A modified Generative Adversarial Network (GAN) architecture for land use classification

Abstract

Land cover and usage classification with satellite images plays an important role in many applications such as land resource management, urban planning, precision agriculture and environmental protection. Faster and easier land cover and usage classification without the prior knowledge of the terrain and training sample assignment can be done using deep learning algorithms. Early works in land use classification are mainly focused on machine learning algorithms. In the past few years some deep learning (DL) architectures are also used in the land cover and usage classification purposes. But these DL architectures need large amounts of training samples to get higher accuracy. In this paper, a modified Generative Adversarial Network (GAN) architecture has been proposed for land use classification. This deep learning model performs physical atmospheric corrections as a pre-processing step. Moreover, the model has shown to perform an efficient classification based on the statistical qualifications performed herein with a limited training dataset acquired from UC Merced land use dataset.