Locally Consistent Transductive Information Maximization for Few-Shot Remote Sensing Scene Classification

Researchers introduced LC-TIM, a transductive few-shot method for remote sensing scene classification.

LC-TIM extends TIM++ with a local consistency regularizer to better exploit relationships among unlabeled query samples.

They also propose a multi-source variant that combines affinity graphs from multiple foundation models.

Across a new open-source benchmark spanning 10 datasets, two remote-sensing vision-language models, and multiple few-shot settings, the method reports state-of-the-art results.