Evaluating Deep Features from Pre-trained Time-Series Deep Learning Model for Pixel-level Tree Species Classification on Forest Inventory

By Takayuki Ishikawa

Abstract
This study explores using deep learning models for tree species classification in National Forest Inventories (NFIs). While traditional NFI updates require labor-intensive fieldwork, this study investigated whether features from pre-trained deep learning models could improve classification accuracy compared to conventional machine learning methods.
Using datasets from Dutch NFI and the Francini dataset, this study compared features extracted from the fine-tuned Presto model (Tseng et al., 2024) against traditional harmonic and seasonal medoid features by Random Forest classifier. The experiment used time-series data from Sentinel-1, Sentinel-2, ERA5 and SRTM satellites (2020) processed through Google Earth Engine.
Results showed that Presto-derived features substantially outperformed traditional hand-crafted features when using the same Random Forest classifier framework. Interestingly, additional domain-specific pre-training on unlabeled Dutch forest data didn't further improve accuracy.
This approach demonstrates that leveraging pre-trained deep learning models offers a cost-efficient method for large-scale tree species classification in NFIs, potentially complementing traditional inventory methods while enhancing both efficiency and scale of forest monitoring.