Rain-Code Fusion: Code-to-Code ConvLSTM Forecasting Spatiotemporal Precipitation

Recently, flood damage has become a social problem owing to unexperienced weather conditions arising from climate change. An immediate response heavy rain is important for the mitigation of economic losses and also rapid recovery. Spatiotemporal precipitation forecasts may enhance accuracy dam inflow prediction, more than 6 h forward mitigation. However, ordinary ConvLSTM limitation predictable range 3-timesteps in real-world forecasting irreducible bias between target prediction ground-truth value. This paper proposes rain-code approach spatiotemporal code-to-code forecasting. We propose novel rainy feature that represents temporal process using multi-frame fusion timestep reduction. perform studies with various term ranges based on standard ConvLSTM. applied region within Japanese hourly data, under 2006 2019 approximately 127 thousands hours, every year May October. apply radar analysis data central broader an area 136 × 148 km2. Finally we have provided sensitivity size several range.