| 摘要: |
| 养殖水体水质的优劣直接影响养殖对象的成长, 准确、快速、全面地掌控养殖水环境的水质参数变化情况具有重要意义。传统的水质指标监测方法都通过人工采样的方式, 不仅耗费时间长, 且只能体现局部水体情况。针对这些问题, 提出了一种乌鸦搜索算法(CSA)结合偏最小二乘回归(PLSR)的高光谱特征波段筛选方法, 快速构建回归模型, 实现光谱数据的精准预测反演。以连片的养殖小区为研究对象, 采集养殖水体样本并拍摄同时期的高光谱影像数据。首先对提取的采样点光谱数据利用多种数据变换方法分别预处理; 其次利用这些数据, 对水质指标总氮(TN)、氨氮(NH4+-N)、总磷(TP)和化学需氧量(COD)分别构建全波段的SVR和AdaBoost回归模型, 同时与提出的CSA-PLS自动筛选波段方法和传统的连续投影算法(SPA)筛选波段后构建的模型进行比较分析; 最后根据决定系数(R2)和均方根误差(REMS)选出适合各水质指标的最优模型。从实验结果可以看出, 所提波段筛选方法的AdaBoost模型预测结果优于SVR和传统SPA方法提取特征波段后构建的模型, 与全波段最优模型相比, 在评价指标R2和RMSE上TN提升了18.32%和10.73%; NH4+-N提升了17.42%和11.19%; COD提升了2.15%和2.54%。结果表明, 基于CSA-PLS算法的光谱波段自动筛选方法结合AdaBoost构建的预测反演模型是有效、可行的, 具有较高的精准度, 为实现养殖水环境实时准确的预警调控提供了一种新的数据预测模型。 |
| 关键词: 高光谱数据 水质预测 乌鸦搜索算法 养殖水环境 集成学习 |
| DOI：10.11693/hyhz20231100245 |
| 分类号:X87;X832 |
| 基金项目:浙江省重点研发计划项目,2022C02027号;国家重点研发计划项目,2019YFD0900302号;浙江省公益技术研究项目,LGN20C190004号;湖州市自然科学资金项目,2021YZ18号。 |
附件 |
|
| FAST HYPERSPECTRAL PREDICTION AND INVERSION MODEL OF AQUACULTURE WATER QUALITY BASED ON CSA-PLS ALGORITHM |
|
MA Qi-Liang1, LIU Mei2, QI Heng-Nian1,3, YANG Xiao-Ming1, YUAN Ju-Lin2
|
|
1.Huzhou University, Zhejiang Provincial Key Laboratory of Smart Management & Application of Modern Agricultural Resources, Huzhou 313000, China;2.Zhejiang Institute of Freshwater Fisheries, Huzhou 313001, China;3.School of Information Engineering, Huzhou University, Huzhou 313000, China
|
| Abstract: |
| The quality of aquaculture water directly affects the growth of aquatic products, and it is of great significance to accurately, quickly and comprehensively control the water quality parameters changes of the aquaculture water environment. The traditional monitoring methods of water quality index are all by means of manual sampling, which not only takes a long time, but also can only reflect the local water conditions. To solve these problems, a crow search algorithm (CSA) combined with partial least squares regression (PLSR) was proposed to select hyper spectral characteristic bands, which can quickly predict and invert spectral data. In this paper, we collected water samples and took hyperspectral image data of the same period in a contiguous aquaculture zones. Firstly, various data transformation methods are applied to preprocess the sampling points spectral data extracted; Secondly, using these data, the SVR and AdaBoost regression models are separately constructed for the water quality indicators: Total Nitrogen (TN), Ammonium Nitrogen (NH4+-N), Total Phosphorus (TP), and Chemical Oxygen Demand (COD) across the entire spectrum. At the same time, the model is compared with the model constructed after the proposed CSA-PLS automatic band screening method and the traditional Successive Projections Algorithm (SPA) band screening method; finally, the best models suitable for each water quality indicator are chosen based on the coefficient of determination (R2) and root mean square error (RMSE). From the experimental results, it could be seen that the TN, NH4+-N, TP and COD prediction models that trained by the proposed waveband selection method and the Adaboost model perform better than SVR in predicting TN, NH4+-N, TP, and COD. The optimal prediction model for these parameters outperforms the traditional SPA band selection method in terms of evaluation criteria R2 and RMSE, compared with the optimal model using full spectra, the TN prediction improved by 18.32% and 10.73%; NH4+-N improved by 17.42% and 11.19%; COD improved by 2.15% and 2.54%. The results indicate that the prediction and inversion model based on CSA-PLS and AdaBoost is effective and feasible, which provides a new data acquisition model for real-time and accurate early warning and regulation of aquaculture water environment. |
| Key words: hyperspectral data water quality model prediction crow search algorithm (CSA) aquaculture water environment integrated learning |
