Accurate water-level monitoring is vital for effective river management, flood prevention, and environmental conservation efforts. The increasing frequency and severity of flooding events, driven by climate change, highlight the critical need for reliable, robust, and sustainable water-level measurement systems. Such systems are particularly necessary in remote and resource-constrained settings where conventional infrastructure is lacking or insufficient.

Traditional water-level measurement methods, including radar, ultrasonic, pressure sensors, and conventional imaging systems, encounter significant limitations. These include high acquisition and maintenance costs, susceptibility to environmental damage, vandalism risks due to conspicuous placement, and dependence on stable, continuous power sources. Consequently, there remains an unmet demand for affordable, resilient, and autonomous monitoring devices capable of functioning reliably in challenging and off-grid environments.

To address this gap, this study presents the development and validation of an innovative water-level monitoring prototype as a prove of concept. The proposed system integrates a low-cost Raspberry Pi-based imaging sensor equipped with infrared illumination to enable accurate measurements both during the day and at night. An onboard processing unit autonomously captures and analyses images in real-time using one of two distinct image-processing algorithms: the mean-difference method and the Kolmogorov–Smirnov (KS) test method. These algorithms quantify water levels by detecting pixel-intensity contrasts, thus eliminating the need for direct physical contact with the water body.

The prototype underwent field validation in a natural river environment, demonstrating robust and consistent performance. The system achieved an accuracy within ±4.6 cm at a 95% confidence interval. Notably, it exhibited stable performance under varying environmental conditions, with moderate bias differences between daytime and nighttime scenarios, and minimal sensitivity to precipitation effects. The mean-difference algorithm demonstrated superior precision, while the KS-test method offered enhanced robustness against environmental variability.

This research underscores the practical feasibility and significant potential of low-cost, autonomous camera-based systems for sustainable water-level monitoring. Such solutions can substantially improve disaster preparedness and environmental management, especially in remote or economically disadvantaged regions lacking traditional monitoring infrastructure. Future research directions include integrating renewable energy solutions such as solar power to enhance operational autonomy, exploring advanced image-processing algorithms for increased accuracy, and conducting extended validations across a broader range of hydrological and environmental scenarios.

This thesis performed a novel approach to shear strength determination using cone penetration tests and weather data to predict undrained shear strength and volumetric water content. Through SHAP feature importance analysis, models were interpreted and improved through feature engineering.

As the frequency and intensity of natural disasters increase, there is growing recognition of the need to address climate change and limit the increase in global average temperature. The shipping industry, which contributes 2.9% to global anthropogenic greenhouse gas emissions, plays a significant role by releasing substantial amounts of CO2, other harmful gases, and fine particles, negatively impacting climate change and the health of those near ports or waterways. As part of a wider initiative to reduce those effects, the International Maritime Organization has set the goal for the shipping industry to achieve net zero emissions by 2050. In order to accomplish this, accurate and comprehensive information on emissions is crucial.

Various methods have been developed to estimate emissions in the shipping industry. Top-down methods are applied using large-scale data to estimate emissions over a wide area. This approach provides a comprehensive overview but lacks the specific details required for local interventions. In contrast, bottom-up methods are applied and start with detailed data at the source and aggregate information to estimate the total emissions. Bottom-up methods offer more precise insights, however, more extensive data and complex modeling is required. To obtain more precise understandings, the modern bottom-up models use Automatic Identification System as input, a globally used system for tracking vessels. AIS data consists, among others, of time-dependent variables such as speed, location, and ship identification number. The data is used to assign an operational mode to the ship (sailing, maneuvering, anchoring, berthing). Based on this mode, it is determined if the main engines of the ship are on. If they are, the engine power is calculated from the resistance force acting on the moving ship. It is possible to convert engine power to fuel consumption and later to emissions with Specific Fuel Oil Consumption and emission factors.

Currently used emission models calculate engine power with the use of empirical formulas. So far, these models have mostly been validated by comparing them either to a noon report, or to another model. As a result, it gives no insight into the capability of a model to predict with a high spatial resolution, which is important in port areas and inland waterways.

