Abstract: ‍ ‍Terahertz （THz） communication is one of the most promising candidate techniques for future ultra-high-speed wireless communications and has attracted wide attentions from academia and industry. In this paper， a novel THz communication system based on simple spatial sparse modulation—THz spatial modulation （THz-SM） was proposed， and the performance degradations caused by signal distortions， which result from transceiver hardware impairments， were investigated. According to the a posteriori maximum criterion， the maximal likelihood （ML） detection method was derived for the proposed THz-SM based on the multi-dimensional noise model. In addition， considering the application of THz to future multi-domain and multi-dimensional communication scenarios， where the conventional detection methods may be unsuitable and the imposed complexity is usually high， an extreme learning machine （ELM）-based intelligent detection algorithm having simple structure and low computational complexity was proposed. Simulation results demonstrated that the proposed ML algorithm outperforms the conventional ML， and the performance of the proposed ELM-based scheme approaches that of the optimal scheme， also are better than these of deep neural networks （DNN）-based and support vector machine （SVM）-based schemes.