Most new satellite platforms, from evolved cubesats to sophisticated large spacecrafts, such as those in the Sentinel series, are equipped with high-resolution cameras that produce detailed images. These images, if properly utilized, represent crucial sources of information for a wide range of applications. However, the volumes of data generated by these space missions require efficient processing methods, making it essential to develop techniques for automatic analysis and interpretation. The use of advanced object detection techniques, such as those based on deep learning, has proven to be key. Among the most promising techniques is the YOLO (You Only Look Once) model, which enables the rapid detection and segmentation of features in images. YOLOv8, in particular, has shown significant performance improvements, increasing detection consistency and reliability. Thanks to its efficient design and real-time processing capabilities, its architecture proves to be well suited for rapid and precise detection, even on large-scale satellite images. This study explores the application of YOLOv8 for the detection and segmentation of geophysical features (mainly lakes) in high-resolution Sentinel-2 satellite images in Italy. The approach demonstrated high precision and robustness, even under partial visibility conditions due to cloud coverage hiding part of the scene, thanks to the use of extended training (from 60 to 100 epochs). Different metrics have been computed and reported to confirm these findings. YOLOv8 also proved to be efficient, suggesting its possible application onboard the imaging spacecraft itself, opening new possibilities in autonomous operations. Overall, the adoption of these deep learning models for automatic satellite image processing offers great potential, improving data management efficiency.