Accurate and up-to-date road information is essential for both effective urban planning and disaster management. Today, very high resolution (VHR) imagery acquired by airborne and spaceborne imaging sensors is the primary source for the acquisition of spatial information of increasingly growing road networks. Given the increased availability of the aerial and satellite images, it is necessary to develop computer-aided techniques to improve the efficiency and reduce the cost of road extraction tasks. Therefore, automation of image-based road extraction is a very active research topic. This thesis deals with the development and implementation aspects of a semi-automatic road extraction strategy, which includes two key approaches: multidirectional and single-direction road extraction. It requires a human operator to initialize a seed circle on a road and specify a extraction approach before the road is extracted by automatic algorithms using multiple vision cues. The multidirectional approach is used to detect roads with different materials, widths, intersection shapes, and degrees of noise, but sometimes it also interprets parking lots as road areas. Different from the multidirectional approach, the single-direction approach can detect roads with few mistakes, but each seed circle can only be used to detect one road. In accordance with this strategy, a RoadModeler prototype was developed. Both aerial and GeoEye-1 satellite images of seven different types of scenes with various road shapes in rural, downtown, and residential areas were used to evaluate the performance of the RoadModeler. The experimental results demonstrated that the RoadModeler is reliable and easy-to-use by a non-expert operator. Therefore, the RoadModeler is much better than the object-oriented classification. Its average road completeness, correctness, and quality achieved 94%, 97%, and 94%, respectively. These results are higher than those of Hu et al. (2007), which are 91%, 90%, and 85%, respectively. The successful development of the RoadModeler suggests that the integration of multiple vision cues potentially offers a solution to simple and fast acquisition of road information. Recommendations are given for further research to be conducted to ensure that this progress goes beyond the prototype stage and towards everyday use.

<p>Amazonia contains the largest remaining tracts of undisturbed tropical forest on earth, and is thus critical to international nature conservation and carbon sequestration efforts. Amazonian forests are notoriously difficult to study, however, due to their species richness and inaccessibility. This has limited efforts to produce the accurate, high-resolution biodiversity maps needed for conservation and development. The aims of the research described here were to identify efficient solutions to the problems of tropical forest inventory; to use these methods to identify floristic patterns and their causes in western Amazonia; and propose new means to map floristic patterns in these forests.</p><p> Using tree inventories in the vicinity of Iquitos, Peru, I and a colleague systematically evaluated methods for rapid tropical forest inventory. Of these, inventory of particular taxonomic groups, or taxonomic scope inventory, was the most efficient, and was able to capture a majority of the pattern observed by traditional inventory techniques with one-fifth to one-twentieth the number of stems and species. Based on the success of this approach, I and colleagues specifically evaluated two plant groups, the Pteridophytes (ferns and fern allies) and the Melastomataceae (a family of shrubs and small trees), for use in rapid inventory. Floristic patterns based on inventories from either group were significantly associated with those based on the tree flora, and inventories of Pteridophytes in particular were in most cases able to capture the majority of floristic patterns identified by tree inventories. These findings indicate that Pteridophyte and Melastomataceae inventories are useful tools for rapid tropical forest inventory.</p><p> Using Pteridophyte and Melastomataceae inventories from 138 sites in northwestern Amazonia, combined with satellite data and soil sampling, I and colleagues studied the causes of vegetation patterns in western Amazonian forests. On the basis of these data, we identified a floristic discontinuity of at least 300km in northern Peru, corresponding to a 15-fold difference in soil cation concentrations and an erosion-generated geological boundary. On the basis of this finding, we assembled continent-scale satellite image mosaics, and used these to search for additional discontinuities in western Amazonia. These mosaics indicate a floristic and geological discontinuity of at least 1500km western Brasil, driven by similar erosional processes identified in our study area. We suggest that this represents a chemical and ecological boundary between western and central Amazonia.</p><p> Using a second network of 52 pteridophyte and soil inventories in northwestern Amazonia, we further studied the role of geology in generating floristic pattern. Consistent with earlier findings, we found that two widespread geological formations in western Amazonia differ eight-fold difference in soil cation concentrations and in a majority of their species. Difference in elevation, used as a surrogate for geological formation, furthermore explained up to one-third of the variation in plant species composition between these formations. Significant correlations between elevation, and cation concentrations and soil texture, confirmed that differences in species composition between these formations are driven by differences in soil properties. On the basis of these findings, we were able to use SRTM elevation data to accurately model species composition throughout our study area.</p><p> I argue that Amazonian forests are partitioned into large-area units on the basis of geological formations and their edaphic properties. This finding has implications for both the ecology and evolution of these forests, and suggests that conservation strategies be implemented on a region-by-region basis. Fortunately, the methods described here provide a means for generating accurate and detailed maps of floristic patterns in these vast and remote forests.</p> / Dissertation

