Accurate characterization of the seafloor is crucial in tasks such as scene understanding, environmental health monitoring and oceanic target detection. While traditional approaches for seafloor segmentation and texture classification rely on hand-crafted features, they often fail to capture information across multiple scales. Additionally, textures defining the seafloor often change according to gradients, making aquisition of precise training and evaluation labels difficult. Our current efforts are two-fold: 1.) develop methods which capture seafloor texture information across multiple scales and 2.) utilize imprecise label information to inform feature learning for seafloor segmentation and texture classification. We aim to develop machine learning methods which can classify seafloor textures across multiple resolutions, while also capturing the level of prediction confidence over regions where textures merge.