Crewed or autonomous flight and lander missions require accurate hazard detection and avoidance (HDA) algorithms. Unsuccessful landing attempts on hazardous, unknown terrains such as on asteroids using unreliable algorithms can jeopardize mission vehicle safety and damage its physical integrity. In this thesis, the performance of current state-of-the-art algorithms like Least Median Square (LMedSq) plane-fitting, that does not rely on machine learning is compared against the performance of deep neural networks (DNN) such as UNET and Convolution-Deconvolution (ConvDeconv). Due to a lack of sufficient asteroid surface data for model training and testing, an environment generator was created where rocks, craters and slope layer were separately generated and added together to produce a final DEM for input. Various tests were carried out on these methods to investigate their performance in terms of run-time and accuracy. The results show that, compared to the conventional approaches, DNNs are more accurate and require shorter run-times to detect hazards. However, memory requirement vs. accuracy trade-off would be a deciding factor when determining which DNN would better suit mission requirements.