| Name: | Shivam Kumaran |
| Affiliation: | Space Applications Centre, ISRO |
| Conference ID: | ASI2025_391 |
| Title: | Segmentation and deblending of filaments and dense cores in the Interstellar medium using CNN and explainable-AI. |
| Authors: | Shivam Kumaran 1, Ushasi Bhowmick 1, Vipin Kumar 1 |
| Authors Affiliation: | 1 Space Sciences Division, Space Applications Centre, Indian Space Research Organisation,
Ahmedabad, 380015, India |
| Mode of Presentation: | Oral |
| Abstract Category: | Stars, Interstellar Medium, and Astrochemistry in Milky Way |
| Abstract: | The cold matter in the interstellar medium resides within complex filamentary structures. Submillimetre/Far-infrared observations reveal close association of filaments with dense-cores in star-forming regions, supporting the theory of filament fragmentation to form pre-stellar cores. Intersection of the filaments forms hub-filament systems acting as the origin of high-mass star formation. Large scale identification and characterisation of ISM filaments is essential to establish the link between diffused ISM and the stellar initial mass function. In this work we aim to generate a comprehensive catalogue of filaments and the embedded dense-cores. We develop CNN based model which utilizes parallel architecture for simultaneous extraction of filaments and dense-cores, requiring no manual parameter tuning. The filament identification branch is a derivative of U-Net model and is trained using DisPerSE filaments on Herschel Gould Belt Survey column density maps. It is necessary to deblend dense-cores from filaments as their sharp extremas create breaks in otherwise continuous filament skeleton. Additionally, the blended nature of cores with no clear boundary, makes segmentation based supervised learning difficult. To overcome this, We investigate the use of R-CNN and Grad-CAM based explainable-AI for dense-cores segmentation. A binary classifier is trained on HGBS dense-cores catalogue, to predict the probability of input region containing a source, achieving a validation accuracy of >84%. Grad-CAM is used to generate segmentation map from the classifier predictions.
We extract filaments-skeleton and dense-cores for Herschel’s HGBS and HOBYS surveys. The radial density profile of model-extracted and DisPerSE filaments have a similar distribution in lower dimension principle-component space. We aim to make the model robust and instrument-agnostic in order to generate filament skeleton map on various galactic -plane surveys in order to create an exhaustive catalogue of hub-filament systems.
|
