In the United States, the Federal Communications Commission has adopted rulespermitting commercial wireless networks to share spectrum with federalincumbents in the 3.5 GHz Citizens Broadband Radio Service band (3550-3700MHz). These rules require commercial wireless systems to vacate the band whencoastal sensor networks detect radars operated by the U.S. military; a keyexample being the SPN-43 air traffic control radar. For such coastal sensornetworks to meet their operating requirements, they require highly-accuratedetection algorithms. In addition to their use in sensor networks, detectionalgorithms can assist in the generation of descriptive statistics for librariesof spectrum recordings. In this paper, using a library of over 14,000 3.5 GHzband spectrograms collected by a recent measurement campaign, we investigatethe performance of three different methods for SPN-43 radar detection. Namely,we compare classical energy detection to two deep learning algorithms: aconvolutional neural network and a long short-term memory recurrent neuralnetwork. Performing a thorough evaluation, we demonstrate that deep learningalgorithms appreciably outperform energy detection. Finally, we apply thebest-performing classifier to generate descriptive statistics for the 3.5 GHzspectrogram library. Overall, our findings highlight potential weaknesses ofenergy detection as well as the strengths of modern deep learning algorithmsfor radar detection in the 3.5 GHz band.