School of Agriculture, Food and Ecosystem Sciences - Theses
Permanent URI for this collection
Search Results
1 - 4 of 4
-
ItemUsing LiDAR for landscape-scale mapping of potential habitat for the critically endangered Leadbeater's Possum( 2019)Leadbeater’s possum (LBP) (Gymnobelideus leadbeateri) is a critically endangered arboreal marsupial located in Victoria’s Central Highlands in southeastern Australia. Populations of hollow-bearing trees, the key nesting habitat resource for Leadbeater’s Possum, are rapidly declining. Disturbances such as bushfires and timber harvesting are believed to be the major causes in the reduction of hollow-bearing trees and thus the extent, quality, and connectivity of LBP habitat. LBP also requires a connected understorey or midstorey for foraging and movement within the forest. Changes in LBP habitat at the landscape-scale are interpolated from plot-level measures and assumed distribution of habitat resources based on forest type and disturbance history. Long-term monitoring at the plot-level has reduced uncertainty around habitat resources required by LBP within the montane ash (Eucalyptus regnans, E. delegatensis, E. nitens) forests of the region; however, it has not addressed uncertainty around the landscape distribution of nesting and foraging habitat resources. Developing effective conservation strategies for LBP requires an understanding of the distribution and quality of available habitat across its potential range. This currently does not exist. I used Light Detection and Ranging (LiDAR) data to develop empirical models that quantify the distribution, quality, and connectivity of LBP habitat in the montane ash forests within a multi-scale framework. LiDAR-derived topographic and vegetation structural metrics were integrated with different habitat analysis approaches to model the distribution of critical LBP habitat features (hollow-bearing trees, foraging and midstorey stratum connectivity) at multiple scales across ~340,000 ha of forest in Victoria’s Central Highlands. In Chapter 2, tree-level individual tree delineation (ITD) algorithms were developed to identify individual trees and analyse tree crown attributes. ITD algorithms for overstorey trees were applied to the whole study area to estimate the abundance and distributions of large live trees with big crowns that have high probability of containing hollows. Based on field surveys in 1939 regrowth mountain ash, the dominant overstorey crowns identified by the ITD algorithms had an 86% success rate in identifying mapped trees in mature forest plots. Our landscape analyses estimated 405,000 large live trees (crown width >15m) and 572,000 (95%CI: 318,000-808,000) potential large live tree with hollows (DBH>150cm), which equate to a mean density of one live large old tree per ha across the Central Highlands. ITD algorithms were also used to map canopies and to calculate the proportion of projection area of midstorey stratum, a surrogate for midstorey coverage, for use in assessments of foraging habitat. Hollow-bearing trees are a key feature of mature forests around the world. They provide critical habitat resources for hollow-dependent animals, many of which are threatened, due to a range of natural and human disturbances. In Chapter 3, to develop more accurate inventories of hollow-bearing trees (both live and dead form classes) across a large landscape, I used LiDAR-based metrics of forest structure and topography, coupled with datasets on environmental conditions to develop statistical models of abundance of hollow-bearing trees using machine learning tools. These provided empirical estimates of HBT density for every hectare of the Central Highlands within the LiDAR footprint. I identified 1,519,000 (95%CI: 1,306,000 – 1,730,000) hollow-bearing trees may occur across the full extent of Victoria’s Central Highlands. This included 833,000 (95%CI: 742,000-923,000) live HBTs and 686,000 (95%CI: 564,000-807,000) dead HBTs. The predictive models provided rigorous, repeatable estimates of tree abundance across a wide range of vegetation classes and forest management zones (with appropriate estimates of uncertainty), as well as a new understanding of the complexity of the structural, topographic, and environmental features associated with abundance of hollow-bearing trees and their spatial variability over large areas. In addition to HBTs for nesting, LBP require a well-connected midstorey stratum to facilitate movement amongst nesting trees and foraging. In Chapter 4, the association between field-assessed multi-storey vegetation connectivity and LiDAR structural metrics was evaluated to develop landscape-scale predictions of connectivity and foraging habitat (density of live wattle) for LBP. Forest structural types with high midstorey connectivity, which are a critical habitat feature for LBP, occupied 17.2% of the study landscape. The landscape-scale predictions of vegetation connectivity provided an understanding of the factors (e.g., topography, fire, logging and interactions among strata) that help shape connectivity and foraging habitat. This understanding is critical for improving the management of the region’s forests and, in particular, ensuring that both key habitat elements, HBTs and dense midstorey vegetation, either co-occur in the same stand or are in close proximity to ensure sufficient high-quality habitat for LBP over space and time. In Chapter 5 I used the landscape-level mapping of habitat resources developed in Chapters 2-4 to predict landscape-scale LBP habitat suitability based on fields records of LBP occurrence. The validated habitat suitability model for LBP had an overall accuracy of 87.3% and an AUC of 0.889. The model predicted that 28,000 ha (95%CI: 17,000ha-40,000ha) of the study landscape (i.e., 9.6%) supported suitable habitat for LBP. Twelve variables were found to have the