Comparison of approaches for spatial interpolation of weather data on a specific date
Date
2020-03
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: This study compares four approaches to spatial interpolation of minimum
and maximum temperature, and rainfall weather variables using data from
92 weather stations in the region of KwaZulu-Natal in South Africa. The approaches
are Kriging with external drift (KED), Gaussian filter (GF), random
forest (RF) and multilayer perceptron (MLP). The comparison was done
against the background that the need for permanent gridded weather data
for the region is important for agricultural and forest management. Also,
there is little information regarding the suitability of methods for prediction
in terms of performance variables for gridded data generation. The
present research addresses these challenges by demonstrating the application
of KED, GF, RF and MLP at a 1km2 spatial resolution accross three
weather variables: minimum and maximum temperature, and rainfall to assess
their performance. Four specific dates were selected to represent both
dry and wet seasons for the years 2016 and 2017. The dates are 15th of January
2016 and of 2017 for the summer season, and 15th of July 2016 and
of 2017 for the winter season respectively. Both years were considered because
from available data, they are on the records as the driest (2016) and
wettest (2017) in the region for the period 2008 to 2018. A cross-validation
scheme was employed to assess the model performances and error evaluations
were compared using RMSE, MAE and R2 measures. The results were
found to be almost similar across the four methods except for the RF model
that outperformed in the periods considered for both years. Particularly, RF
performed with the lowest RMSE and MAE errors for minimum and maximum
temperature for both 15th of July 2016 and of 2017 as against the other models. The performance of RF is explained by the method’s properties of
being an ensemble technique. RF prediction follows from the principle of
random selection of variables with high importance which allows for the
decrease of uncertainty. The result of this research has importance for guiding
decisions regarding forest management and climate driven businesses.
AFRIKAANSE OPSOMMING: Geen opsomming beskikbaar.
AFRIKAANSE OPSOMMING: Geen opsomming beskikbaar.
Description
Thesis (MSc)--Stellenbosch University, 2020.
Keywords
UCTD, Spacial analysis (Statistics), Weather -- Data processing, Interpolation spaces, Ensemble learning (Machine learning), Decision trees