Effects of Orthophoto Band Combinations on Semantic Segmentation

Abstract

Recently classify of the high resolution orthophotos using the convolutional neural network (CNN), which is the popular architecture of image classification applications with deep learning. In this study, trainings were carried out using the DeepLabv3 architecture based on the CNN network named ResNet. Potsdam dataset was selected as the study region, which is presented as an open data set by the International Society for Photogrammetry and Remote Sensing (ISPRS).A total of 2112 images were used, 352 of this images used for verification and another 352 images used for test data. It has been trained with five different spectral band combinations: RG (red-green), RB (red-blue), GB (green-blue), RGB (red-green-blue) and IRRG (infrared-red-green). After the trainings,the classification success was compared on the test data. RG, RB, GB, RGB, IRRG band combinations produced, %91, %85, %91, %92, %91 training accuracy rates, respectively. Results demonstrate that, using different band combinations on trainings give us different accuracy.