Semantic segmentation of single SAR imagery leveraging historical Sentinel-1 and Sentinel-2 data

Xiaoyan LU and Qihao WENG

[Paper]

IT-SS-18K Dataset

Each dataset contains 17,676 training samples and 395 test samples, all at a spatial resolution of 10 m with an image size of 510 × 510 pixels. Since the Image Translation and Semantic Segmentation datasets cover the same geographic regions, we jointly refer to them as IT-SS-18K, where “18K” denotes the approximately 18,000 samples in the combined training and test sets.

IT-SS-18K datasets are released at Quark Drive, Code: 1wTx

Implementation steps

1. Train a SAR-to-optical translation model using 17,676 paired stretched optical and SAR images.

2. Apply the trained translation model to SAR images from the training set (17,676) and the test set (395) to generate synthesized optical images.

3. Train a degradable fusion model using the 17,676 synthesized optical images, the corresponding real SAR images, and their associated semantic labels.

4. Feed the 395 synthesized optical images together with the corresponding real SAR images into the fusion model to obtain the final segmentation results.