[Paper Review] LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation

Junjue Wang, Zhuo Zheng+3 more|arXiv (Cornell University)|Oct 17, 2021

Domain Adaptation and Few-Shot LearningComputer Science55 references161 citations

TL;DR

LoveDA introduces a high-resolution remote sensing land-cover dataset with urban and rural domains to study semantic segmentation and unsupervised domain adaptation, benchmarked across multiple models and UDA methods.

ABSTRACT

Deep learning approaches have shown promising results in remote sensing high spatial resolution (HSR) land-cover mapping. However, urban and rural scenes can show completely different geographical landscapes, and the inadequate generalizability of these algorithms hinders city-level or national-level mapping. Most of the existing HSR land-cover datasets mainly promote the research of learning semantic representation, thereby ignoring the model transferability. In this paper, we introduce the Land-cOVEr Domain Adaptive semantic segmentation (LoveDA) dataset to advance semantic and transferable learning. The LoveDA dataset contains 5987 HSR images with 166768 annotated objects from three different cities. Compared to the existing datasets, the LoveDA dataset encompasses two domains (urban and rural), which brings considerable challenges due to the: 1) multi-scale objects; 2) complex background samples; and 3) inconsistent class distributions. The LoveDA dataset is suitable for both land-cover semantic segmentation and unsupervised domain adaptation (UDA) tasks. Accordingly, we benchmarked the LoveDA dataset on eleven semantic segmentation methods and eight UDA methods. Some exploratory studies including multi-scale architectures and strategies, additional background supervision, and pseudo-label analysis were also carried out to address these challenges. The code and data are available at https://github.com/Junjue-Wang/LoveDA.

Motivation & Objective

• Introduce a challenging HSR land-cover dataset with urban and rural domains to study domain transferability.
• Provide benchmarks for semantic segmentation and unsupervised domain adaptation on LoveDA.
• Analyze challenges such as multi-scale objects, complex backgrounds, and inconsistent class distributions to guide future methods.

Proposed method

• Collect 0.3 m resolution images from 3 Chinese cities across 18 areas and build two-domain (urban vs. rural) splits.
• Annotate at pixel level with 8 land-cover classes and ensure spatially independent train/val/test splits.
• Benchmark 11 semantic segmentation models and 8 UDA methods on LoveDA to establish baselines.
• Investigate multi-scale architectures and strategies, additional background supervision, and pseudo-label analysis to address challenges.
• Provide baseline training details, including data augmentation, optimization, and backbone pretraining information.
• Make code and data publicly available for reproducibility.

Experimental results

Research questions

• RQ1 How well do state-of-the-art segmentation models generalize across urban and rural domains in LoveDA?
• RQ2 How effective are adversarial and self-training UDA approaches on cross-domain land-cover segmentation with LoveDA?
• RQ3 What impact do multi-scale strategies and background supervision have on segmentation accuracy in LoveDA?
• RQ4 How does pseudo-label analysis (e.g., CBST) affect UDA performance under inconsistent class distributions?

Key findings

• LoveDA enables cross-domain evaluation by containing urban and rural domains with inconsistent class distributions.
• HRNet achieves strong segmentation performance among tested architectures and benefits from multi-scale fusion.
• Multi-scale augmentation during training and testing (MSTrTe) significantly improves mIoU across methods.
• Background supervision improves segmentation, particularly for the challenging background class.
• In UDA, self-training methods generally outperform adversarial methods on LoveDA due to class distribution mismatch across domains.
• CBST pseudo-labeling shows robustness to target class proportion and can substantially boost cross-domain performance under proper tuning.