The problem that this thesis deals with is the development of a system for semantic seg- mentation of buildings from satellite images that has the ability of out-of-distribution generalization. The dataset we are working on is high-resolution satellite images that came from the Inria Aerial Image Labeling Dataset competition. The semantic seg- mentation model which we use is based on the resolution pyramid method. For the improvement of out-of-distribution generalization, we use the domain translation and domain adaptation method. The domain adaptation model is based on a refined Cy- cleGAN architecture that is similar to the CyCADA architecture. Initially, we develop software support for learning deep models based on MLOps principles to have the best possible control over the learning process, and then we use it to conduct further expe- riments. The research part of this paper deals with the development and comparison of methods for out-of-distribution generalization. Out-of-distribution data are images of cities whose landscape and architecture differ from the cities where the model was initially trained. In the experiments, we initially measure the covariate shift between different domains. We then try to reduce the covariate shift by using various augmenta- tions aimed at breaking spurious correlations in the data. After that, we use the domain translation method and domain adaptation method to increase the performance of the model on out-of-distribution datasets. The main contribution of this work is the cre- ation of the end-to-end ML pipeline for automatically learning deep models and for process control. Also, we show the results of the experiments with the proposed model for domain adaptation on out-of-distribution domains.