Answers are grounded in this paper and cite their sources.
Remote sensing images are evolving daily for their applications in surveillance, planned urbanization, law enforcement, climate change detection, agriculture, and monitoring catastrophes. Artificial intelligence techniques in this application heavily depend on the quality of remote sensing (RS) images. The low-quality RS images, such as those based on deep learning (DL) architectures trained on the data, impact the AI model's performance. Fusing features and models increases the pattern redundancy, sometimes affecting the model performance and leading to overfitting. This paper proposes a novel batch normalization deep bottleneck residual (BNDBRNet) architecture to classify aerial scene recognition from low-resolution remote sensing images accurately. The proposed architecture is based on the seven-batch normalized residual blocks that accurately extract deeper information from the raw data. These blocks extract contextual and spatial characteristics from satellite image data. In the training of the proposed model, hyperparameters are selected through Bayesian optimization (BO), which makes it smoother than manual selection. Finally, the GradCAM visualization technique tests and interprets the trained model. Four datasets are employed in this work for the experimental process, such as EuroSAT, NWPU-RESISC45, and SIRI-WHU. A customized CoastalD dataset is also prepared to validate the proposed model for predicting coastal regions. The accuracy of these datasets is 94.62%, 92.88%, 99.75%, and 90.0%, respectively. In addition, several ablation studies and comparisons were conducted, and the proposed model outperformed the classification of aerial scenes and coastal regions using low-contrast RS images.
No takes yet. Share an insight, caveat, or question.
Rubab et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: