MCE-RX-ST: A Hybrid Deep Learning Model for Visual Classification in Remote Sensing Using Minimum Cross Entropy, ResNeXt, and Swin Transformer

Visual classification in remote sensing is a challenging task due to satellite images' diverse characteristics, including variations in spatial resolution, spectral properties, and environmental conditions. This paper presents MCE-RX-ST, a hybrid deep learning model that integrates Minimum Cross Entropy Thresholding (MCET), ResNeXt, and the Swin Transformer to enhance classification accuracy and robustness. The model incorporates a preprocessing step using minimum cross-entropy thresholding-based segmentation, which focuses on identifying areas of interest, such as land, water, or urban regions. This helps disregard irrelevant parts of the image. Additionally, the segmentation mask serves as a weighting mechanism guiding the fusion of features extracted by ResNeXt and the Swin Transformer. ResNeXt is responsible for extracting detailed spatial features through its grouped convolutional layers, leveraging high cardinality to capture fine-grained details and rich local representations. These spatial features are then refined by the Swin Transformer, which excels at capturing global and contextual dependencies across the image through its hierarchical self-attention mechanism. This combination of enhanced local feature extraction and global context understanding boosts the discriminative power of the model, resulting in improved classification performance. Experimental evaluations on the UCMerced LandUse and WHU-RS19 datasets demonstrate a promising improvement in accuracy compared to baseline methods, highlighting the model’s effectiveness in remote sensing tasks. MCE-RX-ST shows promising results in applications such as land-cover mapping and disaster management, offering improvements in classification performance while maintaining interpretability.