There is little room for improvement in current semi-empirical bottom-up methods. However, in recent years, machine learning has shown the capability of replacing and often outperforming empirical models in other fields of research. This is due to the ability to model nonlinear behavior, find relations that humans can not, and work in higher dimensions. As a ship's engine power is reliant on a multitude of factors, machine learning might be the solution towards more accurate predictions.

This research aims to assess the capability of machine learning models to predict a ship's engine power with a high spatial resolution, with a focus on LSTM models. To do so, onboard sensor data from two ships was used; one sea-going container vessel and one inland tanker. The measured data was used to validate the semi-empirical method. Afterwards, the machine learning model was trained against the same sensor data. The predictions of the semi-empirical and machine learning models were then compared to one another.

Comparisons showed that the error for total energy used for the container vessel went from +62.87% to -0.82% when using a Bi-LSTM model with speed, acceleration, draught and depth as input. For the assessment of the spatiotemporal predictions, the MAE and RMSE were used. Based on these performance indicators, a normal LSTM network with speed, acceleration, draught and depth as input performed best. The MAE compared to the reference went down from 9542 to 2863, or about 7% of the maximum engine power.

The second case study highlighted the challenges that come with engine power prediction in inland waterways. Firstly, the fact that the speed over ground is used has a bigger influence in this case, as the ship will always sail with or against the current. Secondly, the influence of the blockage factor can
not be ignored without having higher losses. In addition to these two shortcomings, the model was trained on a diesel-electric ship. This ship has a much more constant power profile than ships with a more classical propulsion run on fossil fuels. This likely made the model more robust to changes. The combination of these three aspects caused less accurate predictions. Although there was more data available, errors were higher than for the case study. The best performing models showed a +3.41% error on total energy use and an MAE of 205.65, or about 12.8% of the maximum engine power.

This research has contributed new insights to the field of maritime emission modeling. The potential of machine learning models has been shown in comparison to an existing semi-empirical model. In addition, this model has now been validated using measurement data for a seagoing container vessel. Finally, shortcomings of the method using machine learning were exposed and a solution is proposed for a more complete, generalizable model.

This thesis aims to enhance rainfall nowcasting by improving motion field predictions and ensemble generation within PySTEPS using machine learning techniques. Accurate nowcasting is crucial for flood early warning, agriculture, transportation, and public safety. The steady-state assumptions and ensemble generation in PySTEPS, and many other nowcasting methods, face significant challenges in maintaining accuracy over time, especially during convective weather events characterized by rapid changes in precipitation patterns and their movement. The research focuses on three objectives: identifying the current errors and uncertainties in PySTEPS motion field estimations, constructing dynamic motion field predictions with the development of a new deep-learning model, MotioNNet, and developing motion field ensemble generation methods for MotioNNet.

We introduce a new dataset containing 10,000 motion field events, consisting of events with high rainfall and motion errors in the NL. MotioNNet, trained/validated/tested on 75%/12.5%/12.5% of the dataset, improves motion prediction accuracy by approximately 13%. It provides accurate temporal adjustments (e.g., advecting the motion), especially in structured and stable events. For ensemble generation, probabilistic techniques such as SpatialDropout and Monte Carlo dropout are employed within MotioNNet. The results indicate better ensemble predictions in structured and stable events, accurately identifying problematic areas and generating realistic members. However, it presents a small spread compared to PySTEPS. Challenges remain in accurately capturing uncertainty during highly dynamic convective events, but new paths are open for future enhancement.

The study concludes that integrating machine learning models into traditional nowcasting frameworks can substantially enhance motion field predictions and ensemble generation. Future work should focus on further refining these models or exploring more advanced architectures (e.g., ConvLSTM, diffusion models, or transformers), improving context-dependent perturbations in ensemble generation, and enhancing the references and assessments. These advancements hold potential for improving motion field nowcasting accuracy.

