Abstract: ‍ ‍The emerging artificial neural networks in recent years have shown their great potential in signal processing owing to the strong self-learning adaptability and parallel information processing capability. Reservoir computing is a bio‑inspired paradigm derived from the recurrent neural network， featuring excellent performances in classification and temporal data processing tasks. In this paper， we propose and implement two different reservoir computing paradigms based on the MEMS resonators with different nonlinear responses. Furthermore， orienting to specific radar applications including signal prediction， image recognition， and feature recognition， we employ classic datasets， such as NARMA， MNIST-handwritten digits， LFM pulse waveforms to benchmark the proposed two different reservoir computing paradigms. Meanwhile， the results also demonstrate the potentials of the physical reservoir computing based on the nonlinear MEMS resonators in applications including radar signal prediction， classification and feature extraction. Moreover， this paper also provides a new powerful tool for radar signal processing in complex electromagnetic environment， and paves the way to the interdisciplinary research of MEMS sensing technology and radar signal processing technology.