UHR-BAT: Budget-Aware Token Compression Vision-Language model for Ultra-High-Resolution Remote Sensing

Ultra-high-resolution (UHR) remote sensing imagery couples kilometer-scale context with query-critical evidence that may occupy only a few pixels. Such vast spatial scale leads to a quadratic explosion of visual tokens and hinders the extraction of information from small objects. Previous works utilize direct downsampling, dense tiling, or global top‑K pruning, which either compromise query-critical image details or incur unpredictable compute. In this paper, we propose UHR-BAT, a query-guided and region-faithful token compression framework to efficiently select visual tokens under strict context budget. Specifically, we leverage text-guided, multi-scale importance estimation for visual tokens, effectively tackling the challenge of achieving precise yet low-cost feature extraction. Furthermore, by introducing region-wise preserve and merge strategies, we mitigate visual token redundancy, further driving down the computational budget. The experimental results show that UHR-BAT achieves state-of-the-art performance across various benchmarks.

Ultra-high-resolution remote sensing images, such as satellite or aerial images, cover very large areas while often requiring the model to identify tiny but important objects. This makes automatic understanding difficult because processing the whole image in detail is computationally expensive, while simply reducing the image resolution may remove key visual evidence. We propose UHR-BAT, a method that helps vision-language models focus on the image regions most relevant to a user's question. Instead of treating all image parts equally, UHR-BAT uses the text query to guide which visual information should be kept, merged, or discarded. This allows the model to preserve important fine-grained details while reducing unnecessary visual tokens and computation. Experiments show that UHR-BAT improves the understanding of ultra-high-resolution remote sensing images across multiple benchmarks.