Automated Image Geo-Localization Using Deep Visual Feature Embeddings

Publication Date

2026

College

College of Sciences & Mathematics

Department

Math and Computer Science, Department of

Student Level

Undergraduate

Presentation Type

Poster

Summary

This project investigates automated image geo-localization using deep visual feature representations and machine learning classification methods. The objective is to predict the geographic location where an image was taken using only its visual content. Images were processed using a pretrained self-supervised vision transformer, DINOv2, to extract high-dimensional visual embeddings representing scene structure, environment, and visual context. Several prediction approaches were evaluated, specifically nearest-neighbor matching, regression models predicting latitude and longitude directly, and classification models predicting geographic regions defined by spatial grid cells. Using nearest neighbors as a baseline, results showed that regression performs poorly while classification performs significantly better at identifying the correct geographic region. Dimensionality reduction techniques PCA and UMAP were used to analyze the embedding space, revealing that images cluster according to geographic and environmental similarity. These results demonstrate that pretrained visual representations contain meaningful geographic information and can be used effectively for automated image geo-localization.

This document is currently not available here.

Share

COinS