Minimum Class Confusion based Transfer for Land Cover Segmentation in Rural and Urban Regions

TL;DR

This paper introduces a transfer learning-based semantic segmentation method for land cover mapping in rural and urban areas, demonstrating significant performance improvements using various transfer learning strategies and highlighting the importance of dataset similarity.

Contribution

It presents a novel transfer learning approach for land cover segmentation that compares supervised, semi-supervised, and unsupervised methods, emphasizing dataset similarity and zoom level.

Findings

Transfer learning improves MIoU by 3.4% in rural and 12.9% in urban regions.

Semi-supervised learning is more effective when datasets differ in zoom level or labeling rules.

HRNet outperforms other building segmentation approaches in multi-class segmentation.

Abstract

Transfer Learning methods are widely used in satellite image segmentation problems and improve performance upon classical supervised learning methods. In this study, we present a semantic segmentation method that allows us to make land cover maps by using transfer learning methods. We compare models trained in low-resolution images with insufficient data for the targeted region or zoom level. In order to boost performance on target data we experiment with models trained with unsupervised, semi-supervised and supervised transfer learning approaches, including satellite images from public datasets and other unlabeled sources. According to experimental results, transfer learning improves segmentation performance 3.4% MIoU (Mean Intersection over Union) in rural regions and 12.9% MIoU in urban regions. We observed that transfer learning is more effective when two datasets share a comparable zoom level and are labeled with identical rules; otherwise, semi-supervised learning is more effective by using the data as unlabeled. In addition, experiments showed that HRNet outperformed building segmentation approaches in multi-class segmentation.