Satellites can provide cost-effective remotely sensed images. The usefulness of these images is increasing, as the sensors improve with each new satellite. Furthermore, satellite-acquired imagery is in digital form, suggesting the possibility of automating remote sensing tasks such as mapping. To date however, satellite imagery has only yielded two-dimensional planimetric information. With stereo pairs of satellite imagery, the capability for generating the third dimension, height, exists as well. The French SPOT Satellite (Chevrel, Courtois and Weill, 1981) for example, can image high-resolution stereo pairs. Depths are generated from such stereo pairs by stereo matching, normally the task of the human stereo vision system. This paper describes a computer system which automates the process of stereo matching. With this system, the digital equivalent of a contour map (a digital terrain model [DTM]) can be generated automatically directly from digital satellite images. Digital terrain models are quite useful, serving every purpose that a contour map does, and others as well. For example, terrain-dependent parameters such as volumes can be computed easily from DTMs, and DTMs are used in the production of orthophotos. Unfortunately, generating a DTM manually requires many hours. This cost has motivated many attempts to automatically correlate stereo images, without much success. Recent computational vision research has reported some progress in this area (Barnard and Fischler, 1982). This paper describes work that extends these results and applies them to digital satellite images. The result is a system for generating DTMs with significant capability, demonstrated with results from both simulated SPOT images and real Landsat 5 Thematic Mapper (TM) images. DTMs with height accuracies better than 60 metres have been obtained from Landsat images. Before these results are presented, the problem is discussed in detail, and the algorithm which has been developed is described.