Accurate forecasts are essential for integrating wind energy into the power grid. With wind energy's growing role in the renewable mix, precise short-term generation forecasts are increasingly vital. Turbine-level forecasts are critical for optimal wind farm operation, control, and planning. However, the wind's unpredictability and the complex interactions between turbines make forecasting challenging. Existing methods either depend on resource-intensive physical models or overlook turbine interactions, resulting in isolated, overly simplistic models.

This master's thesis explores the use of Graph Neural Networks (GNNs) in conjunction with Recurrent Neural Networks (RNNs) for wind power generation forecasting in wind farms. The result is an integrated approach that can learn the interactions between turbines through GNNs to handle the temporal dynamics of the data through RNNs.

The research begins with an in-depth overview of wind turbines and their arrangement in wind farms, highlighting the importance and challenges of Wind Power Forecasting (WPF). It establishes that the spatial configuration of turbines complicates power generation estimation, especially over different time horizons. Graphs are introduced as an effective way to represent the interconnections between turbines. Various methods for constructing these graphs are reviewed, focusing on predefined heuristics and their limitations in adapting to changing wind farm characteristics.

The core of this thesis is the AG-LSTM Network, a model that integrates RNNs and GNNs for short-term WPF for turbines within a wind farm. It utilizes historical power generation data, a variety of future and past covariates, and static turbine-specific features. The model is based on an encoder-decoder architecture with an Adaptive Graph Long-Short Term Memory Cell (AG-LSTM), which generates a dynamic adjacency matrix representing the farm's state at each moment. This matrix combines the input features and hidden states across turbines, resulting in an embedding that is transformed into the power forecast.

Through comprehensive experiments and analyses, the AG-LSTM Network proves to be more accurate than state-of-the-art methods on two real-life datasets when future covariates are unavailable. Further experiments demonstrate the quality of the forecast and break down the contribution of the different choices regarding its architecture. The results underscore the potential of hybrid neural network architectures in enhancing WPF, providing a valuable tool for the renewable energy sector.

In conclusion, this research contributes to the renewable energy field by proposing a novel hybrid neural network model that effectively addresses the complexities of short-term WPF. Future work could explore further exploitation of future covariates and the robustness of the model against incomplete input data.

Nowcasting high-intensity precipitation is crucial for emergency services and municipalities when making weather-dependent decisions. This research implements and trains a deep generative model for nowcasting using a cleaned precipitation radar composite dataset spanning 15 years, with a 5-minute temporal and 1 km spatial resolution.

We propose and apply a method for speckle-like clutter removal to enhance data quality, particularly for high-intensity precipitation rates. The deep generative model is trained with two data sampling strategies to balance the data and improve the accuracy of high-intensity precipitation forecasts. Model performance is evaluated using several standard metrics, and we propose an adaptation to one metric to quantify a score to peak anticipation time in precipitation nowcasting. We also compare our model's performance with a state-of-the-art deterministic Lagrangian extrapolation-based nowcasting system.

Our results show that the proposed data quality improvement method effectively removes certain errors from historical radar data. Although the deep generative model currently scores low on the standard metrics, the model trained with a focus on high-intensity precipitation shows an improved score to peak anticipation time. Both deep generative models exhibit less blurring compared to the state-of-the-art model and, in some cases, perform similarly or outperform it.

Conceptual models in hydrology are widely used, allow for easy interpretation and require little data. Machine learning models in hydrology often outperform conceptual models but lack the ease of interpretability, require large amounts of data and and do not obey physical laws. Hybrid approaches aiming to combine the advantages of both approaches are becoming more popular. A Neural Ordinary Differential Equations approach is introduced to combine a differential equation-based conceptual model with a neural network. Additionally, conceptual models and LSTM models are used as benchmarks. The models are tested using the LamaH-CE dataset as well as the E-OBS dataset. In many cases the hybrid models outperform the conceptual model. However, to further improve the performance of hybrid models more research is needed to make the models more computationally efficient and optimized training strategies are required to explore the full potential of the approach.

