This work aims to characterize the accuracy of decoded motor unit activities during multiple contraction conditions based on electromyography (EMG) decomposition techniques, and to evaluate the performance of extracted neural features for the estimation of muscle activation. Twelve healthy undergraduates participated in the experiments to perform the isometric contraction of knee extension with four levels. The high-density EMG signals were decomposed into motor unit spike trains based on convolution kernel compensation. Two neural features were extracted for the cross-correlation analysis with force. On average, (7±4) motor units were identified from the medial vastus muscle (MVM), while (9±5) motor units were identified from the lateralis vastus muscle (LVM). The average pulse-to-noise ratio (PNR) was 30.1 dB, corresponding to the decomposition accuracy of over 90%. The average correlation coefficient between the two neural features of MVM and the force was (0.79±0.08) and (0.80±0.08), respectively, while the average correlation coefficient of LVM was (0.85±0.05) and (0.85±0.06), respectively. These results demonstrate the feasibility of the identification of motor unit activities under various contraction conditions, and the strong correlation between neural features and force indicates the application of decomposition techniques in rehabilitation, exercise training, and human-machine interfacing.