Deep Learning-Based Multimodal Satellite Precipitation Retrieval with HPC-Optimized Inference

Abstract

Accurate precipitation retrieval is important for weather monitoring and hydrological applications, especially where radar coverage is limited or affected by terrain. This thesis adapts a deep-learning precipitation retrieval framework to geostationary Meteosat Third Generation Flexible Combined Imager (FCI) observations and extends it with passive microwave (PMW) observations from the Advanced Technology Microwave Sounder (ATMS) and Arctic Weather Satellite (AWS). The main aim is to evaluate whether sparse PMW observations improve FCI-only precipitation retrieval over the MetCoOp Ensemble Prediction System (MEPS) domain, and to improve the efficiency of inference on a GPU-based system. The retrieval model estimates near-surface precipitation from recent satellite observa tions using a 3D encoder-decoder convolutional neural network. FCI provides dense and temporally continuous geostationary input, while ATMS and AWS provide sparse PMWinput, available only when polar-orbiting swaths pass over the domain. Missing PMW observations are handled with masks, allowing full-domain retrievals even when microwave coverage is absent. The models are evaluated against BALTRAD radar-derived precipitation for selected periods in 2025. The FCI-only model captures the general precipitation structure, but its precipitation amount is not stably calibrated across the evaluation periods. The main result is that adding PMW improves several metrics: mean absolute error decreases from 2.70 to 2.14 mm/day, R2 increases from-0.19 to 0.17, and the Matthews correlation coefficient at the 0.1 mm/day threshold increases from 0.22 to 0.35. The benefit is strongest when PMW observations are available and weaker when microwave coverage is missing. A separate zenith-angle experiment shows that viewing geometry can strongly affect the retrieved precipitation distribution, but did not give a balanced improvement. Profiling shows that the original inference pipeline was limited mainly by data loading, small tile processing, CPU-side operations, and NetCDF writing, rather than by the 3D convolutional network alone. Batched tile inference, asynchronous I/O, and multi-GPU execution improved throughput. The optimized two-GPU pipeline achieved a 1.58× speedup for near-real-time processing, while temporal sharding achieved a 2.27× speedup for offline processing. Overall, the thesis shows that dense FCI input can be combined with sparse PMW observations, and that system-level optimizations can substantially speed up inference without changing the trained model.