In recent years, space technologies have become increasingly integral to various industrial and scientific domains, emphasizing the critical role of remote sensing and Earth observation in guiding decision-making processes. This doctoral thesis aims to enhance the understanding of remote sensing applications by leveraging Convolutional Neural Networks (CNNs) across different convolution degrees (1D, 2D, 3D, and 4D). The primary research question explores how the selection of convolution dimensionality influences CNN performance across diverse tasks. 1D, 2D, 3D, and 4D convolutions are selected for recognizing spectral, spatial, and temporal features of data. A comprehensive systematic literature review was conducted, analyzing publications related to 1D, 2D, 3D, and 4D-CNNs, addressing their domains of application, machine learning techniques used, and corresponding datasets. Findings are synthesized in the form of ontology for CNN architecture selection providing a structured framework, which is further evaluated in four case studies. The 1DCNN proved to be an appropriate method for extracting spectral information sufficient for predicting point parameters such as Secchi disk depth. The 2D-CNN demonstrated success in extracting spatial features from Sentinel-2 images for road detection. The 3D-CNN with its ability to extract spectral-spatial features was successfully applied to the problem of fire propagation potential detection. The 4D-CNN combined temporal with spectral and spatial features, which proved extremely useful in detecting burned areas. Finally, a comparison of all four convolutional neural networks for the task of semantic segmentation of land cover was performed, where the 3D-CNN stood out as the best. Results of all case studies are aligned with the central hypothesis that selecting appropriate convolution degree significantly enhances deep learning performance in environmental monitoring applications. These findings contribute to understanding of convolution strategies in remote sensing, offering insights for researchers and practitioners seeking to optimize neural network architectures for Earth observation technologies.