Scholarlink

This work explores the use of multimodal large language models (MLLMs) for geospatial object detection and enumeration in satellite imagery. Using the publicly available RarePlanes [1] dataset with pre-chipped satellite imagery, we evaluate nine frontier MLLMs on their ability to detect and classify aircraft in overhead imagery. Each model is provided with ontology class definitions and structured output requirements to identify and count aircraft across seven categories. We assess model performance on 300 representative image samples, measuring both object enumeration and classification accuracy. Our methodology examines model sensitivity to object scale, density, and visual context, and probes the extent to which MLLMs can support dynamic, zero-shot search without external embeddings or traditional computer vision models. By isolating the MLLM as a reasoning engine over visual data, this study informs the design of agentic pipelines for automated broad-area search, where adaptability, task generalization, and structured spatial reasoning are essential for defense and intelligence applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating Multimodal Large Language Models for Geospatial Object Identification and Enumeration in Overhead Imagery

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tim Klawa,
  • Frank Tanner

摘要

This work explores the use of multimodal large language models (MLLMs) for geospatial object detection and enumeration in satellite imagery. Using the publicly available RarePlanes [1] dataset with pre-chipped satellite imagery, we evaluate nine frontier MLLMs on their ability to detect and classify aircraft in overhead imagery. Each model is provided with ontology class definitions and structured output requirements to identify and count aircraft across seven categories. We assess model performance on 300 representative image samples, measuring both object enumeration and classification accuracy. Our methodology examines model sensitivity to object scale, density, and visual context, and probes the extent to which MLLMs can support dynamic, zero-shot search without external embeddings or traditional computer vision models. By isolating the MLLM as a reasoning engine over visual data, this study informs the design of agentic pipelines for automated broad-area search, where adaptability, task generalization, and structured spatial reasoning are essential for defense and intelligence applications.