Land scene classification using diversity promoting metric learning-convolutional neural network

The land scene classification by remote sensing images predicts semantic class of image blocks by removing visual primitives in remote sensing images. However, there is a problem of within-class diversity and between-class similarity that degrades a performance of scene classification. In this research, the diversity promoting metric learning–convolutional neural network (DPML-CNN) method is proposed for classifying land scene images. The metric learning with convolutional neural network (CNN) maps the same scene image class closer and the different class scenes as far as possible which makes the method much discrimination. The diversity promoting in metric learning is used to reduce the overlapping of the same scene class by uncorrelation of every parameter and provides unique information for those parameters. The UC Merced, AID, and NWPU RESISC45 datasets are utilized in this research for evaluating the proposed DPML-CNN method with evaluation metrics like accuracy and kappa coefficient. The DPML-CNN method reached highest accuracy of 99.27% and 99.84% for 50% and 80% training ratios on the UC Merced dataset when compared to other existing methods like multi-level semantic feature clustering attention (MLFC-Net) and global context spatial attention (GCSA-Net).