reliable evaluation. The system applies a recall-focused threshold to improve drought detection and delivers warnings through AI-generated voice calls in Twi, Ewe and Dagbani, alongside English SMS alerts. A feasibility test with farmers confirmed the effectiveness of the feedback mechanism for collecting drought verification data and supporting future model improvement. The findings demonstrate the potential of combining artificial intelligence, remote sensing and inclusive communication technologies to support climate resilience among smallholder farmers in Ghana. The study also highlights the need for expanded datasets and further validation before large-scale operational deployment.