Mapping dominant leaf type based on combined Sentinel-1/-2 data – Challenges for mountainous countries

• DOI: 10.1016/j.isprsjprs.2021.08.017
• Publication Date: 2021-08-28T17:23:12Z
• Authors: Lars T. Waser; Marius Rüetschi; Achilleas Psomas; David Small; Nataliia Rehush
• Keywords: Thematic map; Elevation (ballistics); Laser Scanning

Countrywide winter and summer Sentinel-1 (S1) backscatter data, cloud-free Sentinel-2 (S2) images, an Airborne Laser Scanning (ALS)-based Digital Terrain Model (DTM) a forest mask were used to model subsequently map Dominant Leaf Type (DLT) with the thematic classes broadleaved coniferous trees for whole of Switzerland. A novel workflow was developed that is robust, cost-efficient highly automated using reference data from aerial image interpretation. Two machine learning approaches based on Random Forest (RF) deep (UNET) country three sets predictor variables applied. 24 subareas aspect slope categories applied explore effects complex mountainous topography performances. The split into training, validation test spatially stratified 25 km regular grid. accuracies both RF UNET generally highest Kappa (K) around 0.95 when predictors included S1/S2 topographic aspect, elevation DTM. While only slightly lower obtained S2 DTM lowest S1 included, performing worse than UNET. countrywide level performed overall similarly, substantial differences in performances, i.e. higher variances accuracies, found northwest northeast orientations. combined use mitigated these problems related shadows therefore superior single or data. comparison independent National Inventory (NFI) plot demonstrated precisions K 0.6 predictions DLT indicated trend increasing deviations mixed forests. Copernicus High Resolution Layer (HRL) 2018 revealed exception pure forest. Although, spatial patterns DTL similar, better areas distinct stand level, largest occurring used. In contrast, more accurate stands. This study goes beyond case meets requirements sets, particular regarding repeatability, updating, costs characteristics training sets. 10 m maps add complementary explicit information existing NFI estimates are thus relevant forestry practice other fields.