Performance Comparison of Drone and Bird Classifier Trained and Tested on Monostatic and Bistatic L-band Staring Radar Data

Methods of training deep learning models for target classification are proliferating. Data availability in radar is a constant concern despite advancements in low-sample-support training and synthetic data generation. This work explores the effect on classification performance when changing the train and test data that is obtained using networked radar. A dataset of real birds and drone flights was collected with two staring radars simultaneously operating in monostatic and bistatic mode, keeping all else identical. Spectrograms of the same target in bistatic and monostatic mode hold minor differences dependant on the quasi-monostatic geome-try of bistatic operation. The following investigation tests whether an unseen monostatic test set is found to have similar performance when training with an equivalent bistatic dataset, and vise versa, and also with a mix of data. Monostatic training of the classifier had a 13.9% fall off in overall classification accuracy when testing on bistatic data instead of monostatic data. Inversely, training with bistatic data led to only a 0.3% fall off of monostatic test data compared to bistatic test data. Best results were found with mixing bistatic and monostatic data together for training, which is ideal for distributed multi-static radar nodes for simplicity of training protocol.