Abstract: A method using a batch covariance neural network (CovNN) for remote sensing scene classification is proposed in this paper to reduce the influence of local abnormal feature in remote sensing scene datasets. The batch covariance layer in CovNN can effectively minus the influence caused by the mean of the whole channel for the input feature maps, which acts as a 3D covariance layer. The CovNN can solve the problems of local variations in light intensity and irrelevant features in the local region. Experiments on the AID and NWPU RESISC45 datasets are conducted to test the recall rates of CovNN compared with other state of the art (SOTA) convolutional neural network (CNN) models. Both datasets have abnormal features in the local region of the images, and the proposed CovNN improves the recall rates compared with other CNN models, which are demonstrated by experiments.

Key words: Remote sensing satellite, Artificial intelligence, Deep neural network, Batch covariance, Transfer learning