Deconvolution of magnetic remanence which has been continuously measured along a long-core sample is improved by eliminating the disturbance of severe intensity variation confined to a narrow layer. The deconvolution is constrained by the smoothness of the magnetization measured by the L (sub 2) -norm of second order difference. The previous model uses a single smoothness parameter for the entire sample and the optimum smoothness is obtained by minimizing ABIC (Akaike's Bayesian Information Criterion), while an additional parameter is introduced in the modified scheme for a layer where intensity fluctuation is strong. The modified deconvolution scheme was applied to 5-mm interval magnetic remanence data for two sections of ODP Hole 767B. We assumed another smoothness for the uppermost 1.5 cm of Section 767B-6H1 where the coring disturbance is severe. For Section 767B-6H2, the Australasian microtektite layer (108-113 cm), where the magnetization intensity shows a strong peak, is expressed by a different smoothness parameter from other part. In both cases the minimum ABIC was lowered and the calculated error was reduced, demonstrating the improvement of the model by incorporating additional information other than magnetic remanence data in the prior distribution of the data.