基于改进的PSO优化LSSVM参数的松花江哈尔滨段悬浮物的遥感反演

悬浮物是松花江水质和水环境评价的重要参数之一.利用在松花江哈尔滨段江面上29个采样点的实测高光谱和悬浮物浓度数据,用20个采样点数据为训练集,9个采样点数据为测试集.将机器学习和全局优化智能计算方法引入,应用改进的粒子群(PSO)优化最小二乘支持向量机(LSSVM)参数,以均方根误差RMSE为适应度函数,根据迭代得到LSSVM最优参数值,用700 nm和750 nm光谱反射率比值(R700/R750)为特征变量,悬浮物数据为目标变量,用训练集数据训练得到反演模型,使用测试集数据进行验证.结果表明,此模型收敛速度快,精度高,得到预测值的均方根误差RMSE为10.11 mg·L-1,平均绝对百分误差MAPE为10.72%,模型决定系数R2为0.952,该方法可用来对其它水质参数反演预测提供参照.

Remote-sensing retrieval of suspended solids based on improved PSO-LSSVM at the Harbin section of the Songhua River

Suspended solid is one of the most important parameters for evaluating water qualities and water environmental conditions of the Songhua River. In this study, both observed hyperspectral and suspended solids concentration data were used, which were derived from 29 samples at the Harbin section of the Songhua River. Among those data, 20 were served as training set and 9 were designated as testing set. In order to retrieve the suspended solids, machine learning and intelligent calculation method for global optimization were performed. Least squares support vector machine (LSSVM) parameters were optimized by improved Particle Swarm Optimization (PSO). Based on root mean square error (RMSE, as a proxy of fitness function), LSSVM optimal parameters were obtained with permutations. We defined the spectral reflectance ratios of 700 nm and 750 nm (R₇₀₀/R₇₅₀) as feature variables and the concentration data of suspended solids as target variables, and carried out the retrieval model from the training set. Afterwards, the retrieval model was evaluated by the testing set. The results demonstrated that the retrieval model had fast convergence rate and high precision with a low RMSE of predicted values (10.11mg·L^-1), a low MAPE(10.72%) and a high R² (0.952). In a word, the results suggested that the method can be used to provide reference for retrieval and prediction of other water quality parameters.