MultiWeather-ThermalUAV: A Scene-Consistent Multi-Weather Multi-Modal Dataset for UAV Semantic Segmentation

We present MultiWeather-ThermalUAV, a scene-consistent multi-weather multi-modal dataset for UAV semantic segmentation. The dataset extends Safe-UAV, a synthetic aerial dataset, with synthetically generated thermal imagery and four weather conditions: clear, fog, rain, and snow. Its defining property is scene consistency: geometry, object placement, and segmentation annotations are identical across all weather conditions for every scene, enabling controlled isolation of weather-induced domain shift from scene-level variation. This property is absent from existing UAV segmentation and weather robustness benchmarks, which either provide a single modality, a single weather condition, or do not preserve scene content across conditions. The dataset comprises 11,907 unique scenes and 142,884 total images across three modalities (RGB, thermal, segmentation mask) and four weather conditions, organised into train, validation, and test splits. We establish a benchmark protocol in which models are trained exclusively on clear weather and evaluated across all conditions without adaptation. Experiments with U-Net and LR-ASPP MobileNetV3-Large across RGB, thermal, and early fusion modalities reveal that thermal imaging is substantially more robust than RGB under fog and rain, that early fusion via channel stacking collapses under fog despite achieving the highest clear weather performance, and that snow produces the mildest degradation across all configurations. The dataset, generation code, and experiment pipeline are publicly released to support reproducible research on weather-robust multimodal UAV perception.