NASA and IBM introduced an open model for analyzing the lunar surface

NASA and IBM have released a specialized foundational model designed to process geospatial data of the Moon. The project aims to automate the analysis of satellite images and other planetary data collected during space missions.

The model is trained on a multimodal dataset that includes information about the topography and composition of the Earth's satellite. During testing, the neural network showed results that surpass standard architectures in tasks of crater detection and ice deposit prospectivity assessment. The use of specialized weights allows the system to more accurately interpret geological features of the relief in low-light conditions.

The development is part of a strategy to implement machine learning in planetary research. The source code and architecture of the model are available to the scientific community, allowing researchers to adapt the solution to specific mapping tasks. The integration of expert knowledge from USRA ensured high accuracy of object classification on the surface.

The implementation of such tools reduces the time it takes to process terabytes of data coming from orbital vehicles. Previously, similar tasks required manual labeling or the use of less efficient computer vision algorithms. Now researchers can automate the search for resources necessary for planning future landings and creating permanent bases on the Moon.