DescribeEarth: Fine-Grained Image Captioning

• DescribeEarth is a multimodal framework for generating detailed, localized captions in remote sensing images.
• It leverages object-centric inputs to move beyond generic scene descriptions towards precise object-level details.
• The system supports flexible region-of-interest queries, enabling analyses of features such as airports, harbor structures, and damaged buildings.

DescribeEarth is a multimodal LLM framework for fine-grained, object-centric language generation in remote sensing imagery, introduced together with the task of Geo-DLC, the DE-Dataset, and the DE-Benchmark (Li et al., 30 Sep 2025). Its stated purpose is to move remote sensing captioning beyond image-level scene summary toward object-level fine-grained image captioning, where a model receives an image and a user-specified region of interest and produces a detailed description of a localized target such as an airport, harbor, damaged building, ship, pylon, or other object embedded in a large overhead image.

1. Geo-DLC as a task formulation

The paper defines Geo-DLC, short for geospatial detailed localized captioning, as a task in which an image $I$ and a user-specified region of interest $R$ are mapped to a detailed textual description $T$: $\begin{aligned} T = Model(I, R) \label{eq:task_define} \end{aligned}$ The region $R$ may come from a bounding box, point, mask, or interactive query, but the expected output goes beyond naming the object. It is intended to describe intrinsic object features, spatial features, and contextual environmental attributes in a coherent way (Li et al., 30 Sep 2025).

The motivation is tied to the properties of