STPS-NET: Federated Learning for Plant-Specific Spatio-Temporal Modeling in Urban Rooftop Agriculture

Urban rooftop agriculture in tropical megacities has significant potential to improve food security and environmental sustainability. However, large-scale deployment remains constrained by three key challenges: strong microclimate variability across urban rooftops, heterogeneous cultivation requirements across plant species, and farmer reluctance to share operational data with centralized cloud systems. This study introduces STPS-NET, a federated learning framework designed for privacy-preserving yield prediction in rooftop agriculture in Dhaka, Bangladesh. The proposed model employs a plant-specific neural architecture that integrates multi-scale convolutional feature extraction, bidirectional long short-term memory networks, and self-attention mechanisms with 32-dimensional botanical embeddings validated by domain experts. During federated aggregation, a botanical similarity weighting strategy enables knowledge transfer between botanically related species while preserving data locality at individual rooftop sites. The framework was trained using a transparent dataset comprising 143,500 IoT sensor observations collected from 20 rooftop gardens over nine months (March-November 2024). To address limited observations during extreme pre-monsoon heat and early germination stages, physically consistent synthetic data were used exclusively for training, while all evaluation metrics were computed on held-out real measurements. Sensitivity analysis indicates that models trained with real-only data (MAE = 0.48, R 2 = 0.93), 25% synthetic augmentation (MAE = 0.38, R 2 = 0.95), and 50% synthetic augmentation (MAE = 0.30, R 2 = 0.96) generalize effectively to real data. Across 50 federated learning rounds, the proposed framework reduces yield prediction error from 1.47 to 0.30 kg m-2 and achieves R 2 = 0.96, outperforming 17 baseline methods. Field validation across eight independent rooftop sites further demonstrates practical agronomic benefits, including a 21.1% increase in yield, a 15.5% reduction in water usage, and a 16.9% reduction in fertilizer costs. Ablation analysis shows that plant-specific embeddings provide the largest performance gain (+0.12 in R 2), followed by temporal modeling through bidirectional LSTM and multi-scale feature fusion. In addition, differential privacy with ϵ = 1.0 preserves 98.9% of model utility, indicating that strong privacy guarantees can be achieved with minimal impact on predictive performance.