Multispectral remote sensing images are often degraded by clouds, resulting in the reduced efficiency and accuracy of image interpretation. Thin cloud removal is one of the most important and significant tasks for optical multispectral images. In this article, we propose a novel thin cloud removal method for multispectral images, which is a combination of traditional methods and deep learning methods. First, we adopt U-Net to estimate the reference thin cloud thickness map of the cloudy image. Then, a convolutional neural network named Slope-Net is designed to estimate the thickness coefficient of each band relative to the reference thin cloud thickness map to obtain the thin cloud thickness maps of different bands. Finally, the recovered clear image can be obtained by subtracting the thin cloud thickness maps from the cloudy image according to the traditional thin cloud imaging model. To train U-Net and Slope-Net, a wavelength-dependent thin cloud simulation method is presented to generate a labeled dataset composed of synthetic cloudy images, corresponding clear images, reference thin cloud thickness maps, and thickness coefficients. Qualitative and quantitative comparison experiments are conducted on both synthetic cloudy images and real cloudy images from the Landsat 8 Operational Land Imager. The results indicate that the proposed method can effectively remove thin clouds in multispectral images with various land cover types and maintain good color fidelity.