This paper addresses the problem of accurate estimation of geospatial models from a set of groundwater recharge and discharge maps and from auxiliary remote sensing and terrestrial raster measurements. The motivation for our work is driven by the cost of field measurements, and by the limitations of currently available physics-based modeling techniques that do not include all relevant variables and allow accurate predictions only at coarse spatial scales. The goal is to improve our understanding of the underlying physical phenomena and increase the accuracy of geospatial models - with a combination of remote sensing, field measurements and physics-based modeling. Our approach is to process a set of recharge and discharge maps generated from interpolated sparse field measurements using existing physics-based models, and identify the recharge and discharge map that would be the most suitable for extracting a set of rules between the auxiliary variables of interest and the recharge and discharge map labels. We implemented this approach by ranking recharge and discharge maps using information entropy and mutual information criteria, and then by deriving a set of rules using a machine learning technique, such as the decision tree method. The novelty of our work is in developing a general framework for building geospatial models with the ultimate goal of minimizing cost and maximizing model accuracy. The framework is demonstrated for groundwater recharge and discharge rate models but could be applied to other similar studies, for instance, to understanding hypoxia based on physics-based models and remotely sensed variables. Furthermore, our key contribution is in designing a ranking method for recharge and discharge maps that allows us to analyze multiple plausible recharge and discharge maps with a different number of zones. This JAVA based software package, Spatial Pattern To Learn (SP2Learn), is designed to be user-friendly and universal utilities for pattern learning to improve model predictions from sparse measurements by computer-assisted integration of spatially dense geospatial image data and machine learning of model dependencies. The reliability indices from SP2Learn will improve the traditionally subjective approach to initiating conceptual models by providing objectively quantifiable conceptual bases for further probabilistic and uncertainty analyses. This new approach has been tested using the dataset from Buena Vista Groundwater Basin, a thoroughly understood system in the Central Sand Plains of Wisconsin. This project was supported by the National Center for Supercomputing Applications - Faculty Fellows Program and the Illinois Water Supply Planning Project.