DeepAR: Atmospheric River Segmentation

Last updated:
Reading Time: 1 min
Section: Research Projects
Tags: deep-learning climate-science computer-vision test-time-training
Description: In progress - adapting SAM-2 to segment atmospheric rivers and study them in climate model data.

This is a semester project supervised by Dr. Bedartha Goswami (Department of Data Science, IISER Pune), focused on deep learning-based segmentation of atmospheric rivers (ARs).

Atmospheric rivers are narrow corridors of concentrated water vapor transport in the atmosphere, and a major driver of extreme precipitation and flooding. Accurately detecting and segmenting them matters for flood forecasting, water management, and understanding how they may change under climate change.

Traditional tools like TempestExtremes detect ARs using hand-crafted, threshold-based rules, which can be rigid and generalize poorly across climates and resolutions. Deep learning-based segmentation instead learns AR shapes directly from data, adapting more flexibly across climates and resolutions.

The goal is to build DeepAR, a deep learning-based AR segmentation model built on the Segment Anything Model 2 (SAM-2) with custom architectural changes and LoRA fine-tuning, and apply it to the CMIP6 dataset for statistical climate analysis. I’m also exploring test-time training (TTT) to help the model generalize across datasets.

Ground truth atmospheric river mask from the PIKART dataset
Figure 1: Ground truth AR mask (PIKART dataset).
DeepAR predicted atmospheric river probability on the same date
Figure 2: DeepAR's predicted AR probability for the same date.
DeepAR predicted atmospheric river probability on CMIP6 data
Figure 3: DeepAR's predicted AR probability on CMIP6. Notice the AR masks are much noisier now - illustrating the distribution shift problem.

Status: in progress — full write-up, code, and model weights to follow once results are public.