This study aims to apply artificial intelligence models to satellite-based remote sensing, focusing specifically on the use of raw, unprocessed data. We developed a simulation workflow that generates raw-like data based on high-resolution L1 images and applied deep learning models to object detection and classification tasks. We trained YOLOv11n and YOLOX-S models on the raw and L1 datasets, respectively, to compare their performance and analyzed them using explainability tools. While both models demonstrated similar performance at low confidence levels, the model trained on raw data struggled to identify object boundaries at high confidence levels.
