Search

Remote sensing has been used with various computer vision techniques in order to accurately classify different types of land covers from satellite images. One such technique is the Neighborhood Pixels Method, reported to have an overall accuracy of 94.0% in classifying vegetation, water, and built-up land cover from images taken from the Landsat 8 OLI satellite. In this study, we attempt to increase the accuracy of the technique by determining a more appropriate pixel neighborhood size. The previous study which developed the technique was first replicated, including the use of the same Landsat 8 OLI satellite images for training and testing, the building of lookup tables from the medians of 9x9 pixel neighborhoods, and the implementation of the same scoring system for the prediction step of the technique. Experiments on different neighborhood sizes were then conducted, with various statistics recorded for analysis and extracting insights. The overall accuracy of the original Neighborhood Pixels Method was shown to have improved by using other neighborhood sizes, with the highest average accuracy of 95.75% achieved by using 13x13 pixel neighborhoods. The results indicate that finding an appropriate neighborhood size can be an important step in accurately classifying varying land covers of a satellite image. [ABSTRACT FROM AUTHOR]