Transfer Learning for Land Cover Semantic Segmentation

Abstract

In this paper, we describe a transfer learning based semantic segmentation method for generating land cover maps from low quality satellite images. We use level 16 semantic segmentation maps to learn a baseline segmentation model. We compare combined training with other source datasets from different sources in supervised and semi-supervised transfer learning settings. Experiments show that using transfer learning improves recognition performance from 60.2% to 63.6% miou in rural areas and 79.6 % to 92.5 % miou in urban settings. Observations indicate that transfer learning is more advantageous when two datasets share a comparable zoom level and are annotated with identical rules; otherwise, treating the data as unlabeled and employing semi-supervised learning is more effective.