Currently, wildfires threaten the environment, human life, and property, and have an increasing tendency in frequency and extent, as reported in the current literature. Hence, providing accurate wildfire spatial information is critical for post-event assessment to measure the losses. Deep learning models have proved their potential in terms of accuracy and computational robustness for many tasks, including semantic segmentation. In addition, remote sensing data are increasingly available and improved in quantity and quality worldwide. In contrast, most current deep learning models in computer vision are developed for close-range photo applications, and a few work well for overhead remote sensing data while preserving the original geospatial information. Therefore, this study configures the ResNet152, EfficientNet-b7, Timm-RegNetX_320, and Timm-RegNetY_320, the largest backbone in their family, with the popular models of U-Net, Feature Pyramid Network (FPN), Pyramid Scene Parsing Network (PSPNet), DeepLabV3, and DeepLabV3Plus to segment wildfire boundaries from Sentinel-2 images. The advancements of Google Earth Engine (GEE) are leveraged to collect Sentinel-2 surface reflectance images and wildfire masks of 30 current events around the world to create the training and validation dataset. All the models were evaluated for accuracy and performance efficiency, and it was found that the structure of LinkNet with EfficientNet-b7 was the most accurate, achieving a precision of around 90%. The other models had precision variations related to their network complexity and backbones; however, they mostly gained an accuracy greater than 80%, improving around 10% compared to the dNBR thresholding method of the previous studies. Wildfire maps are nicely mapped for selected events in the world (France and South Africa) and four recent events in South Korea. The model accuracy, data collection challenges, and model training efficiency are thoroughly discussed.