Sewage systems hold a critical function within the fabric of urban infrastructure, primarily by ensuring efficient wastewater management. Within this network, sewage pumping stations emerge as important components. The performance of such stations is susceptible to being undermined by system failures and malfunctions, which can cause significant implications for both public health and occurrences of Combined Sewer Overflow (CSO) events. Consequently, discerning the patterns and root causes of these failures is of great importance.
The present study explores and seeks to understand the reliability of sewage pumping stations, particularly in the setting of the Department of Public Works in Rotterdam. The scope of this study is confined to data collected between 2016 and 2020. The research encompasses a diverse range of interconnected aspects, such as examining discrepancies in failure data, categorizing types of failures, tracking changes in failure patterns over time, and selecting a fitting statistical model to accurately represent the findings. The core focus of this study is to formulate an analysis framework to assess the reliability and failure patterns in wastewater pumping stations. In these patterns, various trends can be discerned through interarrival time - the period between two failures. For instance, an increasing trend indicates less frequent failures, suggesting improved reliability of the pumps. The devised framework combines both objective and subjective trend analyses, multiple trend tests, and employs a range of models such as the Homogeneous Poisson Process (HPP), Renewal Process (RP), and Non-Homogeneous Poisson Process (NHPP). This methodological approach is structured to facilitate a better understanding of the patterns of system behavior and shifts. Analyzing pump performance patterns over five years revealed various trends among 447 pumps studied. Of these, 254 had consistent failure patterns; 98 showed a stable trend, 146 improved over time with fewer failures, and 10 experienced more failures, indicating declining performance. The remainingpumps were divided into various trend groups or segments, demonstrating different patterns in their performance over the five years. The application of statistical methods elucidated these failure patterns, contributing to the evaluation of sewage pump station efficiency. Notwithstanding these progressions, my investigation has identified certain domains that could derive advantages from additional enhancements. The current methodology, which entailed manual configuration of kernel parameters and dependence on piecewise linear regression, was deemed insufficient for about 50% of the pumps. The study proposes the integration of sophisticated parameter estimation techniques, including Bayesian optimization or grid search, and alternative modeling methodologies to tackle this issue. It is recommended that forthcoming research expands its scope beyond Rotterdam and investigates a wider variety of pump mechanisms to validate the generalizability of the findings presented in this study. In conclusion, this thesis develops an analytical framework based on examination of failure patterns and system dynamics. It establishes a platform for future progress by outlining a distinct pathway for prospective investigations pertaining to the failures in sewage pumping stations.

This thesis introduces a novel Generative Adversarial Network application called SchemaGAN, which has been adapted from the Pix2Pix architecture to take Cone Penetration Test (CPT) data as a conditional input and generate subsoil schematizations. For training, validation and testing, a database of 24,000 synthetic schematizations of size 32x512 pixels was created, representing a broad spectrum of stratigraphical complexity in the layered models. Each synthetic cross-section was additionally transformed into a CPT-like image with less than 1% of the original data remaining at random locations along the model. After training for 200 epochs, the best-performing SchemaGAN Generator was chosen from the validation, and the effectiveness of SchemaGAN in generating subsoil schematizations was tested against traditional interpolation methods (Nearest Neighbour, Inverse Distance Weight, Kriging, Natural Neighbour) and newer methods such as Inpainting. The evaluation metrics obtained reveal that SchemaGAN outperforms all other methods, with results characterized by clearer layer boundaries and accurate anisotropy within the layers. In contrast, Nearest Neighbour and Kriging are characterized by a lack of continuity and blurry layer boundaries respectively. Inverse Distance, Natural Neighbour and Inpainting fail to come close to the performance of the other methods. The superior performance of SchemaGAN is confirmed through a blind survey, in which SchemaGAN ranked as the top-performing method in 78% of cases according to experts in the field. Results also suggest that SchemaGAN is the least affected method by the location of CPT data along the cross-section. In a real case study, SchemaGAN demonstrates better predictive accuracy for known CPT data than both Nearest Neighbour and Kriging interpolation methods. The future potential lies in refining its performance by considering enhancements such as training with real CPT data, incorporating additional conditional inputs, and exploring larger inputs or specialized databases. All the code related to the project has been made publicly accessible.

