A variety of statistical methods are available to detect sudden changes, or breakpoints, in time series when used as multi-temporal change detection technique. However, these methods are unreliable in the presence of noise. Neural nets might detect breakpoints better. These deep learning models are able to generalize and optimize well, even in the presence of noise. This research tests the feasibility of different neural net architectures to detect breakpoints in generic linear time series. Two relatively simple neural nets are proposed, combined with four different descriptions of breakpoint, and trained on synthetic
data. The neural nets are tested on two datasets: On a separate synthetic dataset and on Australian rainuse-efficieny (RUE) time series, a surrogate for dryland ecosystem functioning. Some of the neural nets built performed exceptionally well on synthetic data, outperforming a benchmark statistical method with
margin. The direct translation to RUE time series was less successful. The results shows great promise for the use of neural nets in change detection. A generalist change detection approach by use of neural nets is likely not optimal. Current developments in deep learning, as well as choosing the right user-case, show
great promise to unlock the full potential of neural nets in time series analysis.

Coastal dunes habitats are in the transition zone of marine and terrestrial ecosystems and therefore have high biodiversity in terms of animal and vegetation species. These species provide important ecosystem services to the human kind. Due to the development and operation of Maasvlakte 2, an increase of nitrogen emission is expected in the province Zuid-Holland of the Netherlands. When this is taken up by the vegetation and soil in the dunes, encroachment by shrubs will dominate. This will lead to a decrease in species diversity of dune habitats in protected areas of the Natura 2000 network, which is an unfavourable situation conforming to the Nature Protection Law (Wet Natuurbescherming). The main loss is expected to be in grey dunes and humid dune slacks of the Natura 2000 areas nearby Maasvlakte 2. This research focuses on the classification of grey dunes, white dunes, humid dune slacks, woody vegetation, water and their sub-classes in the Natura 2000 areas Solleveld & Kapittelduinen, Voornes Duin and Goeree & Kwade Hoek. Remote sensing in combination with machine learning have been proven to be useful in the monitoring of dune habitats. Open access satellite data provide high temporal and geographic coverage, but are often limited in their spatial resolution. Contrarily, high spatial resolution products are limited in their spectral resolution. In this thesis, the effects of spatial resolution, spectral features and feature combinations on the classification of dune habitats using aerial images and a Landsat-7 image (2012) are analysed. The classifications are performed with a Random Forest classifier, using both a pixel-based and region-based approach. In contrast to the traditional pixel-based method, the region-based method incorporates spectral and statistical information of neighbouring pixels. The best performance is achieved with the region-based approach on a 4m resolution aerial image with an accuracy of 87%. The 10m region-based classification is the second best with an accuracy of 83%. The misclassified pixels are mostly classified as the other sub-class of the same main-class. The region-based method should not be applied to satellite images with low spatial resolution as it causes more loss of detail. More spectral bands contribute to a better classification performance, but they contribute less than the spectral bands that were already used. Also, Landsat-7 images were used to create yearly habitat maps for 2012 and 2019 to subsequently analyse the habitat changes between those years, although the accuracies of the Landsat-7 maps are an issue. The main changes in habitat are found in the woody vegetation (shrubs, trees) and humid dune slacks. The woody vegetation decreased in area, causing an increase in grey dunes area in Voornes Duin and Goeree & Kwade Hoek, whereas the opposite has happened in Solleveld & Kapittelduinen. Loss of area in humid dune slacks is detected in all three regions of interest. The local dune manager confirmed that the detected changes in woody vegetation are mostly correct and caused by dune management activities. The perceived changes in humid dune slacks are less reliable. Finally, a modified training dataset is used to classify Sentinel-2 images in 2019. These results compared reasonably well with the few locations we visited during our fieldwork campaign.

