A Review of FITS Data Structures and Scientific Metadata for Astronomical Machine Learning

The Flexible Image Transport System (FITS) is a widely adopted standard for storing, exchanging, and analysing astronomical data. Developed in 1979, formally described in 1981, and endorsed by the International Astronomical Union in 1982, FITS combines multidimensional numerical arrays and tabular data with structured metadata that support scientific interpretation. This narrative review examines the historical development and internal architecture of FITS, including Header/Data Units, data types, and header keywords relevant to photometric calibration, coordinate transformations, and observational provenance. It compares FITS with consumer and general-purpose raster formats, distinguishing the preservation of numerical values from the retention of astronomical metadata and calibration context. Representative applications in major space missions and observatory pipelines illustrate the role of FITS in managing images, spectra, and time-series data. The review also discusses the preparation of FITS-based data for artificial intelligence and machine learning, including calibration, quality filtering, normalization, and metadata integration. Particular attention is given to the distinction between archival data containers and the numerical arrays or tensors used by computational models. FITS provides a well-established foundation for reproducible astronomical analysis, although scientific validity and model performance depend on data quality, calibration, preprocessing, and methodological choices rather than file format alone. The review offers an educational foundation for students and researchers developing computational methods for astronomical classification and physical parameter estimation.