Contextual band addition and multi-look inferencing to improve semantic segmentation model performance on satellite images

Abstract

ABSTRACT In this study, we introduce two spatial approaches to enhance semantic segmentation accuracy in remote sensing imagery in resource-constrained computational environments, with a focus on edge regions. The first approach, Contextual Bands Addition, integrates overview, incorporating essential contextual information. The second approach, multi-look inferencing, utilises multiple spatial perspectives to refine segmentation. Our results show significant improvements: Contextual Band Addition with tile overlaps increases IoU scores by 4–5%. Multi-look inferencing enhances IoU scores by 2.5% post-training. Combined, these strategies yield a 6–7% overall performance boost, valuable for semantic segmentation on limited training samples from satellite imagery datasets.