Geometry and CUDA code pinpoint remote island resort

A recent technical write-up outlines the process of identifying a specific island resort and its coordinates using mathematical geometry and CUDA programming. The author resolved an open-source intelligence challenge, designated as gralhix #004 and created by Sofia Santos, by analysing a drone photograph without relying on image recognition tools such as Google Lens. The methodology focused on extracting geometric fingerprints from the image to determine the relative positions of three visible landmasses.

The process began by filtering global land polygons from OpenStreetMap to identify potential matches. The author utilised the full global coastline vectors, a dataset of 882 MB, and applied heuristic filters to narrow the search to the tropics. This initial filtering reduced the dataset to 141,131 land polygons. Further constraints were applied to identify clusters of at least three landmasses within 20 kilometres of each other, generating 80,690,777 candidate triples for analysis.

To manage the computational load, CUDA programming was employed to assign one thread per candidate triple. These threads calculated angles and distance ratios in parallel to match the spatial arrangement observed in the photograph. This parallel computing approach resulted in 158,784 initial matches. Additional filtering steps included checking for open water via rectangle intersection and analysing the Polsby Popper Score to ensure the shape of the primary landmass resembled a coral cay.

Verification of the remaining candidates involved satellite imagery and digital elevation models. The author used Sentinel-2 imagery via the Earth Search API to analyse vegetation using Normalised Difference Vegetation Index values. Digital elevation models from the Copernicus DEM GLO-30 were then used to confirm elevation constraints. These steps narrowed the candidates to 26 locations situated in southern Asia, Australia, Oceania, and Brazil.

The final location was identified as an island in Micronesia, with coordinates of 7° 21′ 48.4″ N, 151° 45′ 20.7″ E. The analysis also determined that the camera was facing north-west, at a bearing of 324.97 degrees. The write-up, published on a personal blog, notes that the solution was achieved through human-led mathematical and programming efforts rather than large language model generation.

The technical report highlights the utility of combining geographic data with parallel computing for precise geolocation. By leveraging OpenStreetMap data and Copernicus satellite products, the author demonstrated a method for solving geolocation challenges that relies on geometric consistency rather than visual pattern recognition.