Current evaluation methods for spectral similarity are mainly based on the shapes and amplitudes of spectra, but these two can only reflect the outline information of spectra, and cannot well reflect the fingerprint characteristics of spectra of ground objects. In order to better embody the application of spectral characteristics in the evaluation, it is proposed herein a method for evaluating spectral similarity based on first-order gradient information. Firstly, it is proposed an MSAM similarity evaluation method, further, a modified gradient spectral angle matching (MGSAM) method by adjusting the traditional spectral angle similarity evaluation method SAM. MGSAM compares the gradient angle matching degree of the two spectral curves. The gradient information of the spectral curves can highlight the existence of “fingerprint” characteristics such as spectral absorption peaks, so MGSAM can fully reflect the similarity of the spectral characteristics of the two contrast curves. By analyzing the influence of offset information and spectral depth on MSAM and MGSAM, it is pointed out that MGSAM has stronger robustness to offset information, and can objectively reflect the difference in spectral depth, so as to directly reflect the fidelity of spectral features in the photoelectric systems or related algorithms. By applying MGSAM as the evaluation method to the evaluation of compressed sensing imaging system, the simulation results showed that as the change of sampling rate, the MSAM values ranged between 0.998～1, while MGSAM values ranged between 0.72～1, with obvious change and great difference. It objectively reflects the fidelity ability of the compressed sensing system for spectral features and has a stronger differentiation ability, thereby providing a more objective evaluation method for such systems. By applying MGSAM to the classification of ground objects based on spectral similarity, and selecting Salinas, Pavia and Indian Pines for the test data, the results showed that the average classification accuracy based on MSAM was 0.86, while that based on MGSAM was 0.93. This shows that MGSAM can highlight the role of spectral features in classification and greatly improve the classification accuracy.