| Abstrak/Abstract |
Remote Sensing Scene Classification (RSSC) is a fundamental task for understanding high-resolution aerial imagery and supports a wide range of applications such as land-use analysis, environmental monitoring, and urban planning. Despite recent advances in deep learning, many existing studies focus primarily on in-dataset evaluation, whereas the generalization capability of modern convolutional architectures under cross-dataset conditions remains insufficiently explored. To address this gap, this study investigates the effectiveness of ConvNeXt-Tiny as a transfer learning backbone for RSSC and systematically compares its performance with widely used Convolutional Neural Networks (CNNs), namely ResNet50, DenseNet121, and MobileNetV2. Experiments were conducted using two benchmark datasets, NWPU-RESISC45 and AID, with 20 shared scene categories. Four experimental scenarios were designed, including in-dataset evaluation on each dataset and cross-dataset evaluation without fine-tuning to assess robustness under domain shift. All models were pretrained on ImageNet and trained using an identical transfer learning protocol to ensure a fair comparison. Performance was evaluated using accuracy, precision, recall, F1-score, and macro-averaged Area Under the Receiver Operating Characteristic Curve (ROC-AUC). Experimental results demonstrate that ConvNeXt-Tiny achieves strong in-dataset performance, matching or slightly outperforming ResNet50 on NWPU and showing competitive results on AID. More importantly, ConvNeXt-Tiny maintains robust cross-dataset generalization, achieving performance comparable to ResNet50 and significantly outperforming DenseNet121 and MobileNetV2. ROC-AUC analysis further confirms the stable discriminative capability of ConvNeXt-Tiny across different evaluation scenarios. These findings indicate that modern convolutional designs such as ConvNeXt-Tiny offer an effective and robust solution for RSSC, particularly under domain shift conditions. |
