A machine learning algorithm to detect fog from space
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Complex adaptive systems (MPCAS), MSc
Computer science – algorithms, languages and logic (MPALG), MSc
Computer science – algorithms, languages and logic (MPALG), MSc
Publicerad
2024
Författare
Svensson, Kevin
Johansson, Nils
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Fog detection is important for traffic safety. Detecting fog using machine learning
on satellite data has been researched before, but not on a global scale using syn thetic data. The aim of the thesis is to use a synthetic dataset of simulated MODIS
satellite data to determine the viability of machine learning algorithms for detecting
fog in satellite images. The synthetic dataset we use is simulated using a fast ra diative transfer model called RTTOV by inputting various atmospheric information
for different conditions. The dataset is tabular and no spatial or temporal relation ship exists between the data points meaning each pixel is treated independently.
We use the synthetic data to train and evaluate numerous machine learning models
including various implementations of XGBoost and feed forward deep neural net works. We also apply a model trained on synthetic data to a real MODIS image. We
demonstrate that classification models can achieve good recall values on synthetic
data when oversampling fog in the training data, the best being 0.87 recall with a
deep neural network. However, we find that this comes at the cost of a large amount
of false positives evident by the low precision value of 0.27. It is concluded that no
model performed satisfactory results for replacing existing methods of fog detection.
We identify the acquisition of supplemental labeled real satellite images as a possi bility for future improvement, allowing for spatial analysis which is impossible with
the independent pixels of the synthetic dataset alone. However, this is a non-trivial
task due to the challenges in obtaining and labeling a sufficiently large and diverse
dataset of real satellite images.
Beskrivning
Ämne/nyckelord
MODIS, fog, machine learning, nowcasting