Synthetic Aperture Radar (SAR) satellites provide reliable imaging capability under all weather and illumination conditions and are therefore well suited for large-area infrastructure monitoring. Indian missions such as RISAT-1/1A generate a substantial volume of SAR data; however, systematic extraction of infrastructure features from these datasets remains challenging due to speckle noise, geometric distortions, and limited availability of labelled training data.
The proposed work aims to develop a practical machine-learning framework for extracting infrastructure features such as roads, bridges, and built-up areas from RISAT-1/1A C-band SAR imagery. In addition to conventional supervised deep-learning approaches, the project will explore self-supervised learning techniques to utilise large volumes of unlabelled SAR data for feature representation learning.
The representations obtained through self-supervised training will subsequently be fine-tuned using limited annotated datasets for infrastructure mapping tasks.
Achievements
The project being undertaken by ISRO in collaboration with academia is directly or indirectly beneficial to the society.