Understanding Representation Gaps Across Scales in Tropical Tree Species Classification from Drone Imagery

Technical Summary: Understanding Representation Gaps Across Scales in Tropical Tree Species Classification from Drone Imagery

Venue: ICLR 2026 Machine Learning for Remote Sensing (ML4RS) Workshop

1. Problem Statement

Accurate biodiversity monitoring in tropical forests is hindered by extreme species richness and high inter-species visual similarity. While high-resolution crown-view UAV imagery (centimeter-scale) is scalable for large-area monitoring, it often lacks the fine-grained botanical cues (e.g., leaf shape, flower structure) necessary for reliable species identification. Conversely, close-up imagery (sub-millimeter scale)—whether from citizen science (smartphones) or targeted drone flights—provides the necessary detail but is difficult to acquire at scale.

The core research question is: How do existing vision models perform across these two different spatial scales, and can the high-fidelity information from close-up views be leveraged to improve canopy-level classification?

2. Methodology

The researchers utilized a dataset of high-resolution RGB drone imagery collected from Barro Colorado Island (BCI) during 2024–2025.

• Data Modalities:
  • Crown-view: Orthomosaics at ~4 cm Ground Sampling Distance (GSD), segmented into individual tree crowns using the CanopyRS pipeline.
  • Close-up: Targeted drone-acquired images at ~0.4 mm resolution, providing high-detail botanical views.
• Experimental Setup:
  • The study focuses on 84 species, characterized by a long-tailed distribution (severe class imbalance).
  • Models Evaluated: Four distinct architectures: ResNet-50 (supervised CNN baseline), DINOv3 (self-supervised ViT), BioCLIPv2 (biologically informed vision-language model), and Pl@ntNet (a model pre-trained on millions of close-up citizen science plant photos).
  • Evaluation Modes: For crown-view imagery, they compared "individual-image" predictions against "soft-voting" (averaging predictions across 16 temporal snapshots of the same tree to account for phenological changes).
• Proposed Solution (Future Work): The authors propose a teacher-student representation alignment framework. A frozen Pl@ntNet model (teacher) would guide a Pl@ntNet-initialized student model to align crown-view embeddings with close-up embeddings via cosine distillation loss.

3. Key Contributions

• Quantification of the Representation Gap: The paper provides a systematic empirical analysis of how model performance shifts when moving from close-up botanical views to canopy-level views.
• Cross-Scale Benchmarking: It establishes a benchmark for evaluating vision foundation models in the specific context of tropical forest biodiversity.
• Identification of the "Rare Species" Vulnerability: The study demonstrates that the performance gap between scales is not uniform but is significantly wider for rare species.

4. Results

• Scale-Dependent Performance: Classification performance is consistently higher on close-up images than on crown-view imagery. Specifically, close-up images achieved a 77.9% accuracy on a single date, whereas crown-view imagery achieved only 74.3%, even after aggregating data over 16 different dates.
• Model Superiority: Pl@ntNet consistently outperformed other models in both modalities, suggesting that representations learned from large-scale botanical datasets transfer effectively to UAV imagery.
• The Long-Tail Effect: As shown in the F1-score comparisons, while common species perform well in both views, rare species suffer a sharp degradation in crown-view accuracy. Close-up imagery maintains much higher F1 scores for rare taxa, even with very few training samples (fewer than 20).
• Temporal Benefit: For crown-view imagery, "soft-voting" across multiple time points consistently improved accuracy compared to single-image predictions.

5. Significance

This work is significant for the field of ecological remote sensing because it identifies a critical bottleneck in automated biodiversity monitoring: the loss of fine-grained information at the canopy scale. By proving that close-up views hold superior discriminative power—especially for the rare species that are most critical to conservation—the paper provides a roadmap for using cross-scale representation learning to bridge the gap. This approach could enable researchers to use limited, high-resolution "snapshots" to significantly boost the accuracy of large-scale, scalable canopy monitoring efforts.