Stop wasting weeks on prompt engineering for satellite imagery; 8 real images are better than 1,000 prompts.
April 15, 2026
Original Paper
Low-Data Supervised Adaptation Outperforms Prompting for Cloud Segmentation Under Domain Shift
arXiv · 2604.08956
The Takeaway
The hype around 'zero-shot' prompting suggests we can just talk our way into any domain, but in specialized tasks like cloud segmentation, this fails. This paper proves that even the best engineered prompts consistently underperform zero-shot baselines. However, supervised fine-tuning with just 8 images (~0.1% of typical data) significantly outperformed any prompting strategy. For domain-specific computer vision, this is the definitive proof that tiny-data training beats large-scale prompting every time. If you have any real data, use it for training instead of wasting time on the 'prompting trap'.
From the abstract
Adapting vision-language models to remote sensing imagery presents a fundamental challenge: both the visual and linguistic distributions of satellite data lie far outside natural image pretraining corpora. Despite this, prompting remains the dominant deployment paradigm, driven by the assumption that domain-specific language can guide frozen model representations toward specialized tasks. We test this assumption directly on a domain where the mismatch is prominent: cloud segmentation for satelli