An Investigation of Ensemble Learning in Dust Storm Prediction Using Machine Learning Techniques

  • Tariq Almurayziq

Student thesis: Doctoral Thesis

Abstract

The main attributes that are known to contribute to dust storms are wind speed, temperature, air pressure, humidity and the type of surface. Weather experts expend considerable effort to achieve a high degree of forecast accuracy when detecting this weather phenomenon and modern technology means that it is now possible to detect dust storms using satellite images. However, modern applications have yet to be applied using historical data for dust events prediction. In another word, most of current forecasting models are build based on satellite monitor, and in proposed approach it argues that the new techniques such as AI could asset in weather forecasting. The archive of historical dust storm events is a valuable database and it is quite possible that applying one or a combination of artificial intelligence (AI) techniques will help to devise a reliable way of predicting future dust events based on these old cases of dust events. This study examines the process of predicting and identifying dust storms by focusing on a Bayesian network (BN) with a case-based reasoning (CBR) approach and rule-based system (RBS).

The aim of this thesis is to examine trends in CBR by exploring some of the challenges and perceived benefits that accrue from the use of CBR in predicting and identifying dust storms. In addition, this study seeks to determine the applicability of CBR and its appropriateness for predicting dust storms. Indeed, some of the previous studies and related findings will form the basis on which the study’s evaluation of CBR’s validity and applicability will be determined.

The current study’s findings, it is arguing that the CBR with other AI techniques could help to predict future dust storm events. The results reveal that there is similarity between dust cases, and there are many attributes of dust storm play more important roles than others in dust prediction such as wind speed important than air temperature, which could provide opportunities for us to search for the optimal weight of these attributes.

The BN-CBR combination illustrates the degree of accuracy in terms of predicting forthcoming dust storms relative to the pure CBR. BN-CBR is able to forecast future dust storms by making comparisons with similar events by reclassify the old dust events using BN and predict coming new dust storm using Nearest neighbour (NN), the ideal value as demonstrated 3NN.

A real example has been used to test the RBS and the results have been positive, satisfying the BN-CBR prediction in the short term. In addition, suitable actions have been delivered that could usefully serve targeted sectors and inform the wider community about this weather phenomenon. Although this is still in the early stages of being developed, the initial results are encouraging. This suggests that the selected approach could prove useful tool to predict future dust events, by using combinations of AI techniques.
Date of AwardOct 2018
LanguageEnglish
Awarding Institution
  • University of Brighton
SupervisorStelios Kapetanakis (Supervisor) & Miltiadis Petridis (Supervisor)

Cite this

An Investigation of Ensemble Learning in Dust Storm Prediction Using Machine Learning Techniques
Almurayziq, T. (Author). Oct 2018

Student thesis: Doctoral Thesis