MultiWeather-ThermalUAV

A scene-consistent multi-weather multi-modal dataset for UAV semantic segmentation.

Dataset Description

MultiWeather-ThermalUAV extends the Safe-UAV 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.

• Repository: github.com/shashankborse/multiweather-thermaluav
• Paper: Under Review, DMLR 2025

Dataset Structure

final_dataset/
  {train, val, test}/
    {clear, fog, rain, snow}/
      {rgb, thermal, masks}/
        000000.png ... NNNNNN.png

Dataset Statistics

Split	Scenes	Total Images
Train	9,524	114,288
Validation	1,190	14,280
Test	1,193	14,316
Total	11,907	142,884

Modalities

• RGB: Visible light imagery (3 channel, uint8)
• Thermal: Synthetic 16-bit thermal imagery (1 channel, uint16)
• Masks: Segmentation masks with 3 classes (uint8)

Classes

ID	Name	Description
0	Horizontal	Safe landing zones
1	Vertical	Obstacles
2	Other	Sloped or irregular surfaces

Weather Conditions

Condition	Simulation Method
Clear	Original Safe-UAV imagery
Fog	Beer-Lambert depth-based attenuation
Rain	Directional streak simulation with motion blur
Snow	Depth-aware particle distribution

Usage

import cv2
import numpy as np

# Load a sample
split = "train"
weather = "fog"
idx = "000000"

rgb = cv2.imread(
    f"final_dataset/{split}/{weather}/rgb/{idx}.png")
rgb = cv2.cvtColor(rgb, cv2.COLOR_BGR2RGB)

thermal = cv2.imread(
    f"final_dataset/{split}/{weather}/thermal/{idx}.png",
    cv2.IMREAD_UNCHANGED)
thermal = thermal.astype(np.float32) / 65535.0

mask = cv2.imread(
    f"final_dataset/{split}/{weather}/masks/{idx}.png",
    cv2.IMREAD_UNCHANGED)

Source Data

Derived from Safe-UAV by Marcu et al., ECCVW 2018.

Citation

@article{borse2026multiweather,
  title   = {MultiWeather-ThermalUAV: A Scene-Consistent Multi-Weather 
             Multi-Modal Dataset for UAV Semantic Segmentation},
  author  = {Borse, Shashank Dilip},
  journal = {Journal of Data-centric Machine Learning Research},
  year    = {2026}
}

Licence

CC BY 4.0. Derived from Safe-UAV under its original licence terms.