A landform image classification algorithm based on sparse coding and convolutional neural network is proposed. The non-subsampled Contourlet transform is applied to the training samples for multi-scale decomposition. The images are selected in the training samples to learn the local features by using sparse coding, and the feature vectors are sorted. The feature vectors with larger gray-scale mean gradients are selected to initialize the convolutional neural network convolution kernel. The results show that the proposed algorithm can obtain better classification results than traditional underlying visual features, which effectively avoids the problem of network training falling into local optimum, and improves the classification accuracy of unmanned aerial vehicles landing landform in natural scenes.