In this work, the added value of machine learning (ML) molecular force fields (FF) for the community of molecular simulations is showcased by successfully calculating transport properties of aqueous potassium hydroxide (KOH (aq)). Classical FFs use relatively simple interatomic potentials to simulate the nano scale. These simulations can predict macroscopic properties, such as density, heat of evaporation, viscosity, and self-diffusivity of the modeled materials. However, these FFs struggle to model materials in which more complicated interactions are relevant for the macroscopic behavior. Examples of such interactions are three-body interactions and chemical reactions. Quantum scale simulation methods are able to compute properties of materials in which these challenging interactions occur, although these methods are limited in length and time scales that can be modeled with realistic computational costs. Transport properties, such as viscosity, self-diffusivity and electric conductivity need these larger length and time scales to be determined accurately. ML can be used for a multi scale approach, bridging the gap between the quantum and the nano scale by training coefficients of general interatomic potentials. This provides the possibility of reaching the time and length scales of traditional molecular simulations with the accuracy of quantum mechanic models. KOH (aq) is selected to highlight the prospects of these multi scale techniques, as the self-diffusion of OH- in this electrolyte is dominated by proton transfer reactions, which has not been modeled successfully with classical FFs.

Results of structure properties produced with ab initio molecular dynamics (AIMD, at quantum scale) simulations are compared with machine learning molecular dynamics (MLMD, at multi scale) simulations. There are no significant differences in the calculated shortest typical atomic distances and coordination numbers for both KOH (aq) and pure water systems. The determined transport properties are in the same order of magnitude as experimental results, although the calculated viscosity is overestimated and the self-diffusion of H2O and K+ are underestimated. This is because the system is simulated at a higher than experimental density and hydrogen bonding is overestimated with the selected quantum mechanics model. The proton transfer reactions are captured in the MLMD simulations, calculating the enhanced self-diffusion of OH- to be (6±2)e-9 m squared per second, which matches experimental results at infinite dilution.

Research pertaining to end-use water analysis plays a pivotal role in enabling local communities to enhance their management of pipelines, water resources, and associated policies. Nowadays, various end-use models have been developed based on diverse databases and measurements. Nonetheless, a predominant drawback prevalent in most of these models is their limited spatial scope and sluggish computational speed. This thesis endeavors to address these challenges through the proposition of a generative adversarial network (GAN) based stochastic end-use demand model. The SIMDEUM model, a stochastic end-use model, was first published in 2010. Since its inception, it has garnered substantial recognition and validation from numerous researchers. Within this thesis, the GAN model utilizes SIMDEUM as the training set and undergoes validation utilizing a comprehensive measurement dataset, encompassing over 1000 households from the Netherlands and the United States. Remarkably, the GAN model attains an error rate of 12% for end uses, coupled with an R2 value exceeding 0.8 for the overall model. In contrast to SIMDEUM, the GAN model significantly enhances computational speed by more than 500%. Furthermore, the GAN model can be tailored to specific requirements and seamlessly processes raw data.It is concluded that the GAN-based stochastic water use model presented in this thesis adeptly simulates end-use water demand.

The edge flow reconstruction task improves the integrity and accuracy of edge flow data by recovering corrupted or incomplete signals. This can be solved by a regularized optimization problem, and the corresponding regularizers are chosen based on prior knowledge. However, obtaining prior information is challenging in some fields. Thus, we consider exploiting the learning ability of neural networks to acquire prior knowledge. In this paper, we propose a new optimization problem for the simplicial edge flow reconstruction task, the simplicial ElasticNet, which is a regularized optimization problem that combines the advantages of the L1 and L2 norm. It is solved iteratively by the multi-block ADMM algorithm, and the convergence conditions are illustrated. By unrolling the simplicial ElasticNet's iterative steps, we propose a neural network with high interpretability and low requirement for the number of training data for the reconstruction task of simplicial edge flows. The unrolling network replaces the fixed parameters in the iterative algorithm with the learnable weights in the neural networks, thus exploiting the neural network's learning capability while preserving the iterative algorithm's interpretability. The core component of this unrolling network is simplicial convolutional filters with learnable weights to aggregate information from the edge flow neighbors, thus enhancing the learning and expressive ability of the network. We conduct extensive experiments on real-world and artificial datasets to validate the proposed approach. It is demonstrated that the simplicial unrolling network is significantly more advantageous than the traditional iterative algorithms and standard non-model-based neural networks in the case of limited training data.

