Forest stand segmentation with time series optical satellite imagery and superpixels

Date

2023-6-21

Author

Demirpolat, Caner

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

481
views

1057
downloads

Cite This

Forest stands are the basic units of forest management whose delineation is a fundamental task of creating stand level field inventories. Traditional methods for forest stand delineation require commercial imagery and usually have limited spatial coverage. Sentinel-2 has special bands aimed for vegetation mapping and provides the highest resolution imagery of globe among publicly available. This thesis aims to take the first step for using solely the Sentinel-2 imagery for forest stand delineation. Supervised and unsupervised superpixel segmentation methods are explored for initial segmentation of forest units. In unsupervised methodology, Simple Linear Iterative Clustering (SLIC) superpixel segmentation is employed with different subsets of multi-temporal Sentinel-2 spectral bands and vegetation indices. A dimension reduction technique on various combinations of time series features is used to construct the three-channel input image for the SLIC algorithm with weights optimized. In supervised methodology named as Gaussian Processes Regressor SLIC and its twin method Random Forests SLIC, a forest stand map which represents stand development classes is used for training the GPR and RF regressor which replaces the spectral distance measure in the SLIC algorithm. In comparison to the segmentation based on three LiDAR-derived features the unsupervised technique demonstrated improved performance in homogeneity measures. The supervised methodology significantly outperformed reference segmentation approaches in terms of boundary adherence and undersegmentation error metrics. The findings of supervised technique show that the new approach of using a superpixel segmentation algorithm based on machine learning with time series Sentinel-2 images yield promising results in the delineation of forest stands.

Subject Keywords

Forest stand delineation, SLIC superpixels, Sentinel-2, Random forest regression, Gaussian processes regression, Modified spectral distance

URI

https://hdl.handle.net/11511/104817

Collections

Graduate School of Natural and Applied Sciences, Thesis