Land cover classification using multi-fusion based dense transpose convolution in fully convolutional network with feature alignment for remote sensing images

With advances in social development and economic growth, remote sensing technology has been attracted greater attention in monitoring the earth data using radar and optical sensors on satellite platforms for a wide range of applications in different fields such as coastal, hazard and natural resources. Satellite images could play a greater role in improving classification accuracy with high spatial resolution and rich spectral information for land cover classification. However, existing image fusion methods achieves low accuracy due to large-scale feature space. To focus on these issues, a deep learning network structure needs to classify different classes with high spatial resolution and rich spectral information to obtain higher accuracy. In this paper, a feature-based classification approach is proposed namely Multi-Fusion based Dense Transpose Convolutional layer in Fully Convolutional Network with Feature Alignment framework (MF-DTCFCN) to label and categorizes the label region in Remote Sensing Images (RSI). Initially, a multi-fusion feature framework is designed by adding a point-wise addition structure to handle large-scale feature space for high-resolution images. Secondly, the optimized features are pre-trained to classify the labels comprised of the most discriminative features in the pre-training network. The density of output label maps are improved by introducing dense transpose convolution in the network. Then combine the output to the feature alignment with point-wise addition is employed to balance the different features and similarities to achieve additional performance for classification. Here, the Land Use/land Cover (LULC) satellite image dataset namely, Sentinel-2 were used to classify the urban areas of Hyderabad city, India. Experimental results depict that the MF-DTCFCN approach outperforms an accurate improvement in classification accuracy than existing methods.