Semantic-Guided Remote Sensing Image Captioning With Dual-Path Visual Feature Fusion

Abstract

Remote sensing image captioning aims to generate natural language descriptions for complex aerial and satellite images. Unlike natural image captioning, remote sensing scenes often contain small objects, dense spatial structures, large-scale background variation, and domain-specific semantic categories. This article proposes a semantic-guided dual-path feature fusion framework for remote sensing image captioning. The framework separately encodes visual patch features and semantic category cues, then fuses them through cross-attention to improve object grounding and sentence generation. Inspired by recent developments in vision-language modeling, the framework emphasizes early multimodal interaction rather than late-stage feature concatenation. The model is designed to improve caption accuracy, semantic completeness, and spatial consistency across benchmark datasets such as UCM-Captions, Sydney-Captions, and RSICD. Attention visualization is used to analyze whether generated words correspond to relevant image regions. The study contributes to remote sensing vision-language research by presenting a structured approach for combining semantic guidance, visual attention, and transformer-based caption decoding.