Combined particle swarm optimization and linear discriminant analysis for landslide image classification: application to a case study in Taiwan

Abstract

In Taiwan many reservoirs are constructed in mountain areas. Unfortunately, several earthquakes shook the soil, and typhoons brought a huge amount of water to the reservoir zone. In the past studies, remote-sensing image data were used to effectively monitor the landslide near reservoirs. In recent years, linear discriminant analysis (LDA) has become a well-known method for image classification. However, there are few studies to optimize the linear classification function. While the ancillary information has been adopted easily by new methodologies, the ancillary information must be examined by a landslide image classification system. To explore the effects of optimization on the LDA equations, three approaches were compared: (a) conventional LDA; (b) combined discrete rough sets and LDA (DRS + LDA), which identify the core factors and the corresponding thresholds of landslide occurrence; and (c) combined particle swam optimization algorithm and LDA (PSO + LDA), which optimizes the parameters of LDA equation to attain the best classification outcomes. The above methods were applied to a reservoir region in Taiwan, and the following classification results were obtained. The application of DRS + LDA in our case study reduced the number of ancillary attributes from 14 to 5, and resulted in an accuracy rate of 0.83. On the other hand, the application of PSO + LDA resulted in the same accuracy rate as that of DRS + LDA, whereas the accuracy rate of conventional LDA was found to be 0.78.