支持向量机在湖库营养状态识别中的应用

Application of support vector machine to lake and reservoir trophic status recognition

依据我国湖库富营养化评价标准和支持向量机(SVM)原理及方法,构建基于交叉验证(CV)的CV-SVM湖库营养状态识别模型,采用随机内插的方法在各分级标准阈值间生成训练样本和测试样本,在达到预期识别精度后将模型运用于全国24个湖库营养状态的识别,并与投影寻踪法、评价指标法和神经网络评价法的识别结果进行比较。结果表明:基于线性核函数的CV-SVM模型对于随机生成的训练样本和测试样本的正确识别率分别达到97.8%和97.3%(5次平均),对全国24个湖库营养状态的识别结果与采用投影寻踪法、评价指标法和神经网络评价法的识别结果基本相同,模型具有泛化能力强、识别精度高、收敛速度快、不易陷入局部极值等特点。

According to China＇ s lake and reservoir eutrophication assessment standards and the support vector machine （SVM） theory and method, a CV-SVM lake and reservoir trophic status recognition model was constructed based on cross-validation （CV）. With the interpolation method, the training samples and testing samples were randomly generated within the classification threshold. After the model＇ s desired accuracy was achieved, it was applied to the recognition of trophie status of 24 lakes and reservoirs nationwide and compared with the projection pursuit, evaluation index, and neural network evaluation methods. The results are as follows： the recognition rate using the CV-SVM model based on a linear kernel function reached 97. 8% and 97. 3% （for five times on average） for the randomly generated training and testing samples, respectively. The trophic status recognition results of the 24 lakes and reservoirs were basically consistent with those obtained by the projection pursuit, evaluation index, and neural network evaluation methods. The mode/ has the advantages of high generalization ability and recognition accuracy, fast convergence, and the unlikely occurrence of a local minimum.