Machine Learning Based Drone Detection: Differentiating sUAS from Birds for Base Security

The U.S. Air Force needs an efficient method to differentiate small Unmanned Aerial Systems (sUAS) from birds, as the current detection system lacks the capability to make rapid and accurate threat assessments near military bases. sUAS pose potential security risks, carrying payloads or surveillance equipment that could compromise operations. To address this challenge, this project aimed to develop a machine learning algorithm for drone detection. The initial phase of the project focused on selecting appropriate sensors to provide high-quality data for the algorithm. However, midway through development, the focus shifted to optimizing classification algorithms, as the dataset was externally provided rather than self-collected. The primary approach involved implementing a smoothness metric algorithm alongside a Random Forest Classifier. The Random Forest Classifier was particularly effective due to its ability to prevent overfitting in a supervised learning context and handle a dataset comprising 116 features, later refined using the Select K-Best method for improved training efficiency. The performance of the system was evaluated using a confusion matrix, yielding a recall rate of 94%, indicating a high detection capability for drones. However, some false positives persisted due to the dataset’s inherent variance. Objects were classified as drones based on multiple criteria, including a specific dataset field indicating drone presence and a calculated smoothness metric within a defined range (0.2 to 0.5). While this variance occasionally led to false positives, it ensured that potential threats were not overlooked, with final classification decisions left to human operators. Future work includes the development of a graphical user interface (GUI) to enhance usability and further testing with an expanded dataset to improve classification accuracy.