To improve angular resolutions and precision in radar object detection and estimation, designs of radar systems tend to employ more antenna elements or modules, for instance the millimeter wave (mmWave) radar for autonomous driving. However, as the number of antennas increases, the computational complexity for radar signal processing also increases drastically. An effective way to resolve such a dilemma is to select optimal antenna subarrays for individual objects under detection and estimation, which not only can constrain the computational complexity for subsequent radar signal processing, but also can maintain the precision of angular estimation. In view of the complexity of optimal subarray selections, existing results typically use greedy approaches to search for suboptimal solutions. More recently, deep learning (DL) methods are proposed to improve the correctness of subarray selection. Motivated by the recent advancements in DL, we propose herein a novel DL method of subarray selection for radar multi-target detection and angles estimations based on a one-dimensional Saak transformation. Compared to the existing DL method based on convolutional neural networks (CNN), the proposed method not only has a higher correctness in subarray selection, but also has a lower computational complexity. Simulation results also show that signal collections for subarray selection can be limited to two chirp signal durations, which significantly reduce the time delays for subsequent radar signal processing.