Feature extraction and selection algorithm based on self adaptive ant colony system for sky image classification
Sky image classification is crucial in meteorology to forecast weather and climatic conditions. The fine-grained cloud detection and recognition (FGCDR) algorithm is use to extract colour, inside texture and neighbour texture features from multiview of superpixels sky images. However, the FGCDR prod...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | eng eng |
| Published: |
2023
|
| Subjects: | |
| Online Access: | https://etd.uum.edu.my/10897/1/permission%20to%20deposit-grant%20the%20permission-s902039.pdf https://etd.uum.edu.my/10897/2/s902039_01.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Sky image classification is crucial in meteorology to forecast weather and climatic conditions. The fine-grained cloud detection and recognition (FGCDR) algorithm is use to extract colour, inside texture and neighbour texture features from multiview of superpixels sky images. However, the FGCDR produced a substantial amount of redundant and insignificant features. The ant colony optimisation (ACO) algorithm have been used to select feature subset. However, the ACO suffers from premature convergence which leads to poor feature subset. Therefore, an improved feature extraction and selection for sky image classification (FESSIC) algorithm is proposed. This algorithm consists of (i) Gaussian smoothness standard deviation method that formulates informative features within sky images; (ii) nearest-threshold based technique that converts feature map into a weighted directed graph to represent relationship between features; and (iii) an ant colony system with self-adaptive parameter technique for local pheromone update. The performance of FESSIC was evaluated against ten benchmark image classification algorithms and six classifiers on four ground-based sky image datasets. The Friedman test result is presented for the performance rank of six benchmark feature selection algorithms and FESSIC algorithm. The Man-Whitney U test is then performed to statistically evaluate the significance difference of the second rank and FESSIC algorithms. The experimental results for the proposed algorithm are superior to the benchmark image classification algorithms in terms of similarity value on Kiel, SWIMCAT and MGCD datasets. FESSIC outperforms other algorithms for average classification accuracy for the KSVM, MLP, RF and DT classifiers. The Friedman test has shown that the FESSIC has the first rank for all classifiers. Furthermore, the result of Man-Whitney U test indicates that FESSIC is significantly better than the second rank benchmark algorithm for all classifiers. In conclusion, the FESSIC can be utilised for image classification in various applications such as disaster management, medical diagnosis, industrial inspection, sports management, and content-based image retrieval. |
|---|