Swin-ResU-Net: Hybrid Transformer CNN for Spatial Segmentation of Satellite Imagery

View at publisher Summarise with VIRAT
Authors: Tejas Chauhan, K. Rajakumar, Charan Rajana, Kaviti Sravan, Sai Bhargav Kasetty
Cited by: 0
References: 29

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The resolution of satellite imagery has significantly improved due to the unprecedented growth in Earth observation technologies, necessitating the use of suitable and efficient semantic image segmentation techniques to perform spatial analyses. Most traditional CNN-based segmentation models fail to model long-range spatial dependencies and multiscale contextual relationships, which are inherent in satellite imagery. In this paper, the authors present a hybrid deep learning framework that combines the Swin Tiny Transformer with a ResNet50 backbone within a UNet-style architecture. The suggested model maintains spatial localisation through encoder-decoder skip connections and models global dependencies by utilising Swin Transformer's hierarchical selfattention mechanism. With the ResNet50 backbone, this enhances feature reuse by enabling precise gradient flow. The OpenEarthMap dataset, which featured a varied collection of globally dispersed land-cover samples, was then used to train and validate the model. With a peak validation IoU of 0.54 and an accuracy of 0.67, the experimental results show a progressive improvement over epochs, outperforming a number of baseline architectures in terms of both boundary precision and spatial consistency. The Swin-ResU-Net framework that was suggested brings together hierarchical transformer representations and convolutional feature learning so that the satellite imagery can be segmented spatially with precision. The results of the experiment show that the proposed method is not only robust but also generalizable, as it has consistently improved its performance across different evaluation metrics.

Fields of Study & Subjects

Remote-Sensing Image Classification Advanced Neural Network Applications Remote Sensing and Land Use

Publication details

Journal / source
International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT)
Pages
1133–1138
Publication date
1 February 2026
Language
EN