Multi-modal semantic segmentation with RGB and Thermal (RGB–T) inputs is critical for robust aerial robotic perception, yet existing methods face challenges in effectively adapting pretrained representations and balancing complementary cues under variable illumination.To overcome these challenges, we introduce Segment Anything Guided Multi-Modal Fusion Network (SAM-FuseNet), a novel RGB–T framework that leverages modality-specific encoders built on the Segment Anything Model 2 (SAM2) Vision Transformer (ViT) backbone, enhanced with Low-Rank Adaptation (LoRA) modules for efficient modality-specific fine-tuning without full model retraining. Cross-modal fusion is achieved via a pixel-wise spatial attention mechanism combined with channel recalibration, allowing adaptive weighting of contributions from each modality. A dual-pathway segmentation head further improves prediction quality: the first pathway progressively refines coarse-to-fine masks under SAM2 guidance, while the second pathway performs hierarchical feature fusion to capture long-range contextual information. Extensive experiments on the Caltech RGB–T and PST900 benchmarks demonstrate the effectiveness of SAM-FuseNet, achieving improvements of up to 3.93% in mean Intersection-over-Union (mIoU) while reducing the trained model parameters by more than 50%. Qualitative results further demonstrate sharper boundaries, robustness under underexposed conditions, and resilience to modality degradation.

Hyperspectral image (HSI) classification is often challenged by cross-scene domain shifts and limited target annotations. Existing approaches relying on class-agnostic moment matching or confidence-based pseudo-labeling tend to blur decision boundaries, propagate noise, and struggle with spectral overlap and class imbalance. We propose Spectral Prototype Attention Domain Adaptation (SPADA), a framework that integrates an attention-guided spectral–spatial backbone with dual prototype banks and distance-based posterior modeling. SPADA performs global and class-conditional alignment through source supervision, kernel-based distribution matching, and prototype coupling, followed by diversity-aware active adaptation and confidence-calibrated refinement via prior-adjusted self-training. Across multiple cross-scene benchmarks in urban and inter-city scenarios, SPADA consistently outperforms strong baselines in overall accuracy, average accuracy, and Cohen’s $\kappa$, achieving clear gains on classes affected by spectral overlap or imbalance and maintaining low variance across runs, demonstrating robust and stable domain transfer.