Information extracted from aerial photographs has found applications in a wide range of areas including urban planning, crop and forest management, disaster relief, and climate modeling. At present, much of the extraction is still performed by human experts, making the process slow, costly, and error prone. The goal of this thesis is to develop methods for automatically extracting the locations of objects such as roads, buildings, and trees directly from aerial images. We investigate the use of machine learning methods trained on aligned aerial images and possibly outdated maps for labeling the pixels of an aerial image with semantic labels. We show how deep neural networks implemented on modern GPUs can be used to efficiently learn highly discriminative image features. We then introduce new loss functions for training neural networks that are partially robust to incomplete and poorly registered target maps. Finally, we propose two ways of improving the predictions of our system by introducing structure into the outputs of the neural networks. We evaluate our system on the largest and most-challenging road and building detection datasets considered in the literature and show that it works reliably under a wide variety of conditions. Furthermore, we are releasing the first large-scale road and building detection datasets to the public in order to facilitate future comparisons with other methods.

This thesis estimated vegetation biochemical properties at multiple spatial scales and investigate vegetation temporal dynamics under climate influences in a heterogeneous tallgrass ecosystem in Southern Ontario using remote sensing data. Ground hyperspectral and space multispectral remote sensing data derived Normalized Difference Vegetation Index (NDVI) and Simple Ratio (SR) were used to estimate biochemical properties at the species, canopy and landscape level. Both vegetation indices explained 32% to 56% of the variations in biochemical properties at the species level, 16% to 53% at the canopy level, and over 60% at the landscape level. MODIS NDVI and climate data were also collected to investigate the vegetation-climate relationships during the growing season and the lag effects of climate factors on vegetation at the peak growing season. The findings indicate that temperature is the key climate factor that drives the annual cycle, and there is a time lag effect of climate factors on vegetation.

Accurate determination of solar flux at the top of the atmosphere (TOA), which can only be made from satellite observations, is essential for climate studies. In the present study, we developed a new technique to derive the solar fluxes at the TOA and at the surface from operational meteorological satellites. Two key steps in the technique are the narrowband to broadband (NTB) conversion and deriving the net solar flux at the surface from TOA observations. We developed a new NTB conversion algorithm from ScaRaB observations and radiation transfer model simulations. In deriving the net solar fluxes at the surface from TOA measurements, the effects of absorbing aerosols have been investigated. / The technique described above has been applied to AVHRR data to derive a dataset of solar flux in the Mackenzie River Basin (MRB) for the Canadian GEWEX Enhanced Study (CAGES) period from June 1998 to September 1999. The derived net solar fluxes at the surface were evaluated with the surface measurements in the basin and good agreement was achieved. / Radiation fields from two atmospheric models used in the Mackenzie GEWEX Study (MAGS) project, the Canadian Regional Climate Model (CRCM) and the Global Environmental Multiscale (GEM) model, were evaluated against satellite retrievals of radiation fluxes in the present study. It was found that the CRCM simulated the TOA reflected flux well in the MRB for the summer of 1994, but large biases were found in the partition of absorbed solar radiation between the atmosphere and the earth's surface. The net surface solar radiation was found to be overestimated by about 15% in the CRCM. Evaluation of the preliminary output from the new version of CRCM shows substantial improvement. Evaluation of radiation fields from the GEM model shows good agreement under clear skies, but under cloudy skies, the TOA albedo simulated by the GEM model in the MRB was about 30% lower than observations for the summer of 1999.

