Deep features from Pre-trained CNNs and a vision transformer for amazonian timber species recognition using classical machine learning and random image patches

Description

This dataset contains 864 macroscopic transverse-section wood images of six commercially harvested timber species from the Brazilian Amazon: Eschweilera grandiflora, Lecythis chartacea, Lecythis pisonis, Pseudopiptadenia suaveolens, Pouteria ramiflora, and Tachigali myrmecophila. The dataset comprises samples from 24 trees and 72 wood blocks. Macroscopic images were acquired using a Bioptika L60 stereomicroscope at 10× magnification, with image dimensions of 2569 × 1920 pixels. Wood samples were collected from the Forest Management Area of Fazenda Rio Flores I, located in Vitória do Xingu, Pará, Brazil (SIRGAS 2000; 52°13′19.42″ W, 02°52′06.12″ S). The property covers a total area of 286.5913 ha, of which 157.9396 ha corresponds to the Legal Reserve Area. Forest harvesting activities were authorized by the competent environmental agency under Forest Exploitation Authorization (FEA) No. 273936/2021 (SEMAS-PA, 2021). The six timber species were randomly selected for sampling, and access to genetic heritage was registered in CGEN/SISGEN under registration No. A599206. This dataset is associated with the article: Deep features from Pre-trained CNNs and a vision transformer for Amazonian timber species recognition using classical machine learning and random image patches, published in the Journal of the Indian Academy of Wood Science. DOI: 10.1007/s13196-026-00425-5.

Steps to reproduce

Wood samples from commercially harvested timber species in the Brazilian Amazon were collected from the Forest Management Area of Fazenda Rio Flores I, located in Vitória do Xingu, Pará, Brazil (SIRGAS 2000; 52°13′19.42″ W, 02°52′06.12″ S). The property covers a total area of 286.5913 ha, of which 157.9396 ha corresponds to the Legal Reserve Area. Forest harvesting activities were authorized by the competent environmental agency under Forest Exploitation Authorization (FEA) No. 273936/2021. Six timber species were randomly selected for sampling, and access to genetic heritage was registered in CGEN/SISGEN under registration No. A599206. After species selection, wood discs approximately 15 cm thick were removed from the stumps of recently felled trees during forest harvesting operations, with at least two trees sampled per species. From each disc, a radial strip approximately 15 cm wide and 15 cm thick was obtained from bark to bark through the pith. Three solid wood blocks measuring 10 × 10 × 10 cm were subsequently prepared from each strip. The transverse surfaces of the blocks were sanded with 120-grit sandpaper, and macroscopic images were acquired using a Bioptika L60 stereomicroscope at 10× magnification, with image dimensions of 2569 × 1920 pixels. For each wood block, six macroscopic images were acquired at different positions on each of the two transverse surfaces, yielding 12 images per block. Because three blocks were sampled from each tree, 36 images were obtained per tree, resulting in a total of 864 images from 24 trees. The dataset represents six species belonging to five genera and three botanical families. Image acquisition covered the entire radial extent of the stem, from bark to bark through the pith, without distinguishing between heartwood and sapwood.

Institutions

• Universidade Federal do Parana

  PR, Curitiba
• Universidade Federal do Para

  PA, Belem
• Universidade Federal do Recôncavo da Bahia

  Bahia, Cruz das Almas

Categories

Funders

• National Council for Scientific and Technological Development

  Federal District, Brazil

  Grant ID: 303352/2022-1
• Office to Coordinate Improvement of University Personnel

  Grant ID: 001

Related Links

Licence