Accurate, fast detection of clouds in satellite imagery has many applications, for example Numerical Weather Prediction (NWP) and climate studies of both the atmosphere and of the Earth’s surface temperature. Most operational techniques for cloud detection rely on the differences between observations of cloud and of clear-sky being more or less constant in space and in time. In reality, this is not the case - different clouds have different spectral properties, and different cloud types are more or less likely in different places and at different times, depending on atmospheric conditions and on the Earth’s surface properties. Observations of clear sky also vary in space and time, depending on atmospheric and surface conditions, and on the presence or absence of aerosol particles. The Bayesian approach adopted in this project allows pixel-specific physical information (for example from NWP) to be used to predict pixel-specific observations of clear sky. A physically-based, spatially- and temporally-specific probability that each pixel contains a cloud observation is then calculated. An advantage of this approach is that identification of ambiguously classed pixels from a probabilistic result is straightforward, in contrast to the binary result generally produced by operational techniques. This project has developed and validated the Bayesian approach to cloud detection, and has extended the range of applications for which it is suitable, achieving skills scores that match or exceed those achieved by operational methods in every case. High temperature gradients can make observations of clear sky around ocean fronts, particularly at thermal wavelengths, appear similar to cloud observations. To address this potential source of ambiguous cloud detection results, a region of imagery acquired by the AATSR sensor which was noted to contain some ocean fronts, was selected. Pixels in the region were clustered according to their spectral properties with the aim of separating pixels that correspond to different thermal regimes of the ocean. The mean spectral properties of pixels in each cluster were then processed using the Bayesian cloud detection technique and the resulting posterior probability of clear then assigned to individual pixels. Several clustering methods were investigated, and the most appropriate, which allowed pixels to be associated with multiple clusters, with a normalized vector of ‘membership strengths’, was used to conduct a case study. The distribution of final calculated probabilities of clear became markedly more bimodal when clustering was included, indicating fewer ambiguous classifications, but at the cost of some single pixel clouds being missed. While further investigations could provide a solution to this, the computational expense of the clustering method made this impractical to include in the work of this project. This new Bayesian approach to cloud detection has been successfully developed by this project to a point where it has been released under public license. Initially designed as a tool to aid retrieval of sea surface temperature from night-time imagery, this project has extended the Bayesian technique to be suitable for imagery acquired over land as well as sea, and for day-time as well as for night-time imagery. This was achieved using the land surface emissivity and surface reflectance parameter products available from the MODIS sensor. This project added a visible Radiative Transfer Model (RTM), developed at University of Edinburgh, and a kernel-based surface reflectance model, adapted here from that used by the MODIS sensor, to the cloud detection algorithm. In addition, the cloud detection algorithm was adapted to be more flexible, making its implementation for data from the SEVIRI sensor straightforward. A database of ‘difficult’ cloud and clear targets, in which a wide range of both spatial and temporal locations was represented, was provided by M´et´eo-France and used in this work to validate the extensions made to the cloud detection scheme and to compare the skill of the Bayesian approach with that of operational approaches. For night land and sea imagery, the Bayesian technique, with the improvements and extensions developed by this project, achieved skills scores 10% and 13% higher than M´et´eo-France respectively. For daytime sea imagery, the skills scores were within 1% of each other for both approaches, while for land imagery the Bayesian method achieved a 2% higher skills score. The main strength of the Bayesian technique is the physical basis of the differentiation between clear and cloud observations. Using NWP information to predict pixel-specific observations for clear-sky is relatively straightforward, but making such predictions for cloud observations is more complicated. The technique therefore relies on an empirical distribution rather than a pixel-specific prediction for cloud observations. To try and address this, this project developed a means of predicting cloudy observations through the fast forward-modelling of pixel-specific NWP information. All cloud fields in the pixel-specific NWP data were set to 0, and clouds were added to the profile at discrete intervals through the atmosphere, with cloud water- and ice- path (cwp, cip) also set to values spaced exponentially at discrete intervals up to saturation, and with cloud pixel fraction set to 25%, 50%, 75% and 100%. Only single-level, single-phase clouds were modelled, with the justification that the resulting distribution of predicted observations, once smoothed through considerations of uncertainties, is likely to include observations that would correspond to multi-phase and multi-level clouds. A fast RTM was run on the profile information for each of these individual clouds and cloud altitude-, cloud pixel fraction- and channel-specific relationships between cwp (and similarly cip) and predicted observations were calculated from the results of the RTM. These relationships were used to infer predicted observations for clouds with cwp/cip values other than those explicitly forward modelled. The parameters used to define the relationships were interpolated to define relationships for predicted observations of cloud at 10m vertical intervals through the atmosphere, with pixel coverage ranging from 25% to 100% in increments of 1%. A distribution of predicted cloud observations is then achieved without explicit forward-modelling of an impractical number of atmospheric states. Weights are applied to the representation of individual clouds within the final Probability Density Function (PDF) in order to make the distribution of predicted observations realistic, according to the pixel-specific NWP data, and to distributions seen in a global reference dataset of NWP profiles from the European Centre for Medium Range Weather Forecasting (ECMWF). The distribution is then convolved with uncertainties in forward-modelling, in the NWP data, and with sensor noise to create the final PDF in observation space, from which the conditional probability that the pixel observation corresponds to a cloud observation can be read. Although the relatively fast computational implementation of the technique was achieved, the results are disappointingly poor for the SEVIRI-acquired dataset, provided by M´et´eo-France, against which validation was carried out. This is thought to be explained by both the uncertainties in the NWP data, and the forward-modelling dependence on those uncertainties, being poorly understood, and treated too optimistically in the algorithm. Including more errors in the convolution introduces the problem of quantifying those errors (a non-trivial task), and would increase the processing time, making implementation impractical. In addition, if the uncertianties considered are too high then a PDF flatter than the empirical distribution currently used would be produced, making the technique less useful.

