Publication: Cloud Classification Using Convolutional Neural Networks
Loading...
Identifiers
Publication date
2019-09-23
Defense date
2019-10-08
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
The aim of this thesis is to analyze the performance of Convolutional Neural Networks
(CNN) in the domain of automatic cloud classification. The study makes use of
a set of images never before used on CNNs that was provided by the MATRAS (Modelización
de la ATmósfera y RAdiación Solar) group [1] of the University of Jaén. The
dataset is formed by a total of 1290 of ground-based pictures of the sky taken from Seville
(Spain). In this dataset, cloud types are not represented equally in terms of the number
of instances; it is unbalanced. Two datasets are created from the original dataset: a simple
dataset with the two most prominent cloud types and a complex dataset with 6 cloud
types. The types of cloud considered are stratocumulus and cirrus in the first case and,
stratucumulus, cirrus, stratus, altocumulos, cumulus and clear sky in the second case. A
total of 20 experiments are presented, activating or deactivating a presubsampling layer
(with average pooling) placed before the first convolutional layer. The experiments also
introduce the use of the adagrad optimizer instead of the default SGD (Stochastic Gradient
Descent) optimizer. The resulting models for the 2 cloud type problem provide an
accuracy rate of 91%, showing a positive response towards the presubsampling layer and
the change in the optimizer. As for the complex problem, accuracy ranges from 68% to
76% depending on the model. In this domain, the presubsampling layer and adagrad optimizer
proved to be better for the overall accuracy and the classification of stratocumulus
while being detrimental to the classification of clear sky. Finally, in the 6 cloud types
domain, some limitations were found as the difference in number of instances introduced
a bias in some of the resulting models.
Description
Keywords
Convolutional Neural Networks (CNN), Automatic Cloud Classification, Solar energy