ARTFEED — Contemporary Art Intelligence

LithoBench: A New Benchmark for AI-Driven Lithology Interpretation

ai-technology · 2026-05-11

A new multi-level benchmark called LithoBench has been developed by researchers to assess large multimodal models (LMMs) specifically for interpreting remote-sensing lithology. This interpretation is a complex, knowledge-driven process essential for geological surveys, mineral exploration, and regional mapping, as it involves deducing rock types from nuanced visual, spectral, textural, geomorphological, and contextual indicators. Unlike typical land-cover classification, this task poses significant challenges for automation. LithoBench fills the gap in benchmarks that include lithological annotations, multi-level geological semantics, and assessments informed by experts. It features 10,000 instances of expert-annotated interpretations across 12 lithological categories. This standardized evaluation framework is intended to propel the advancement of geological knowledge-guided LMMs, paving the way for improved automated lithology interpretation and future research in remote sensing and geological AI.

Key facts

  • LithoBench is a multi-level benchmark for evaluating geological semantic understanding in remote sensing lithology interpretation.
  • It contains 10,000 expert-annotated interpretation instances across 12 representative lithological categories.
  • The benchmark addresses the lack of evaluation frameworks for large multimodal models in lithology interpretation.
  • Lithology interpretation is knowledge-intensive, requiring inference from visual, spectral, textural, geomorphological, and contextual cues.
  • The task is fundamental to geological surveys, mineral exploration, and regional geological mapping.
  • Geological knowledge-guided large multimodal models offer new opportunities for automated interpretation.
  • The benchmark captures lithological annotations, multi-level geological semantics, and expert-informed assessment.
  • LithoBench aims to advance reliable automated lithology interpretation.

Entities

Institutions

  • arXiv

Sources