Fine-tune OpenAI CLIP on RSICD satellite imagery using Flax/JAX on TPUs

AI Impact Summary

This work fine-tunes the OpenAI CLIP model on satellite imagery (RSICD, UCM, Sydney) using Flax/JAX to improve cross-modal retrieval for geospatial data. Training runs on TPUs with 8 cores, 1024-batch size, and uses image/text augmentation plus backtranslation (Marian MT) with Adam or Adafactor optimizers, delivering a meaningful uplift over the baseline CLIP in retrieval metrics. The result enables natural-language or caption-driven search over large satellite-image catalogs, accelerating geospatial analytics, but raises dual-use risk and governance considerations for surveillance and policy compliance.