ReCXTUnet: A Novel Framework for Remote Sensing Image Classification using TransUnet and XGBoost

Abstract

Remote sensing image classification (RSIC) is crucial for many environmental and urban applications. RSIC can be difficult due to the high variability and dimensionality in remote sensing image data. This paper presents a novel framework that combines Transformer-based U-Net (TransUNet) and eXtreme Gradient Boosting (XGBoost) for RSIC. TransUNet, known for its powerful feature extraction capabilities, efficiently captures contextual and spatial information from remote sensing images. Additionally, XGBoost improves classification accuracy by efficiently managing high-dimensional data. TransUNet was originally designed for image segmentation tasks, instead of classification. Its architecture is designed to excel at segmenting complex details within images. In our proposed framework, we have adapted TransUNet by adding a classification layer. The fully connected layer of TransUNet serves as the base learner for XGBoost, forming a robust framework for efficient RSIC. This hybrid approach, which combines TransUNet and XGBoost, offers multiple benefits. TransUNet maintains complex details and spatial relationships in images, which improves feature representation. XGBoost provides high predictive accuracy and prevents overfitting with the help of gradient boosting algorithm. This combination tackles challenges in RSIC, such as variations in image quality and noise. We evaluated the proposed approach using high-resolution remote sensing images from the RSI-CB 256 and NWPU-RESISC45 datasets. Our findings show that our framework has outperformed other existing baseline models, attaining an impressive classification accuracy of 91% in RSIC. The experimental results indicate that our approach not only enhances classification accuracy but also remains robust against variations in image quality and noise.