Coastal and Land Use Land Cover Area Recognition From High-Resolution Remote Sensing Images Using a Novel Multimodal Attention Inception Residual Deep Network

초록

As an important problem in earth observation, aerial scene classification tries to assign a specific semantic label to an aerial image. The land use land cover (LULC) classification in the aerial scene through remote sensing (RS) data is a key research area due to the important applications such as climate change, agriculture, urban structure, and water resources. However, classification from low-resolution remote sensing images causes accuracy and precision rate degradation. Deep learning models have shown significant performance in recent years for aerial scene classification. In this work, we proposed a novel end-to-end deep learning framework for LULC classification from RS images. The proposed framework is based on two novel deep learning architectures: Super-resolution residual attention-based network (SR-RAN(5)) and multimodal inception attention-based convolutional neural network (M(2)IAN). In the first stage, a novel SR-RAN5 architecture is designed to generate high-quality RS images from the original datasets. M(2)IAN architecture is proposed for the RS image classification in the second stage. The hyperparameters of the proposed M(2)IAN model are selected through the Redfox optimization algorithm utilized in the training phase. After the training phase, the proposed model is tested on a test set of the selected datasets that are finally utilized for the classification. In addition, the proposed model is also interpreted through the LIME XAI technique, which localizes the important features of the input image. Extensive experiments are conducted on three publically available aerial scene datasets: NWPU, MLRSNet, and Mixed Coastal. On these datasets, the proposed model obtained improved accuracy of 91.8, 90.6, and 90.0%, respectively. Based on the results obtained, ablation studies, and comparison with existing techniques, it is observed that the proposed model achieves state-of-the-art performance. Also, it can be helpful in the real-time RS applications.

키워드