Co-Training Transformer for Remote Sensing Image Classification, Segmentation and Detection

Authors- Mrs.E.Nivaditha, S.Dhinesh, M.Gokulanand, S.Manojkumar

Abstract-The rapid advancement of remote sensing technologies has significantly enhanced the ability to monitor and analyze the Earth’s surface, facilitating applications in environmental monitoring, urban planning, and disaster management. Remote sensing image classification, segmentation, and object detection are fundamental tasks in this domain, enabling automated interpretation of satellite and aerial imagery. Traditional machine learning models, including Convolutional Neural Networks (CNNs), have shown promise but often struggle to handle the diverse and complex nature of remote sensing data, which may include variations in scale, resolution, and lighting conditions. In this work, we propose a novel approach for remote sensing image analysis by cotraining a Transformer-based architecture with a state-of-the-art detection framework, specifically Detectron2. The Transformer model excels at capturing long-range dependencies and global context in image data, while Detectron2, a high-performance object detection framework, is capable of fine- grained instance segmentation and detection tasks. By co-training these two models, we aim to leverage the strengths of both architectures to improve classification accuracy, segmentation precision, and detection performance for remote sensing images. The co-training strategy involves joint learning where the Transformer model is used to capture semantic features from the remote sensing images, while Detectron2 refines spatial boundaries and detects objects of interest. We demonstrate the effectiveness of this approach through a series of experiments on benchmark remote sensing datasets, showing significant improvements in classification accuracy, segmentation masks, and detection performance compared to traditional methods, solutions for real-world applications such as land cover classification, urban monitoring, and disaster response.

DOI: /10.61463/ijset.vol.13.issue2.191