ResMergeNet: A Residual Learning Based U-Net for Building Segmentation using Multi-Resolution Data Fusion

Abstract. Data fusion in remote sensing is a critical task for integrating diverse datasets to enhance the accuracy of geospatial analysis. This research aims at building segmentation on a merge data combining the WHU building dataset and Massachusetts Building Dataset, leveraging deep learning for effective feature extraction. A model, ResMergeNet, based on Residual U-Net, is proposed to address challenges such as spatial resolution mismatch, complex building structures, and environmental diversity. The model successfully resolves issues like dataset heterogeneity, noise interference, and occlusions caused by trees and other objects. It also handles variations in building sizes, shapes, and boundaries across different datasets. The model achieves strong performance, with an IoU of 90.63%, accuracy of 95.13%, and an F1-score of 81.00%. The proposed architecture is also compared with other state-of-the art models and can be used in future in applications such as land use monitoring and large-scale building footprint mapping for improved geospatial analysis and smart city development.