Evaluating Temporal Dimensionality in Sentinel-2 Data for Invasive Alien Tree Mapping

Invasive alien trees can transform ecosystems by altering ecosystem processes, modifying habitats, and reducing biodiversity. Accurate maps of invasive alien trees are required for restoration programmes, and for cost-effectiveness the efficient use of freely available datasets is essential. We ask whether increasing temporal dimensionality improves invasive alien tree classification accuracy in dynamic, fire-adapted ecosystems. We systematically tested the benefit of increasing temporal dimensionality of Sentinel-2 imagery from single scenes through annual and monthly composites to weekly timesteps (full stack). We used field-collected observation datasets to train a Random Forest classifier. Ground-truthing was performed to assess transferability and stand-level cohesiveness in classification results. Accuracy improved significantly with temporal dimensionality until a saturation point was reached. This occurred at the monthly composite level, where accuracy improved relative to the annual composite (9%), but showed no meaningful difference compared to the full stack (0.2%). Moderate to high temporal dimensionality (monthly composite and full stack) achieved ~10% higher overall accuracy than single-scene approaches. For alien tree taxon-specific discrimination, moderate to high temporal dimensionality achieved >17% higher accuracy. In terms of feature importance, Enhanced Vegetation Index (EVI) and Normalized Difference Vegetation Index (NDVI) consistently ranked among the top features across treatments, followed by shortwave infrared and visible bands. Invasive alien trees in temporally dynamic, fire-driven grasslands were best discriminated using moderate to high temporal dimensionality. We recommend the inclusion of temporal information in the form of monthly aggregates, while finer temporal resolution may be unnecessary for invasive alien tree mapping in southern African grasslands.