Neural Network Classification of Mangrove Species from Multi-seasonal Ikonos Imagery

Abstract

Tropical forests in many areas of Central and South America experience strong seasonality in climatic variables such as rainfall, solar radiation, wind speed, and relative humidity. Such seasonality is typical of the mangrove forests we study along the Caribbean coast of Panama. Tied to this environmental variation are changes in leaf phenology and physiology that can affect the spectral properties of leaves and thus our ability to discriminate canopies of differing species composition. The goals of this study were two-fold. First, we compared the efficacy of three different classification methods for discriminating mangrove canopies, including a back-propagation, feed-forward neural network classifier with two hidden layers of 24 and 12 neurons (hereafter, BP:24:12), a newly developed clusteringbased neural network classifier (CBNN), and a maximum likelihood classifier (MLC). Comparisons were made with and without added textural information. Our second aim was to compare the absolute and relative discrimination abilities of these methods when applied to images of the same forest acquired in different seasons. Two sets of Ikonos images acquired in February (dry season) and May (early wet season) 2004 were analyzed in this study. When only spectral information was considered, MLC and CBNN discriminated differences in canopy species composition with higher accuracy than the BP:24:12 method. When second-order textural information was also taken into account, CBNN outperformed MLC and presented the best classification accuracy, i.e., kappa value equaled 0.93. Analyses of the wet season (May) image were consistently more accurate in discriminating mangrove canopies of differing species composition than analyses of the dry season (February) image, regardless of the classification method or the inclusion of textural information.