strongest influence on estimating habitat suitability for LBP. Of these, four related to foraging habitat, three to nesting habitat, three to climatic and topographic factors, and two to disturbance history. Our approach highlights the potential for using high-resolution, spatially explicit data on forest structure at a landscape-scale to map the distribution and abundance of suitable habitat for a critically endangered species. Habitat suitability models generally account for the selection of suitable habitat and their appropriate geographical extents. However, they rarely account for the accessibility of this habitat and connectivity among habitat patches. Effective conservation of species requires that patches of their habitat are connected in space and time. In Chapter 6, graph-theoretic connectivity networks based on resistance surfaces were generated from the LiDAR mapped cover of vertical stratum. The least-cost links between patches from networks were integrated with a kernel density estimator to identify functionally connected regions. The mapping of landscape connectedness identified a potential landscape-scale metapopulation structure for LBP within the Central Highlands. This provides a tool that could be used to expand the existing protected area network to support the metapopulation processes of LBP. The identification of functionally connected regions could enhance the conservation planning for long-term population persistence. The conservation and management of endangered species is major focus of forest and land managers around the world. High-quality data on the distribution and quality of habitat is critical to the development of effective conservation and recovery strategies for species. This thesis developed a set of habitat analysis models for quantifying habitat resources, habitat suitability and connectedness for the critically endangered Leadbeater’s Possum across a range of spatial scales. These analyses highlight the benefits of using high-resolution, spatially explicit LiDAR data from the whole landscape to identify and map (1) the distribution and abundance of hollow-bearing trees, (2) midstorey stratum connectivity, (3) the distribution and abundance of suitable habitat, and (4) metapopulation structures within functionally connected regions. Dynamic modelling, informed by these LiDAR-derived models, can be used to forecast the likely consequences of changing habitat abundance under different management scenarios and evaluate both the short-term and long-term effectiveness of different conservation strategies.
-
ItemUnravelling spatio-temporal water balance patterns in topographically complex landscapes( 2017)The Budyko framework for understanding how precipitation (P) is partitioned into evapotranspiration (ET) and streamflow has been shown to be remarkably robust at large spatial scales. The Budyko model simply uses P and potential ET as variables. However, at smaller spatial scales additional predictor variables are required to partition precipitation. In steep uplands, topography appears to exert strong control on the water balance at the hillslope scale. Organization of vegetation suggest heterogeneity in the water balance at these scales. This topographic control though is poorly characterized in most environments and therefore not well represented in models. The aim of this thesis is to quantify the effect of local topography (aspect and drainage position) on forest water balance as a first step towards a down-scaling of the Budyko model in steep upland terrain. Six intensively instrumented sited were established on three drainage positions and two aspects in mixed species eucalypt forests (MSEF), with all other variables remaining constant. Continuously monitored water balances were extrapolated across a ~70 ha catchment using a Random Forest model and LiDAR characterization of stand density and structure. The study demonstrated that spatial vegetation patterns emerged in response to topographic control on water-availability via soil depth, water redistribution and sub-canopy radiation loads. Moreover, short-term variations of overstory transpiration (To) were driven by atmospheric forcing, whereas seasonal and annual To patterns were explained by sapwood area index (As, R^2:0.89). Understory and forest floor evapotranspiration (ETu) was controlled by sub-canopy short-wave radiation. Further, the combined effect of aspect and drainage position on water balance partitioning markedly diverged along the south and north-facing transect. All plots on northern aspects had a positive water balance (P> ET), whereas only the ridge plot on the south-facing slope had a positive water balance, while the lower hillslope had higher ET rates than rainfall inputs.ET measurements from the distributed plots could be up-scaled using terrain and vegetation information derived from LiDAR and unveiled strong spatial variability of To (4.5-fold), ETu (3.5-fold) and total ET rates (2-fold) over as little as 200 m distance. The observed ET range corresponds to eucalypt forests typically located >100s of kilometers apart, with the lower end similar to arid open woodlands in Western Australia and the upper end to tall mountain ash forests in the Victorian highlands. Predicting ET using the Budyko framework revealed strongly biased ET estimates in relation to landscape position, where ~18% of the catchment area plotted above the theoretical water limit, confirming the importance of topographic water redistribution. Further, model residuals were explained well by As and terrain patterns. My thesis presented strong links between vegetation patterns, topography, soil depth and energy and water fluxes in upland MSEF. Ultimately, the study demonstrated the potential of remotely sensed vegetation and terrain patterns to infer and scale water-balance patterns in heterogeneous upland forests.
-
ItemThe potential use of LiDAR and digital imagery in selection of suitable forest harvesting systems( 2013)Major factors affecting the productivity and efficiency of mechanized forest harvesting systems include stand conditions (e.g. tree form, tree size, crown size and the type and density of trees), terrain conditions (e.g. slope, ground roughness, ground strength, road and drainage features, etc.), yield, operator performance (e.g. experience, skill and work technique), and machinery limitations or design. The purpose of the study was to examine whether ‘tree size’ and ‘slope’ information derived from LiDAR (Light Detection And Ranging) data and multispectral imagery, could be used to predict the productivity and efficiency of forest harvesting equipment. Tree size is known to be the biggest influence on harvester productivity. The study aimed at developing a productivity model for a harvester operating in a 35-year-old radiata pine (Pinus radiata) plantation in South Australia using data obtained from low density LiDAR (Light Detection And Ranging) (2.6 points / m2) and high resolution Quickbird imagery (60 cm). Tree size extracted from a harvester onboard computer system (OBC) was used to estimate tree size impact on harvester productivity by conducting a time study. LiDAR-derived tree heights were not found to be significantly different (p < 0.05) from field measured tree heights and the absolute mean underestimation of LiDAR-height was 1.3 m. LiDAR-derived tree height estimates were found to be poorly related to tree volume and hence to harvester productivity. This was believed to be the result of the stand’s thinning history reducing the range of tree sizes i.e. removal of trees in the thinning operations. An attempt, therefore to estimate crown diameter from Quickbird imagery and or low density LiDAR was made which, in combination with LiDAR height, might be used to estimate tree volume and hence harvester productivity. Slope is a major terrain factor affecting harvester productivity. A study in Tasmania examined the ability of LiDAR to derive terrain slope over large areas and to use the derived slope data to model the effect of slope on the productivity of a self-levelling feller-buncher in order to predict its productivity for a wide range of slopes. Low intensity LiDAR (>3 points / m2) flown in 2011 over the study site was used to derive slope classes. A time and motion study carried out for the harvesting operation was used to evaluate the impact of tree volume (estimated from manual tree measurements) and slope on the feller-buncher productivity. The study found that productivity of the feller-buncher was significantly greater on moderate slope (11-18°) than on steep slope (18-27°). This difference in productivity resulted from operator technique differences related to felling. The productivity models were tested using LiDAR-derived slope and trees not used in the model development and were found to be able to accurately predict the effect of slope on the productivity of the feller-buncher. To better understand the drivers of harvesting productivity, a detailed comparative study of two single-grip harvesters was carried out in Australian Pinus radiata clearfell harvesting operations. Significant differences in productivity between the harvesters were found to be largely due to operator working technique differences. This factor cannot be determined through remote sensing. However, its influence can be reduced by using a general productivity model obtained from multiple operators.
-
ItemLiDAR estimation of aboveground tree biomass in native sclerophyll forest( 2011)Accurate estimation of aboveground tree biomass is a fundamental aspect of studies on carbon stocks of forest ecosystems. Destructive sampling is the most accurate method of estimation of biomass. However, because of its destructive nature and being both labor intensive and time consuming, destructive sampling cannot be applied for large areas. Alternatively, allometric equations developed for a particular species of trees or a general allometric equation for a specific type of forest can be used to estimate aboveground biomass for larger areas. This requires massive fieldwork, which itself is problematic, and it is not always possible to carry out field inventory in forests which are remote and inaccessible. Advanced remote sensing technology is now in the process of being established as the best and most practical alternative of the field-based methods of biomass estimation for large areas and is being used in the study of forests at the regional and national levels of a growing number of countries. Light Detection And Ranging (LiDAR) is a relatively new, active, remote sensing technology, which is capable of providing three-dimensional structural information of forests and, therefore, can be used to estimate various structural and biophysical parameters of forest stands more accurately than by other optical and RADAR-based remote sensing technologies. The development of hardware and software for the LiDAR system has rapidly advanced during the last decade and has matured to a degree that it is now possible to analyze LiDAR points, which are from individual tree crowns. As a result, LiDAR has now been used as an operational tool in European and North American forestry. In Australia, the use of LiDAR is still in an initial, research phase and there are only a few studies that have investigated its applicability in the broadleaf evergreen forests that dominate the forested lands of Australia. The main focus of this study was the LiDAR-estimation of aboveground tree biomass in two different types of Eucalyptus-dominated sclerophyll forests of the Central Highlands of Victoria, Australia. The applicability of LiDAR remote sensing in predicting stem density, canopy height indices (mean dominant height, Lorey's mean height and quadratic mean canopy height) and basal area was also explored in this research. Furthermore, the scaling-up of LiDAR estimates of biomass across the landscape and biomass mapping for large areas were also demonstrated. Using LiDAR data for the respective sampling plots, the mean dominant height for the Central Highlands Ash Regrowth (CHAR) forest was estimated with an R2 of 87.1 % and an RMSE of 3.9 m (9.5 %) and for the Black Range Mixed Species (BRMS) forest with an R2 of 92.1 % and an RMSE of 1.9 m (6 %). The R2 of the model predicting Lorey's mean height for the CHAR forest had 84.6 % with an RMSE of 4.03 m (11.1 %) and for the BRMS forest, it was 94.6 % with an RMSE of 1.7 m (5.9 %). Similarly, the quadratic mean canopy height was estimated with an R2 of 48.4 % and an RMSE of 4.9 m (17.9 %) for the CHAR forest and with an R2 of 92.7 % and RMSE of 1.9 m (7.4 %) for the BRMS forest. New methods to estimate the number of trees and the basal area from LiDAR data were developed in this study. When these methods were used, the number of trees was predicted with a mean prediction error of - 64.1 trees/ha (- 7.6 %) with a predicted value of 776.5 trees/ha for the calibration plots and a mean prediction error of 105.3 trees/ha (14.4 %) with a predicted value of 838.2 trees/ha for the validation plots in the CHAR forest. The mean prediction error for the basal area in the CHAR forest was 9.7 m2/ha (16.4 %) with a predicted value of 68.9 m2/ha for the calibration plots and 0.2 m2/ha (0.32 %) with a predicted value of 66.4 m2/ha for the validation plots. In the BRMS forest, the mean prediction error for the number of trees was 80 trees/ha (8.6 %) with a predicted value of 1010.5 trees/ha for the calibration plots and 5 trees/ha (0.9 %) with a predicted value of 584.4 trees/ha for the validation plots. The mean prediction error for the basal area in the BRMS forest was 8.3 m2/ha (13.9 %) with a predicted value of 68.4 m2/ha for the calibration plots and 6.13 m2/ha (9.3 %) with a predicted value of 71.9 m2/ha for the validation plots. LiDAR metrics such as the mean height, quadratic mean height, 90th percentile height and standard deviation of heights of LiDAR points were used as predictors of biomass. Three additional metrics, which have not been used in previous studies, were also derived and used in the regression analysis. These metrics were: scale parameter of the 2-parameter exponential distribution, largest extreme value distribution and smallest extreme value distribution of elevation (above sea level) data of the LiDAR points. Six prediction models for the CHAR forest and seven prediction models for the BRMS forest were developed. All the models predicted the biomass quite accurately. The models for the CHAR forest had R2 values that ranged from 58 % to 64 % and the R2 values of the models for the BRMS forest ranged from 58 % to 79.8 %. The results of the validation of the models in this study showed that the range of the average prediction bias of the models for the CHAR forest ranged from - 13.6 tons/ha to 6.1 tons/ha, and the range of the prediction bias of the models for the BRMS forest was from - 30.6 tons/ha to 8.1 tons/ha. In this study, I also developed a new multistage processing technique of LiDAR data to isolate individual trees. In this technique, LiDAR data for the sampling plots were first split vertically into four separate data sets representing reflections from the canopy layers of the dominant, co-dominant, intermediate and understory trees. Each data set was then processed using the LiDAR software, Toolbox for Lidar Data Filtering and Forest Studies (TiFFS) to isolate and obtain information on individual trees. By using this new multistage processing technique, about 52 % of the trees were isolated correctly. Individual tree information extracted from LiDAR data was then used to estimate the aboveground biomass of the trees. The aboveground biomass of individual trees isolated from LiDAR data was estimated quite accurately using the LiDAR-derived DBH from the LiDAR-estimated height of the trees. First, the LiDAR-derived DBH was used to estimate the aboveground biomass of individual trees from the general allometric equation for the native sclerophyll forest. A linear regression equation (R2 of 66.4 %) was then developed to estimate the individual tree biomass with the LiDAR-derived biomass as the predictor. To validate the linear model, the aboveground biomass of 76 new trees was estimated, and the average prediction bias obtained was - 78.1 kg, 9.6 % lower than the average biomass estimated from the field-measured DBH and the general allometric equation. Finally, based on the predictive model developed in this study, the aboveground tree biomass was estimated for 100,000 plots of 20 m × 20 m in size stretching across the landscape for an area of 4,000 hectares of the Central Highlands Ash Regrowth (CHAR) forest. The GPS coordinates and biomass data for these 100,000 plots were then used to map the biomass, which was of high (20 m) resolution and provided accurate information on the spatial distribution of the aboveground biomass across the landscape. The very simple and quite accurate methods of estimating the stem density and the basal area from LiDAR data developed in this study could be very useful in various aspects of forest management in Australia. Similarly, new LiDAR metrics (scale parameters of distribution of LiDAR elevation) developed and used in estimating the aboveground tree biomass could be very useful in predicting biomass of hilly forest areas. The multistage processing technique of LiDAR data developed in this study was very effective in detecting a greater number of intermediate and understory trees in forests that have a multistory structure. The scaling-up demonstrations of this study showed that LiDAR can be used to estimate the aboveground biomass across the landscape and for biomass mapping of eucalypt forests. The results of this study can have a great impact on the application of LiDAR as an operational tool for sustainable forest management and in estimating forest biomass and monitoring its change over time for climate change mitigation and adaptation research in Australia. This is the first thesis focused on application of LiDAR remote sensing in estimating various structural as well as biophysical parameters in the evergreen temperate forests having a multistory and heterogeneous structure dominated by Eucalyptus species in the moist and dry temperate region of the Central Highlands of Victoria, southeastern Australia. The thesis not only explored the applicability of LiDAR in estimating various canopy height indices, stem density, basal area and aboveground tree biomass but also demonstrated that LiDAR estimates of the plot-level biomass can be scaled-up to the landscape level quite easily and can then be used to produce very high resolution biomass maps for large areas. Three new LiDAR metrics derived in this study from the analysis of distribution of elevation of LiDAR points can play a very important role in accurately estimating the aboveground tree biomass of hilly forest areas. The accuracy of LiDAR-generated DEM and DSM and the other LiDAR height metrics for forests having a hilly terrain could be low compared to those variables for forests spread over a plain terrain. This might introduce some error in the predicted biomass from LiDAR metrics for hilly-terrain forests, which are related to the DEM and DSM generated. On the other hand, distributional parameters derived from the elevation data of LiDAR points do not depend on the DEM and DSM and, therefore, could provide more accurate prediction of biomass for forests in mountainous regions. This study has for the first time developed indirect methods of estimating stem density and basal area using LiDAR data. This is very important because direct methods (developing regression models with various LiDAR metrics as the predictors) of estimating these attributes usually do not produce good results even for pine forests for which LiDAR estimates are more accurate compared to the estimates for non-uniform and mixed species broad-leaved forests. The multistage processing technique of LiDAR data developed in this study could have great impact on the application of LiDAR as an operational tool in forestry because the process is able to detect more intermediate and understory trees in multistory mixed species forests, which have a great influence in the overall functioning of forest ecosystems.