Unsupervised cross-sensor domain adaptation using adversarial network for land cover classification

Remote sensing data gathered by various satellites (cross-sensor data having) are used to have a significant impact to lower down the performance of the stand-alone land cover classification model. The collected data from source and target regions have different probability distributions due to the different resolution of images and different geographical locations. To deal with this problem, an adversarial network-based unsupervised cross-sensor domain network has been investigated by considering two source -> target scenarios using hyperspectral and aerial image datasets. Initially, an unsupervised generative adversarial network (GAN) has been implemented to minimize the distribution between both domains. Following that, the transformed target images are obtained using the trained GAN architecture. Thereafter, a deep convolutional neural network (DCNN) has been trained using the source images and finally, the trained DCNN is used to predict the land cover classes under a multi-sensor framework. The effectiveness of the proposed scheme has been compared with the state-of-the-art techniques, and the results are found to be promising to handle the issues under an unsupervised cross-sensor environment.