Per-pixel and sub-pixel classifications of high-resolution satellite data for mangrove species mapping

Abstract

High spatial resolution sensors such as IKONOS and QuickBird, are expected to classify mangrove species more accurately than coarse spatial resolution satellite images. Conventional per-pixel classification techniques could not improve the classification accuracy when such high-resolution images are applied. Such failure has encouraged the invention of more sophisticated and deterministic techniques i.e. subpixel classifications. In this study, the mangrove forest at Sungai Belungkor, Johor, Malaysia was classified using IKONOS data. Two classification approaches were applied, namely per-pixel and sub-pixel techniques. The conventional per-pixel classifiers used in this study were Maximum Likelihood (ML), Minimum Distance to Mean (MDM) and Contextual Logical Channel (CLC) while the Linear Mixture Model (LMM) was selected as the sub-pixel classification approach. The classification results revealed that the CLC classification with a contrast texture measure at window size 21 x 21 yielded the highest accuracy (82%) in comparison to the ML (68%) or MDM (64%). The spatial distribution of the classified mangrove species and classes coincided with the common mangrove zones in Malaysia. For the results of the LMM, the fraction of pixels measured from the satellite imagery and observed in the field gave a good correlation with an R2 value of 0.83 for Bakau minyak, a moderate correlation with an R2 of approximately 0.71 for Bakau kurap and an R2 of 0.75 for the ‘Others’ type of mangrove species. An error image was also created to compare the best fitting spectrum produced by the inversion of the LMM with the original observed spectrum, where the maximum RMS error was only 5%.