Advanced Large Language Models for Satellite Image Processing

Pretrained advanced vision-language models (VLMs), as large language models (LLMs), integrate computer vision and natural language processing (NLP) to perform tasks like image captioning and visual question answering (VQA) in remote sensing (RS). The rapid growth in high-throughput acquisition of multispectral satellite imagery in RS has highlighted the limitations of traditional image analysis techniques because of the computational cost of hand-crafted measurements based on manual annotation and domain adaptation. This chapter responds to the increased demand for intelligent, scalable solutions in satellite image processing by studying the role of VLMs in reducing the semantic gap between RS images and human vision. The state-of-the-art analysis of VLMs, EarthGPT, novel object transformers with contrastive language–image pretraining (CLIP) (NOTC), and large VLMs are tailored for RS-specific scenarios. Core contributions include early development of model architectures, training strategies, multimodal fusion methods, and their applications, including object detection, scene classification, image captioning, VQA, and change detection. This chapter also points out key research gaps, such as a lack of datasets, domain generalization issues, and the requirement of interpretable and efficient models. From this perspective, the chapter highlights the necessity for domain-specific category pretraining, lightweight architectures for real-time deployment, and a cross-modal learning architecture. Incorporating new VLMs in RS processes would have the potential to bring earth observation, disaster response, climate monitoring, and sustainable development to the next level.