In order to better preserve high-frequency information in demosaicing multispectral images, we propose a new demosaicing method for multispectral images based on an improved guided filter. Firstly, the strong correlation between adjacent pixels based on the autoregressive model is constructed, gradually estimates the model parameters at each pixel, and the optimal estimation value is obtained by minimizing the estimation error in the local window, interpolates the sampling dense band G, and generates high-quality guide images. The windowed intrinsic variation coefficient is then introduced into the penalty factor to obtain a weighted guide filter with edge sensing ability and to reconstruct the remaining sparse sampling bands. Finally, the CAVE dataset and the TokyoTech dataset are used for simulation. The experimental results show that compared with the mainstream five-band multispectral image demosaicing method, the peak signal-to-noise ratio and structure similarity of the reconstructed image in the CAVE dataset and the TokyoTech dataset are improved by 3.40%, 2.02%, 1.34%, 0.30% and 6.11%, 5.95%, 2.28%, 1.42%, respectively. The local structure and color information of the original image are also better preserved, and the edge artifacts and noise are reduced.