Espeletia Scene Classification and Canopy Height Dataset from UAV Multispectral Imagery in a Colombian Páramo

This dataset contains multimodal UAV-derived image patches and metadata for ecological monitoring of Espeletiopsis corymbosa in a páramo ecosystem in Villapinzón, Cundinamarca, Colombia. It was created to support the development and benchmarking of machine learning methods for scene classification and related regression tasks in high-mountain vegetation monitoring.

The dataset is derived from a UAV survey acquired with a MicaSense Altum sensor and associated photogrammetric products. For each annotated tile, the dataset provides:

• a multispectral-plus-thermal image patch,
• a canopy height model (CHM) patch,
• a digital surface model (DSM) patch,
• and a tile-level class label.

The primary benchmark task is scene classification, where each tile is assigned one of three labels describing the presence and health condition of Espeletiopsis corymbosa:

0: no Espeletiopsis corymbosa present in the tile,
1: healthy Espeletiopsis corymbosa present,
2: unhealthy Espeletiopsis corymbosa present.

In addition to classification, the inclusion of CHM and DSM layers makes the dataset suitable for regression and multitask learning experiments, for example canopy height estimation or structure-aware classification.

The dataset was built from polygon annotations defined on a geospatial grid. Each polygon corresponds to a tile of approximately 256x256 pixels, although the exact native crop size may vary slightly depending on raster alignment and geospatial bounds. Exported patches were standardized to a fixed size for machine learning use.

This release focuses on the larger UAV flight and is intended as a benchmark-ready dataset for multimodal learning in ecological remote sensing.

Dataset Structure

espeletia_scene_cls_chm/
├── train/
│   ├── ms/
│   ├── chm/
│   ├── dsm/
│   └── labels.csv
├── val/
│   ├── ms/
│   ├── chm/
│   ├── dsm/
│   └── labels.csv
├── test/
│   ├── ms/
│   ├── chm/
│   ├── dsm/
│   └── labels.csv
├── dataset_summary.csv
├── split_statistics.csv
└── dataset_config.json