The measurement accuracy of magnetic gradient tensor system is limited by the system errors such as zero drift, sensitivity difference and the three axis nonorthogonality of the magnetic sensor, as well as the misalignment errors among the axis systems of different sensors. In this paper, an error parameter linear model of the planar cross magnetic gradient tensor system is constructed, and a twostep linear calibration method is proposed. Firstly, the linear equation set of single magnetic sensor system error is constructed by using two nonlinear variable conversions, and the error parameters are estimated by least squares method. The actual output of the sensor is calibrated to its ideal orthogonal output. Secondly, the linear equation set of the misalignment errors among the ideal orthogonal axes of the sensors is constructed by using the rotation matrices and the least square solution is obtained. The outputs of sensors are calibrated to the orthogonal coordinate system in reference platform frame. Both processes have no any mathematical simplification. Simulations and experiments show that the proposed twostep linear calibration method of the tensor system is more accurate compared with the conventional linear calibration method neglecting the second or higher order small quantities. The simulated estimation accuracy of the error parameters is better than 93%, the root mean square error of the actually tested total field intensity is less than 13 nT and the root mean square error of the tensor components is less than 90 nT/m in the experiments after calibration.