古尔班通古特沙漠灌木冠幅预测模型

古尔班通古特沙漠是中国第二大沙漠，也是中国固定和半固定沙丘主要分布区，固沙灌木种较多。冠幅不仅是反映固沙灌木可视化的重要参数，也是反映沙漠植被生长情况的重要变量。以3种沙丘（固定沙丘、半固定沙丘和流动沙丘）上主要固沙灌木为研究对象，利用12种基础模型、BP（Backpropagation Neural Network）神经网络和支持向量机（Support Vector Machine，SVM）机器学习算法建立了基于固沙灌木株高和冠长率的冠幅预测模型，同时将两种机器学习算法拟合结果与基础模型进行比较，最终选出了适合研究区的冠幅预测模型。结果表明：（1）不同沙丘类型和不同灌木种类的最优冠幅预测模型不同，且固定沙丘和半固定沙丘模型优于流动沙丘。3种沙丘类型最优拟合为M2（Quadratic Model）模型；（2）白梭梭（Haloxylon persicum）在半固定沙丘和流动沙丘上拟合的最优模型分别为M2、M7（Gompertz），沙拐枣（Calligonum mongolicum）最优模型为M2，蛇麻黄（Ephedra distachya）和油蒿（Artemisia ordosica）在半固定沙丘和流动沙丘上拟合较优模型分别为M2、M7。总体来说，基础模型M2和M7可以较好地预测不同类型的灌木冠幅值；（3）基于径向基（Radial Basis Function）核函数的支持向量回归机的冠幅预测模型明显优于BP神经网络模型。

Canopy width prediction models for the Gurbantunggut Desert

The Gurbantunggut Desert， the main distribution area of fixed and semi-fixed dunes， is the second largest desert in China with a relatively large variety of sand-fixing shrubs. The canopy width is an essential parameter for the visualization of sand-fixing shrubs， which is also a necessary variable for the growth of desert vegetation. The purpose of this study was to assess the main sand-fixing shrubs on three types of sand dunes （fixed， semi-fixed， mobile dunes） in the sand zone. The experiment uses 12 base models， Backpropagation Neural Network （BP）， Support Vector Machine （SVM） machine learning algorithms to develop a canopy width prediction model based on the height and crown length rate of sand-fixing shrubs. To compare the two machine learning algorithm fits，canopy prediction models suitable for the experiment was selected. The results were as follows： （1） The different optimal canopy width prediction models had a difference in dune types and shrub species， and the fixed and semi-fixed dune models outperform the mobile dune models. The three dune types were optimally fitted to the M2 （Quadratic Model） model. （2） The optimal models fitted for Haloxylon persicum on semi-fixed and mobile dunes were M2 and M7（Gompertz） model， respectively. The best-fitting model for Calligonum mongolicum was M2 model. In contrast， the better-fitting models for Serpentine and Artemisia ordosica were M2 and M7 model on semi-fixed and mobile dunes， respectively. Generally， the models of M2 and M7 had a better predictors of different types of shrub crown width values. （3） The Radial Basis Function （RBF） kernel-based support vector regression machine was better than the BP neural network model in crown width prediction.