A novel two-step framework for bush biomass change-detection and quantification in Namibia using ground- and drone-based LiDAR - a pilot study

Abstract

Across the African continent, shifts in bush vegetation patterns are increasingly reported, challenging local economies and communities, while their ecological implications remain subject to debate. Although remote sensing approaches capture the phenomenon well at larger scales, tracking bush vegetation changes locally is crucial, since local drivers explain most of the variation. Here, we aimed to quantify bush biomass on a local scale in north-central Namibia by deploying mobile laser scanning (MLS) and unmanned aerial vehicle laser scanning (UAV-LS). To achieve this, we implemented a novel two-step voxelization framework that first applies a change-detection algorithm to identify harvested biomass, followed by a subsequent quantification step. In a cross-validation approach, we were able to predict bush biomass by an average relative mean absolute error (øRMAE) of 17.02 % and 20.52 % using MLS and UAV-LS data respectively. Although the small difference between platforms was surprising, the narrower range of MLS RMAE values (11.04 - 23.55 %) relative to UAV-LS (11.28 - 31.23 %) indicated a greater robustness of the MLS approach. Depending on the chosen voxel-size, to detect change between two point clouds, the analysis revealed a trade-off between limiting the registration error and capturing detail. The results presented here demonstrate the feasibility of the approach while highlighting substantial potential for improving the change-detection algorithm, model complexity, and the experimental setup, ultimately allowing more accurate biomass estimations and broader scalability. Thereby, local changes in bush biomass could be reliably quantified, paving the way for future research on its ecological implications and consequences. • Mobile- and UAV laser scanning data can predict bush biomass in Namibia. • A change detection algorithm identifies harvested biomass. • Since voxel-based, the algorithm balances registration error with captured detail. • Prediction errors were minor across platforms, yet MLS yielded more robust results.