[Paper Review] PolSF: PolSAR image dataset on San Francisco

Xu Liu, Licheng Jiao+1 more|arXiv (Cornell University)|Dec 16, 2019

Synthetic Aperture Radar (SAR) Applications and TechniquesEngineering8 references18 citations

TL;DR

This paper introduces PolSF, a newly curated and publicly available PolSAR image dataset for San Francisco, comprising five multi-temporal, multi-sensor polarimetric SAR images from AIRSAR, ALOS2, GF3, RISAT, and RS2 satellites. The dataset provides pixel-level annotations for land cover classification across diverse urban, vegetated, water, and mountainous regions, enabling advanced research in PolSAR image classification and fusion.

ABSTRACT

Polarimetric SAR data has the characteristics of all-weather, all-time and so on, which is widely used in many fields. However, the data of annotation is relatively small, which is not conducive to our research. In this paper, we have collected five open polarimetric SAR images, which are images of the San Francisco area. These five images come from different satellites at different times, which has great scientific research value. We annotate the collected images at the pixel level for image classification and segmentation. For the convenience of researchers, the annotated data is open source https://github.com/liuxuvip/PolSF.

Motivation & Objective

• Address the scarcity of annotated PolSAR datasets for scientific research in remote sensing.
• Provide a diverse, multi-source PolSAR dataset covering different satellites, imaging times, and spatial resolutions.
• Enable pixel-level classification and segmentation tasks by offering detailed ground truth annotations for urban, vegetated, water, and mountainous areas.
• Support research in single-source and multi-source PolSAR image classification and fusion by making high-quality, labeled data publicly accessible.
• Facilitate benchmarking and reproducibility in PolSAR image analysis through open-source data distribution.

Proposed method

• Acquired five PolSAR images from the IETR website, covering San Francisco from different satellites (AIRSAR, ALOS2, GF3, RISAT, RS2) and imaging times (1989–2018).
• Processed raw PolSAR data using ESA PolSARpro v6.0 (Biomass Edition) to generate PauliRGB pseudo-color images for visualization.
• Aligned each PolSAR image with high-resolution Google Earth imagery from the same time period for accurate spatial reference.
• Used Labelme software to manually annotate pixel-level land cover categories, including urban, water, vegetation, bare soil, and mountainous regions.
• Converted the annotated color maps into structured label files with consistent encoding for machine learning and analysis.
• Cropped and standardized the images to focus on the San Francisco region, preserving spatial and spectral fidelity for research use.

Experimental results

Research questions

• RQ1 How can multi-sensor, multi-temporal PolSAR data from diverse satellites be effectively curated and annotated for scientific research?
• RQ2 What are the key challenges in creating consistent, high-quality pixel-level annotations for PolSAR images across different sensors and imaging conditions?
• RQ3 To what extent can a multi-source PolSAR dataset support robust evaluation of pixel-level classification and segmentation models?
• RQ4 How do variations in spatial resolution and imaging time affect the performance and generalization of PolSAR classification models?
• RQ5 What is the potential of open-access, annotated PolSAR datasets in advancing research in remote sensing and SAR image analysis?

Key findings

• The PolSF dataset includes five PolSAR images from five different satellites, covering a wide temporal span (1989–2018) and varying spatial resolutions (2.33m to 18m).
• Pixel-level annotations were successfully created for all five datasets, with land cover categories ranging from 5 to 6 classes depending on the image.
• The dataset is publicly available via GitHub at https://github.com/liuxuvip/PolSF, enabling open and reproducible research.
• The use of PauliRGB decomposition enabled effective visualization of polarimetric information, aiding in accurate annotation and data interpretation.
• The dataset supports a range of research applications, including single-source and multi-source PolSAR image classification and fusion tasks.
• The inclusion of high-resolution Google Earth imagery as a reference improved the accuracy and consistency of the manual annotation process.