Combining satellite imagery with machine learning presents an opportunity for assembling globally comprehensive observations of many variables simultaneously. However, current approaches require custom systems, expert knowledge, access to imagery, and extensive computational resources. Moreover, while such measurements provide a potential treasure trove of data for applied economists, imperfect mappings from images to outcomes pose threats to parameter identification in standard regression models. Across two papers, we address both problems. First, we develop a generalized system that enables researchers with basic statistical training to use satellite imagery and machine learning to study any variable visible from space at high resolution, global scale, and low computational cost. We demonstrate the generalizability of our system by constructing high resolution estimates for seven domains across the globe (forest cover, population density, elevation, nighttime lights, household income, road length, and housing prices); we find comparable performance to a state-of-the-art convolutional neural network at a fraction of the cost. Second, we derive the conditions under which error in these remotely sensed predictions biases parameters of interest when they are used in linear regression. We develop and validate a procedure to determine the extent and direction of this bias, and to correct for it.