Modelling vegetation index after droughts in South Africa using time series spatial data

Abstract

As climate change intensifies, understanding the relationship between drought and vegeta tion dynamics is critical. This study trains a Long Short Term Memory Model (LSTM) to pre dict vegetation greenness when given static and dynamic features as inputs. The aim is to as sess it’s performance, both across time and in specific cases, and explore alternative weather conditions. The model was trained on spatially split data and achieved an overall perfor mance of R2 = 0.87, highlighting its capacity to capture complex spatio-temporal dynamics. Feature ablation showed that precipitation is the most influential predictor, especially during drought and recovery periods. A Brier skill score analysis showed the model’s generally high performance during periods of poor vegetation growth. Counterfactual simulations involv ing perturbed precipitation and temperature inputs during drought events allowed for an ex ploration of alternative climate scenarios. This research shows the versatility and robustness of the LSTM in modelling vegetation dynamics.