Skip navigation

Use of soft computing pattern recognition techniques to analyse tropical cyclones from meteorological satellite data

Use of soft computing pattern recognition techniques to analyse tropical cyclones from meteorological satellite data

Khalid, Fakhar (2013) Use of soft computing pattern recognition techniques to analyse tropical cyclones from meteorological satellite data. PhD thesis, University of Greenwich.

[img] PDF
Fakhar_Khalid_2013.pdf - Published Version
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (38MB)


Tropical cyclones are potentially the most destructive of all natural meteorological hazards. When these cloud systems make landfall, they cause significant amounts of human death and injury as well as extensive damage to property, therefore reliable cyclone detection and classification form important weather forecasting activities.

The work carried out in this research has focused on developing a unique fuzzy logic based patter recognition algorithm for identifying key recognisable structural elements of North Atlantic basin hurricanes. Due to vaguely defined boundaries and cloud patterns associated to cyclones and hurricanes, existing hurricane detection and classification techniques such as Advanced Dvorak’s T-classification technique and Objective Dvorak’s techniques fail to deal with the pattern uncertainties. Therefore, a fuzzy logic pattern recognition model was developed and implemented to overcome the shortcomings of manual and subjective algorithms for tropical cyclone detection and intensity classification.

Three key storm patterns are recognised during the cyclogenesis of any hurricane: Central Dense Overcast (CDO); the eye of the storm; and the spiral rain bands. A cognitive linguistic grammar based approach was used to semantically arrange the key structural components of hurricanes. A fuzzy rule based model was developed to recognise these features in satellite imagery in order to analyse the geometrical uncertain shapes of clouds associated with tropical storms and classify the detected storms’ intensity. The algorithms were trained using NOAA AVHRR, GOES, and Meteosat satellite data at spatial resolution of ~4km and ~8km. The training data resulted in fuzzy membership function which allowed the vagueness of cloud patterns to be classified objectively. The algorithms were validated by detecting and classifying the storm cloud patterns in both visible and infrared satellite imagery to confirm the existence of the storm features. Gradual growth of hurricanes was monitored and mapped based on 3 hourly satellite image dataset of ~8km spatial resolution, which provided a complete temporal profile of the region. 375 North Atlantic storms were processed comprising of a period of 31 years with over 112,000 satellite images of cloud patterns to validate the accuracies of detection and intensity classification of tropical cyclones and hurricanes.

The evidence from this research suggests that the fusion of fuzzy logic with traditional pattern recognition techniques and introduction of fuzzy rules to T-classification provides a promising technique for automated detection of tropical cyclones. The system developed displayed detection accuracy of 81.23%, while the intensity classification accuracy was measured at 78.05% with an RMSE of 0.028 T number. The 78.05% intensity classification accuracy was based on storm being recognised from a preliminary stage of tropical storms. The accuracy of hurricane or tropical cyclone intensity estimation from T1 onwards was recorded as 97.12%. While the storms were recognised, their central locations were also estimated because of their importance in tracking the hurricane. Validation process resulted in an RMSE of 0.466 degrees in longitude and RMSE of 0.715 degrees in latitude. The high RMSE which averages 38.4 km suggests that the estimated centre of the storms were around 38.4 km away from the actual centre measured by NOAA. This was an anomalous figure caused by 9 wrongly georeferenced images. Correcting this error resulted in an RMSE of 0.339 degrees in longitude and RMSE of 0.282 degrees in latitude, approximating an average shift of 19km. In 1990s these forecasting errors hovered around 100km area, while according to NOAA the current research averages the track accuracies around 50km, making this research a valuable contribution to the research domain.

This research has demonstrated that subjective and manual hurricane recognition techniques can be successfully automated using soft computing based pattern recognition algorithms in order to process a diverse range of meteorological satellite imagery. This can be essential in the understanding of the detection of cloud patterns occurring in natural disasters such as tropical cyclones, assisting their accurate prediction.

Item Type: Thesis (PhD)
Additional Information:
Uncontrolled Keywords: weather forecasting, meteorological satellite imagery, satellite data, algorithms
Subjects: Q Science > QC Physics
Pre-2014 Departments: School of Science
School of Science > Department of Pharmaceutical, Chemical & Environmental Sciences
Last Modified: 17 Mar 2017 10:39

Actions (login required)

View Item View Item


Downloads per month over past year

View more statistics