Vision-Language Modeling Meets Remote Sensing: Models, Datasets and Perspectives

TL;DR

This paper reviews recent advances in vision-language models for remote sensing, covering architectures, datasets, and future directions, highlighting their ability to improve data analysis and user interaction in the domain.

Contribution

It provides a comprehensive taxonomy, detailed analysis of existing models, datasets, and discusses future research challenges in remote sensing vision-language modeling.

Findings

VLM models achieve strong performance across remote sensing tasks.

Contrastive, instruction tuning, and generative models are key categories.

Large-scale datasets and architectural innovations drive progress.

Abstract

Vision-language modeling (VLM) aims to bridge the information gap between images and natural language. Under the new paradigm of first pre-training on massive image-text pairs and then fine-tuning on task-specific data, VLM in the remote sensing domain has made significant progress. The resulting models benefit from the absorption of extensive general knowledge and demonstrate strong performance across a variety of remote sensing data analysis tasks. Moreover, they are capable of interacting with users in a conversational manner. In this paper, we aim to provide the remote sensing community with a timely and comprehensive review of the developments in VLM using the two-stage paradigm. Specifically, we first cover a taxonomy of VLM in remote sensing: contrastive learning, visual instruction tuning, and text-conditioned image generation. For each category, we detail the commonly used network architecture and pre-training objectives. Second, we conduct a thorough review of existing works, examining foundation models and task-specific adaptation methods in contrastive-based VLM, architectural upgrades, training strategies and model capabilities in instruction-based VLM, as well as generative foundation models with their representative downstream applications. Third, we summarize datasets used for VLM pre-training, fine-tuning, and evaluation, with an analysis of their construction methodologies (including image sources and caption generation) and key properties, such as scale and task adaptability. Finally, we conclude this survey with insights and discussions on future research directions: cross-modal representation alignment, vague requirement comprehension, explanation-driven model reliability, continually scalable model capabilities, and large-scale datasets featuring richer modalities and greater challenges.