FMANet: Super-Resolution Inverted Bottleneck-Fused Self-Attention Architecture for Remote Sensing Satellite Image Recognition

Abstract

The remote sensing (RS) image classification task has been studied widely in the RS and geoscience community. The important applications of RS are landslides, earthquakes, land-use, and land cover classification. Landslides and earthquakes are some of the most dangerous natural disasters that frequently occur. High-resolution RS images can be useful for accurately classifying landslide and earthquake regions. The deep learning technique has improved performance compared with the traditional methods; however, these techniques are reliable on large-scale datasets. In this work, we proposed a novel architecture based on super-resolution and fused bottleneck self-attention called (FMANet) convolutional neural network. A new custom deep super-resolution network is designed as the first step to improve the quality of RS images. In the next step, a new fused bottleneck self-attention architecture is proposed that learns the features in two distinct networks: residual and inverted. Both models are trained on the resultant super-resolution images, whereas the hyperparameters are initialized using Bayesian optimization. In the testing phase, features are extracted from the self-attention layer and passed to the shallow narrow neural network for classification. The experimental process of the proposed architecture is conducted on three datasets, MLRSNet, Bijie Landslide, and Turkey Earthquake, and improved the accuracy of 91.0%, 92.8%, and 99.4%, respectively. Results are also compared with state-of-the-art techniques and show significant improvement and the model is also evaluated using the lime for the interpretation of the outcomes proposed model.

Keywords

• Classification
• deep learning (DL)
• networks fusion
• remote sensing (RS)
• super-resolution