The document describes a project using deep learning to classify satellite imagery patches as land, water, or ice. A convolutional neural network called PolarNet was developed and tested on two datasets of varying sizes. PolarNet V1 achieved 70.24% validation accuracy on the smaller dataset, while PolarNet V2 achieved slightly lower accuracy. Expanding the training set improved performance, with PolarNet V1 achieving 68.33% accuracy on the larger dataset compared to 68.64% for PolarNet V2. Confusion matrices are presented to evaluate model performance.