Haze frequently taints multispectral remote sensing photos, resulting in poor image quality. This study proposes a unique dehazing technique for multispectral remote sensing images based on a deep convolutional neural network (CNN) with residual structure. Before training a regression from the hazy image to the clear image, the residual structure is initially connected in parallel among several CNN individuals. The ultimate dehazing result is obtained by combining the outputs of CNN users and weight maps. The established CNN network's employees receive training using a variety of haze sample levels in order to acquire a variety of dehazing skills through the multi-scale convolutional mining of multi-scale haze features. The CNN individuals are fused together in an adaptive way, and the weight maps also adjust to the fuzzy distribution. The clear scene can be regained by adding a fuzzy image to the end-to-end network as intended. It is suggested to use a wavelength-dependent haze simulation technique to create labelled data for training the network and to create extremely accurate hazy multispectral images. The results of the trials show that, depending on the situation, the suggested strategy can effectively remove haze from each band of multispectral photos.