基于高光谱与改进BP神经网络的水体生化需氧量(BOD)估算

Estimation of Biochemical Oxygen Demand(BOD)Content in Water Bodies Based on Hyperspectral and Improved BP Neural Network

生化需氧量(BOD)是用于评估水质及其污染程度的重要指标,水体BOD含量对于保护水资源、改善水环境以及维护生态平衡具有重要意义。为解决BOD传统测量存在设备昂贵易损、操作专业性强、时间周期长等问题,分别使用7种不同的高光谱预处理方法对100组水体BOD光谱透射率数据进行处理,再将高维度的光谱数据通过PCA降维,将主成分数降低至10,建立BP神经网络模型,通过比较模型的精度,拟合程度和运行效率筛选最优模型,建立了一种结合不同高光谱预处理技术和PCA降维技术的改进BP神经网络模型来实现对BOD值精准且快速的估算。结果显示,相比于基于原始数据建立的BP神经网络模型,7种结合预处理方法建立的模型在保证估算精度提高的情况下,训练时间上都有一定程度的下降,在经过PCA降维处理后又有进一步下降,其中PCA降维后的SNV-BP模型训练时间相较于未改进模型的7.4658 s,下降至3.9341 s,证明了高光谱预处理和PCA降维对高光谱数据中冗余信息有良好的去除能力,能够同时对光谱数据起到优化作用,使得BP神经网络回归模型在整体上得到性能提升,其中PCA降维后的VN-BP改进模型估算效果最为显著,保证效率的同时,均方根误差RMSE低至0.02425,R 2高达0.9929,说明改进模型为高效准确地监测估算水体BOD含量提供了一种可能的思路。

Biochemical oxygen demand(BOD)is an important indicator used to assess water quality and its pollution level,and the BOD content of water bodies is important for protecting water resources,improving the water environment and maintaining ecological balance.In order to solve the problems of expensive and fragile equipment,professional operation and long time cycles in traditional BOD measurement,seven different hyperspectral preprocessing methods were used to process 100 sets of BOD spectral transmittance data in water bodies.Then the high-dimensional spectral data were dimensionally reduced by PCA,reducing the principal fraction to 10.A BP neural network model was established,and the optical model was selected by comparing the accuracy,fitting degree,and operational efficiency of the model.An improved BP neural network model combining different hyperspectral pre-processing techniques and PCA dimensionality reduction techniques was proposed to achieve accurate and fast estimation of BOD values.The results showed that compared with the BP neural network model based on the original data,the training time of the seven models combined with the preprocessing method has been reduced to a certain extent while ensuring the improvement of the estimation accuracy,and further reduced after the PCA dimensionality reduction process.Specifically,the training time of the PCA dimensionality reduced SNV-BP model reduced to 3.9341 s compared with 7.4658 s of the unimproved model.This demonstrated that the hyperspectral preprocessing and PCA dimensionality reduction can remove redundant information from hyperspectral data and optimize the spectral data at the same time,so that the overall performance of the BP neural network regression model can be improved.The VN-BP improved model after PCA dimensionality reduction has the most significant estimation effect,and the RMSE was as low as 0.02425 and R 2 as high as 0.9929.The improved model provides a possible idea for efficient and accurate monitoring and estimation of BOD content in water bodies.