This dataset is the research data for: Accurately Wildlife Censusing by Using Deep Learning for UAV-Based Thermal Imageries

The dataset (00_UAV-derived Waterfowl Thermal Imagery Dataset) captured in a playa wetland area of Nebraska, the United States, during the spring season which contain 355 drone-derived thermal images at 200 feet AGL flight height with more than one wildlife present in each image (512x640 px, 7.5cm/pixel GSD) and 187 images without wildlife presented. The corresponding label was given in a CSV file with five columns where the first column indicates the image name, and the last four columns indicate the position of each waterfowl in the corresponding image (every row indicates a single wildlife). Every wildlife was represent by a bounding box defined by the last four columns (the position of the upper left corner the bounding box was expressed as (x,y) which is equal to the (column, row) of the image, the width and height indicate the pixel additions based on the upper left corner (x, y) in the column and row direction respectively).

Corresponding RGB images were also given in the '01 RGB Images' file. The dataset can be used as training data for automated detection with machine learning or deep learning algorithm.

The orthomosaic of Smith WPA, which contains 2915 wildlife, was given in '02 Test Orthomosaic'. The orthomosiac can be used as test data. The corresponding Label was also given in the same format as above.