Assessing the value of temporal information for invasive alien tree mapping using Sentinel-2 imagery in southern Africa

This repository contains the data and code developed in the MAPWAPS Project, where methods to improve the accuracy of invasive alien tree mapping were explored.

This code implements the processing workflow used to evaluate how different levels of temporal information influence invasive alien tree (IAT) mapping accuracy using Sentinel-2 imagery. The workflow constructs five feature sets of increasing temporal complexity, ranging from single-date mosaics to a full annual stack of acquisition-date mosaics.

Sentinel-2 imagery for a given calendar year (in this case, 2023) is first composited into weekly scenes, with images of lower cloud cover given preference within each week. These weekly composites are then filtered to retain only those with less than 40% cloud cover. The resulting scenes form the base temporal dataset from which all treatments are derived. Weekly compositing is used to ensure consistent spatial coverage across large study areas, where multiple satellite overpasses with different acquisition dates may be required to fully cover the region of interest. Four commonly used spectral indices relevant to vegetation condition are subsequently calculated and appended as additional bands.

Two treatments consist of single-date mosaics selected to represent different phenological windows (Treatments 1 and 2). In addition, three multitemporal scenarios are generated from this base dataset: a single annual composite derived from all scenes (Treatment 3), twelve monthly composites representing seasonal dynamics (Treatment 4), and the full stack of weekly mosaics (Treatment 5).

For each treatment, the resulting imagery—either a single composite or a stacked multitemporal feature image—is converted into a predictor dataset suitable for machine learning. Training samples derived from field observations are extracted and used to train a Random Forest classifier using a 70–20–10 cross-validation approach. Accuracy metrics, together with visual inspection of classification outputs, are then compared across treatments to assess whether incorporating temporal information improves invasive alien tree mapping performance.