Electromyogram (EMG) pattern-recognition (PR) is the most widely adopted prostheses/rehabilitation robots control method that seamlessly support multi-degrees of freedom (MDF) function in an intuitive fashion. The feature extraction framework applied in such PR-based control essentially determines the control performance of the prosthetic device. Based on the drawbacks of the commonly utilized feature extraction methods, this study proposed a spatio-temporal-based feature set (STFS) that might optimally characterize EMG signal patterns even in the presence of white Gaussian noise (WGN) to realize consistently accurate and stable decoding of multiple classes of limb-movements. For benchmark evaluation, the performance of the proposed STFS method was examined in comparison to notable existing popular methods using high density surface EMG recordings from 8 amputees, with metrics such as classification error (CE) and feature-space separability index. Compared to the existing methods, the STFS recorded substantial reduction of up 16.73% even in the presence the inevitable WGN at p<0.05. Also, with principal component analysis concept, the proposed STFS feature-space indicates obvious class separability compared to the previous methods. Therefore, the newly proposed STFS method could potentially facilitate the realization of consistently accurate and reliable PR-based control for MDF prostheses/rehabilitation robots.