Accessing velocity data in the Gulf of Mexico is critical for environmental conservation and predicting debris and oil spill movements. This data can provide valuable insights for cleaning the ocean and mitigating marine pollution. Traditionally, researchers have relied on physics models to reconstruct and predict velocity fields at desired spatial and temporal resolutions. However, obtaining this data is not only computationally expensive but also error-prone. While accurate measurements can be obtained using ocean drifters, their sparsity necessitates extensive extrapolation to create comprehensive velocity fields. We propose applying a deep learning model called Physics-Informed Neural Networks to reconstruct ocean surface velocity fields using sparse measurements obtained from drifters. With data from 200 drifters, we successfully reconstructed the surface velocity field in the Gulf of Mexico, achieving a Correlation Coefficient of more than 0.91. Notably, this performance surpasses that of classical data extrapolation methods, including Inverse Distance Weighted and Universal Kriging algorithms.