In this study, a multilayer perception convolutional neural network (MPCNet) was proposed for the pixel-level classification of multispectral remote sensing images, which combines the spectral information and spatial structure features of pixels. The performance of a land-cover-classification algorithm was tested based on the Jilin-1 spectral satellite (Jilin-1GP) images in the Nashik research area, India. To ensure high reliability of the experiment, the Landsat8, Sentinel-2A, and HJ-1A images were used within the same time interval for synchronized classification to perform qualitative and quantitative evaluations. Moreover, three current popular algorithms, namely, support vector machine(SVM), LightGBM, and shallow convolutional neural network(CNN), were selected to compare the algorithm performance. The experimental results indicate that the overall classification accuracy on the Jilin-1GP images can reach 94.0%-95.8%, and the Kappa coefficient can reach 0.932-0.948. The overall classification accuracy of the MPCNet increase by 3.7 percentage compared with that of the shallow CNN, which exhibits high accuracy.