Regional drought prediction from Sentinel-2 time series using Random Forest, DNN, and 1D-CNN: a case study in Marchfeld, Austria

In recent years, the integration of machine learning (ML) with earth observation data has improved early warning systems and drought management strategies through advanced drought monitoring and prediction. This study proposes an operational framework that synergistically combines multiple ML models with high-resolution Sentinel-2 data to estimate agricultural drought conditions, while focusing on maize fields in Eastern Austria. The study area includes both irrigated and non-irrigated fields, allowing for comparative performance analysis under different water regimes. A comprehensive network of reference points was established using over 20 satellite-derived indices integrated with ground data including field capacity, precipitation, and soil composition. The integrity of the reference points was further tested with temporal analysis and expert validation. The study employed Random Forest (RF), and two deep learning models, Deep Neural Network (DNN) and One-Dimensional Convolutional Neural Network (1D-CNN), to generate pixel-level drought maps from pre-processed time-series sentinel-derived variables. The prediction accuracy is evaluated over multiple years (2018–2023). The results reveal distinct differences in model performance, with non-irrigated fields demonstrating higher prediction accuracy (on average 79%) and lower error metrics (on average 0.15), likely due to the clearer drought signals they present. In contrast, irrigated fields present more significant drought patterns, which increase the complexity of prediction, reflected in lower prediction accuracy (on average 77.4%) and error metrics (on average 0.16). Among all models, DNN demonstrated the best overall performance. The results highlight the advantage of deep learning with Sentinel-2 in operational drought monitoring for improved agricultural drought management.