Vision-Based UAV Altitude Estimation Using Deep Learning: A ResNet50 Approach on Nadir Images

In this study, a vision-based deep learning approach is proposed for altitude estimation for unmanned aerial vehicles (UAVs) as an alternative to traditional methods—such as GPS, barometric sensors, and laser altimeters—that are susceptible to environmental limitations. A comprehensive dataset comprising over 300,000 Nadir images, acquired using Mavic 2 Pro and Mavic 2 Zoom platforms under diverse weather, illumination, and terrain conditions, was employed. The images underwent extensive preprocessing, including data augmentation (e.g., rotations and zooming), the integration of GPS data extracted from EXIF metadata with Digital Elevation Model (DEM) information, coordinate transformations, and scaling adjustments to account for differences in camera systems. A pre-trained ResNet50 model, originally trained on the ImageNet dataset, was adapted for regression tasks and trained over 200 epochs using the Adam optimization algorithm and mean squared error (MSE) loss. Experimental results demonstrate that the proposed method achieves high accuracy, with a mean absolute error (MAE) of 4.09 meters in urban areas and 6.06 meters in rural settings, alongside a high overall coefficient of determination (R² = 0.9981 in urban and 0.9884 in rural environments). These findings indicate that the vision-based deep learning approach can serve as a reliable alternative or complement to conventional sensor technologies for UAV altitude estimation.