Objectives  Micro landform classification of grid digital elevation model (DEM) is the foundation of digital landform refinement application and has broad application prospects in pedology, urban planning, civil engineering, military affairs, diplomacy, etc. However, problems, e.g., low degree of automation and incomplete classification, are still existed in the micro landform classification method of grid DEM based on regular knowledge. With the advantages of convolutional neural network (CNN), a CNN method for grid DEM micro landform classification is constructed and its automated implementation process and approach are created.

  Methods  Taking the advantage of CNN that can automatically learn the sample data and mine the hidden knowledge of the data set, it can automatically learn the features of the micro landform data set and realize the automatic classification of the micro landform combining with the micro landform data. Firstly, the plane curvature, section curvature, slope and elevation of grid DEM in the experimental area are extracted as data sets. According to the decision table and prior knowledge, the hill-position of the sample data set is divided into 6 types, including summit, shoulder, back-slope, foot-slope, toe-slope and alluvium. Among which, 2/3 data are used as the training set, and the remaining sample data are used as the test set. According to the data requirements of the CNN and the characteristics of the data set, a convolution layer is constructed, which can automatically learn the data characteristics in training set, determines the training model, and uses the test set to evaluate the model?s accuracy. After the model meets requirements of the accuracy, the new micro landform factor generalizes the model are entered and the results of landform classification are outputted, i.e., the model is used to automatically classify hill-position of the grid DEM data.

  Results  CNN is used for grid DEM micro landform classification. The experimental results show that the overall accuracy of the model test dataset established by the selected typical sample points in the experimental area reaches 97.95%, and the Kappa index is 0.974 9, which demonstrates the applicability of the CNN algorithm in micro landform classification. CNN method has a strong dependence on landform category knowledge contained in typical sample points. Incomplete description of local shape category feature knowledge will affect the CNN algorithm in mining implicit knowledge directly, resulting in inconsistent adaptability to 6 types of micro landform. The highest classification accuracy of alluvium is 99.64%, while the lowest are toe-slope with 92.95%, and the accuracy rates of summit, shoulder, back-slope and foot-slope are 95.89%, 99.58%, 99.08% and 98.87%, respectively. The incomplete degree based on superposition analysis is 35.91%, while the method based on CNN does not exist incomplete. The reason is that superposition analysis is a classification method of multi-factor system for each grid node. For different regions, the landform complexity is inconsistent, and the incomplete degree range is uncertain. The CNN method takes its advantage of back propagation to solve the problem of incomplete classification. The classification accuracy of CNN method for micro landform classification of grid DEM data with different pixel sizes is different, among which, the average classification accuracy of the pixel size of 5 m is the highest with 97.67%. The classification accuracy rate with the pixel size of 10 m is relatively lower with 96.09%. The grid DEM with the pixel size of 5 m has the lowest accuracy dispersion of classification results of each category among the 6 types of hill-position, and shows strong adaptability to all type of landform. In summary, grid DEM with the pixel size of 5 m is more suitable for automatic classification of micro landform based on CNN.

  Conclusions  Automatic classification of hill-position realized by CNN model is suitable for grid DEM micro landform classification. Compared to the classification method based on rule knowledge, the complicated data overlay analysis process is avoided and the completeness is improved. The overall accuracy of the model reaches 97.95%, among which, the highest classification accuracy is alluvium with 99.64%, while the lowest is toe-slope with 92.95%. Considering the influence of the pixel size of the grid DEM data on the classification accuracy, the grid DEM of 5 m × 5 m is more adaptable than 2.5 m × 2.5 m and 10 m × 10 m.