TreeSp2Vec 1

TreeSp2Vec 1 is the foundation stone of Latent Forest. It is the initial and simplest approach for tree species representation learning.

It was initially developed for learning a latent space of tree species existing in Spanish forests using exclusively national forest inventory data and environmental cartography. Trees are represented in a latent space of 16 dimensions basing on tree size, phytosociology and habitat features.

TreeSp2Vec 1 implements a non-variational non-generative architecture, a multi-layer perceptron, focused on tree species classification for producing two latent spaces of 16 dimensions tree- and species-level embeddings (E and W, respectively).

Besides of including tree- and species-level embeddings, there are also two available versions of TreeSp2Vec 1 basing on the development focus. "P" embeddings are trained seeking maximum performance and, consequently, only include species with high classification accuracy (Matthew's Correlation Coeficient > 0.5). In contrast, "A" embeddings are applicability-oriented, meaning that more species are considered at the expense of lower accuracy. For guiding user applications of TreeSp2Vec 1, the classification accuracy for each species is included in the downloadable files (column "MCC").

TreeSp2Vec 1 by Latent Forest project is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

TreeSp2Vec 1 features

Scale	tree, species
Tree size features	✓
Phytosociological features	✓
Habitat features	✓
Anatomophysiological features	-
Variational	-
Generative	-
Structure-aware	-
Dynamic	-
Download	✓ (species)
* tree-level embeddings are not available yet... but coming soon :)

Download TreeSp2Vec 1

Changelog

TreeSp2Vec-1.00-W-P (07/2021)

- Version 1.00, Performance (P), species-level (W). Species: 40. Countries: Spain.