Spatial and temporal classification and prediction of aspen probability in boreal forests using machine learning algorithms

Abstract

Mapping and classifying the probability of occurrence of Populus tremula L. (aspen) in boreal forests is a complex task for sustainable forest management and biodiversity conservation. As a key broadleaved species in the taiga region, aspen supports both forestry and local biodiversity habitats. This study introduces a methodology to predict aspen presence using Sentinel- 2 satellite data and machine learning, combining binary classification (presence/absence) with probability estimation. We utilized spectral features (e.g., NDVI, EVI, SAVI) extracted from Sentinel- 2 imagery, employing a logistic regression model to classify aspen occurrence. To assess feature importance, we applied Permutation Importance (PI) and SHAP, enhancing model interpretability and ensuring transparency in identifying influential factors for forest management applications. Results revealed the significant role of spectral features in determining aspen growth probability. SAVI exhibited a strong effect on classification accuracy due to its soil correction capability, while EVI and NDVI proved highly important in summer, reflecting seasonal vegetation dynamics. High EVI values often indicate complex vegetation and conifer biomass, whereas aspen, with its distinct canopy and phenology, shows lower EVI compared to conifers. NDVI, tied to aspen’s photosynthetic activity, remained a reliable indicator in mixed taiga forests. The model achieved an overall accuracy of 94.77% with XGBoost and 95.03% with Random Forest across all seasons, demonstrating robust performance. This reliable aspen distribution data aids forestry planning, such as harvesting, and the algorithm automates inventorying and mapping aspen stands, reducing reliance on labor-intensive ground surveys.