Temperature Time Series Prediction Using Spatiotemporal Deep Learning Model

Time-series remote sensing data is a valuable source of information that may be applied to different tasks, including urban expansion tracking, crop monitoring, flood monitoring and coastline management. Despite the increasing availability of spatiotemporal data, accurately modeling its complex dynamics remains a significant challenge due to the interplay of numerous intricately interconnected factors across both spatial and temporal domains. Conventional approaches often leverage Convolutional Neural Networks (CNNs) to capture spatial correlations. However, these CNN-based methods typically rely on a single-channel data structure to represent diverse periodic patterns. This reliance on a singular channel renders the model susceptible to over-parameterization when attempting to capture intricate periodic dependencies, while simultaneously introducing information loss through the convolution process. To address these limitations, we propose a novel model that leverages a video-shaped multi-channel data structure to efficiently represent diverse patterns in spatiotemporal data. To mitigate the limitations of existing approaches, a mixed-pointwise convolution was integrated, enabling the capture of periodic dependencies without incurring over-parameterization or information loss. Experiments were conducted on Coordinated Regional Climate Downscaling Experiment (CORDEX) dataset for the Egypt for the 20-year period from 2001 to 2020. The obtained results demonstrate its superiority of the proposed approach over the current state-of-the-art, achieving a reduction in root mean squared error of up to 3%.