Toward Interpretable, Data-Driven Frameworks for 3D Point-Cloud Analysis in Forest Ecology
Loading...
Authors
Stone, Gunner
Issue Date
2025
Type
Dissertation
Language
en_US
Keywords
Deep Learning , L-Systems , Latent Space Regularization , Pointclouds , Remote Sensing , Synthetic Data
Alternative Title
Abstract
Three-dimensional point clouds have transformed our ability to capture natural environments at scale, offering detailed representations of forest structure, canopy architecture, and understory composition. Yet despite this promise, the use of deep learning on ecological point clouds remains limited by three challenges: the scarcity of annotated data, reliance on models developed for non-ecological domains, and feature representations that lack biological interpretability. This dissertation addresses these challenges through three contributions. First, it introduces a modular pipeline for generating synthetic vegetative point clouds at scale, combining procedural tree models, randomized environmental factors, and simulated LiDAR acquisition. This pipeline alleviates data scarcity by enabling the creation of diverse, labeled datasets spanning species, growth stages, and canopy conditions. Second, it advances point-cloud architectures with two innovations: semantics-aware diffusion conditioning, which generates high-fidelity point clouds with per-point labels, and PointRTD, a noise-robust transformer whose regularizer improves stability under occlusion while enhancing class separation. This design accelerates convergence during training and improves discriminative accuracy on standard benchmarks. Third, it proposes an interpretable embedding framework inspired by L-systems, aligning latent spaces with parametric growth descriptors such as branching angles and crown spread. This framework regularizes reconstructions while producing embeddings that map directly to ecologically meaningful traits. Together, these contributions provide a unified framework for ecological machine learning that mitigates data scarcity, reduces model-domain mismatch, and grounds latent representations in biological structure. The result is a set of methods that both advance the state of 3D machine learning and open new possibilities for ecological analysis.