Nowadays Earth observation satellites provide forest fire authorities and resource managers with spatial and comprehensive information for fire stabilization and recovery. Burn severity mapping is typically performed by classifying bi-temporal indices (e.g., dNBR, and RdNBR) using thresholds derived from parametric models incorporating field-based measurements. Analysts are currently expending considerable manual effort using prior knowledge and visual inspection to determine burn severity thresholds. In this study, we aim to employ highly automated approaches to provide spatially explicit damage level estimates. We first reorganize a large-scale Landsat-based bi-temporal burn severity assessment dataset (Landsat-BSA) by visual data cleaning based on annotated MTBS data (approximately 1000 major fire events in the United States). Then we apply state-of-the-art deep learning (DL) based methods to map burn severity based on the Landsat-BSA dataset. Experimental results emphasize that multi-class semantic segmentation algorithms can approximate the threshold-based techniques used extensively for burn severity classification. UNet-like models outperform other region-based CNN and Transformer-based models and achieve accurate pixel-wise classification results. Combined with the online hard example mining algorithm to reduce class imbalance issue, Attention UNet achieves the highest mIoU (0.78) and the highest Kappa coefficient close to 0.90. The bi-temporal inputs with ancillary spectral indices work much better than the uni-temporal multispectral inputs. The restructured dataset will be publicly available and create opportunities for further advances in remote sensing and wildfire communities.