Unlocking Cross Domain Classification in High Resolution Remote Sensing with Deep Learning Models

Remote sensing images and their applications have been very much relevant in society since ancient times. An increasing number of sensors are deployed alongside new satellites, capable of generating remote sensing photos with exceptionally high spatial resolution. Most importantly these very high resolution images are freely available. The classification and processing of remote sensing images are increasing day by day. New sophisticated algorithms are developed to process the remote sensing images which seems to be very challenging. The goal of deep learning algorithms is to extract even the smallest features from remote sensing images. Classifying images from remote sensing across domains is a challenging task. The proposed approach utilises ten pre-trained models for cross domain classification of remote sensing images without employing any domain adaptation strategies. The cross domain dataset comprises three challenging land use datasets associated with remote sensing image classification. The results indicate that the introduction of additional deep learning algorithms with reduced parameters is necessary for effectively handling cross domain classification in the absence of domain adaptation techniques. The maximum accuracy obtained is 83.66 % while the minimum accuracy is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 5. 1 4 \%}$</tex>. DenseNet-201 outperforms among other deep learning models where ShuffleNet performs least among them.