Supporting ecological restoration: leveraging satellite data and machine learning to map invasive alien tree biomass

Abstract

Invasive alien trees pose significant ecological and economic threats globally, including biodiversity impacts, and significant water consumption. However, the biomass of invasive alien trees presents an opportunity to offset the high costs of clearing as part of restoration programmes. This study presents one of the first repeatable, regional-scale modelling frameworks for estimating invasive alien tree biomass at a 25 m resolution. The framework evaluated several machine learning algorithms (Random Forest, Neural Networks, Support Vector Machine, Gradient Tree Boost and Classification and Regression Tree) within various model configurations using different combinations of predictors from satellite sensors/data products (Sentinel-1, Sentinel-2, GEDI, Global Canopy Height, and SRTM). Field-measured biomass and airborne lidar data were used for model training and validation. The model with the best performance was also the simplest, a model integrating predictors from Sentinel-1, Sentinel-2, and SRTM and using the Random Forest classifier (validation: r s = 0.72; R 2 = 0.52; RMSE = 0.89 t/pixel; independent ground-truthing: r s = 0.96; R 2 = 0.70; rRMSE = 59.95%). This model was applied across three bioclimatically distinct catchments in southern Africa to demonstrate how this framework can inform the prioritisation of clearing of invasive alien trees based on the maximisation of biomass extraction. This research supports ecological restoration through alien tree clearing by bridging the gap between restoration practitioners, decision-makers and the private sector in relation to biomass extraction, for green biofuels or the biochar industry. This work could support the leveraging of alternative finances for restoration, assisting the Global South with reducing reliance on public sector funding. • Regional invasive alien tree (IAT) biomass maps are required to support restoration. • Machine learning with satellite predictors were used to model IAT biomass (t/25 m 2 ). • Different model configurations were validated and ground-truthed in southern Africa. • A RF model using predictors from Sentinel-1, Sentinel-2, and SRTM performed best. • Biomass maps of IATs for three catchments were produced to illustrate model utility.