玉米叶片重金属铅含量的SVD-ANFIS高光谱预测模型

重金属污染农作物后可通过食物链进入人体从而严重危害身体健康。如何快速准确地监测农作物中重金属含量已成为当今生态与粮食安全等领域的重要研究内容。常规的生化监测方法存在操作繁琐、过程长、具有破坏性等缺点，而高光谱遥感具有光谱分辨率高、信息量大、生化反演能力强、方便快捷、对监测对象无损伤等优势，因此利用高光谱遥感技术监测农作物中重金属含量已成为遥感领域的热点研究之一。以不同浓度Pb(NO₃)₂溶液胁迫下盆栽玉米植株为研究对象，基于不同铅离子(Pb2+)胁迫梯度下玉米叶片的反射光谱及其中Pb2+含量的测定数据，结合奇异值分解(SVD)理论和自适应模糊神经网络推理系统(ANFIS)结构，建立了一种Pb2+含量预测的SVD-ANFIS模型。首先对各胁迫梯度下玉米的老叶(O)、中叶(M)、新叶(N)三种叶片的反射光谱数据进行SVD处理，获取原始光谱信息的奇异值；然后选择O，M和N叶片对应的奇异值来寻求ANFIS结构的最佳输入组合，最终选定O-M（双输入）组合作为ANFIS结构的输入量，通过训练和学习获得最优模糊规则库后，ANFIS结构的输出量即为叶片中Pb2+含量，从而实现了SVD-ANFIS模型的预测性能。研究结果表明，该模型的输出误差值较小，精度较高，在模糊训练过程中隶属函数选为钟型函数时预测效果最佳。利用多参数的反向传播(BP)神经网络预测模型对SVD-ANFIS模型的预测优越性进行验证时，得到BP模型和SVD-ANFIS模型的决定系数(R2)分别为0.977 6和0.988 7，均方根误差(RMSE)分别为2.455 9和0.601 3，可见SVD-ANFIS模型的拟合度更高，预测效果更好。同时选取不同年份的Pb污染玉米叶片等光谱数据对SVD-ANFIS模型进行可行性检验，其R2和RMSE分别为0.986 4和0.887 4，说明SVD-ANFIS模型能较好的用于玉米叶片中Pb2+含量预测且具有较高的鲁棒性，可作为预测玉米叶片中重金属含量的一种方法。

SVD-ANFIS Model for Predicting the Content of Heavy Metal Lead in Corn Leaves Using Hyperspectral Data

Heavy metals can enter the human body through the food chain after the crops had been polluted by them and can seriously harm the body health. Therefore, how to quickly and accurately monitor the content of heavy metals in crops has become important research in the fields of ecology and food security. The conventional biochemical monitoring methods have the disadvantages of cumbersome operation, long implementation process and destructiveness, while the hyperspectral remote sensing has the advantages of high spectral resolution, a large amount of information, strong biochemical inversion ability, convenience and fast, and no damage to the monitored object, so using hyperspectral remote sensing to monitor of heavy metal content in crops has become one of the hotspots in the field of remote sensing research. The potted corn plants stressed by different concentrations of Pb(NO₃)₂ solution were used as the research object in the paper, based on the data of the reflectance spectra of corn leaves under different lead ion (Pb2+) stress gradients and the measured Pb2+ contents in the leaves and combined with the Singular Value Decomposition (SVD) theory and Adaptive Network-based Fuzzy Inference System (ANFIS) structure, an SVD-ANFIS model was established for predicting the Pb2+ content in corn leaf. Firstly, SVD was used to process the reflectance spectra of Old leaves (O), Middle leaves (M), New leaves (N) under different stress gradients so that the singular values of the original spectral information were obtained. Then, the singular values corresponding to O, M, N leaves were selected to seek the optimal input combination of the ANFIS structure. Finally, the singular values of the spectra of the O-M (double-input) combination were selected as the input quantity of the ANFIS structure. After obtaining the optimal fuzzy rule base through training and learning, the output quantity of ANFIS structure was the content of Pb2+ in the leaves. Thus the SVD-ANFIS model achieved its predictive performance. The results showed that the model’s output error value was small and the prediction accuracy was high, and the prediction effect was best when the membership function was chosen as bell function in the fuzzy training process. When the multi-parameter Back Propagation (BP) neural network prediction model was used to verify the superiority of the prediction of the SVD-ANFIS model, the determination coefficient (R2) of the BP model and SVD-ANFIS model were 0.977 6 and 0.988 7, and the root means square error (RMSE) were 2.455 9 and 0.601 3 respectively, so the SVD-ANFIS model was shown to has a higher fit degree and better prediction effect. At the same time, spectral data of the corn leaves polluted by Pb2+ in different years were selected to test the feasibility of the SVD-ANFIS model, and its R2 and RMSE were 0.986 4 and 0.887 4, respectively, it indicated that the SVD-ANFIS model could be better used to predict the content of Pb2+ in corn leaves with high robustness and could be used as a method to predict the content of heavy metals in corn leaves.