Skip to content

Semantic enrichment and dynamic refinement: prototype-guided remote sensing image segmentation

Oct 2026 · Journal of Applied Remote Sensing · 0 citations

Abstract

Semantic segmentation of remote sensing imagery is a pivotal task in computer vision. Despite remarkable progress driven by deep learning—particularly convolutional neural networks and Transformers—existing methods still struggle with complex backgrounds, multiscale targets, and limited annotated data, leading to suboptimal performance for small objects and fine-grained categories. To address these challenges, we propose ProtoFiNet, a semantic segmentation model that integrates prototype learning with foundation-model knowledge to enhance feature representation and segmentation accuracy. Unlike conventional approaches relying on randomly initialized prototypes, ProtoFiNet enriches prototypes with semantic information from the pretrained Scale-MAE model, fine-tuned via low-rank adaptation (LoRA) for efficient adaptation to remote sensing scenarios. The model incorporates a contextual prototype refinement branch for dynamic prototype updating, coupled with two loss functions—prototype disentanglement optimization and prototype compactness contrastive loss—to enhance intra-class compactness and inter-class separability. Here, we show that ProtoFiNet achieves state-of-the-art performance across three benchmark datasets: ISPRS Vaihingen (mIoU=84.68%), ISPRS Potsdam (mIoU=87.64%), and LoveDA (mIoU=52.81%), outperforming existing methods by 1.56% to 3.1% in mIoU. These results highlight the model’s effectiveness in segmenting small targets and complex backgrounds.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.