Unlocking the potential of vision language models on satellite imagery through fine-tuning