Land use and land cover (LULC) maps are essential tools in various fields, including environmental studies, urban planning, agriculture, disaster management, and policy development. These maps are also critical for estimating spatially dependent parameters through land use regression models. In the context of urban air quality studies, LULC information is particularly valuable, as pollution dispersion is influenced by both the proximity of emission sources and the urban fabric’s ability to trap pollutants under certain meteorological conditions. However, different applications require tailored LULC classifications relevant to specific research questions. This study proposes a methodology for generating customizable LULC maps, allowing users to define classification categories based on their needs. Using Sentinel-2 satellite imagery and building height data, we employ machine learning models to achieve high-resolution (10 m2) LULC classification. Specifically, we compare the performance of the random forest algorithm, and the U-Net deep learning architecture based on standard classification metrics and visual assessment of spatial patterns. Additionally, we evaluate our approach against other state-of-the-art segmentation models, including PSPNet, MANet, and Segformer, as well as variations with pretrained encoder architectures. The results demonstrate that our U-Net-based model is particularly effective in capturing spatial patterns, making it a promising tool for LULC classification in various applications.