Published September 11, 2026 | Version 1.0

Dealing with the unknown: open-set deep learning enables realistic tree species mapping in a Brazilian Woodland Savanna formation

Description

This repository contains the datasets used to perform all analyses presented in the paper "Tree species mapping in a dense Brazilian Cerrado formation based on UAV imagery and open-set deep learning models".

The repository includes the following data:

  • Orthomosaics: Two orthomosaics used for patch extraction. Both mosaics were used to generate training patches, while orthomosaic.tif was also used during the large-scale inference stage.

  • Shapefiles: Polygon layers representing individual tree crowns. These include manually delineated crowns used for training and validation patch extraction, and automatically delineated crowns used for large-scale inference.

  • Extracted image patches:

    • Training and validation patches extracted from manually delineated crowns and split into 70% training / 30% validation subsets.

    • Unknown species (unseen by modls) patches used for open-set calibration and evaluation.

    • CanopyRS-derived patches extracted from automatically delineated crowns for large-scale inference.

Files

adjacent_mosaic.tif

Files (9.0 GB)

Name Size
md5:87ac68731dafeff59cc4912c74807b36
4.4 GB Preview Download
md5:360335110e503a797dc4a25a5c21442d
2.7 MB Preview Download
md5:464952d1614bbc41bbff1621f7c5e146
470.7 kB Preview Download
md5:f3df165cd49262e2efd5c132b43b12a6
2.2 GB Preview Download
md5:6617c682061f264a2a5a72d559d0e850
2.5 GB Preview Download

Additional details

Funding

Fundação de Amparo à Pesquisa do Estado de São Paulo
National Council for Scientific and Technological Development

Dates

Created
2026-09-10