Investigation on Dimensionality Reduction methods for Tree-Crown Segmentation in Hyperspectral imagery using Segment Anything Model

Abstract. Forests play a vital role in global ecosystems, and accurate monitoring of tree crowns is essential for forest management and biodiversity conservation. This study investigates the use of hyperspectral imagery and dimensionality reduction methods for individual tree-crown (ITC) segmentation, a crucial task in forest monitoring. Traditional LiDAR-based methods are often expensive and computationally intensive, making hyperspectral imagery a promising alternative due to its data-richness. However, since most deep learning segmentation methods accept only 3-channel images, we adapt hyperspectral images from a benchmark dataset by applying dimensionality reduction techniques such as Principle Component Analysis (PCA), Factor Analysis, and Uniform Manifold Approximation and Projection (UMAP) to transform high-dimensional data into 3-channels, before performing segmentation using Segment Anything Model (SAM). The results show significant improvements over RGB imagery with dimensionality reduction methods, however the overall segmentation accuracy remains poor. With an average F1-score of 0.26, some methods achieved up-to 0.38 at specific sites. The results varied between sites due to different density and tree types in the image data. Factor Analysis and an approach with UMAP utilising vegetation indices produced the most promising results.