Mixing Real and Synthetic Data in Neural Network Training: A Case Study Regarding Runway Detection

The landing phase is one of the most critical and accident-prone stages of a flight. Automating this phase has the potential to significantly increase aviation safety by reducing pilot workload. A promising research approach involves using deep neural networks for runway detection for autonomous landing. However, effectively training these models requires large, high-quality datasets—a costly and logistically challenging task, especially when real aerial imagery is required. To mitigate these costs, data augmentation is often employed using synthetic images. However, the optimal balance between real and synthetic data remains an open question. The objective of this work is to evaluate the impact of replacing real images with synthetic ones on the performance of deep neural networks in runway segmentation and corner point extraction tasks. The contribution of this work is to provide an estimate of the size and composition of the dataset that combines real and synthetic images for training these networks. The findings will support future efforts in runway detection, offering insights that can aid in the creation of datasets and accelerate the development of future technologies, not only for aircraft-related technologies, but also for neural network training in general.