This study presents a self-calibration method for optical Earth observation satellites equipped with matrix sensors. Precise geolocation of each pixel in a satellite image requires accurate modelling of the acquisition geometry, typically achieved through refinement that corrects the geometry using measurements such as correspondences between image pixels and ground coordinates, or between pixels in different images.</p> <p>A critical aspect of this modelling involves the sensor's internal geometry, which defines the line-of-sight (LOS) vector for each pixel in the focal plane. The calibration method proposed in this article eliminates the need for exogenous data (e.g., higher-resolution satellite images or airborne sensor imagery) by relying solely on a set of acquisitions in a specific configuration.</p> <p>The method is evaluated and validated using CO3D imagery. Several evaluation criteria were developed for this purpose, including the reliability of the refinement process, the quality of tie-point intersections, inter-site result comparisons, and alignment with absolute references. This paper does not include the in-orbit performances due to confidentiality agreement.