Scene segmentation of remotely sensed images with data augmentation using U-net++

Abstract

Deep learning is the current advanced solution for remote sensing segmentation. Massive high-quality training datasets are the basic inputs to deep learning networks for solving the segmentation problems. Most of the existing remotely sensed image datasets have low segmentation accuracy due to their coarse spatial resolution and the susceptibility to image noise. Image augmentation is a technical means of effectively solving deep learning trainings in small and/or low-quality training datasets, which has continuously accompanied the development of deep learning and machine vision. Many augmentation techniques and methods have been proposed to enrich and augment the training datasets and to improve the generalization ability of neural networks. Common image augmentation methods are based mainly on image transformations, such as photometric changes, flips, rotations, dithering and blurring. In this paper, the segmentation task of multispectral remote sensing data is validated by augmentation methods. The segmentation accuracy was found to be 96.10%, which is higher than that (92.36%) of the corresponding un-augmented data.