Maximizing the Potential of Vision Language Models (VLM) for Satellite Imagery Through Fine-tuning

Unlike ordinary photos, satellite imagery has a nadir view perspective, varying scales, and multi-spectral characteristics. Due to these differences, existing Vision Language Models (VLM) struggle to accurately interpret satellite data.

To address this, a fine-tuning strategy that reflects the unique characteristics of satellite imagery is key. High-quality satellite datasets are used to optimize training so that the model can understand the patterns and terrain features unique to satellite imagery.

Going through this fine-tuning process significantly improves performance in object detection, terrain classification, and complex scene captioning within satellite imagery. This serves as an important technical foundation for increasing the utility of VLMs in the field of remote sensing.