loading page

Cloud Detection Algorithm for Hydrometeorological Applications Using Computer Vision Techniques and a Ground-Based Whole-Sky Camera Network
  • Carlos Toro,
  • Carlos D. Hoyos
Carlos Toro
Sistema de Alerta Temprana de Medellín y el Valle de Aburrá - SIATA, Universidad Nacional de Colombia

Corresponding Author:catoror@unal.edu.co

Author Profile
Carlos D. Hoyos
Universidad Nacional de Colombia,Universidad EAFIT,Corporación Clima
Author Profile


The Aburrá Valley is a narrow highly complex mountainous terrain located in the Colombian Andes. Due to topographical features of the region, and the tropical setting, the meteorological variability is very high and in specific periods of the year limit the atmospheric pollutant vertical dispersion, resulting in high concentrations within the valley. The presence of prevalent low-level clouds in these periods reduce incoming solar radiation to the surface thus diminishing surface sensible heat flux to the lower atmosphere. Therefore, the spatial distribution and temporal variability of cloud coverage play a crucial role in the surface energy balance in the region. Cloud variability is also relevant to study the local hydrological cycle and long-term climate variability and change. The most widespread techniques for cloud observations are human observations, which strongly depends on the objectivity and commitment of the observer, and satellite observations, that has the disadvantage of having a relatively low temporal and spatial resolution for some applications. This research focuses on the implementation of an operational system for in-situ clouds detection based on a ground-based network of whole-sky visible cameras. The operational system uses computer vision techniques and image classification algorithms such as Support Vector Machine (SVM), K-Nearest Neighbors (KNN) among others to obtain. The algorithm starts with a group of training images with set attribute tags such as cloudy and cloud-free skies, to identify and calculate the percentage of clouds coverage in untagged images. The methodology then projects the images from the natural polar coordinates system of whole-sky cameras to 2D cartesian coordinates. Following the reprojection, overlapping images from each camera are combined using a panorama stitching technique to generate a single regional cloud fraction map. Clear-sky combined direct and indirect incoming solar radiation are adjusted using the regional cloud map to generate an estimate of the spatial distribution of cloud-forced incoming solar radiation. Cloud height from ceilometers and in-situ pyranometer measurements of incoming solar radiation provide the additional required information to generate the radiation maps.