Advanced time series forecasting of vegetation health using deep learning models: A remote sensing approach to analyzing climate change impact

The growing consequences of climate change on vegetation ecosystems require advanced predictive tools for environmental monitoring and adaptive management. This research explored a new application of hybrid deep learning models to forecast the Normalized Difference Vegetation Index (NDVI) time series, using Sentinel-2 high-resolution satellite images. Specifically, this research investigated vegetation dynamics in four climatically different regions of Northern Iraq from 2016 to 2024, developing and comparing eight deep learning models, including traditional recurrent networks (Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Gated Recurrent Unit (GRU)) and Convolutional Neural Networks (CNN), resulting in unique hybrid models that combine spatial and temporal feature extraction mechanisms. The study utilized a large dataset of 43,200 images with a spatial resolution of 10 m, employing systematic data preparation that included NDVI processing (NDVI calculations, normalization, and time-series sequence construction) necessary for model training and learning. The model performance was rigorously evaluated, where hybrid models were demonstrated to outperform other models, with BiLSTM-GRU appearing to deliver high accuracy (coefficient of determination scores R 2 of up to 0.851) and low prediction errors (Mean Squared Error (MSE) as low as 6.04 × 10 −4 ). In terms of ecological region, model performance was assessed across regions, as well as across different regions, finding general trends in performance, particularly in regions with homogeneous vegetation cover at each time sampling period. The Monte Carlo dropout method offered the opportunity to infer uncertainty, which in turn helped build confidence in predictions. The predictions for the future periods of 2025–2028 show promising seasonal patterns and long-term trends, which are important with respect to climate-adjusted planning.