Graduation date: 1997

Satellite imagery such as Landsat has been in use for decades for many landscape and regional scale mapping applications, but has been too coarse for use in detailed forest inventories where stand level structural and compositional information is desired. Recently available high spatial resolution satellite imagery may be well suited to mapping fine-scale components of ecosystems, however, this remains an area of ongoing research. The first goal of this thesis was to assess the capacity of high spatial resolution satellite imagery to detect the variability in late seral coastal temperate rainforests in British Columbia, Canada. Using an object-based classifier, two hierarchical classification schemes are evaluated: a broad classification based on structural (successional) stage and a finer classification of late seral vegetation associations. The finer-scale classification also incorporates ancillary landscape positional variables (elevation and potential soil moisture) derived from Light Detection and Ranging (LiDAR) data, and the relative contribution of spectral, textural and landscape positional data for this classification is determined. Results indicate that late seral forests can be well distinguished from younger forests using QuickBird spectral and textural data. However, discrimination among late seral forest associations is challenging, especially in the absence of landscape positional variables. Classification accuracies were particularly low for rare forest associations. Given this finding, the objective of the third chapter was to explicitly examine the caveats of using high spatial resolution imagery to map rare classes. Classification accuracy is assessed in several different ways in order to examine the impact on perceived map accuracy. In addition, the effects on habitat extent and configuration resulting from post-classification implementation of a minimum mapping unit are examined. Results indicate that classification accuracies may vary considerably depending on the assessment technique used. Specifically, ignoring the presence of fine-scale heterogeneity in a classification during accuracy assessment falsely lowered the accuracy estimates. Further, post-classification smoothing had a large effect on the spatial pattern of rare classes. These findings suggest that routinely used image classification and assessment techniques can greatly impact mapping of rare classes.

Australia is one of the leading mineral resource extraction nations in the world. It is one of the world’s top producers of nickel, zinc, uranium, lithium, coal, gold, iron ore and silver. However, the complexity of the environmental issues and the potentially damaging consequences of mining have attracted public attention and political controversy. Other types of underground natural resource exploitation, such as ground water, gas or oil extractions, also cause severe land deformation on different scales in space and time. The subsidence due to underground mining and underground fluid extractions has the potential to impact on surface and near surface infrastructure; as well as water quality and quantity, that in turn has the potential to impact on threatened flora and fauna, and biodiversity conservation. Subsidence can also impact natural and cultural heritage. To date, most of land deformation monitoring is done using conventional surveying techniques, such as total stations, levelling, GPS, etc. These surveying techniques provide high precision in height at millimetre accuracy, but with the drawbacks of inefficiency and costliness (labour intensive and time consuming) when surveying over a large area. Radar interferometry is an imaging technique for measuring geodetic information of terrain. It exploits phase information of the backscattered radar signals from the ground surface to retrieve the altitude or displacements of the objects. It has been successfully applied in the areas of cartography, geodesy, land cover characterisation, mitigation of natural or man-made hazards, etc. The goal of this dissertation was to develop a system which integrated differential interferometric synthetic aperture radar (DInSAR), ground survey data and geographic information systems (GIS) as a whole to provide the land deformation maps for underground mining and water extraction activities. This system aimed to reinforce subsidence assessment processes and avoid or mitigate potential risks to lives, infrastructure and the natural environment. The selection of suitable interferometric pairs is limited to the spatial and temporal separations of the acquired SAR images as well as the characteristics of the site, e.g. slope of terrain, land cover, climate, etc. Interferometric pairs with good coherence were selected for further DInSAR analysis. The coherence analysis of both C- and L-band spaceborne SAR data was studied for sites in the State of New South Wales, Australia. The impact of the quality of the digital elevation models (DEM), used to remove the static topography in 2-pass DInSAR, were also analysed. This dissertation examined the quality of the DEM generated using aerial photogrammetry, InSAR, and airborne laser scanning (ALS) against field survey data. Kinematic and real-time kinematic GPS were introduced here as an efficient surveying method for collecting ground truth data for DEM validation. For mine subsidence monitoring, continuous DInSAR mine subsidence maps were generated using ERS-1/2, Radarsat-1 and JERS-1 data with the assumption of negligible horizontal displacement. One of the significant findings of this study was the results from the ERS-1/2 tandem DInSAR, which showed an immediate mine subsidence of 1cm occurred during a period of 24 hours. It also raised the importance of SAR constellations for disaster mitigation. In order to understand the 3-D displacement vectors of mine deformation, this dissertation also proposed a method using the SAR data acquired at 3 independent incidence angles from both ascending and descending orbits. Another issue of the high phase gradient, induced by the mine subsidence, was also addressed. Phase gradient was clearly overcome by having the L-band ALOS data with an imaging resolution of 10m, which is better than the imaging resolution of 18m of the previous generation of the Japanese L-band SAR satellite, JERS-1. The ground survey data over a similar duration was used for validation. Besides mine subsidence monitoring the land deformation caused by groundwater pumping were also presented. In contrast to mine subsidence, the underground water extraction induced subsidence has the characteristics of a slow rate of change and less predictable location and coverage. Two case studies were presented. One was at the geothermal fields in New Zealand and another was the urban subsidence due to underground water over exploitation in China. Both studies were validated against ground survey data. Finally, SAR intensity analysis for detecting land deformation was demonstrated when DInSAR was not applicable due to strong decorrelation. The region of land surface change, which may be caused by human activities or natural disasters, can be classified. Two cases studies were given. The first study was the surface change detection at an open-cut mine. The second one was the 2004 Asian tsunami damage assessment near Banda Aceh. The results presented in this dissertation showed that the integrated system of DInSAR, GIS and ground surveys has the potential to monitor mine subsidence over a large area. The accuracy of the derived subsidence maps can be further improved by having a shorter revisit cycle and better imaging resolution of the newly launched and planned SAR satellites and constellation missions. The subsidence caused by groundwater pumping can be monitored at an accuracy of millimetre by utilising the technique of persistent scatterer InSAR.

Some mounted ill. / Includes bibliography / xii, [468] leaves : ill. (some col.) ; 30 cm. / Title page, contents and abstract only. The complete thesis in print form is available from the University Library. / Thesis (Ph.D.)--Dept. of Engineering, University of Adelaide, 1984

This thesis is concerned with the dynamics of the middle atmosphere over the southern-most continent on our planet, Antarctica. Building on previous observational and theoretical efforts to understand the dynamics of this region of the atmosphere, the work encompasses instrument improvements as well as data analysis studies of gravity waves in the lower and middle stratosphere, tides in the upper mesosphere and lower thermosphere, and short-period planetary waves in the stratosphere and mesosphere. An upgrade of the Medium-Frequency Spaced Antenna (MFSA) radar at Scott Base, Antarctica, was carried out in 2004/5 in order to ensure continued operation and enhance its capabilities to measure gravity waves. As a result, the quality of the wind measurements was greatly enhanced and the amount of data collected is now greater by a factor of approximately 15 compared with before the upgrade. Analysis of over two decades of wind velocity data from Scott Base yields a reliable climatology of the dynamics of the upper mesosphere and lower thermosphere in this area, namely the tidal oscillations and the background winds. In addition, interannual variability is assessed and periodicities of approximately 11 years and strong positive trends in tidal amplitudes are reported. Mechanisms that could explain the observed behaviour are proposed. The data are then combined with wind measurements from Halley, the longitudinal conjugate site, in order to study the zonal character of the semi-diurnal tide. Zonal wavenumber 1 and 2 waves are both found and mechanisms that could explain the generation of a wavenumber 1 component are suggested. Two further sets of MFSA radar wind measurements are used to investigate the behaviour of planetary waves with periods of between two and four days in the Antarctic middle atmosphere. Satellite temperature measurements further help to create a more complete picture of these waves. Baroclinic and barotropic instabilities, which result from shears of the zonal wind, appear to be responsible for much of the observed wave activity. In addition, a quasi-to day wave event in mid-May 2005 with unusually large amplitudes is examined and suggested to be linked to a solar proton event. Gravity wave activity over Antarctica is studied using temperature profiles obtained through the satellite radio occultation technique. Although the measurements are restricted to below 35 km altitude, high-resolution temperature profiles allow conclusions to be drawn about the seasonal, geographical, and height distribution of gravity wave activity. Mountain waves are found to be important over the Antarctic Peninsula and the Transantarctic mountains where they contribute more than 20% of the observed wave activity in the lower stratosphere. In addition, the analysis indicates the importance of critical-level filtering and Doppler-shifting.

Thesis (Ph. D.)--University of Adelaide, Dept. of Electrical and Electronic Engineering, 1985. / Includes bibliographical references (leaves xvii-xxvi).

Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2006. / Source: Dissertation Abstracts International, Volume: 68-02, Section: B, page: 1171. Adviser: Farzad Kamalabadi. Includes bibliographical references (leaves 108-114) Available on microfilm from Pro Quest Information and Learning.