Abstract Accurate maps of tree species distribution are essential for forest science and management but remain difficult to generate over large regions. New algorithms from the field of deep learning may have better abilities to extract species-specific signatures from complex time-series signals of satellite imagery. Here, we propose a transformer-encoder model, derived from SITS-BERT that integrates Sentinel-2 time-series with forest inventory data to classify dominant tree species in Baden-Württemberg, Germany. The model distinguishes eight classes of major tree species with an overall accuracy of 77%, based on a conservative validation approach that relies on completely independent test samples. In direct comparison, our classification clearly outperforms alternative map products. We further analyse the influence of training data quality and quantity, and particularly examine whether the canopy cover at the location of the reference sample points has an effect on the obtained results. We demonstrate that transformer architectures remain relatively robust even with limited or noisy references but at the same we show that the model improves if only samples with a high canopy cover are used during training. Our results highlight the potential of deep learning for large-scale forest species mapping.
Költzow et al. (Thu,) studied this question.