Title: Enhancing Cross-Domain Dead Tree Identification in Aerial Photography Through Knowledge Distillation

Abstract

As global tree mortality rates rise due to climate change, the accurate detection of dead trees in aerial imagery has become crucial for evaluating forest health. However, the generalization of detection models is frequently hindered by domain variability and a lack of labeled training data. To address these challenges, this research builds upon the TreeMort-1T-UNet (Tree Mortality 1-Task U-Net) architecture, which was originally trained on Finnish aerial imagery as the source domain. We employ knowledge distillation (KD) to facilitate the model’s adaptation to multiple target domains, specifically incorporating datasets from Poland, Germany, and Estonia that reflect varied forest ecosystems.

The study evaluates four distinct KD strategies—Basic, Self, Feature-level, and Ensemble—comparing them against a standard fine-tuning baseline. Performance was measured using Mean Tree IoU, Instance F1-score, Instance Precision, and Mean Centroid Error. Furthermore, we conducted representational analyses, including cosine similarity, Centered Kernel Alignment (CKA), Structural Similarity Index Measure (SSIM), t-SNE visualization, and linear probing, to assess domain invariance.

Among the tested methods, Feature-level KD demonstrated superior performance. On the Polish dataset, it achieved a Mean Tree IoU of 0.106, an Instance F1-score of 0.63, an Instance Precision of 0.55, and a Mean Centroid Error of 3.039. The method also maintained robust precision across other target domains, recording scores of 0.15 for Finland, 0.67 for Poland, 0.60 for Germany, and 0.59 for Estonia. Notably, the Feature-level approach excelled in low-data environments by reducing false positives and exhibited strong representational invariance, evidenced by higher deep-layer CKA and SSIM values, improved domain mixing in t-SNE plots, and a linear probing AUC of 0.95. These characteristics make it particularly well-suited for forestry applications where precision is paramount. Additional ablation studies indicated that components such as feature alignment play a critical role in balancing performance across various metrics. Ultimately, these results highlight the potential of knowledge distillation to improve transfer learning in remote sensing, providing a scalable and domain-robust solution for sustainable forest management and ecological monitoring.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC