Assessing the Effectiveness of Deep Embeddings for Tree Species Classification in the Dutch Forest Inventory

TL;DR

This work investigates whether deep embeddings from pre-trained remote sensing models can outperform traditional hand-crafted features for Dutch NFI tree species classification, addressing data scarcity and scalability challenges. By evaluating Presto, Tessera, and AlphaEarth embeddings against harmonic features, and by fine-tuning Presto with an MLP, the study shows consistent performance gains (2–9 percentage points) across datasets, with the largest improvements on balanced, larger samples. The findings reveal that fine-tuned embeddings capture richer phenological and structural information than fixed hand-crafted features, and that MLP classifiers can leverage these representations effectively, particularly in data-rich settings. Operationally, the approach offers a computationally efficient path to large-scale, wall-to-wall forest monitoring, highlighting the potential to enhance NFIs while reducing reliance on labor-intensive field campaigns.

Abstract

National Forest Inventory serves as the primary source of forest information, however, maintaining these inventories requires labor-intensive on-site campaigns by forestry experts to identify and document tree species. Embeddings from deep pre-trained remote sensing models offer new opportunities to update NFIs more frequently and at larger scales. While training new deep learning models on few data points remains challenging, we show that using pre-computed embeddings can proven effective for distinguishing tree species through seasonal canopy reflectance patternsin combination with Random Forest. This work systematically investigates how deep embeddings improve tree species classification accuracy in the Netherlands with few annotated data. We evaluate this question on three embedding models: Presto, Alpha Earth, and Tessera, using three tree species datasets of varying difficulty. Data-wise, we compare the available embeddings from Alpha Earth and Tessera with dynamically calculated embeddings from a pre-trained Presto model. Our results demonstrate that fine-tuning a publicly available remote sensing time series pre-trained model outperforms the current state-of-the-art in NFI classification in the Netherlands, yielding performance gains of approximately 2-9 percentage points across datasets and evaluation metrics. This indicates that classic hand-defined features are too simple for this task and highlights the potential of using deep embeddings for data-limited applications such as NFI classification. By leveraging openly available satellite data and deep embeddings from pre-trained models, this approach significantly improves classification accuracy compared to traditional methods and can effectively complement existing forest inventory processes.

Figures (4)

• Figure 1: Overview of the study area and the three datasets used in this study. The simplified, class-balanced dataset is identical to related work by francini_DutchForestSpeciesMap_2024.
• Figure 2: Confusion matrices for the best performing models with deep embeddings by MLP classifier. Complex & Imbalanced (COMB)(left), Simplified & Balanced (SIBA) (right). Each cell has average count and recall percentage.
• Figure 3: Confusion matrices for Complex & Imbalanced (COMB) datasets. Top left: Fine- tune Presto with MLP Classifier, Top right: Presto (Frozen Presto features), Middle left: Fine-tune Presto (Presto features), Middle right: Harmonic + medoid feature, Bottom left: AEF, Bottom right: TESSERA. Each cell has average count and recall percentage.
• Figure 4: Confusion matrices for Simplified & Imbalanced (SIMB) datasets. Top left: Fine- tune Presto with MLP Classifier, Top right: Presto (Frozen Presto features), Middle left: Fine-tune Presto (Presto features), Middle right: Harmonic + medoid feature, Bottom left: AEF, Bottom right: TESSERA. Each cell has average count and recall percentage.