In recent years, UAV remote sensing has been used to estimate water quality parameters, such as suspended solids (SS), turbidity (TUB), and chlorophyll-a (chl-a) levels, due to its low cost, convenience, and high resolution. The matching pixel-by-pixel (MPP) algorithm is one of the methods to find the optimal regression equation for retrieving water quality parameters from UAV images. However, MPP has a high computational burden and commonly experiences overfitting problems. Therefore, we propose an improved MPP algorithm (called IMP-MPP) to solve the above problems by sampling pixels based on clustering results and selecting models with more filtering conditions. In this study, Qingshan Lake, Hangzhou City, Zhejiang Province, China, is taken as the study area to evaluate the suspended solids and turbidity indicators. A total of 45 in situ samples and the UAV images around the sampling points are analyzed and processed by the proposed IMP-MPP algorithm, along with the average value method and MPP for comparison purpose. The experimental results show that the determination coefficient, average relative error and comprehensive error of the best inversion model derived by the IMP-MPP algorithm for SS are 0.8255, 15.08%, and 0.1981, respectively. The determination coefficient, average relative error and comprehensive error of the best inversion model for TUB are 0.8311, 16.49% and 0.2033, respectively. The results suggest that the IMP-MPP algorithm is promising in finding more accurate inversion models for SS and TUB. Finally, based on the optimal models, suspended solids and turbidity distribution maps are generated for further research.

近年来，由于其低成本，方便和高成本，无人机遥感技术已用于估算水质参数，例如悬浮固体（SS），浊度（TUB）和叶绿素a（chl-a）水平解析度。逐像素匹配（MPP）算法是找到用于从无人机图像中检索水质参数的最佳回归方程的方法之一。但是，MPP具有较高的计算负担，并且通常会遇到过拟合问题。因此，我们提出了一种改进的MPP算法（称为IMP-MPP），通过基于聚类结果对像素进行采样并选择具有更多过滤条件的模型来解决上述问题。本研究以浙江省杭州市青山湖为研究对象，对悬浮物及浊度指标进行了评价。提出的IMP-MPP算法以及平均值方法和MPP进行了分析和处理，总共对45个原位样品和采样点周围的无人机图像进行了处理。实验结果表明，由IMP-MPP算法推导的SS最佳反演模型的确定系数，平均相对误差和综合误差分别为0.8255、15.08％和0.1981。最佳反演模型的确定系数，平均相对误差和综合误差分别为0.8311、16.49％和0.2033。结果表明，IMP-MPP算法有望为SS和TUB找到更准确的反演模型。最后，基于最佳模型，生成了悬浮固体和浊度分布图，以供进一步研究。