TRACEVE Datasets for Mapping Evergreen Broad-Leaved Cover

This dataset collection was developed in a side study of the TRACEVE project (https://www.uibk.ac.at/en/projects/traceve/) and consists of three datasets: s2_vpo2024.nc, s2_vdb2000.nc, and s2_upd.nc, used to train and evaluate deep learning models for mapping evergreen broad-leaved (EVE) cover in Italian forests. s2_vpo2024.nc: A reference dataset of vegetation plot observations from 1440 plots across five protected areas in Italy. It includes species-level cover values across multiple vertical layers, with a focus on broad-leaved forests. s2_vdb2000.nc (Forest Vegetation Database Italy): A large-scale collection of 4807 plot observations from various studies across Italy (not older than 2000), used for supervised pretraining. It provides broader ecological diversity for improved model generalization. s2_upd.nc (Unlabeled Pretraining Dataset): A dataset of 92869 randomly sampled points across Italian forests, derived from Sentinel-2 time series, used for self-supervised learning. Each plot is assigned the corresponding preprocessed Sentinel-2 L2A time series: cloud removal based on Sen2Cor Scene Classification offset correction according to latest baseline DN scaled to reflectance These datasets support research on deep learning-based forest mapping and pretraining strategies. They can be used in studies on transfer learning, time-series analysis, and large-scale vegetation monitoring in Italy. For more details, see the related publication: Advancing forest mapping: pretraining strategies and deep-ensemble based uncertainty for predicting evergreen broad-leaved cover.