Introduction
Polarimetric Synthetic Aperture Radar (PolSAR) imagery is an essential tool in remote sensing, offering the unique capability to capture high-resolution data regardless of weather conditions or time of day. Unlike optical sensors, PolSAR can penetrate clouds and operate in total darkness, making it invaluable for disaster monitoring and environmental surveillance. Despite these advantages, the field faces a significant hurdle: the scarcity of pixel-level annotated datasets. Training modern deep learning models for tasks such as image classification and segmentation requires vast amounts of ground-truth data, which is notoriously difficult to produce for radar imagery due to its complex nature.
To address this bottleneck, this paper introduces PolSF, an open-source dataset featuring five multi-sensor PolSAR images of the San Francisco Bay Area. Spanning nearly three decades of satellite and aerial observations, PolSF provides high-quality pixel-level annotations, serving as a vital resource for research in multi-sensor data fusion and the development of robust classification algorithms.
Data Curation and Methodology
The authors developed a systematic pipeline to ensure the accuracy and usability of the PolSF dataset. The process began with the collection of raw PolSAR data from the Institute of Electronics and Telecommunications of Rennes (IETR). This raw data represents the complex backscatter signals captured by the sensors, which require significant processing before they can be used for visual interpretation or machine learning.
Preprocessing and Annotation
Raw data files were processed using the ESA PolSARpro v6.0 software. This tool was used to perform polarimetric decomposition and generate PauliRGB pseudo-color images. PauliRGB is a standard visualization technique where different polarimetric channels (representing different scattering mechanisms) are mapped to the Red, Green, and Blue color channels. This allows human annotators to distinguish between surface scattering, volume scattering (e.g., from vegetation), and double-bounce scattering (e.g., from urban structures).
To create reliable ground-truth masks, the authors cross-referenced these radar images with high-resolution optical imagery from Google Maps corresponding to the same time periods. This was necessary because radar imagery can be counter-intuitive to interpret due to geometric distortions. Manual pixel-level annotation was performed using the open-source LabelMe software. Finally, these annotations were remapped and encoded into standardized ground-truth mask files.
The PolSF Dataset
The PolSF dataset consists of five primary PolSAR images of the San Francisco area, captured by different platforms between 1989 and 2018. This multi-sensor approach allows researchers to compare how different radar systems—operating at different frequencies and resolutions—interpret the same geographic features.
The dataset includes:
- SF-AIRSAR (1989): 10m spatial resolution with 5 classes (Mountain, Water, Urban, Vegetation, Bare soil).
- SF-RS2 (2008): 8m spatial resolution from Radarsat-2, featuring 5 classes including High-Density and Low-Density Urban areas.
- SF-ALOS2 (2015): 18m spatial resolution with 6 classes.
- SF-RISAT (2016): The highest resolution in the set at 2.33m, featuring 6 classes.
- SF-GF3 (2018): 8m spatial resolution from the Gaofen-3 satellite, with 6 classes.
Key Insights and Applications
Multi-Sensor Synergy
By providing co-located data from five different sensors (AIRSAR, ALOS2, GF3, RISAT, RS2), PolSF enables the direct study of sensor-specific noise profiles and imaging characteristics. This is crucial for developing models that can generalize across different hardware platforms, a major challenge in the remote sensing community.
Temporal and Structural Analysis
The 30-year span of the dataset offers a unique perspective on urban development and topographical changes in the San Francisco Bay Area. Furthermore, the standardized labeling of up to six classes—including varying urban densities—lowers the barrier for researchers working on complex land-cover classification and multi-source data fusion.
Conclusion and Future Work
PolSF fills a critical gap in the remote sensing community by providing a high-quality, annotated multi-sensor dataset. While currently limited to a single geographical region and lacking baseline performance benchmarks, it serves as a foundational tool for future research in pixel-level classification and multi-source image fusion. Future work will likely involve expanding the dataset to other regions and establishing benchmark metrics using state-of-the-art machine learning architectures.
Summary updated