This thesis presents and evaluates different methods to semantically segment 3D-models by rendered 2D-views. The 2D-views are segmented separately and then merged together. The thesis evaluates three different merge strategies, two different classification architectures, how many views should be rendered and how these rendered views should be arranged. The results are evaluated both quantitatively and qualitatively and then compared with the current classifier at Vricon presented in [30]. The conclusion of this thesis is that there is a performance gain to be had using this method. The best model was using two views and attains an accuracy of 90.89% which can be compared with 84.52% achieved by the single view network from [30]. The best nine view system achieved a 87.72%. The difference in accuracy between the two and the nine view system is attributed to the higher quality mesh on the sunny side of objects, which typically is the south side. The thesis provides a proof of concept and there are still many areas where the system can be improved. One of them being the extraction of training data which seemingly would have a huge impact on the performance.

In this report I summarize my master’s thesis work, in which I have investigated different approaches for fusing imaging modalities for semantic segmentation with deep convolutional networks. State-of-the-art methods for semantic segmentation of RGB-images use pre-trained models, which are fine-tuned to learn task-specific deep features. However, the use of pre-trained model weights constrains the model input to images with three channels (e.g. RGB-images). In some applications, e.g. classification of satellite imagery, there are other imaging modalities that can complement the information from the RGB modality and, thus, improve the performance of the classification. In this thesis, semantic segmentation methods designed for RGB images are extended to handle multiple imaging modalities, without compromising on the benefits, that pre-training on RGB datasets offers. In the experiments of this thesis, RGB images from satellites have been fused with normalised difference vegetation index (NDVI) and a digital surface model (DSM). The evaluation shows that the modality fusion can significantly improve the performance of semantic segmentation networks in comparison with a corresponding network with only RGB input. However, the different investigated approaches to fuse the modalities proved to achieve similar performance. The conclusion of the experiments is, that the fusion of imaging modalities is necessary, but the method of fusion has shown to be of less importance.

