Dataset for Tree Species Detection in Heterogeneous Forests Using Aerial RGB Imagery

Dataset Description This dataset contains labeled and unlabeled aerial RGB image patches developed for individual tree detection and tree species classification using deep learning methods. The labeled imagery was derived from the SWISSIMAGE orthophoto product provided by the Swiss Federal Office of Topography (swisstopo) [1]. SWISSIMAGE provides high-resolution orthorectified RGB aerial imagery with a ground sampling distance (GSD) of approximately 10 cm. The imagery was acquired using the Leica ADS100 push-broom scanning system and covers forest areas in northern Switzerland, including parts of the Swiss Central Plateau and Jura regions across nine cantons. The image acquisition period was between 2017 and 2021 during the vegetation season. The labeled dataset contains 4,438 aerial RGB image patches with 10,128 expert-provided bounding-box annotations representing four tree species: C1: Norway spruce (Picea abies)C2: Silver fir (Abies alba)C3: Scots pine (Pinus sylvestris)C4: European beech (Fagus sylvatica) The original tree reference records were compiled from three independent sources: the Arborizer dataset, the Swiss Long-term Forest Ecosystem Research programme (LWF), and the Swiss National Forest Inventory (NFI) [2]. The dataset was obtained from the labeled image collection introduced in the reference study [2]. Since the original file-level membership of the training, validation, and test subsets was not publicly available, the experimental partitions were reconstructed from the complete labeled collection. To prevent information leakage, all partitions were generated at the image level. Each source image and all associated annotations were assigned exclusively to a single subset. Filename-based verification confirmed that no image overlap exists among training, validation, and test subsets. Two image-disjoint partitioning strategies are provided: Initial image-disjoint partitionDesigned for baseline Faster R-CNN experiments and preliminary pseudo-labeling studies.Distribution-aware reconstructed partitionConstructed to preserve image-level independence while approximating the class distribution and annotation density characteristics reported in the reference study [2]. In addition, this dataset includes 7,059 unlabeled RGB image patches extracted from publicly available SWISSIMAGE imagery [1]. Each patch has a size of 256 × 256 pixels and the same spatial resolution and acquisition characteristics as the labeled images. These unlabeled images were prepared for semi-supervised learning and pseudo-label generation experiments. All labeled and unlabeled datasets were checked for filename overlap to ensure independence between training, validation, test, and unlabeled subsets. This dataset can support research in: individual tree detection,tree species classification,weakly supervised learning,semi-supervised learning,pseudo-label generation,and deep learning-based forest remote sensing.References [1] Swiss Federal Office of Topography (swisstopo).SWISSIMAGE: High-resolution aerial orthophotos of Switzerland.Available at:https://www.swisstopo.admin.ch/ [2] Reference dataset study:Tree species detection and classification from aerial RGB imagery using deep learning approaches.The labeled dataset introduced in this study is publicly available through Zenodo: