Automatic 3D building reconstruction has been a hot research area; a task which has been done manually even up today. Automating the task of building reconstruction enables more applications where up to date information is of great importance. This thesis proposes a system to extract parametric buildings and trees from dense aerial stereo image data. The method developed for the tree identification and parameterization is a totally new approach which have yielded great results. The focus has been to extract the data in such a way that small flying platforms can use it for navigational purposes. The degree of simplification is therefor high. The building parameterization part starts with identifying roof faces by Region Growing random seeds in the digital surface model (DSM) until a coverage threshold is met.For each roof face a plane is fitted using a Least Square approach.The actual parameterization is started with calculating the intersection between the roof faces. Given the nature of 2.5D DSM data there is no possibility to perform wall fitting. Therefor all the walls will be constructed with a 2D line Hough transform of the border data of all the roof faces. The tree parameterization is done by searching for possible roof face topologies resembling the signature of a tree. For each possible tree topology a second degree polynomial surface is fitted to the DSM data covered by the faces in the topology. By looking at the parameters of the fitted polynomial it is then possible to determine if it is a tree or not. All the extraction steps were implemented and evaluated in Matlab, all algorithms have been described, discussed and  motivated in the thesis.

Automation of urban tree inventory using remote sensing is needed not only to reduce inventory costs but also to support carbon accounting for urban planners and policy-makers. However, urban areas are heterogeneous and complex, and a more sophisticated approach is needed for using remote-sensing technology like lidar for tree inventory in urban areas than is required for forested environments. Based on remote sensing and field data from a suburban residential area in the central United States, this dissertation presents a methodology for utilizing airborne small-footprint lidar data to inventory urban trees. This dissertation proposes approaches that have the potential to automate three main activities of urban tree inventory -- identifying the locations of trees, classifying the trees into taxonomic categories, and estimating biophysical parameters of individual trees -- using airborne lidar data. Mathematical morphological operations followed by a marker-controlled watershed segmentation were found to perform well (r = 0.82 to 0.92) to delineate individual tree crowns in urban areas, especially when the trees occur in relatively isolated conditions. Using various distribution metrics of lidar returns, random forests were used to classify individual trees into different taxonomic classes (broadleaves/conifers, genera, and species). A classification accuracy of 80.5% was obtained when separating trees only into broadleaf and conifer classes, 50.0% for genera, and 51.3% for species. Using spectral metrics from high-resolution satellite imagery in addition to lidar-derived predictors improved the classification accuracies by 10.4% (to 90.9%) for broadleaf and conifer, 8.4% (to 58.4%) for genera and 8.8% (to 60.1%) for species compared to using lidar metrics alone. Prediction models to estimate several biophysical parameters such as height, crown area, diameter at breast height, and biomass were developed using lidar point cloud distributional metrics from individual trees. A high level of accuracy was attained for estimating tree height (R²=0.89, RMSE=1.3m), diameter at breast height (R²=0.82, RMSE=9.1cm), crown diameter (R²=0.90, RMSE=0.7m) and biomass (R²=0.67, RMSE=1.2t). Our results indicate that, while using lidar data alone can achieve the automation of major urban forest inventory tasks to an acceptable level of accuracy, a synergistic use of lidar data with other spectral data such as hyperspectral or orthoimagery, which are usually available at least in the United States for most urban areas, can considerably improve the performance of the lidar-based method. / Ph. D.

The automated Tree Identification and Delineation Algorithm (TIDA) was developed for application to high spatial resolution digital imagery of Australian native eucalypt forest. The algorithm is based on contiguous, threshold-based spatial clustering of pixels and was designed to cope with the complex asymmetric crowns typical of eucalypts. / To facilitate systematic algorithm evaluation, a forest scene simulation model was created for the simulation of visually realistic remotely sensed images. The model is based on the principles of ray-tracing and the geOll1etric description of scene objects and background. The model simulates the appearance of a forest scene viewed and illuminated from specific directions and under known atmospheric conditions. The distinctive 'modular' structure of eucalypts was represented by modelling crowns as small (branch-scale) spheroids distributed over a larger spheroidal envelope. / Using the simulation model, TIDA performance was evaluated in terms of forest structure (canopy cover, crown cover and canopy structural variability) and the remote sensing environment (view zenith angle, solar zenith angle and aerosol optical thickness). Prior to the evaluation, a methodology was developed for objectively estimating the optimum spatial resolution for TIDA application in a given image. The methodology was based on incremental Gaussian smoothing and exploited TIDA's sensitivity to changes in image spatial resolution. This process demonstrated the importance of individual crown cover, rather than crown size, as the main factor determining the optimum spatial resolution for tree delineation. / Results indicate that TIDA is most suited for application in forests with high canopy cover and high crown cover. The structural complexity of forest canopies, represented by the diameter and overlap of crowns and tree height, had a relatively small impact on TIDA performance. Increasing view zenith angle consistently caused a decrease in TIDA performance. A small phase angle between the sun and sensor produces optimum TIDA performance when both canopy and crown cover is high. As crown or canopy cover decrease, high positive and negative sun zenith angles yield superior TIDA results by decreasing the brightness of the background relative to the canopy and improving the identification of tree peaks. For both dense and sparse canopies, back-scattered radiation from the forest canopy was more suited to automated tree crown delineation than forward-scattered radiation. Imagery acquired under an optically thick atmosphere was found to increase TIDA performance compared to scene illumination under strong direct light. The advantage stemmed from a strengthening of the relationship between geometric and radiometric crown shape. / Through an awareness of limitations imposed by the remote sensing environment, the potential for manipulation of image characteristics, and preferential selection of acquisition conditions, TIDA performance can be optimised to suit various structural forest types. Canopy cover, crown cover, view zenith angle, sun zenith angle, background brightness and image spatial resolution are key criteria in assessing the suitability of automated tree crown delineation as an image interpretation procedure.

The region covered by this study reaches from the valley floor of the San Joaquin-Sacramento Valley, with an altitude of about 150 feet, to the top of the Sierra Nevada Mountains, whose peaks, in Amador County, reach to more than 9000 feet. Mokelumne Peak has an altitude of 9371 feet. The north and south distance, across the county at Ione, is 22 miles, while at Cook's Station, it is barely 5 miles. Amador County has an area of about 600 square miles. Thus, a study of trees in this county, gives a fairly representative cross-section picture of the Sierras at this latitude of 39 degrees.

The importance of single-tree-based information for forest management and related industries in countries like Sweden, which is covered in approximately 65% by forest, is the motivation for developing algorithms for tree detection and species identification in this study. Most of the previous studies in this field are carried out based on aerial and spectral images and less attention has been paid on detecting trees and identifying their species using laser points and clustering methods. In the first part of this study, two main approaches of clustering (hierarchical and K-means) are compared qualitatively in detecting 3-D ALS points that pertain to individual tree clusters. Further tests are performed on test sites using the supervised k-means algorithm in which the initial clustering points are defined as seed points. These points, which represent the top point of each tree are detected from the cross section analysis of the test area. Comparing those three methods (hierarchical, ordinary K-means and supervised K-means), the supervised K-means approach shows the best result for clustering single tree points. An average accuracy of 90% is achieved in detecting trees. Comparing the result of the thesis algorithms with results from the DPM software, developed by the Visimind Company for analysing LiDAR data, shows more than 85% match in detecting trees. Identification of trees is the second issue of this thesis work. For this analysis, 118 trees are extracted as reference trees with three species of spruce, pine and birch, which are the dominating species in Swedish forests. Totally six methods, including best fitted 3-D shapes (cone, sphere and cylinder) based on least squares method, point density, hull ratio and slope changes of tree outer surface are developed for identifying those species. The methods are applied on all extracted reference trees individually. For aggregating the results of all those methods, a fuzzy logic system is used because of its good reputation in combining fuzzy sets with no distinct boundaries. The best-obtained model from the fuzzy system provides 73%, 87% and 71% accuracies in identifying the birch, spruce and pine trees, respectively. The overall obtained accuracy in species categorization of trees is 77%, and this percentage is increased dealing with only coniferous and deciduous types classification. Classifying spruce and pine as coniferous versus birch as deciduous species, yielded to 84% accuracy.

Drivgods utgör ett hot mot dammsäkerheten vid vattenkraftsverk eftersom det ökar risken för blockering av kraftverkens utskov vilket i sin tur kan resultera i överströmning av dammen och dammbrott. Fram till idag har drivgods inte ansetts vara ett problem i Sverige. Flera allvarliga drivgodsincidenter utomlands i kombination med en prognostiserad ökning av flödesintensiteter och översvämningstillfällen i Svenska vattendrag har dock lett till ett ökat intresset hos Svenska dammägare. Det saknas idag tillförlitliga metoder för prognostisering av potentiella drivgodsmängder vid extremflöden, vilket försvårar planering av preventiva åtgärder. Denna studie har som mål att utveckla specifika metoder för automatiserad kvantifiering och geografisk analys av drivgods. En modell har utvecklats i ArcGIS Modelbuilder för (i) avgränsning och klassificering av riskområden längs ett vattendrag baserat på identifierade riskfaktorer för drivgods-bildning, (ii) kvantitativ estimering av trädvolym utifrån höjdrasterdataset framtagna från LIDAR punktmoln, och (iii) identifiering av trädindivider baserat på lokalisering av lokala maxpunkter i en digital ytmodell skapad från laser-data. I syfte att utvärdera metodens tillämpbarhet samt undersöka förekomsten av eventuella geografiska mönster i utbredningen av potentiella drivgodskvantiteter appliceras modellen på utvalda dammanläggningar längs med älven Ljungan i Sverige. Modellens noggrannhet i uppskattningen av trädindivider valideras i en fältkontroll. Resultaten indikerar en mycket god överrensstämmelse mellan modellerade och fältkontrollerade värden på trädantal, dock med något underestimerade modellerade värden. Ett större antal träd som riskerar att falla i älven identifierades vid högre vattenflöden som ett resultat av en generellt större geografisk omfattning på översvämmade ytor. Stora vegetationsvolymer identifierades främst i riskområden med en stor andel höga träd men även i riskområden med ett stort antal träd. I vissa riskområden erhölls större relativa vegetationsvolymer trots ett betydligt mindre relativt antal träd vilket antyder att både trädhöjd och trädantal styr vegetationsvolymers storlek. Resultaten av utförda statistiska och geografiska analyser förbättrar överblicken över drivgodsets volym och geografiska fördelning längs vattendrag vilket i sin tur även leder till en bättre förståelse för drivgodsproblemets magnitud för specifika dammanläggningar. Den automatiserade metoden genererar reproducerbara resultat som är jämförbara mellan olika dammanläggningar och vattendrag över tid. Metoden har stor potential för framtida drivgodsanalyser då den erbjuder möjligheten att identifiera riskområden och estimera drivgodskvantiteter inom dessa. / Floating debris poses a threat to dam safety at hydropower dams as it increases the risk of spillways becoming blocked, which can in turn result in dam overflow and failure. Until now it has not been considered a problem in Sweden. However, the occurrence of several major international incidents, together with a projected future increase in streamflow intensities and flood events in Swedish rivers has raised the interest of Swedish dam owners. Presently there is a lack of robust methods for forecasting potential magnitudes of floating debris in extreme streamflow scenarios, which limits planning of preventive measures. This study aims to develop explicit methods for automated quantification and geographical analysis of floating debris. A model is developed in ArcGIS Modelbuilder for (i) identification and classification of risk areas along a river based on identified risk factors for floating debris formation, (ii) quantitative estimation of vegetation volume from elevation raster datasets derived from LIDAR point clouds, and (iii) identification of trees within risk areas based on locating local maxima in a digital surface raster derived from laser-data. In order to evaluate the applicability of the method as well as to investigate the existence of geographical patterns in potential floating debris quantities, the model is applied on selected dam facilities along the Ljungan River in Sweden. The accuracy of modeled tree amounts is validated in a field study. The results indicate high correlation between modeled number of trees and ground truth data, though modeled values are slightly underestimated. A greater number of trees at risk of falling into the river were identified during higher streamflow events as a result of larger areas being flooded. Large volumes of vegetation were identified in risk areas with a high proportion of tall trees as well as in risk areas with a large number of trees. In some risk areas, greater relative vegetation volumes were obtained despite a significantly smaller relative number of trees. This suggests that vegetation volume as a factor depends both upon the number of trees as well as tree height. The results of performed statistical and geographical analyses provide a better overview of floating debris distribution along rivers, leading to a better understanding of potential magnitudes of the problem for specific dam facilities. The automated method generates reproducible results that are comparable between different dam facilities and rivers over time. The method has significant potential for future floating debris analyses as it offers the possibility to identify risk areas and estimate floating debris quantities within them.

For scientific or socio-economic reasons it is often necessary or desirable that biological material be identified. Given that there are an estimated 10 million living organisms on Earth, the identification of biological material can be problematic. Consequently the services of taxonomist specialists are often required. However, if such expertise is not readily available it is necessary to attempt an identification using an alternative method. Some of these alternative methods are unsatisfactory or can lead to a wrong identification. One of the most common problems encountered when identifying specimens is that important diagnostic features are often not easily observed, or may even be completely absent. A number of techniques can be used to try to overcome this problem, one of which, the Self Organizing Map (or SOM), is a particularly appealing technique because of its ability to handle missing data. This thesis explores the use of SOMs as a technique for the identification of indigenous trees of the Acacia species in KwaZulu-Natal, South Africa. The ability of the SOM technique to perform exploratory data analysis through data clustering is utilized and assessed, as is its usefulness for visualizing the results of the analysis of numerical, multivariate botanical data sets. The SOM’s ability to investigate, discover and interpret relationships within these data sets is examined, and the technique’s ability to identify tree species successfully is tested. These data sets are also tested using the C5 and CN2 classification techniques. Results from both these techniques are compared with the results obtained by using a SOM commercial package. These results indicate that the application of the SOM to the problem of biological identification could provide the start of the long-awaited breakthrough in computerized identification that biologists have eagerly been seeking. / Dissertation (MSc)--University of Pretoria, 2011. / Computer Science / unrestricted