Measurements of a tachymetric network in the geothermal area of Bjarnarflag in North-East Iceland are performed during four consecutive years (2015-2018). For the yearly adjustment of the measurements an alternative iteration scheme to Baarda’s ’B-method of testing’ is proposed, because the level of significance of the F-test performed in the detection step is too large due to the large redundancy of the measurements. The proposed alternative method detects outliers and verifies the stochastic model simultaneously, resulting in 3
to 4% rejected measurements per year. From the yearly solutions absolute and relative horizontal velocities are estimated. The relative velocities are significant, but have large uncertainties. Adding three years of data reduces the mean standard deviation of the estimated velocities from 2.4 mm in east-direction to 0.9 mm and 1.8 mm in north-direction to 0.8 mm. The improvement of precision can be accelerated by reevaluating the stochastic input parameters, adding more GNSS measurements and reconsidering the network design.
Improved horizontal relative velocities can be used to understand the horizontal deformation patterns in the area of study due to extraction of water or steam by the geothermal powerplant, due to the instability of the benchmaks or due to natural processes.

In 2017 the Netherlands had 7146km of railways [Ramaekers et al., 2009], which are owned and managed by ProRail and have to be frequently surveyed, for quality inspection. Surveying is done using a variety of surveying techniques, many of which require surveyors to walk on or near the tracks, which could cause injuries. An alternative surveying technique would be photogrammetry using images acquired with a camera mounted on a drone. In this technique a 3D point cloud is constructed out of images, which is georeferenced using GPS measurements on the drone and Ground Control Points (GCP). The positions of the rail rods can then be found by modelling the rail rod using the point cloud. The quality of the point cloud, defined here as a combination of accuracy, precision, point density, completeness, and the rail rod feature detectability, is an important factor in deciding if photogrammetry is a sufficient surveying technique for this project.

In this research the influence of the distance between the camera and the track, the baseline, and the number and spread of GCPs on the quality of the resulting point cloud. The distance to the track was set to 25, 30 and35m and the baseline was varied between 2 and 10m. The requirement set for the accuracy is 15mm, while the precision has to be below 10mm. For the point density an K-nearest neighbors (K-nn) of at least 10 is required, the completeness requires less than 10% missing cells and finally the feature detectability requires that at least once every 2m the location of the rail can be determined. The features are detected using a 2D model of the rail that is fitted in a flattened slice of the point cloud using a modified version of the iterative closest point algorithm.

It is shown that the accuracy and precision are influenced by the distance and baseline, but no clear relation was found, while the accuracy improves when using GCPs. For the point density it is shown that increasing the distance between the camera and the track linearly decreases the point density, while the baseline has no impact and for the completeness a decreasing trend was found when increasing the baseline, while no relation could be found between the completeness and the distance between the camera and the track. For the rail rod feature detectability it was shown that the percentage of accurately found rail rod positions goes down when increasing the distance to the track and the baseline. Overall it is proven that when using a 100MP camera, 5 GCPs, a distance of 25m, and a baseline of 2.17m the rail rods could be surveyed so that the requirements set for this research were met.

The contemporary trend shows a shift from rule-based algorithms to deep learning. In the last few years, this field has been developing rapidly and its popularity has increased to a large extent. This happened for a good reason, since deep learning was able to solve some of the hardest problems in fields like image recognition, natural language processing and speech recognition. A large proportion of its success has to be credited to the explosion of big data in the past two decades. Structured data sets are essential to deep learning systems. The rising amount of automatic identification system data is a key example in the current big data boom. The automatic identification system produces spatiotemporal movement data of vessels. It was designed as a collision avoidance system, but researchers have been looking into ways to leverage its data for other tasks. Analyzing behavior in the movement data of vessels can help policymakers and monitoring operators with decision making processes. Improving these processes lead to safer and more resilient marine environments. Unfortunately, the possibility of applying deep learning on vessel movement data is an underexposed topic. This project attempts to explore the research gap in this topic. The objective is therefore to give an overview of the possibilities, complications and opportunities given the current state of the art. Ultimately, this project may serve as a rough guide for those who wish to explore the crossings where deep learning and vessel movement data meet.