TranSegNet: A Hybrid Transformer Model for Satellite Imagery Segmentation with Performance Benchmark Against U-Net Variants

To address the crucial demand for reliable satellite data analysis, particularly in disaster-affected regions and poor countries, we introduce TranSegNet, a sophisticated hybrid transformer based deep learning network intended for satellite image segmentation. TranSegNet efficiently solves critical difficulties such as road identification, building detection, and land cover categorization. Its performance is examined using the DEEPGLOBE - CVPR18 dataset. The model includes a new deep learning architecture that merges Transformer-based approaches with classic convolutional neural networks, harnessing the benefits of both to boost segmentation precision and generalization. When contrasted with top models like EfficientNet, SE-ResNeXt, and SE-ResNet, which attained Intersection over Union (IoU) scores between 0.47 and 0.51, TranSegNet excelled with an IoU score of 0.68. This improvement is ascribed to the model’s successful extraction of spatial and contextual information through its hybrid architecture. TranSegNet’s performance illustrates its potential to increase satellite picture segmentation, delivering major benefits for disaster management, urban planning, and environmental monitoring, and setting a baseline for future study in satellite data processing.