Change detection (CD) of Land Use and Land Cover (LULC) is a valuable field of study for resource monitoring and land planning. The progress of remote sensing and the availability of satellite data allowed for the development of LULC monitoring through bi-temporal orthoimages processing. However, the annotation of changes based on geospatial data is a time-consuming and laborious task. Thus, the automation of this process is in current research and development through the implementation of CD models. Change Vector Analysis (CVA) emerged as the first principal model for CD, consisting of pixel-based image differentiation. The main drawback of pixel-based techniques like CVA is their sensitivity to radiometric variations and misregistration. Recently, supervised deep-learning-based methods achieved great success in the CD task due to their ability to extract deep features in a pair of images and their robustness to inner-class variance. Unsupervised deep-learning-based models still are significantly less performant than supervised ones but need to be developed in order to avoid being dependent on the laborious production of labeled datasets for training. In addition, the lack of labeled CD datasets was pointed out in previous papers. Several binary change datasets filled this gap, but the lack of semantic datasets, containing information about the nature of changes, still persists. These datasets are crucial for the training and evaluation of CD models. We tackle the enunciated issues in two parts. We propose a very high resolution (VHR) 0.25 m/px semantic CD dataset comprised of 127 paired tiles of 4000×4000 pixels covering 1 km2. All the geospatial data utilized for the dataset is provided by Swisstopo and the changes are reported along Swiss railway tracks. In addition, the labeling is partly automated by a script processing vectorized data analysis on vectorized geospatial data. It has been assessed that the automated processing was useful and efficient for building change detection. However, manual labeling is still necessary for the other classes of changes and the semantic classification of the changes. The dataset is proposed as a benchmark for two unsupervised models: a CVA acting as a baseline for the evaluation of a promising novel unsupervised CD deep learning model based on reconstruction loss (CDRL). CDRL outperformed CVA when setting appropriate parameters, with an F1-score of 0.24 versus 0.19 respectively. However, we struggled to get the same accuracy as the original authors.