Advancing forest mapping: Pretraining strategies and deep-ensemble based uncertainty for predicting evergreen broad-leaved cover from Sentinel-2 time series

Abstract

The distribution changes of evergreen broad-leaved tree and shrub species (EVE) at the border of Mediterranean and temperate forests due to climate change and land-use changes necessitates accurate mapping techniques to support biodiversity monitoring and climate adaptation strategies. Remote sensing time series provide valuable data for vegetation analysis, yet functional-level mapping in mixed forests remains challenging due to limited observation data. To tackle this limitation, the present study investigates pretraining strategies for mapping EVE cover in selected Italian forest areas using Sentinel-2 time series and a probabilistic Convolutional Neural Network. Additionally deep ensemble-based uncertainty estimation is used to further enhance the interpretability of the output. We compare three model training strategies: (i) direct training on up-to-date but limited field data, (ii) supervised pretraining on a larger, diverse forest vegetation database before fine-tuning, and (iii) self-supervised pretraining on large-scale unlabeled time series before fine-tuning. Our results demonstrate that pretraining on contextually similar datasets in combination with a spatial split of training and validation data significantly enhances predictive performance and generalization to unseen regions in a cross-validation experiment. Additionally, we assess epistemic and aleatoric uncertainty to improve interpretability and identification of out-of-distribution predictions. This study highlights the benefits of pretraining and uncertainty quantification for large-scale remote sensing applications with limited availability of labeled data.