The Alps are experiencing a gradual reduction in snow cover due to rising temperatures, impacting the landscape and dependent ecosystems. While several models have been developed to study snow cover in the region, there is a lack of visual representations. This research employs a Conditional Generative Adversarial Network (cGAN) to generate a multispectral Landsat-8 image of the Alps using environmental data inputs. The study utilizes elevation, monthly precipitation, and monthly temperature data from November to March to produce an end-of-winter Landsat-8 image. The resulting multispectral image is then used to calculate the Normalised Difference Snow Index (NDSI) and determine the snow extent by counting pixels with NDSI values above 0.4. Two climate scenarios are considered, and the generated images are compared to actual Landsat-8 imagery. Findings indicate that while the generated imagery closely resembles the real imagery, the snow extent is generally underestimated in the current model configuration, and the snow reflectance is consistently overestimated across all training steps of the cGAN. Additionally, it is observed that increasing the spatial distance between the training and testing locations leads to increased error in the results. The thesis demonstrates the feasibility of using a cGAN to generate snow extent, but suggests that enhancements to the training dataset and cGAN architecture are necessary for improved accuracy. By leveraging the cGAN, future climate scenarios can be visually represented through multispectral imagery, enabling a more detailed understanding of potential future landscapes under different climate conditions.

The use of deep learning in global weather forecasting has shown significant promise in improving both forecasting accuracy and speed. Traditional numerical weather prediction models have gradually improved forecasting skills but at the cost of increased computational complexity. In contrast, new deep learning models, trained directly on reanalysis data, have demonstrated significant gains in forecasting accuracy, achieving competitive levels of performance.

However, the potential of deep learning in predicting spatiotemporal chaotic systems, such as weather patterns, remains unexplored. To address this gap, we investigate the efficacy of a data-driven Fourier neural operator Markovian forecaster to replicate the intrinsic predictability and the characteristic Lyapunov spectrum of the Kuramoto-Sivashinsky system.
FNO reproduces intrinsic predictability and precisely estimates the characteristic Lyapunov spectrum, even with a small dataset.
They cannot represent one of the invariant symmetries, a zero characteristic Lyapunov exponent in the spectrum.

Our findings suggest that deep learning can not only enhance the speed and accuracy of traditional numerical forecasting models but also replicate the weather's chaotic nature.
This has significant implications for generating large ensembles and improving overall probabilistic forecasts.

The research is limited to a deterministic system defined on a single process time scale surrogated by FNO.
The future merits a similar study to investigate if FNO models can perform similarly on larger, stochastic, coupled, and/or multiple time-scale spatiotemporal chaotic systems.

Machine learning models offer promising potential in precipitation nowcasting. However, a common issue faced by many of these models is the tendency to produce blurry precipitation nowcasts, which are unrealistic. Previous research on the deep learning model - TrajGRU (Shi et al., 2017) indicated that data imbalance in radar images and the double-penalty effect of pixel-wise loss functions are underlying causes for this blurriness.
In this thesis, we continue to explore various approaches to improve the predictive performance of TrajGRU. Our research has first investigated spatially and temporally enhanced loss functions to address the two remaining issues: data imbalance and double penalty. The second part of our research focuses on manifold optimizations within the model network, such as incorporating additional model inputs or increasing batch size, to understand the model’s limitations.
Our results reveal that enhanced loss functions did not lead to predictive improvements and even resulted in undesired checkerboard patterns. Changes to the model network make a difference in the image sharpness and predictive rain evolution. Our visual analysis indicates that a larger batch size generates sharper rain field edges; predictions by using multiple parameter groups exhibit more rain dynamics. The incorporation with other transformed datasets introduces finer structures within rain fields. Although the blurriness has not been completely resolved, our study recommended future work can continue exploring the optimization in the TrajGRU network.

