AI models map satellite features better after training on Earth-specific data

An Earth-observation model trained for language-guided segmentation achieved more than twice the region-overlap score of the strongest general-purpose comparator on the authors’ benchmark.

Big Tech

Benedikt Blumenstiel · Hugues Devimeux · Johannes Jakubik · Konrad Schindler

IBM Research · ETH Zurich

Research Digest·6 November 2026·2 min read

Illustration: AI models map satellite features better after training on Earth-specific data

In one line

TerraMesh-Masks significantly boosts open-vocabulary segmentation on satellite imagery versus general-purpose models.

Before you read this

Three ideas this paper builds on

benchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.322 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.240 papers →large language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.279 papers →

Read it as

Existing segmentation models struggle with satellite images because they are trained on everyday photos. They fail to recognize features like industrial zones or wetlands described in natural language. This limits how flexibly users can analyze Earth observation data without retraining models for each new category.

TerraMesh-Masks is a new dataset containing 16 million image segments paired with text descriptions sourced from OpenStreetMap. The dataset covers diverse geography and object types. Using it, the authors trained an open-vocabulary version of the TerraMind model.

The model achieved 35.18% Intersection over Union (IoU), a measure of segmentation accuracy, compared to 16.29% for a top general-purpose model. This more than doubles performance, showing that domain-specific training is crucial for satellite image analysis and can enable flexible, language-driven queries.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

open-vocabulary segmentation benchmark multimodal learning vision-language model large language model