Federated Split Learning Optimizes IoT Rainfall Prediction with Adaptive Compression

Researchers have developed an innovative federated split learning (FSL) framework specifically for IoT rainfall prediction, designed to overcome communication bottlenecks in bandwidth-constrained environments. Traditional FSL approaches often optimize either activation compression or synchronization frequency in isolation. This new framework, however, jointly manages both activation compression and the synchronization interval using a latency-driven scheduler on the server side, incorporating per-client Exponential Moving Average (EMA) smoothing. Evaluated through extensive simulations and a real-world Raspberry Pi deployment, the system demonstrated remarkable efficiency. It achieved an 87% reduction in activation upload payload and a 54% reduction in synchronization traffic compared to a float32 baseline, while also significantly reducing runtime jitter from +/-688 seconds to +/-10 seconds. Crucially, these substantial communication savings were achieved with only minor variations in predictive quality (AUPRC), confirming that aggressive quantization and sparser aggregation do not materially degrade performance in this context.