Land cover mapping is an important part of resource management, planning, and economic predictions. Improvements in remote sensing, machine learning, image processing, and object based image analysis (OBIA) has made the process of identifying land cover types increasingly faster and reliable but these advances are unable to utilize the amount of information encompassed within ultra-high (sub-meter) resolution imagery. Previously, users have typically reduced the resolution of imagery in an attempt to more closely represent the interpretation or object scale in an image and rid the image of any extraneous information within the image that may cause the OBIA process to identify too small of objects when performing semi-automated delineation of objects based on an images’ properties (Mas et al., 2015; Eiesank et al., 2014; Hu et al., 2010). There have been few known attempts to try and maximize this detailed information in high resolution imagery using advanced textural components. In this study we try to circumnavigate the inherent problems associated with high resolution imagery by combining well researched data transformations that aid the OBIA process with a seldom used texture transformation in Geographic Object Based Image Analyses (GEOBIA) known as the Gabor Transform and the hierarchal organization of landscapes. We will observe the difference made in segmentation and classification accuracy of a random forest classifier when we fuse a Gabor transformed image to a Normalized Difference Vegetation Index (NDVI), high resolution multi-spectral imagery (RGB and NIR) and Light Detection and Ranging (LiDAR) derived canopy height model (CHM) within a riparian area in Southeast Iowa. Additionally, we will observe the effects on classification accuracy when adding multi-scale land cover data to objects. Both, the addition of hierarchical information and Gabor textural information, could aid the GEOBIA process in delineating and classifying the same objects that human experts would delineate within this riparian landscape.

Cette thèse est consacrée aux forêts aléatoires, une méthode d'apprentissage non paramétrique introduite par Breiman en 2001. Très répandues dans le monde des applications, les forêts aléatoires possèdent de bonnes performances et permettent de traiter efficacement de grands volumes de données. Cependant, la théorie des forêts ne permet pas d'expliquer à ce jour l'ensemble des bonnes propriétés de l'algorithme. Après avoir dressé un état de l'art des résultats théoriques existants, nous nous intéressons en premier lieu au lien entre les forêts infinies (analysées en théorie) et les forêts finies (utilisées en pratique). Nous proposons en particulier une manière de choisir le nombre d'arbres pour que les erreurs des forêts finies et infinies soient proches. D'autre part, nous étudions les forêts quantiles, un type d'algorithme proche des forêts de Breiman. Dans ce cadre, nous démontrons l'intérêt d'agréger des arbres : même si chaque arbre de la forêt quantile est inconsistant, grâce à un sous-échantillonnage adapté, la forêt quantile est consistante. Dans un deuxième temps, nous prouvons que les forêts aléatoires sont naturellement liées à des estimateurs à noyau que nous explicitons. Des bornes sur la vitesse de convergence de ces estimateurs sont également établies. Nous démontrons, dans une troisième approche, deux théorèmes sur la consistance des forêts de Breiman élaguées et complètement développées. Dans ce dernier cas, nous soulignons, comme pour les forêts quantiles, l'importance du sous-échantillonnage dans la consistance de la forêt. Enfin, nous présentons un travail indépendant portant sur l'estimation de la toxicité de certains composés chimiques. / This is devoted to a nonparametric estimation method called random forests, introduced by Breiman in 2001. Extensively used in a variety of areas, random forests exhibit good empirical performance and can handle massive data sets. However, the mathematical forces driving the algorithm remain largely unknown. After reviewing theoretical literature, we focus on the link between infinite forests (theoretically analyzed) and finite forests (used in practice) aiming at narrowing the gap between theory and practice. In particular, we propose a way to select the number of trees such that the errors of finite and infinite forests are similar. On the other hand, we study quantile forests, a type of algorithms close in spirit to Breiman's forests. In this context, we prove the benefit of trees aggregation: while each tree of quantile forest is not consistent, with a proper subsampling step, the forest is. Next, we show the connection between forests and some particular kernel estimates, which can be made explicit in some cases. We also establish upper bounds on the rate of convergence for these kernel estimates. Then we demonstrate two theorems on the consistency of both pruned and unpruned Breiman forests. We stress the importance of subsampling to demonstrate the consistency of the unpruned Breiman's forests. At last, we present the results of a Dreamchallenge whose goal was to predict the toxicity of several compounds for several patients based on their genetic profile.

In this thesis, two different machine learning models have been applied on road traffic data from two large cities in Sweden: Gothenburg and Stockholm. The models have been evaluated with regard to classification of street types in urban environments. When planning and developing road traffic systems it is important that there is reliable knowledge about the traffic system. The amount of available traffic data from urban areas is growing and to gain insights about historical, current and future traffic patterns the data can be used for traffic analysis. By training machine learning models that are able to predict what type of street a measuring location belongs to, a classification can be made based on historical data. In this thesis, the performance of two different machine learning models are presented and evaluated when street types are predicted and classified. The algorithms used for the classification were K-Nearest Neighbor and Random Forest which were applied to different combinations of attributes. This was done in order to identify which attributes that lead to the optimal classification of street types in Gothenburg. For training the algorithms the dataset consisted of traffic data collected in Gothenburg. The final model was applied on the traffic data in Stockholm and hence the prediction of street types in that area were obtained. The results of this study show that a combination of all tested attributes leads to the highest accuracy and the model that obtained these results was Random Forest. Even though there are differences between topography and size of the two cities, the study leads to relevant insights about traffic patterns in Stockholm.

