Shadow Correction in UAV Imagery Using a Deep Convolutional U-Net Neural Network

Shadows in unmanned aerial vehicle (UAV) imagery introduce significant visual and radiometric distortions that adversely affect feature extraction, classification, and three-dimensional reconstruction accuracy. To mitigate these effects, this study proposes a deep learning–based shadow correction framework built upon a customized U-Net architecture enhanced with a VGG19-based encoder. The model was trained on paired UAV images (shadowed and shadow-free) normalized to the [0,1] intensity range. Network parameters were optimized using the ADAM optimizer and a hybrid loss function combining mean absolute error (MAE) and mean squared error (MSE) to balance pixel-level accuracy and structural preservation.

Quantitative evaluation using Root Mean Square Error (RMSE) and Peak Signal-to-Noise Ratio (PSNR) demonstrated that the proposed model achieved an RMSE of 0.0404 in the normalized [0,1] range (equivalent to 10.31 in the 8-bit [0–255] scale) and a PSNR of 27.87 dB, indicating accurate reconstruction of shadowed regions while preserving fine structural and textural details. Qualitative assessments further confirmed stable performance across high-resolution short-range imagery without introducing noticeable artifacts.

The results suggest that enhanced convolutional architectures remain highly effective for shadow correction in high-detail UAV scenarios, providing a favorable balance between reconstruction accuracy and computational efficiency. Integrating such deep convolutional frameworks into UAV image preprocessing pipelines can significantly improve radiometric consistency and the analytical reliability of photogrammetric datasets.