Deep learning the droughts: Spatiotemporal monitoring of protracted dry conditions in Queensland, Australia

Protracted droughts, defined as successive dry events occurring before full vegetation recovery, pose an emerging challenge to assessing ecosystem resilience. Conventional drought analyses often treat these events as isolated, overlooking the temporal dependencies that govern vegetation stress. This study addresses this limitation by proposing an integrated spatiotemporal drought assessment approach that links Standardised Precipitation Index (SPI) based identification of individual drought events with vegetation recovery dynamics derived from Normalised Difference Vegetation Index (NDVI) and Vapour Pressure Deficit (VPD), enabling the identification of protracted droughts across Queensland, Australia. Results reveal a noticeable seasonal pattern, with NDVI consistently decreasing following drought periods (SPI-1 < −1) and showing a negative-lagged relationship with VPD, indicating delayed vegetation recovery. The mean recovery duration across Queensland was estimated at 3.38 months, quantified as the time lag between drought termination (SPI ≥ −0.5) and the return of NDVI to baseline conditions. Based on this definition, region-specific recovery lags were derived and used to generate ground-truth labels for protracted drought events. Spatiotemporal drought patterns were analysed using a hybrid deep learning framework integrating a convolutional neural network (CNN) and recurrent layers, including long short-term memory (LSTM) and bidirectional-LSTM (BiLSTM), with and without multi-head attention (MHA). The CNN-LSTM-MHA model achieved the highest performance (accuracy 95.39% and F1-score 0.9569) in detecting protracted drought events across space and time, demonstrating the significance of integrating temporal memory with attention mechanisms. This study shows that coupling remote sensing indices with deep spatiotemporal learning enhances protracted drought detection by explicitly modelling vegetation recovery-dependent temporal dependencies, capturing lagged NDVI responses to atmospheric dryness and enabling more consistent identification of successive drought events before full recovery, thereby improving the interpretation of vegetation resilience and supporting adaptive management in climate-sensitive regions. • Identifies overlapping droughts (protracted drought) before vegetation fully recovers • Uses NDVI and VPD as derived measurements to account for atmospheric dryness • Reveals an average vegetation recovery duration of 3.38 months across Queensland • Deep spatiotemporal model improved protracted drought classification