Vision-based UAV Altitude Estimation using Deep Learning: A ResNet50 Approach on Nadir Images

This study proposes a vision-based deep learning approach for unmanned aerial vehicle (UAV) altitude estimation as an alternative to traditional methods such as GPS, barometric sensors, and laser altimeters, which are often sensitive to environmental disturbances. A large-scale dataset of 303,710 nadir images was collected using Mavic 2 Pro and Mavic 2 Zoom platforms under diverse weather, illumination, and terrain conditions. Each image was labeled with above-ground-level (AGL) altitude by integrating EXIF-based GPS altitude with a 30 m Digital Elevation Model (DEM) through coordinate transformation and terrain subtraction. A pre-trained ResNet50 model, originally designed for image classification, was reconfigured as a regression network and fine-tuned for 200 epochs using the Adam optimizer and mean squared error (MSE) loss. The proposed model achieved a mean absolute error (MAE) of 4.09 m in urban areas and 6.06 m in rural areas, with R^2 scores of 0.9981 and 0.9804, respectively. Comparative experiments with alternative CNN architectures show that the adapted ResNet50 provides favorable accuracy versus complexity trade-off. These results indicate that the proposed monocular, nadir-image-based framework can serve as a reliable complement or alternative to conventional altitude sensors in UAV operations.