Fine-scale detection of vegetation in semi-arid mountainous areas with focus on riparian landscapes using Sentinel-2 and UAV data

Abstract

Sparse vegetation such as riparian forests and trees outside forests (TOF) cover only small areas but present various ecological advantages. The detection of these vegetation types in semi-arid mountainous areas is challenging as trees are heavily mixed with other land cover types. Their mapping requires therefore high-resolution imagery. We propose to leverage the advantages and synergies of freely available Sentinel-2 data and a light-weight consumer-grade unmanned aerial vehicle (UAV) with a simple red–greenblue (RGB) camera to detect these vegetation types. In our approach, an object-based random forest land cover classification is first developed over smaller sites using very high-resolution UAV data. The resulting maps are afterwards used as training data for multi-temporal Sentinel-2 based classifications at regional scale. We tested the approach in five different riparian landscapes of a semi-arid mountainous area in Iran. For comparison, mono- and multi-temporal Sentinel-2 data were also used alone – without support from UAV data – to build pixel-based random forest classification models at regional scale. Our results show that compared to the best mono-temporal results, the multi-temporal classification approach improved the overall accuracy and Kappa values of Sentinel-2 classifications from 77.0% to 83.9% and 0.72 to 0.81, respectively. The producer’s and user’s accuracy of the riparian forest class were also improved from 64.0% to 70.0% and 57.1% to 73.7%, respectively. Combining UAV and Sentinel-2 data improved the overall accuracy only slightly, but enabled a much better detection of Persian oak stands – for this class, the producer’s accuracy increased by 13.0 percentage points. Overall, we recommend the combined use of UAV and multi-temporal Sentinel-2 data to detect Persian oak forest stands.