Our project, Emergency Landing Airport Locator is going to address the issue of finding the nearest, valid airport to land at for airplanes (commercial or not) in distress to take into consideration when deciding where to emergency land. Currently, pilots use a Flight Management System (FMS) to plan their route and which airport they will land at before they take off. Then, when they approach the airport, Air Traffic Control (ATC) will direct them to the exact runway they need to land on [6]. The problem arises when, during an emergency, a pilot needs to manually reconfigure their FMS whilst under extreme stress. The goal of our project is to build a model to assist them with this part of the process.

We have identified two datasets that we have used in conjunction to build this model. The first is a list of airports, their IDs, and their locations. The second is runways by airport ID, their length, the types of aircraft that can land there, and more. By matching these two sets by Airport ID, we can build a list of runway locations to feed into our model.

Overall, this exploration into the viability of using machine learning to help decide landing locations proved extremely insightful and highlighted a multitude of benefits and drawbacks that the use of ML in the decision-making process can bring. In particular, while our original models were able to predict a viable location for landing, there were many instances where the model would suggest something entirely infeasible, potentially confusing pilots should they use it. Much of this inaccuracy came from our original, incorrect application of KNNs as well as some issues in the testing set since some test planes had no viable airports at all given they were in the middle of the ocean. Furthermore, while the exploration into separating the collection of airports into regions of viability for a plane using K-Means did reduce the numerical input under certain clustering numbers, the accuracy of the model’s predictions thereafter only suffered. Additionally, though perhaps able to be engineered in the future, SVMs also proved a dead-end for us as they introduced a significant amount of computational complexity to our system and, to our knowledge, were unable to be trained properly on our dataset.

Read more here: https://github.gatech.edu/pages/tdeshmukh8/CS4641-Group-104/