Implemented state-of-the-art ConvLSTM and TrajGRU models for precipitation nowcasting, leveraging satellite data with higher spatial (4 km x 4 km) and temporal resolution (15 min) . Created datasets of 1,560 and 779 sequences to train and test. Achieved RMSE of 3.69 mm/hr with 2-hour lead time and RMSE of 8.06 mm/hr with 4-hour lead-time
Precipitation nowcasting for short-term storm forecasting (0–6 hours) is essential for timely severe weather warnings. Traditional methods such as numerical weather prediction (NWP) and radar extrapolation, often lack accuracy at short scales and are computationally intensive. Recent deep learning models, such as ConvLSTM and TrajGRU have offered promising advances by capturing complex spatiotemporal dynamics. This paper aims to evaluate these models on satellite data, addressing the limitation posed radar’s limited global coverage, while focusing on the region of Sindh, Pakistan — a region with minimal meteorological infrastructure. Thus, by contributing towards the improvement of global nowcasting capabilities this work addresses critical forecasting needs heightened by climate change.
REPORT: Final_Report.pdf
Sample Test Result: Target (Left) vs Prediction (Right)
