Leveraging Sentinel-1/2 time series and deep learning for accurate forest tree species mapping

Accurate mapping of tree species is critical for forest management, carbon sequestration estimates and ecosystem assessment. Remote sensing provides an efficient approach using satellite image time series (SITS), but complex data poses challenges for classifiers and feature analysis. This study presents a deep learning-based classification method using Sentinel-1/2 SITS for mapping forest tree species and tree species biodiversity. Specifically, temporal data from unlabeled forest pixels were used for pretraining the model through self-supervised learning, followed by fine-tuning with species samples, enhancing model performance. Various configurations of temporal data were tested for classification, and their impact was evaluated. To address species maps accuracy overestimation caused by homogeneous pure-species stands, a pseudo-labeling approach was employed to incorporate mixed-species scenarios. Additionally, statistical and visualization methods were applied to SITS and model analysis. The results showed that longer time series tended to improve species identification and model confidence, with OA increasing from 0.496 (6–7 months) to 0.795 (1–12 months), macro-F1 from 0.384 to 0.779, and a significant improvement in predicted scores. As data from subsequent year was incorporated, accuracy growth slowed and stabilized, reaching OA of 0.847 and macro-F1 of 0.836, compared to 0.764 and 0.737 for the non-pretrained model. Certain vegetation indices, such as NDre and NDVIre, which are sensitive to physiological changes, highlight species differences during key phenological stages, especially between deciduous and evergreen species. This study demonstrates the potential of combining SITS with deep learning for species classification and provides a comprehensive analysis, contributing to ecological research and sustainable forest management.