Detecting and localizing archaeological monuments and historical man-made terrainstructures is essential for learning and preserving our cultural heritage. With theadvancement of laser scanning technology, it is possible to acquire Airborne LaserScanning (ALS) point clouds and create Digital Terrain Models (DTMs), which can beanalyzed by archaeologists for interesting monuments and structures. However, manually inspecting high volumes of DTM data is a time-consuming task. The goal of this research is to utilize deep learning for automated detection of archaeological monuments and historical man-made terrain structures in DTMs. Southern Lower Saxony, i.e. specifically the Harz mining region, was chosen as the study region because a significant number of monuments can be found here. Due to the limited amounts of annotated data and the large amounts of unlabeled data, the focus is on Self Supervised Learning (SSL).SSL involves two steps: pretext and downstream. In the pretext, a model is trained onunlabeled data to learn intrinsic characteristics and interesting patterns in the input.Downstream is the second step, which involves learning patterns from annotated datasets.In the downstream step, the trained model from the pretext step is either used a fixed feature extractor or directly finetuned for supervised tasks on annotated datasets.In this research, convolutional encoder-decoder networks and Generative AdversarialNetworks (GANs) are trained on unlabeled DTM data in the SSL pretext. The trainedmodels are then customized for downstream tasks such as classification, instancesegmentation, and semantic segmentation. They are then finetuned on small amounts of annotated data for detection of archaeological monuments and man-made terrainstructures in the Harz region in Lower Saxony.Experiments are conducted on three different datasets from the Harz region. The firstdataset contains areal structures which includes archaeological monuments such ascharcoal kilns, burial mounds and mining holes and other man-made terrain structuressuch as bomb craters. The second dataset contains linearly elongated structures which includes archaeological monuments such as ditches and hollow ways and other man-made structures such as paths and roads. The third dataset from Harz includes annotated examples of historical stone quarries. Results of the experiments indicate the positive impact of SSL pretraining on the downstream tasks. The best classification algorithm performs similar with and without SSL pretraining. However, for instance and semantic segmentation tasks which are much more complex, SSL pretraining improves the Mean Average Precision (MAP) score by 5.28 % and the Mean Intersection Over Union (MIOU) score by 4.72 %, respectively, on the Harz areal dataset. On the linear structures dataset, the increase in MAP and MIOU scores are 6.18 % and 1.22 %, respectively. Finally, SSL pretraining leads to an increase of 3.02 % in the MIOU score in the stone quarries dataset.