Accurate short term rain predictions are important for flood early warning systems, emergency services, energy management and other services that that make weather dependent decisions. Recently introduced machine learning models suffer from blurry and unrealistic predictions at longer lead times, causing poor performance on the rarer heavy rainfall events. The objective of this research was to explore how the loss function in a recurrent, convolutional neural network (TrajGRU, X. Shi et al. (2017)) can be modified, to get sharper and more realistic predictions, without worsening its performance. Six perceptive loss functions, which should better represent how an image is perceived, are thoroughly analysed to understand the reasons behind their functioning in the context of precipitation nowcasting. It was shown that these perceptive losses can lead to an improvement in sharpness for the first lead time, but that the blurriness for longer lead times arises due to the imbalance in the dataset and the uncertainty of the model with respect to the location of the rain. This thesis gets concluded with recommendations on how to deal with these two problems.

Rain-fed agriculture is the main source of food in Ghana therefore improving quantitative rainfall estimates is essential for local farmers to predict crop growth using vegetation models. Rainfall dynamics in the tropics is an ongoing topic of research due to their complexity and sub-grid precipitation variability. At the same time, tropical areas such as Ghana are the most affected by a lack of proper rain gauge network coverage. Traditional methods rely on empirical assumptions and statistical theories that require continuous calibration and still struggle to accurately represent local variability. The aim of this paper is to demonstrate the potential of a Deep Learning (DL) approach using bi-spectral information of water vapor imagery (WV) and thermal infrared (TIR) as a starting point to develop an effective alternative to the Cold Cloud Duration (CCD)
method which is a widely applied statistical technique by satellite rainfall products like Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) and Tropical Applications of
Meteorology using SATellite data (TAMSAT) that are specifically designed for Africa.
WV inhibition of low-level features assures the correct depiction of strong convective motions usually related to heavy rainfall which is crucial in tropical areas where convective rainfall is dominant. The addition of WV 7.3m is particularly beneficial in North Ghana as tropical systems
are advecting dry air from the nearby Sahara desert creating discontinuities in precipitation events which translates into dry intrusions and dry slots seen in the images of the WV channel.
The developed Deep learning model showed strong performances in binary classification where it outperformed IMERG-Final false alarms count resulting in lower rainfall overestimation (FBias < 2.0), although further research is needed to overcome the very poor relation between GEO-IR images and actual rainfall estimates at the surface.

This research aims to analyse the sensitivity of the YOLOv5 object detection algorithm to current issues related to the tracking of icebergs in SAR imagery. To this end a sensitivity study was done on (1) the sensitivity of the algorithm to variations in input image resolution, (2) the sensitivity of the algorithm to variations in contrast between an iceberg and its surroundings and (3) the sensitivity of the algorithm to variations in icebergs size. The results show that the algorithm is very robust against variations in contrast between iceberg and surroundings, but is significantly sensitive to iceberg size. Furthermore, it seems that only by using high resolution images, the spatial features of icebergs can be well distinguished from features of other objects in the ocean. The YOLOv5 algorithm thus shows great potential for iceberg detection applications, but it should be explored if the
current sensitivity to size can be overcome if a more evenly distributed training dataset is used. On top of this, it should be noted that this research only serves as an exploratory analysis on the application of the algorithm and it should thus still be explored if our results based on augmented data, also apply on real data.

