Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning from Noisy Pseudo-labels for All-Weather Land Cover Mapping

Wang Liu, Zhiyu Wang, Xin Guo, Puhong Duan, Xudong Kang, Shutao Li

TL;DR

This work tackles all-weather land-cover mapping with SAR by addressing noisy pseudo-labels derived from optical segmentation. It introduces a two-stage approach: (1) domain-adaptive optical segmentation with Resolution Alignment Augmentation (RAA) to generate reliable pseudo-labels, and (2) SAR segmentation trained with a noise-robust objective that combines standard cross-entropy and a symmetric cross-entropy loss, aided by a DACS-style self-training framework and training tricks. Key contributions include the image-level alignment via RAA, a threshold-free pseudo-label weighting, and the LCE+LSCE objective for SAR; the method achieves first place in the GRSS 2025 Data Fusion Contest and outperforms official baselines by several percent in mean IoU. The approach has practical impact for robust all-weather land-cover mapping using SAR, and the authors suggest extending with vision foundation models for SAR encoders in future work.

Abstract

Semantic segmentation of SAR images has garnered significant attention in remote sensing due to the immunity of SAR sensors to cloudy weather and light conditions. Nevertheless, SAR imagery lacks detailed information and is plagued by significant speckle noise, rendering the annotation or segmentation of SAR images a formidable task. Recent efforts have resorted to annotating paired optical-SAR images to generate pseudo-labels through the utilization of an optical image segmentation network. However, these pseudo-labels are laden with noise, leading to suboptimal performance in SAR image segmentation. In this study, we introduce a more precise method for generating pseudo-labels by incorporating semi-supervised learning alongside a novel image resolution alignment augmentation. Furthermore, we introduce a symmetric cross-entropy loss to mitigate the impact of noisy pseudo-labels. Additionally, a bag of training and testing tricks is utilized to generate better land-cover mapping results. Our experiments on the GRSS data fusion contest indicate the effectiveness of the proposed method, which achieves first place. The code is available at https://github.com/StuLiu/DFC2025Track1.git.

Learning from Noisy Pseudo-labels for All-Weather Land Cover Mapping

TL;DR

This work tackles all-weather land-cover mapping with SAR by addressing noisy pseudo-labels derived from optical segmentation. It introduces a two-stage approach: (1) domain-adaptive optical segmentation with Resolution Alignment Augmentation (RAA) to generate reliable pseudo-labels, and (2) SAR segmentation trained with a noise-robust objective that combines standard cross-entropy and a symmetric cross-entropy loss, aided by a DACS-style self-training framework and training tricks. Key contributions include the image-level alignment via RAA, a threshold-free pseudo-label weighting, and the LCE+LSCE objective for SAR; the method achieves first place in the GRSS 2025 Data Fusion Contest and outperforms official baselines by several percent in mean IoU. The approach has practical impact for robust all-weather land-cover mapping using SAR, and the authors suggest extending with vision foundation models for SAR encoders in future work.

Abstract

Semantic segmentation of SAR images has garnered significant attention in remote sensing due to the immunity of SAR sensors to cloudy weather and light conditions. Nevertheless, SAR imagery lacks detailed information and is plagued by significant speckle noise, rendering the annotation or segmentation of SAR images a formidable task. Recent efforts have resorted to annotating paired optical-SAR images to generate pseudo-labels through the utilization of an optical image segmentation network. However, these pseudo-labels are laden with noise, leading to suboptimal performance in SAR image segmentation. In this study, we introduce a more precise method for generating pseudo-labels by incorporating semi-supervised learning alongside a novel image resolution alignment augmentation. Furthermore, we introduce a symmetric cross-entropy loss to mitigate the impact of noisy pseudo-labels. Additionally, a bag of training and testing tricks is utilized to generate better land-cover mapping results. Our experiments on the GRSS data fusion contest indicate the effectiveness of the proposed method, which achieves first place. The code is available at https://github.com/StuLiu/DFC2025Track1.git.

Paper Structure

This paper contains 13 sections, 11 equations, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Overview
  4. Self-training with Resolution Alignment Augmentation
  5. Symmetric Cross Entropy Loss
  6. A Set of Tricks
  7. Experiment Setting
  8. Datasets
  9. Experiment Configs
  10. Results and Analysis
  11. ablation study
  12. Comparing with Previous Work
  13. Conclusion

Figures (1)

  • Figure 1: Illustration of the proposed method for the all-weather land-cover mapping task. RAM indicates the resolution alignment augmentation. CE is the cross-entropy loss. SCE represents the symmetric cross-entropy loss. Self-training is one of the most popular domain adaptation techniques.