Remote Sensing of Invasive Species in the Galapagos Islands: Comparison of Pixel-Based, Principal Component, and Object-Oriented Image Classification Approaches

Abstract

The application of remote sensing techniques in threatened ecosystems such as the Galapagos Islands has shown to be a powerful tool for decision-making. Specifically in the case of San Cristobal Island, it will allow accurate mapping and modeling techniques at relatively low costs for battling invasive species such as guava and wax apple. This research evaluates the performance of three classification techniques for land cover mapping in the agricultural area of San Cristobal in the Galapagos: (a) pixel-based hybrid (supervised/unsupervised classification), (b) principal component pixel-based hybrid, and (c) object-oriented image hybrid classifications. An evaluation of three parametric classification algorithms (maximum likelihood, Mahalanobis distance, and minimum distances) for classification technique was also performed. The goal was to compare and identify the best approach for determining LULC with a focus on invasive species such as guava in the highland territory. The results for both pixel-based approaches are superior than the object-based approach. Nevertheless, it was evident that the principal component classifications tend to mix signature responses and did not show the same discrimination ability. Per-pixel/hybrid classification with maximum likelihood and Mahalanobis distance performs a superior kappa index of 0.8640 and 0.8610, respectively, proving to be more sensitive toward identifying invasive species such as guava and wax apple fields.