A Hybrid IoT–EfficientNet–Temporal Fusion Transformer Framework for Predictive Management of Pests and Diseases in Maize and Rice Production

Sustainable cereal production in tropical agro-ecological zones, particularly in sub-Saharan Africa, where monitoring infrastructure is limited, is increasingly threatened by climate-driven pest and disease outbreaks. To address this challenge, this paper presents a field-validated hybrid artificial intelligence framework that integrates IoT-based environmental sensing with an EfficientNet-B0 deep convolutional architecture for image-based disease classification and a Temporal Fusion Transformer (TFT) for multi-horizon (3‑day and 7‑day) outbreak forecasting. The study was conducted under natural conditions at the World Bank-sponsored CEADESE Cereal Farm of the Federal University of Agriculture, Abeokuta, Nigeria. A total of 4,340 raw field images spanning seven classes (six disease classes including maize rust, leaf blight, northern leaf spot, gray leaf spot, downy mildew, streak virus and one healthy) using a high-resolution Canon digital camera, alongside hourly measurements of temperature, relative humidity, rainfall, and leaf wetness recorded from vegetative growth to early reproductive stages. Preprocessing included normalization to 224 × 224 resolution and median filtering for noise reduction. EfficientNet-B0 achieved 96.2% classification accuracy (weighted F1 = 0.958), outperforming ResNet‑50 and VGG‑16 baselines, while the TFT captured complex climatic interactions and achieved 95.4% outbreak prediction accuracy (AUC = 0.978), a statistically significant improvement over LSTM and GRU forecasting models (p = 0.0035). Five‑fold cross‑validation confirmed model stability with low variance (±0.8%). Unlike simulation‑driven studies, this study is grounded in real tropical field data, enhancing ecological validity. The developed framework enables both reactive disease detection and proactive outbreak forecasting, demonstrating potential as a scalable solution pending further validation on larger multi‑farm datasets.