Deep Change Monitoring: A Hyperbolic Representative Learning Framework and a Dataset for Long-term Fine-grained Tree Change Detection

Yante Li, Hanwen Qi, Haoyu Chen, Xinlian Liang, Guoying Zhao

Introduction

Monitoring individual tree health is a cornerstone of modern ecosystem management. However, traditional remote sensing techniques often fail to capture the necessary detail. Satellite imagery lacks the spatial resolution to identify individual branches or blossoms, while LiDAR, though precise, is prohibitively expensive and fails to capture color-based physiological cues.

Furthermore, tree change detection is inherently a hierarchical problem. Changes are either intrinsic (biological processes like flowering or leaf senescence) or extrinsic (environmental factors like lighting, shadows, or fog). Standard deep learning models, which operate in "flat" Euclidean space, struggle to represent these complex, nested relationships without significant distortion. The authors propose that because tree structures and their life cycles are inherently hierarchical, they should be modeled using hyperbolic geometry.

The UAVTC Dataset

To address these limitations, the authors introduce the UAV Tree Change (UAVTC) dataset, the first high-resolution UAV dataset for long-term, fine-grained tree change detection.

Tree samples across the entire timeline in the UAVTC dataset, illustrating seasonal transitions for White pine and Campsis grandiflora.

Key characteristics of UAVTC include:

  • Scale and Resolution: 245,616 image pairs with an unprecedented resolution of 0.5 cm/pixel.
  • Temporal Density: Data from 85 flight days over a year, capturing changes nearly every week.
  • Diversity: Monitoring of 68 individual trees across evergreen, deciduous, and fruit-bearing species.
  • Fine-grained Annotation: Labels for six distinct states: green leaf, yellow leaf, branch, fallen leaves/sprout, blossom, and destroyed.

Methodology: Hyperbolic Siamese Networks

The core innovation of this work is the Hyperbolic Siamese Network (HSN). This framework combines Siamese architectures with hyperbolic embeddings to better represent hierarchical biological structures.

The HSN architecture utilizes:

  1. Siamese Framework: Dual identical networks compare images from different time points.
  2. Hyperbolic Embedding: Euclidean features are projected into a Poincaré ball via an exponential map.
  3. Hyperbolic Operations: Classification is performed using Hyperbolic Binary Logistic Regression (Hyp-BLR) and trained with a Hyperbolic Binary Cross-Entropy loss.
  4. Feature Clipping: A mechanism to prevent embeddings from reaching the boundary of the Poincaré ball, mitigating the vanishing gradient problem common in hyperbolic networks.

Key Insights and Results

The shift to hyperbolic space yields significant benefits in both accuracy and robustness.

Examples of intrinsic tree changes caused by physiological factors, such as shifts in color, flowering, and leaf loss.
  • Geometry Matters: The authors found that tree image embeddings naturally exhibit high hyperbolicity, justifying the move away from Euclidean models.
  • Performance: HSN achieved an accuracy of 86.83% and an F1-score of 0.7372, significantly outperforming Euclidean baselines like ResNet101 (by 4.7% in accuracy and 24% in F1-score).
  • Robustness to Noise: The hierarchical nature of hyperbolic space allows the model to better distinguish between "intrinsic" physiological changes and "extrinsic" environmental noise, such as shifting shadows or weather conditions.
  • Generalization: Beyond forestry, HSN outperformed state-of-the-art methods in face anti-spoofing tasks, suggesting that many complex visual changes share a hierarchical structure that hyperbolic networks can exploit.

Conclusion and Future Work

While hyperbolic operations introduce a slight computational overhead, the gains in accuracy and robustness for fine-grained change detection are substantial. Future research will focus on expanding the dataset to more diverse forest environments and extending the framework to tasks like tree segmentation and more complex classification.

Summary updated

BibTeX
@inproceedings{li2025deep,
  author = {Yante Li and Hanwen Qi and Haoyu Chen and Xinlian Liang and Guoying Zhao},
  title = {{Deep Change Monitoring: A Hyperbolic Representative Learning Framework and a Dataset for Long-term Fine-grained Tree Change Detection}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.02547},
}
Figure