>Magister Scientiae - MSc / The Cape Floristic Region, situated at the southern tip of Africa, is one of the world’s most botanically diverse regions. The biodiversity of this region faces various types of threats, which can be divided into three main categories, namely increasing urbanisation, agriculture expansion, and the spread of invasive alien vegetation. It has been shown that botanically diverse areas are more prone to invasion by invasive alien plant (IAP) species. The Hawequa conservation area, in the south-western Cape, is particularly botanically diverse, such that it is very prone to aggressive invasion by IAP species. Therefore, conservation management of the Hawequa conservation area urgently need to map, prioritise and clear IAP species. Due to the topographical complexity of this mountainous area, it is not possible to map the distribution of IAP species throughout the protected area by conventional field methods. Remote sensing may be able to provide a suitable alternative for mapping. The aim of this research was to assess various image classification methods,using two types of high-resolution imagery (colour aerial photography and WorldView-2 satellite imagery), in order to map the distribution of IAP species, including small stands and individuals. Specifically, the study will focus on mapping Pinus and Acacia spp. in a study site of approximately 9 225ha in the Hawequa conservation rea. Supervised classification was performed using two different protocols, namely per-pixel and per-field. For the per-pixel classification Iterative Self-Organising Data Analyses Technique (ISODATA) was used, a method supported by ERDAS Imagine. The per-field (object-based) classification was done using fractal net evolution approach (FNEA), a method supported by eCognition. The per-pixel classification mapped the extent of Pinus and Acacia spp. in the study area as 1 205.8 ha (13%) and 80.1 ha (0.9%) respectively, and the perfield classification as 1 120.9 ha (12.1%) and 96.8 ha (1.1%) respectively. Accuracy assessments performed on the resulting thematic maps generated from these two classification methods had a kappa coefficient of 0.700 for the per-pixel classification and 0.408 for the per-field classification. Even though the overall extent of IAP species for each of these methods is similar, the reliability of the actual thematic maps is vastly different. These findings suggest that mapping IAP species (especially Pinus spp.) stands and individuals in highly diverse natural veld, the traditional per-pixel classification still proves to be the best method when using high-resolution images. In the case of Acacia spp., which often occurs along rivers, it is more difficult to distinguish them from the natural riverine vegetation. Using WorldView-2 satellite images for large areas can be very expensive (approximately R120 per km2 in 2011), but in comparison with the cost of mapping and the subsequent clearing, especially in inaccessible areas, it might be a worthwhile investment. Alternative image sources such as the high resolution digital colour infrared aerial photography must be considered as a good source for mapping IAP species in high altitude areas.

M.Ing. / Globally, hazardous weather phenomena such as violent storms, oods, cyclones, tornadoes, snow and hail contribute to signi cant annual xed property damages, loss of movable property and loss of life. The majority of global natural disasters are related to hydro-meteorological events. Hazardous storms are destructive and pose a threat to life and property. Forecasting, monitoring and detecting hazardous storms are complex and demanding tasks, that are however essential. In this study automatic hazardous weather detection utilizing remotely sensed meteorological data has been investigated. Image processing techniques have been analyzed and applied to multispectral meteorological satellite image data obtained from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) instruments on-board the Meteosat Second Generation (MSG) geostationary meteorological satellites Meteosat-8 and Meteosat-9. The primary focus of this study is the detection of potentially hazardous hydrometeorological phenomena in South Africa. A methodology for detecting potentially hazardous storms over South Africa using meteorological satellite imagery from MSG/SEVIRI is presented. An index indicative of the hazardous potential of a storm is de ned to aid in the identi cation of a ected geographical areas and to quantify the destructive potential of the detected storm. The Hazardous Potential Index (HPI) is generated through the use of image processing techniques such as cloud masking, cloud tracking and an image-based analysis of the constituent elements of a severe convective storm. A retrospective review was performed with respect to 20 case studies of documented storms which had adversely a ected areas of South Africa. A red-green-blue (RGB) composite image analysis technique, that may be utilized in the identi cation of severe convective storms using SEVIRI image data, was also applied to these case studies.

