Probing Frozen Vision Foundation Model Features for Tropical Cyclone Intensity: A Leakage-Free Study on the Digital Typhoon Benchmark
Sole author. Under review at IEEE Geoscience and Remote Sensing Letters, 2026.
Code and results
I study how much tropical cyclone intensity information is linearly decodable from the frozen features of vision foundation models. A linear probe on frozen DINOv2 features classifies the four ordinal JMA intensity grades at 62.7% accuracy across 189,364 infrared images and 1,099 storms, evaluated under typhoon-grouped cross-validation so that no storm appears in both training and test.
The evaluation protocol is the main contribution. The benchmark's standard five-class "intensity" task is not an intensity scale: one of its classes is a storm type rather than a higher intensity, and another is an administrative code. Separately, splitting by image rather than by storm inflates accuracy by roughly three points through storm-identity leakage. The attention analysis is reported as a negative result, and the paper claims no state of the art, since frozen features remain behind supervised models.
Research interests: self-supervised and foundation-model representations, remote sensing, label-efficient learning, and evaluation validity.
I'm an AI Engineer at TCS with a data engineering background, from Tamil Nadu, India📍. Python, open source, and code that runs fast 💻.