A2054
Title: Two-stage deep learning framework for spatiotemporal interpolation and probabilistic forecasting of air temperature
Authors: Nouman Iqbal - University of Salento (Italy) [presenting]
Monica Palma - University of Salento (Italy)
Abstract: Accurate spatiotemporal interpolation and forecasting of environmental variables are important for climate analysis and environmental monitoring. Daily air temperature exhibits nonlinear dynamics, seasonal variability, and spatial nonstationarity that are difficult to capture using traditional geostatistical methods. Classical kriging approaches rely on stationarity assumptions and predefined covariance structures, which may reduce flexibility and increase computational costs for large spatiotemporal datasets. A two-stage deep learning framework is proposed for spatiotemporal interpolation and probabilistic forecasting of daily mean air temperature. In the first stage, a spatiotemporal deepkriging model with multi-resolution basis function embeddings captures spatial and seasonal dependencies without explicit covariance estimation. In the second stage, QMLP, QLSTM, and QConvLSTM models are evaluated for multi-step probabilistic forecasting. Model performance is assessed using average MAE, RMSE, and $R^2$ across spatial locations. The framework is applied to NASA POWER MERRA-2 daily meteorological data from 1982-2023 at 30 locations in the Apulia region of Italy. Results indicate improved forecasting accuracy and computational efficiency compared with covariance-based geostatistical approaches.
