Multi-algorithm system for land cover classification in tropical dry forest at Guanacaste Conservation Area, Costa Rica

Abstract

For decades the detection and monitoring of land cover have been performed by aerial-transported remote sensing and satellites. Detecting and quantifying land cover relays on sensor capacities and classification techniques, such as supervised, unsupervised and mixed. The supervised method is the most accurate and depends on the ability of the algorithm used to discriminate the categories. On the land cover associated to the tropical dry forest at Guanacaste Conservation Area, Costa Rica, supervised classifications were used, using the Minimum Distance, Mahalanobies, Maximum Likelihood, Neural Network, Support Vector Machine and Parallelepiped algorithms to determine which algorithm classified the coverages better (late forest, early forest, intermediate forest, gallery forest, pasture, mangrove) according to control points taken through field work. According to the kappa index and precision values   the performance of Maximum Likelihood and Neural Network was highlighted in the classification of land coverages. This study demonstrates that a hierarchical and multi-algorithm scheme can propel the results of the classification of land cover by considering the advantages and limitations of each classification algorithm, especially when considering aspects related to sample size, sensor resolutions (temporal, spatial, radiometric, spectral), atmospheric conditions and composition of vegetation and landscape.