# Deep Pre-trained Time Series Features for Tree Species Classification in the Dutch Forest Inventory

**Authors:** Takayuki Ishikawa, Carmelo Bonannella, Bas J. W. Lerink, Marc Rußwurm  
**arXiv:** [2508.18829](https://arxiv.org/abs/2508.18829)  
**Date:** 2025-08-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.18829v1/

## Key points

- The National Forest Inventory (NFI) is a crucial source of forest information, providing essential data on tree species distribution.
- Traditional on-site campaigns for maintaining NFI are labor-intensive and pose challenges in terms of frequency and scale.
- Remote sensing approaches combined with machine learning offer promising solutions for updating NFIs more efficiently and at larger scales.
- Satellite Image Time Series have been effective in distinguishing tree species based on seasonal canopy reflectance patterns.
- Leveraging deep features from pre-trained remote sensing models can significantly enhance classification accuracy compared to traditional methods.
- A recent study focusing on the Dutch Forest Inventory achieved remarkable results by fine-tuning a publicly available remote sensing time series foundation model, surpassing the current state-of-the-art NFI classification by up to 10% across all datasets.

## Summary

The National Forest Inventory (NFI) is a crucial source of forest information, providing essential data on tree species distribution. However, the traditional method of maintaining these inventories through labor-intensive on-site campaigns poses challenges in terms of frequency and scale. To address this issue, remote sensing approaches combined with machine learning offer promising solutions for updating NFIs more efficiently and at larger scales. Satellite Image Time Series have been effective in distinguishing tree species based on seasonal canopy reflectance patterns. Current approaches typically rely on Random Forest classifiers with hand-designed features and phenology-based metrics. However, leveraging deep features from pre-trained remote sensing models presents a complementary strategy that can significantly enhance classification accuracy. In a recent study focusing on the Dutch Forest Inventory, researchers extracted time-series data from Sentinel-1, Sentinel-2, ERA5 satellites, and SRTM data using Google Earth Engine. By fine-tuning a publicly available remote sensing time series foundation model, the research team achieved remarkable results surpassing the current state-of-the-art NFI classification by up to 10% across all datasets. The study highlights that classic hand-defined harmonic features may be too simplistic for complex tasks like tree species classification in NFIs. By harnessing deep AI features from pre-trained models and openly available satellite data, this approach demonstrates significant improvements in classification accuracy compared to traditional methods. This innovative methodology not only enhances the efficiency of existing forest inventory processes but also showcases the potential of utilizing advanced technologies to overcome data limitations and achieve more accurate results in environmental monitoring and management.