Abstract: High accuracy soil classification and model-based mapping can provide a better evaluation for sustainable management in intelligent agriculture. Nevertheless, most previous studies on the type classification and mapping were derived mainly from hyperspectral images that captured indoor measurement, or airborne hyperspectral and spaceborne multispectral images. Only a few studies were associated with the satellite hyperspectral images in soil classification and mapping, due partly to the lack of satellite hyperspectral data. Thanks to the Gaofen-5 (GF-5) satellite, the high-resolution hyperspectral data can currently be returned, and thereby can make it possible to realize soil classification and mapping using satellite hyperspectral images. However, it has become a great challenge to achieve a high accuracy mapping only by using soil spectral characteristics. Alternatively, the introduction of soil formation factors can effectively improve the accuracy of soil classification. In this study, taking the Baiquan and Mingshui county as the study areas, the GF-5 satellite hyperspectral data was selected to verify whether the combination of satellite hyperspectral data and terrain data (one of the five factors of soil formation) can enhance the performance accuracy of soil classification and mapping. The continuum removed (CR) analysis was performed on the hyperspectral reflectance data (OR), and then the obtained CR and OR data were processed by a principal component analysis, in order to gain the inputs of OR-PCA (Principal Component Analysis) and CR-PCA. Subsequently, some terrain (TA) factors were added to the program of OR-PCA and CR-PCA, particularly including elevation, slope, aspect, curvature, and relief degree of land surface. As such, the soil hyperspectral remote sensing classification models were established in combination with a random forest classifier. The results are as follows: 1) The spectral phase differences among various soil types increased significantly after the continuum removal treatment, indicating a great improvement on the accuracy of the hyperspectral soil classification. Compared with that of OR as input data, the correct classification number of Phaeozems, Chernozems and Cambisols increased by 20, 17 and 27, respectively. The total accuracy and Kappa coefficient were also improved by 5.48% and 0.12, respectively. The findings demonstrated that the CR treatment can be used to greatly increase the accuracy of soil classification and mapping. 2) The PCA method can be used to dramatically reduce the redundancy data of hyperspectral images, while improve the classification efficiency and the accuracy of model. In the mapping data, the misclassification of pixels was significantly reduced after the PCA treatment. Compared with that of OR as the input, the total accuracy of soil classification in the OR-PCA method was improved by 1.71%, and the Kappa coefficient was enhanced by 0.02. Compared with that of CR as the input, the total accuracy of soil classification in the CR-PCA method was improved by 3.67%, and the Kappa coefficient was improved by 0.02. 3) The obtained results indicated that the addition of TA factors can significantly improve the accuracy of soil classification. Among the different input combinations, the CR-PCA-TA group showed the highest accuracy of soil classification when using as the input data, with the classification total accuracy of 81.61% and the Kappa coefficient of 0.72. Compared with that of CR-PCA as the input, the total accuracy of soil classification based on the combination of CR-PCA-TA was improved by 13.01%, and the Kappa coefficient was improved by 0.20, indicating that the high precision soil classification model and soil mapping were realized during this time. The findings can provide new insightful ideas for soil classification and mapping with a high accuracy in a wide range for intelligent agriculture.