Cracks are an outcrop's stress fingerprint etched by a region's tectonic forces and weathering patterns. To deduce the stress history, the cracks of outcrops must be detected and then interpreted by the structural geologist. This task can be expedited by using camera-equipped drones to take many high-resolution photographs that can used to identify the crack distributions. The problem with detecting cracks in petabytes of digital images is that it demands a fast computerized method for highly accurate detection in the presence of noise such as shadows, vegetation, and linear discolorations due to weathering and deterministic edge detectors will typically fail. To mitigate this problem, we develop a modified U-Net architecture with transfer learning to detect cracks with an accuracy over 97.5%. We train this crack detector on labeled drone photos from large sandstone massifs in Saudi Arabia. We also found that this detector with transfer training can be transformed into a semi-universal crack detector that can detect lineaments in orbiter photos of Martian channels and images of volcanic outcrops at a failed dam site in Idaho. We believe our semi-universal crack detector can accurately detect cracks in photos of many different geological environments.