Semantic Prompt and Graph-Convolution-Structure Distillation Framework for Semantic Segmentation of Remote Sensing...

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• High-resolution remote sensing semantic segmentation is vital for land-use monitoring, urban planning, and disaster response.
• Current deep learning models face challenges due to modality heterogeneity, fine-scale object structures, and high computational costs.

Purpose of the Study:

• To propose a compact and effective architecture for remote sensing semantic segmentation.
• To address challenges of modality heterogeneity and high computational cost in deep learning models.

Main Methods:

• Developed a semantic prompt and graph-convolution-structure distillation framework (SPGSNet-S*).
• Integrated multimodal feature enhancement with dual-path knowledge distillation (KD).
• Designed auxiliary spatial feature extraction (ASFE) and RGB representation modules for feature alignment and fusion.
• Introduced graph-convolution-based structure distillation and semantic prompt distillation (SPD).

Main Results:

• SPGSNet-S* achieves competitive performance on Vaihingen and Potsdam datasets, outperforming state-of-the-art methods.
• The model demonstrates high efficiency with only 8.89 M parameters and 2.29 G FLOPs.
• Successfully fused noisy normalized digital surface model (nDSM) features with RGB imagery.

Conclusions:

• SPGSNet-S* offers an effective and computationally efficient solution for high-resolution remote sensing semantic segmentation.
• The proposed framework demonstrates the potential of integrating multimodal feature enhancement and dual knowledge distillation.
• Publicly available code facilitates reproducibility and further research.