Self-supervised fully-convolutional neural networks for segmentation in the object-based image analysis workflow

Abstract

The collection of object-based image analysis literature prescribes a particular workflow for achieving semantic segmentation in small satellite image data sets, relying on non-learned algorithmic segmentation techniques for object-proposal. This research proposes a novel, learned approach to object-proposal in object-based image analysis using fully convolutional neural networks trained with self-supervised learning and synthetic data. Fully-convolutional neural networks are a state-of-the-art approach to segmentation but have limited spatial reasoning and receptive fields, and they also require large amounts of semantically labelled data for end-to-end training. This research shows that using a fully convolutional neural network within object-based image anal-ysis can effectively incorporate learning into the workflow and alleviate concerns regarding neural networks generalising from synthetic data.