# Low Dimensional Embedding of Climate Data for Radio Astronomical Site Testing in the Colombian Andes

## TL;DR

This study uses unsupervised learning to identify regions in the Colombian Andes with low water vapor suitable for millimeter-wave astronomy, based on climate data analysis over 30 years.

## Contribution

Introduces a novel method for correlating climate data with astronomical site suitability using low-dimensional embedding and Bayesian analysis.

## Key findings

- Identified 6 regions with dry, clear-sky conditions.
- Two regions have high probability of low precipitable water vapor.
- Method shows good correlation with global water vapor maps.

## Abstract

We set out to evaluate the potential of the Colombian Andes for millimeter-wave astronomical observations. Previous studies for astronomical site testing in this region have suggested that nighttime humidity and cloud cover conditions make most sites unsuitable for professional visible-light observations. Millimeter observations can be done during the day, but require that the precipitable water vapor column above a site stays below $\sim$10 mm. Due to a lack of direct radiometric or radiosonde measurements, we present a method for correlating climate data from weather stations to sites with a low precipitable water vapor column. We use unsupervised learning techniques to low-dimensionally embed climate data (precipitation, rain days, relative humidity, and sunshine duration) in order to group together stations with similar long-term climate behavior. The data were taken over a period of 30 years by 2046 weather stations across the Colombian territory. We find 6 regions with unusually dry, clear-sky conditions, ranging in elevations from 2200 to 3800 masl. We evaluate the suitability of each region using a quality index derived from a Bayesian probabilistic analysis of the station type and elevation distributions. Two of these regions show a high probability of having an exceptionally low precipitable water vapor column. We compared our results with global precipitable water vapor maps and find a plausible geographical correlation with regions with low water vapor columns ($\sim10$ mm) at an accuracy of $\sim20$ km. Our methods can be applied to similar datasets taken in other countries as a first step toward astronomical site evaluation.