Dottorato in "Transizione Digitale e Sostenibilità": le imprese e le Amministrazioni pubbliche nell'economia globalizzata – Dipartimento di Scienze dell'Economia dell'Università del Salento
Knowledge Sharing and Scientific Interaction
Titolo: DeepKriging and Deep Learning Framework for Spatiotemporal Predictions of Air Temperature in a Multivariate Context
Abstract: Accurate temperature prediction is essential for agriculture, water management, and climate adaptation. This study examines univariate spatiotemporal forecasting of daily mean temperature (T2M) using Holt–Winters, SARIMA, neural network autoregression (NNAR), and feed-forward artificial neural networks (ANN). The best-performing models among these are combined with residual spatiotemporal kriging, yielding a hybrid that improves spatial prediction but remains constrained by its univariate scope and simplifying statistical assumptions. To address these limitations, a multivariate framework based on Spatiotemporal DeepKriging is developed. The model integrates auxiliary meteorological covariates (Tmin, Tmax, relative humidity, and surface pressure) with spatial (Wendland radial basis functions) and temporal (Gaussian kernel) embeddings, enabling the capture of complex nonlinear dependencies. The network is trained and validated on station data using RMSE, MAE, and R², with results showing that DeepKriging significantly outperforms the univariate and hybrid baselines. In parallel, deep learning architectures—multilayer perceptron (MLP), long short-term memory (LSTM), and Conv1D-LSTM—are employed for probabilistic forecasting, capturing nonlinear dependencies, long-term dynamics, and local sequential features.
For spatial prediction, auxiliary covariates are first estimated at unsampled grid points through Random Forest Spatial Interpolation (RFSI). These estimated fields, together with spatiotemporal embeddings, are then provided to the trained DeepKriging network to generate high-resolution maps of daily mean temperature. Using daily observations from 30 meteorological stations in Apulia, Italy (1982–2023), the framework demonstrates superior forecasting performance and produces accurate spatiotemporal temperature surfaces, thereby offering a robust tool for climate-sensitive applications.
.
Relatore: Dott. Nouman Iqbal
Data: 07.10.25 ore 13
Location: Ecotekne, Building C, I piano, Aula 'Mario Signore'