Remote sensing image classification is critical for a wide range of real-world applications, including urban planning and environmental monitoring. However, recent deep learning approaches have predominantly focused on designing increasingly complex architectures equipped with attention mechanisms. While effective, these methods introduce substantial computational overhead, posing significant challenges for practical deployment. To overcome these limitations and enhance generalization without increasing model complexity, we propose a novel and efficient data augmentation strategy termed cat-and-crop . Operating at the semantic level, this method concatenates multiple images from the same category into a composite grid and subsequently applies random cropping to generate novel training samples. The proposed approach effectively expands the training data distribution and strengthens the model’s capacity to recognize objects under diverse spatial contexts. Extensive experiments on the NWPU-RESISC45 and UCMerced-LandUse datasets demonstrate that cat-and-crop consistently improves the performance of various state-of-the-art and widely used models, including CNN-based architectures and Transformer-based models. Specifically, when training ResNet18 from scratch on NWPU-RESISC45, the proposed strategy yields accuracy improvements of 3.51%, 2.85%, and 0.98% over the baseline, CutOut, and CutMix, respectively, underscoring its effectiveness in boosting classification performance. Furthermore, Grad-CAM and feature map analyses verify that the method effectively captures global contextual dependencies and enhances rotational feature representation. As a purely data-level intervention, it introduces zero inference latency and incurs only negligible training overhead. Overall, cat-and-crop provides an efficient, plug-and-play data augmentation solution that is fully compatible with existing pipelines and significantly advances remote sensing scene classification. • Proposes a cat-and-crop augmentation for remote sensing image classification. • Enhances spatial coverage and rotational invariance effectively. • Boosts accuracy across CNNs and ViTs with zero inference cost.
Wenyuan Wang (Fri,) studied this question.