<p> </p> <p>Satellite imagery is becoming increasingly available due to a large number of commercial satellite operators. Many fields use satellite images, including meteorology, forestry, natural disaster analysis, and agriculture. These images can be changed or tampered with image manipulation tools that can cause issues in many applications. Manipulation detection techniques designed for images captured by ``consumer cameras'' tend to fail when used on satellite images. In this thesis we examine methods for detecting splices where an object or area is inserted into a satellite image. Three semi-supervised one-class methods are proposed for the detection and localization of manipulated images. A supervised and supervised fusion approach are also describe to detect spliced forgeries. The semi-supervised one-class method does not require any prior knowledge of the type of manipulations that an adversary could insert in the satellite imagery. First, a new method known as Satellite SVDD (Sat-SVDD) which adapts the Deep SVDD technique is described. Another semi-supervised one-class one-class detection technique based on deep belief networks (DBN) for splicing detection and localization is then discussed. Multiple configurations of the Deep Belief network were compared to other common one-class classification methods. Finally, a semi-supervised one-class technique that uses a Vision Transformer to detect spliced areas within satellite images is introduced. The supervised method does not require prior knowledge of the type of manipulations inserted into the satellite imagery. A supervised method known as Nested Attention U-Net, to detect spliced. The supervised fusion approach known as Sat U-Net fuses the results of two exiting forensic splicing localization methods to increase their overall accuracy. Sat U-Net is a U-Net based architecture exploiting several Transformers to enhance the splicing detection performance. Sat U-Net fuses the outputs of two semi-supervised one-class splicing detection methods, Gated PixelCNN Ensemble and Vision Transformer, to produce a heatmap highlighting the manipulated image region. The supervised fusion approach trained on images from one satellite can be lightly retrained on few images from another satellite to detect spliced regions. In this thesis I introduce five datasets of manipulated satellite images that contain spliced objects. Three of the datasets contains images with spliced objects generated by a generative adversarial network (GAN).</p>

Throughout history, geographical information has been vital in different contexts, such as national security matters, economics, geopolitics, military, and natural resources. Due to the various applications, geographical information has been handled as valuable and sensitive information. As technology evolves, geographical information is becoming increasingly more available. This thesis investigates the data attributes relevant to its sensitivity and creates an information security classification model suitable for the satellite imagery produced, analyzed, and maintained by Maxar Technologies Ltd. All geographical information is of value. Everything from terrain information to protected areas where features such as roads, critical infrastructure, and buildings are of extra interest. Other factors that affect the sensitivity of the imagery are the resolution, amount of information, type of files (3D or other processed data), legislation, and more. The methodology used to achieve this consisted of two major parts, a risk assessment procedure and translating risk contexts and parameters into a classification model. The classification levels identified are PUBLIC, VALUABLE, SENSITIVE, and CLASSIFIED. A classification model is defined for individual imagery and a separate model for projects. A project gets at least the same classification as the highest classed file and other contexts that may affect the sensitivity.   Lastly, the thesis explore automation possibilities and a supervised learning approach is tested on the model created for the classification of files. Various machine learning models are fitted to a dataset that is collected from the satellite imagery products of Maxar and manually classed using the defined classification levels. F-score and MCC are used to evaluate the automation. These are metrics based on the occurrences of false positives and negatives. Furthermore, the thesis discusses topics related to the sensitivity of geographical information and how to handle such information. This thesis tries to lay the foundation for many future work possibilities.

<p>Urban models are of growing importance today for urban and environmental planning, geographic information systems, urban simulations, and as content for entertainment applications. Various methods have addressed aerial or ground scale image-based and sensor-based reconstruction. However, few, if any, approaches have automatically produced urban models from satellite images due to difficulties of data noise, data sparsity, and data uncertainty. Our key observations are that many structures in urban areas exhibit regular properties, and a second or more satellite views for urban structures are usually available. Hence, we can overcome the aforementioned issues obtained from satellite imagery by synthesizing the underlying structure layout. In addition, recent advances in deep learning allow the development of novel algorithms that was not possible several years ago. We leverage relevant deep learning techniques for classifying/predicting urban structure parameters and modeling urban areas that address the problem of satellite data quality and uncertainty. In this dissertation, we present a machine learning-based procedural generation framework to automatically and quickly reconstruct urban areas by using regularities of urban structures (e.g., cities, buildings, facades, roofs, etc.) from satellite imagery, which can be applied to not only multiple resolutions ranging from low resolution (e.g., 3 meters) to high resolutions (e.g., WV3 0.3 meter) of satellite images but also the different scales (e.g., cities, blocks, parcels, buildings, facades) of urban environments. Our method is fully automatic and generates procedural structures in urban areas given satellite imagery. Experimental results show that our method outperforms previous state-of-the-art methods quantitatively and qualitatively for multiple datasets. Furthermore, by applying our framework to multiple urban structures, we demonstrate our approach can be generalized to various pattern types. We also have preliminary results applying this for flooding, archaeological sites, and more.</p>