For the traditional hyperspectral image classification algorithm’s problems of insufficient utilization of feature information and inability to reduce the spatial redundancy of the feature map effectively, an improved hybrid convolution-based multiscale model, MH-CNN, is proposed, which uses a multiscale 3DCNN module for the initial extraction of spatial and spectral features of hyperspectral images, and then adopts a multiscale 2DCNN network embedded with a spatial reconstruction module to the deep spatial features of the feature map is further extracted and optimized. Finally, the fully connected layer accurately calculates the hyperspectral remote sensing images. In this paper, the experiments are carried out on three open source datasets, including Indian Pines, Pavia Centre, and Pavia University, and seven classical classification methods are selected as comparisons and the overall accuracies of this paper’s MH-CNN algorithm on the three datasets reach 97.7%, 99.2%, and 98.5%, respectively. The experimental results show that the MH-CNN algorithm makes full use of both spatial and spectral features in hyperspectral images and, at the same time, effectively reduces the spatial redundancy of the feature maps, improves the classification accuracy compared with other models, and has better comprehensive performance.