NASA's observational and modeled data products encompass petabytes of earth science data available for analysis, analytics, and exploitation. Unfortunately, these data are underutilized due to vast computational resource requirements; disparate formats, projections, and resolutions that hinder data fusion and integrated analyses; complex and disjoint data access and retrieval protocols; and task-specific and non-reusable code development processes that hinder algorithm sharing and collaboration. Due to these limitations, recent advances in unsupervised machine learning using deep neural nets (DNNs) have not been widely adopted for applications such as pixel-based classification, image preprocessing, feature recognition, and scene understanding. Creare proposes to develop an open-source, standards-based Python software framework that removes major barriers to widespread exploitation of geospatial earth science data. This will be achieved through development of PODPAC (Pipeline for Observational Data Processing, Analysis, and Collaboration), a pipeline-based architecture that (1) enables multi-scale and multi-windowed access, exploration, and integration of available earth science data sets to support both analysis and analytics; (2) automatically accounts for differences in underlying geospatial data formats, projections, and resolutions; (3) simplifies implementation and parallelization of geospatial data processing routines; (4) seamlessly integrates with DNN machine learning frameworks; and (5) unifies access, processing, and sharing of data and algorithms via interfaces to existing NASA repositories. To demonstrate the impact of these innovations, we will use PODPAC to derive an on-demand, high-resolution global soil moisture data product from the Soil Moisture Active/Passive (SMAP) satellite radiometer observational data to support applications in hydrology, agriculture, and humanitarian response missions involving flooding, drought, and water resources.