Automated classification and mapping for alluvial geomorphic units.

Interesting paper exploring how knowledge-graphs can be integrated with Large Language Models (LLM) and Vision Language Models (VLMs) to improve object identification and segmentation in remote sensing applications Dawson and Lewin (2025).

This image (Fig 3) shown is output from ChatGPT 4.5 to identify and analyse geomorphic features. 3a Image is of the River Teme, UK near Leintwardine (Source Digimap – ©Get Mapping).. 3b General scene description in response to prompt requesting the identification of fluvial features. 3c: Response when asked to annotate specific features. 3d: Relative elevation model and superimposed channel positions derived from historical Ordnance Survey mapping (Dawson and Lewin, 2024). Digimap: High Resolution (25 cm) Vertical Aerial Imagery 2021 Using: EDINA Aerial Digimap Service.

Abstract

This paper critically reviews existing methodologies for classifying alluvial landform units, emphasizing the semantic frameworks and historical evolution of taxonomies that currently underpin identification and mapping efforts. It highlights the inconsistencies and ambiguities inherent in existing classification schemes, underscoring the need for clearer semantic definitions.

Subsequently, the paper examines automated and semi-automated approaches, including geomorphometry and Geographic Object-Based Image Analysis (GEOBIA), for analyzing remote sensing imagery, with particular attention to their efficacy within fluvial environments.

Recognizing recent advancements in remote sensing and computer vision, especially the increased adoption of taxonomies and ontologies to enable consistent, shareable, reusable, and interoperable geographic data, we advocate the systematic development of domain-specific ontologies for alluvial geomorphic units. We reference internationally accepted standards for ontology creation (ISO/IEC 21838–1:2021) and discuss methodologies for encoding these ontologies into machine-readable schemas suitable for machine learning implementations.

From a multidisciplinary perspective, the paper assesses the potential of ontologies and derived knowledge graphs (KGs) to enhance the semantic segmentation of remote sensing imagery. It also explores emerging techniques integrating KGs with large language models (LLMs) and vision-language models (VLMs). Finally, we outline opportunities and considerations for applying and refining Vision Foundation and Language Models to improve object identification and segmentation in remote sensing applications.

Paper here: https://www.sciencedirect.com/science/article/pii/S0012825225002533