Citation

Sim, Brenda Ru Yi and Lee, Sue Han and Choo, Chung Siung and Loh, Yuen Peng (2025) Canopy to Canopy: Evaluating Model Generalization in 3D Tropical Forest Semantic Segmentation. In: 17th Asia Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2025, 22 October 2025 - 24 October 2025, Singapore.

Text IEEE Xplore Full-Text PDF_33.pdf - Published Version Restricted to Repository staff only Download (694kB)

Abstract

Accurate 3D forest semantic segmentation from LiDAR data is fundamental for crucial applications such as biomass estimation and ecological analysis. However, robust performance remains challenging due to inherent forest diversity. The FORInstance dataset offers a large-scale benchmark with diverse UAV LiDAR scans from five forest environments. However, there is a notable lack of research assessing how well models trained on FOR-Instance generalize to complex and highly diverse tropical forest environments. As state-of-the-art models in this domain are primarily built upon PointNet-based architectures, this study evaluates the semantic segmentation performance and transferability of three foundational models: PointNet, PointNet++ (SSG), and PointNet++ (MSG). Our results reveal a catastrophic failure across all models when transferred to the unseen tropical Sarawak Forest dataset, highlighting the severe domain shift. Furthermore, models that perform the best in in-distribution (IID) scenarios exhibit significantly degraded performance under out-of-distribution (OOD) conditions, highlighting the need for architectures that better balance generalization and discriminative power. Despite ongoing advances in deep learning, the lack of ecologically diverse training data particularly from tropical forests remains a major bottleneck. This study underscores the pressing need for benchmark datasets that span a broader range of forest types to support the development of more generalizable and robust models for real-world 3D forest analysis.

Item Type:	Conference or Workshop Item (Paper)
Uncontrolled Keywords:	Forest semantic segmentation
Subjects:	S Agriculture > S Agriculture (General) S Agriculture > SD Forestry
Divisions:	Faculty of Artificial Intelligence & Engineering (FAIE)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	17 Mar 2026 07:14
Last Modified:	06 Apr 2026 04:13
URII:	http://shdl.mmu.edu.my/id/eprint/15523

Downloads

Edit (login required)