David Austin This presentation outlines the winning solution for the detection of ships vs icebergs from satellite imagery from the most popular image classification competition to date in the history of Kaggle. Insights gained through unsupervised learning of the raw data led to unique solution architecture choices as well as a custom post-processing pipeline based on clustering of the initial data. The solution architecture demonstrates a unique multi-level ensemble of over 200 customized CNN and VGG architectures which are combined through greedy blending and two-level stacking. Through combination of many weaker learners, the objective loss function can be minimized vs the use of fewer stronger learners while still keeping computational cost reasonable.