Friction stir welding is an advanced welding process that is being investigated for use in many different industries. One area that has been investigated for its application is in healing critical nuclear reactor components that are developing cracks. However, friction stir welding is a complicated process and it is difficult to predict what the final properties of a set of welding parameters will be. This thesis sets forth a method using finite element analysis and a random forest model to accurately predict hardness in the welding nugget after processing. The finite element analysis code used and ALE formulation that enabled an Eulerian approach to modeling. Hardness is used as the property to estimate because of its relationship to tensile strength and grain size. The input parameters to the random forest model are temperature, cooling rate, strain rate, and RPM. Two welding parameter sets were used to train the model. The method was found to have a high level of accuracy as measured by R^2, but had greater difficulty in predicting the parameter set with higher RPM.

Diploma thesis deals with the land cover classification in Sidama region of Ethiopia and 2 kebeles, Chancho and Dangora Morocho. High resolution Sentinel-2 and very high resolution PlanetScope satellite images are used. The development of the classification algorithm is done in the Google Earth Engine cloud based environment. Ten combinations of the 4 most important parameters of the Random Forest classification method are tested. The defined legend contains 8 land cover classes, namely built-up, crops, grassland/pasture, forest, scrubland, bareland, wetland and water body. The training dataset is collected in the field during the fall 2020. The classification results of the two data types at two scales are compared. The highest overall accuracy for land cover classification of Sidama region came out to be 84.1% and kappa index of 0.797, with Random Forest method parameters of 100 trees, 4 spectral bands entering each tree, value of 1 for leaf population and 40% of training data used for each tree. For the land cover classification of Chancho and Dangora Morocho kebele with the same method settings, the overall accuracy came out to be 66.00 and 73.73% and kappa index of 0.545 and 0.601. For the classification of Chancho kebele, a different combination of parameters (80, 3, 1, 0.4) worked out better...

The main goal of this paper is the identification of comments on internet websites. The theoretical part is focused on artificial intelligence, mainly classifiers are described there. The practical part deals with creation of training database, which is formed by using generators of features. A generated feature might be for example a title of the HTML element where the comment is. The training database is created by input of classifiers. The result of this paper is testing classifiers in the RapidMiner program.

The thesis deals with the comparison of conventional and heuristic methods in data mining used for binary classification. In the theoretical part, four different models are described. Model classification is demonstrated on simple examples. In the practical part, models are compared on real data. This part also consists of data cleaning, outliers removal, two different transformations and dimension reduction. In the last part methods used to quality testing of models are described.

The presence of amyloid plaques in the hippocampus highlights the incidence of Alzheimer’s disease. Manual segmentation of amyloid plaques is very time consuming and increases the time that can be used to monitor the distribution of amyloid plaques. Distribution carries significant information about disease progression and the impact of potential therapy. The automatic or semi-automatic segmentation method can lead to significant savings in the time which are required when the disease has rapid progression. The description of amyloid plaques and the computed tomography are included in this work. In this diploma thesis are three implemented algorithms, two of them are based on published articles and one’s own methodological solution. The conclusion of the thesis is a quantitative evaluation of the accuracy of implemented segmentation procedures.

This thesis deals with a classification of glioma grade in high and low aggressive tumours and overall survival prediction based on magnetic resonance imaging. Data used in this work is from BRATS challenge 2019 and each set contains information from 4 weighting sequences of MRI. Thesis is implemented in PYTHON programming language and Jupyter Notebooks environment. Software PyRadiomics is used for calculation of image features. Goal of this work is to determine best tumour region and weighting sequence for calculation of image features and consequently select set of features that are the best ones for classification of tumour grade and survival prediction. Part of thesis is dedicated to survival prediction using set of statistical tests, specifically Cox regression

Main goal of classifying proteins into families is to understand structural, functional and evolutionary relationships between individual proteins, which are not easily deducible from available data. Since the structure and function of proteins are closely related, determination of function is mainly based on structural properties, that can be obtained relatively easily with current resources. Protein classification is also used in development of special medicines, in the diagnosis of clinical diseases or in personalized healthcare, which means a lot of investment in it. I created a new hierarchical tool for protein classification that achieves better results than some existing solutions. The implementation of the tool was preceded by acquaintance with the properties of proteins, examination of existing classification approaches, creation of an extensive data set, realizing experiments and selection of the final classifiers of the hierarchical tool.