Optimized loss and self attention for enhanced domain adaptation in remote sensing image classification

Primitive remote sensing (RS) image classification algorithms primarily rely on labeled images to train the model. However, acquiring labeled data in remote sensing is often expensive and labor-intensive, requiring extensive domain expertise, especially when large and diverse datasets are used. Traditional methods, such as maximum likelihood classification and k-nearest neighbors, depend on manually crafted features and struggle to handle large-scale, high-dimensional data, underperforming in these scenarios. These limitations highlight the need for domain adaptation techniques, which can transfer knowledge from labeled datasets to new, unlabeled domains. Domain adaptation approaches have been developed to address this scenario, utilizing existing labeled images for training and classifying unknown images from different scenes. The distribution variability issue can arise due to variations in acquisition environment conditions, scenes, times, or changing sensors. Existing domain adaptation approaches consider one or more types of losses, such as primary losses (e.g., center and triplet losses), secondary losses (e.g., Maximum Mean Discrepancy (MMD), CORAL, and entropy), by extracting features from backbone networks like VGG or ResNet. However, none of the existing work incorporates an attention mechanism alongside all these losses within a unified framework. In this framework, we integrate primary, secondary, and entropy losses along with a self-attention mechanism. We systematically review the performance of these losses on state-of-the-art Neural Network models, including VGG, ResNet, AlexNet, GoogLeNet, EfficientNet, MobileNet, and ViT. Extensive experiments conducted on the RSSCN7, NWPU-RESISC45, AID, and UCMerced datasets, including a comparison of features extracted from the classification layer and the penultimate layer of the fully connected network, validate the effectiveness of the proposed methodology, paving the way for a more robust and accurate remote sensing system capable of handling domain shifts.