Preprocessing Pipeline for Time Series-Based Satellite Image Forecasting

The paper proposes a novel preprocessing pipeline framework for forecasting satellite images using historical timeseries data. While traditional remote sensing approaches predict scalar metrics or class labels, this method aims to generate fullresolution satellite images that represent the likely state of a region over a multi-year horizon. The core objective is to model and predict complex spatio-temporal patterns, such as urban growth, water body changes, land cover changes. The research uses a high-resolution, pixel-annotated satellite dataset FLAIR#2 to train and evaluate the proposed models. The result is a robust pipeline that aids in the creation of an interpretable, visual forecasting tool that helps stakeholders anticipate and understand long-term landscape changes, bridging the gap between cuttingedge generative modelling, time series analysis and responsible artificial intelligence in the domain. The large data volume is managed by compute optimization techniques, and a preprocessing pipeline was built successfully, using techniques like aligned resampling and temporal aggregation. The efficacy of the pipeline was validated by evaluating the dataset after each stage using a Vision Transformer (ViT) model, and measuring key performance metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Peak Signal-to-Noise Ratio (PSNR).