基于离散粒子群和偏最小二乘的湖库型水源地水体悬浮物浓度和浊度遥感反演方法

Remote-sensing retrieval method of suspended solid concentration and turbidity in lakes and reservoirs based on discrete particle swarm and partial least squares

用离散粒子群优选反演悬浮物浓度和浊度的特征波段,可减少水质参数偏最小二乘反演模型的输入参数个数和建模过程的不确定性,提高偏最小二乘反演模型的预测精度。以2014年7月21—23日在南四湖获取的水体实测光谱数据和同步水质分析数据为例,分别建立南四湖水体悬浮物浓度和水体浊度的偏最小二乘(PLS)反演模型和离散粒子群-偏最小二乘(NDBPSO-PLS)反演模型并进行验证。结果表明:经NDBPSO优选之后,反演悬浮物浓度和浊度的NDBPSO-PLS模型输入特征波段由PLS模型的370个分别减少到127个和134个,输入特征变量由PLS模型的60个减少到21个,反演悬浮物浓度和浊度的NDBPSO-PLS模型建模精度和预测精度均优于PLS模型。该方法可以有效提高PLS模型反演水体悬浮物浓度和浊度的精度。

Discrete binary particle swarm optimization （DBPSP） has been used in this study to prioritize the characteristic bands of suspended solid concentration and turbidity. This method can reduce the number of input parameters and uncertainty in partial least squares （PLS） modeling, and improve the prediction accuracy of PLS models. In a case study of water quality of the Nansi Lake, we have developed and verified a PLS model and a NDBPSO-PLS model separately for retrieval calculations of suspended solid concentration and turbidity using the hyper-spectral remote sensing data and synchronous sampling data measured on July 21 to 23, 2014. The results indicate that for the NDBPSO-PLS calculations of suspended solid concentration and turbidity, the number of characteristic bands required by it was only 137 and 134 respectively, a great reduction relative to the number of 370 required by the PLS model. And its input feature variables were reduced to 21 from 60 PLS variables. Its prediction accuracy is better than the PLS model. Thus, NDBPSO-PLS would provide an effective method for improvement of the retrieval calculations of suspended solid concentration and turbidity.