Open Access Repository

Mapping species composition and structure in wet eucalypt forest using multi-source remote sensing data


Downloads per month over past year

Yadav, BKV ORCID: 0000-0002-1076-6099 2019 , 'Mapping species composition and structure in wet eucalypt forest using multi-source remote sensing data', PhD thesis, University of Tasmania.

[img] PDF (Whole thesis)
Yadav_whole_the...pdf | Document not available for request/download
Full text restricted until 11 September 2021.


Tasmanian wet eucalypt forests are internationally important for wood and paper production, carbon storage and biodiversity conservation. These forests contain tall eucalypts over dense understories of rainforest and wet sclerophyll species. This research was motivated by a need for tools to replace costly aerial photo interpretation (PI-type) mapping for describing forest species composition and stand structure. Overall, I aimed to develop approaches for assessing and mapping tree species distribution and forest structure of wet eucalypt forest in a 5 km by 5 km area of the Warra Supersite, Tasmania, using multi-source remote sensing data.
My first study used an airborne LiDAR-derived canopy height model (CHM) and hyperspectral imagery to classify up to five dominant tree species of the forest. I used random forest classifiers on objects generated using data segmentation under a range of scenarios. Fused CHM and Minimum Noise Fraction (MNF) datasets yielded the highest segmentation accuracy (88.71%). The fusion of hyperspectral imagery, CHM and vegetation indices produced the best classifiers (overall accuracy (OA) of 66.7%) followed by the fused dataset of MNF and CHM (OA = 66.0%). Hyperspectral imagery alone provided the lowest classification accuracy (OA = 59.0%). Accuracy for the dominant canopy species (Eucalyptus obliqua) was 90.86% for four vegetation classes and 86.11% for five classes. Classification accuracies for the important understory species, Dicksonia antarctica, were also high under the best models (~84%). Accuracies for other species were low. Thus, fused hyperspectral and LiDAR data were robust and capable of spatially discriminating several important forest species.
My second study utilised LiDAR-derived topographic attributes and mapped geological strata to develop a model for predicting three understory layers of the forest (≥2 to ≤10 m, >10 to ≤30 m and >30 to ≤50 m as proxies for the lower, middle and upper layers, respectively) using five different spatial resolutions using random forest regression. Overall, the 30 m resolution provided the best model for predicting understory layers compared to 1 m, 5 m, 10 m and 20 m resolutions. The predictive power for the upper layer was greatest (R\(^2\) = 0.82), followed by the lower layer and the middle layer. Geology had the highest variable importance score for 5 m, 10 m, 20 m and 30 m resolutions, whereas terrain position index had the highest variable importance score for 1 m resolution. This research demonstrated that LiDAR-derived topographic attributes and geology data could be used to predict the understory vegetation structure.
My third study developed robust and cost-effective approaches for predicting the densities of vertical structural layers of the forest based on multispectral satellite data and simulated operational LiDAR datasets. I assessed the robustness of forest structure models based on thirteen schemes of derivatives (vegetation indices, texture features, and topographic attributes) from three different data sources (Airborne LiDAR downscaled to operational density, WorldView-3 and Landsat-8 (OLI)) at spatial resolutions (1.60 m, 7.5 m and 30 m). Models for the upper and middle layers were better than those for the lower layer. The 30 m Landsat-8 data provided the best results for all three-pixel sizes (R2 values ranged 0.15 to 0.65). Fused data from Landsat-8 and the simulated low-density LiDAR showed modest accuracy for predicting the density of three vertical layers and could be adopted by forest managers and planners. The WorldView-3 data of 1.6 m pixel size did not produce useful models.
In conclusion, the fusion of remote sensing datasets may help assess and map woody plant species composition and structure of wet eucalypt forests with opportunities to replace the traditional, subjective and time-consuming mapping technique of aerial photo interpretation. My results highlight the potential of freely available Landsat-8 (OLI) and operational LiDAR data, and random forest machine learning techniques for predicting and mapping forest species and vertical structural layers of wet eucalypt forests. This thesis addressed data complexities, including multidimensionality and nonlinearity in multi-source data, and provided a robust approach for the assessment of wet eucalypt forest composition and structure.

Item Type: Thesis - PhD
Authors/Creators:Yadav, BKV
Keywords: LiDAR and hyperspectral remote sensing, forest species composition, understory structure, machine learning
Copyright Information:

Copyright 2019 the author

Item Statistics: View statistics for this item

Actions (login required)

Item Control Page Item Control Page