Analyzing Performance Discrepancies: U-Net vs TransUNet for Aircraft Emergency Landing Site Detection

Abstract

In the context of aviation, forced landings are unwanted events that can happen to an aircraft during its flight trajectory. They can be due engine malfunctions, adverse weather conditions and other sudden situations. For this reason, and to ensure passengers' safety, it is imperative to develop methods and procedures to detect potential sites that can be used as emergency landing areas during these crisis situations. Traditionally, pilots use visual indicators to detect such landing sites, this ability can varry from a pilot to another depending on experience, aircraft altitude and other environmental conditions. Such circumstances can make this visual detection task highly difficult. Image segmentation is one of the possible solutions that can be implemented in identifying potential emergency landing sites for aircraft. Precise segmentation should improve on the effective identification of safe landing areas, thereby enhancing aviation safety protocols in general. In this context, the traditional U-Net \cite{unet} architecture has shown exceptional results regarding segmentation tasks. However, a new approach derived from U-Net and incorporating transformers in its encoder, known as TransUNet, has demonstrated promising results, surpassing in some cases those of U-Net. This study investigates the performance of TransUNet compared to traditional U-Net for aircraft emergency landing site detection. Both architectures were implemented, trained, and evaluated using our novel dataset tailored for this purpose. Our work demonstrate that U-Net outperforms TransUNet in terms of accuracy and computational efficiency in this specific segmentation task. In particular, U-Net exhibited superior performance by improving segmentation precision from 80% up to 88% in the testing set. Moreover, the mean Intersection-Over-Union, a metric for segmentation accuracy, have also seen an improvement of 77% for U-Net over 73% for TransUNet. These results emphasise the power of the traditional U-Net architecture for this critical application, underlying its practical relevance in enhancing aviation safety.