科学预测耕地保有量是耕地保护的基础，对缓解用地矛盾、保证粮食安全具有重要指导意义。为探讨不同核函数支持向量回归机（support vector regression，SVR）对耕地面积预测的影响，该文以惠州市为例，分别采用多元回归、BP神经网络及3种不同核函数SVR建立耕地面积预测模型并进行对比试验。预测结果精度分析显示，RBF核函数SVR预测结果平均相对误差为0.54%，均方根误差为0.007，精度最高；Sigmoid核函数SVR预测结果对应误差分别为1.12%及0.012，精度次之；多项式核函数SVR预测结果对应误差为分别为2.71%及0.032，高于BP神经网络模型，但低于多元回归模型。研究表明，在现有3种常用核函数SVR耕地面积预测模型中，基于RBF核函数SVR模型预测能力最强，其次是sigmoid核函数；而多项式核函数则效果较差。%Scientific prediction of cultivated land reserved quantity is important for cultivated land protection. Moreover, it provides guiding significance for relieving the contradictions of land use and ensuring food security. The purpose of this paper is to compare the prediction accuracy of different kernel based support vector regression (SVR), and provide a guideline for cultivated land area prediction. Taking Huizhou city for example, we apply different kernel based SVR to simulate the relationships between cultivated land area and impact factors of land use change. Seven impact factors, including population, socio-economics, industrial structure, living level, agricultural technique and policy, were selected by using the grey correlation method. With the socio-economic statistics of Huizhou city from the statistical yearbook and the policies which have been enacted during 1991 to 2010, corresponding cultivated land areas and influence factors were generated. Using data from 1991 to 2005 as training, SVR based on different kernel functions were employed to build the prediction model for cultivated land areas from 2006-2010. Finally, we apply multiple regressions, BP neural network and SVR based on different kernel functions to predict the cultivated land areas of 2006-2010. According to the predicted values and corresponding actual values, the average relative error and correlation coefficient and the root mean square error were used to validate the performance of different models. The analysis of prediction accuracies showed that the correlation coefficient of multiple regressions stayed at a high level, which reached to 0.970. But the average relative error (13.17%) and the root mean square error (0.173) were biggest. The accuracy of SVR based on polynomial kernel was greatly improved, especially for the average relative error and the root mean square error by comparison with that of multiple regressions. The accuracy of BP neural networks is between that of SVR based on polynomial kernel and SVR based on sigmoid kernel. However, for the BP neural networks based cultivated land area prediction model, the predicted value is difficult to ascertain and is prone to over fitting. The average relative error, the root mean square error and the correlation coefficient of SVR based on RBF kernel is 0.54%, 0.963 and 0.007, respectively. Therefore, it is obvious to find that the model of SVR based on RBF kernel can obtain best accuracy in predicting the cultivated land area, the predicting model of SVR based on sigmoid kernel follows. It is concluded that, in the existing three widely used kernel based SVR models, SVR based on RBF kernel is most suitable to be applied to predict the cultivated land areas, and SVR based on sigmoid kernel follows, while SVR based on polynomial kernels is worst.
展开▼