This work provides a water balance-based approach to improve understanding of drought trends by calculating root storage deficits in the United States. To do this, data is ultimately compiled from 1125 watersheds for the years 2001 to 2016. The root storage deficit is then determined using iteratively optimized transpiration values minus the water available to plants and compared to the latent heat of the FLUXCOM energy fluxes data. The stimulus for this work is a better understanding of desiccation processes. Answers to this are given only indirectly. For example, no concrete consequences can be derived from the calculated root storage deficits. Further-more, it turns out that the water cycle-based approach provides clearly interpretable results only for catchments with sufficient water supply. Thus, increasingly arid catchments do not satisfy the general conditions required for the calculations. Catchments in which the root storage deficit exceeds one meter are declared as ever-growing deficit. In these catchments an increased dry-ing process is to be assumed. In all other catchments, temporally congruent developments are shown for both applied methods. Due to the increased dependence on incoming precipitation, calculations performed with the CHIRPS dataset were repeated with the CRUNCEP dataset. This turns out to be generally smaller but confirms the drying processes suggested by the CHIRPS dataset. In contrast, the extents of drying suggested for both methods differ by up to 972 mm for the CHIRPS precipitation dataset, with holistically calculated root storage deficit maxima differing by an average of 178mm. This work is nonetheless useful because otherwise only grid dataset-based and thus statistically adjusted approaches are common in climatological science. Thus, the water cycle-based approach provides a concrete application of the raster values to point measurements.

Roads in modern cities facilitate different types of users, including car drivers, cyclists, and pedestrians. These different users often have a designated section of the road to operate on. Road management, e.g., by municipalities, needs to take this sectioning into account, preferably in an efficient way. Mobile laser scanning (MLS) point clouds provide accurate and dense three-dimensional (3D) measurements of road scenes, showing strong mapping capabilities, although their massive data volume and lack of structure still bring difficulties in automatic processing. Methods for the automatic classification of road surface types are still largely lacking, and the existing methodology did not consider the potential of MLS point clouds yet. In recent years, point cloud understanding through deep neural networks has achieved breakthroughs. However, perceiving large-scale point clouds by deep learning depends on aggregating local features and progressive downsampling, to extract rich contextual information. As a consequence, low-level features that reveal details in point clouds may not be well preserved, possibly resulting in ambiguous delineation in point cloud classification. For road mapping, inaccurate classification of points on road boundaries hinders the generation of high-quality map products. Some existing deep learning methods propose to mitigate the fuzzy classification near boundaries, either by utilizing refinement for network predictions, or by indirectly modifying neighboring weights when summarizing local information. Approaches to achieving a satisfying overall performance, while maintaining accurate delineation, still need investigation. In this study, we propose a novel approach for road type classification of MLS point clouds in dense urban areas based on a deep neural network. We follow the main architecture of RandLA-Net, a point-wise neural network designed for large-scale point cloud processing. To alleviate the ambiguous delineation of point cloud classification, we propose two strategies. The first is to refine predictions of RandLA-Net by conditional random field. In the second strategy, we incorporate boundary constraints in the network by introducing a novel distance label for each road surface point to represent the distance to its closest boundary. The distance prediction task is combined with road type classification by adding another branch in RandLA-Net to formulate multi-task learning. Through experiments, we show that 3D point cloud semantic segmentation by deep learning is applicable for road type classification. Also, the multi-task learning strategy is verified to be more effective in improving the delineation performance. Using MLS point clouds acquired from 5 German cities (Hamburg, Delmenhorst, Bremerhaven, Hannover, and Oldenburg), we classify road points separately into different usages (sidewalk, cycling path, rail track, parking area, motorway, green area, and island without traffic) and materials (cobblestone, asphalt, plates, unpaved, and railway). When adopting Hannover and Oldenburg for testing, and the other three cities for training, we obtain a mean intersection over union of 46.1% for usage type and 52.0% for material type with the multi-task learning strategy and input features (x, y, z, R, G, B, intensity), outperforming the original RandLA-Net by approximately 4%. Moreover, from the point cloud classification results, we achieve lightweight polygon representations of road objects in different types through post-processing, which is demonstrated to perform better than an image semantic segmentation-based solution quantitatively and qualitatively.