Mental Workload Level (MWL) is an important indicator reflecting the human’s cognitive state in human-computer interaction, such as the applications of vehicle driving and mental state assessment. A useful method of MWL assessment is by utilizing functional near-infrared spectroscopy (fNIRS) to monitor the brain tissue blood oxygen concentration during human’s cognitive process. However, existing fNIRS-based MWL classification methods still have limited performance in terms of classification accuracy and interpretability, since they have not precisely extracted the feature describing human’s cognitive behavior, e.g., cerebral hemodynamics. To address this issue, this paper proposes a new dynamical feature extraction and pattern recognition method for MWL classification by using Dynamic System Theory (DST) and Deterministic Learning (DL) technique. A so-called dynamical feature is extracted by modeling the cerebral hemodynamics using fNIRS in human’s cognitive process, which is mathematically interpretable and specific to the human’s MWL according to DST. Dynamical pattern recognition scheme is then proposed by combining DL and Support Vector Machine, aiming to identify high and low levels of MWL with the dynamical feature of fNIRS. Compared to methods such as EEGNet, DeepConvNet, DCNN, SVM, fNIRS-PreT, and Logistic Regression, our method achieved a performance improvement of 2% to 6% for classifying mental workload levels.