Feature Selection for Mangrove Species-based Classification
WONG Kwan Kit, Frankie
FUNG Tung
Mangrove habitat is one of the highly productive ecosystems. The inventory and distribution of mangrove species is a significant piece of information to understand their habitats and to formulate effective conservation management plans. This study took advantage of hyperspectral (EO1-Hyperion) and multi-temporal SAR (ENVISAT-ASAR) data coupled with in situ field survey for mangrove species classification. Hyperspectral imaging captures a number of narrow contiguous spectral bands ranging from 350 to 2500 nm. With wider spectral coverage and higher spectral resolution, hyperspectral data brings a number of benefits towards vegetation analyses compared to traditional broadband data. The all-weather radar allows continuous data acquisition which in turn enhances the temporal observations and can act as complementary data to the optical one. Besides, the penetration of radar signal captures canopy characteristics such as density, structure, roughness and moisture, which provide additional data dimension to differentiate the species. The aggregation of the two types of data has potential to enhance the discriminability of species. However, given a large number of bands but limited number of field samples, the "curse of dimensionality" which leads to Hughes phenomenon and eventually decline in accuracy is always experienced in a classification problem. Dimension reduction is therefore necessary to retain only salient and relevant feature subset with a view to improve the classification accuracy on one hand and to understand the chemical properties and canopy structure effective in discriminating the species on the other.
The research first used three suboptimal search methods including sequential forward selection (SFS), sequential forward floating selection (SFFS) and oscillating search (OS) developed from the field of computer science to select spectral and radar features that are crucial for mangrove species discrimination. Second, based on various combinations of selected spectral and radar features, classification accuracy were compared and evaluated. Third, the research also assessed the accuracy using different classification algorithms including maximum likelihood (ML), decision tree (DT), neural network (ANN) and support vector machines (SVM).
|