基于荧光光谱法与深度信念网络的稻种发芽率检测方法研究

针对传统稻种发芽率检测效率低、精度差、专业化要求高等问题，通过荧光光谱法结合深度信念网络(DBN)建立稻种发芽率预测模型。首先，将连粳7号和武运粳均分别老化0~7 d后，以5 min为间隔在纯净水中分别浸泡5~30 min。然后用荧光光谱仪检测浸泡液的荧光光谱，光谱数据经中心化后用集合经验模态分解(EEMD)去噪，并通过主成分分析法提取441.5 nm的特征荧光波长。最后，利用偏最小二乘回归(PLSR)，反向传播神经网络(BPNN)，径向基函数神经网络(RBFNN)和深度信念网络(DBN)建立水稻种子发芽预测模型。比较后得出，DBN模型在少数据、弱信号情况下的预测精度最高，预测集相关系数Rp和均方根误差RMSEP最大可达0.979 2和0.101。同时，通过分析混合稻种荧光数据R_p的变化趋势，得到最佳浸泡时间为22.1 min，实际上，精确度超过0.95(R_p)需要5 min左右。研究结果表明，结合荧光光谱法和EEMD-DBN模型，非破坏性地预测水稻种子发芽率具有可行性和高准确性，并且适用于不同颜色和污染水平的水稻种子的检测。

Rice Germination Rate Detection Based on Fluorescent Spectrometry and Deep Belief Network

Traditional rice seed germination rate detection methods have low efficiency, poor accuracy and high specialization. The paper proposed a novel method by using fluorescent spectrometry combined with Deep Belief Network (DBN) to establish forecasting model for rice seed germination rate. Firstly, two varieties of seeds, Lianjing 7 and Wuyunjing, with 0～7 artificial aged days separately were soaked into purified water for 5～30 minutes with every 5 minutes’ interval. Then the fluorescence spectrums of the soak solutions were detected using fluorescence spectrometer. In addition, the spectrum data were centralized and then denoised with Ensemble Empirical Mode Decomposition (EEMD). The characteristic fluorescence wavelength of 441.5nm was extracted using Principal Component Anamysis (PCA). Finally, the rice seed germination predicting models were establishee with Partial Least Squares Regression (PLSR), Back Propagation Neural Network (BPNN), Radial Basis Function Neural Network (RBFNN) and Deep Belief Network (DBN), respectively. The results showed that the accuracy of DBN model was the highest in the case of less data and weak signal, and R_p=0.979 2, RMSEP=0.101. At the same time,we got the best soaking time is 22.1 min by analyzing the changing trend of mixed rice seed fluorescent data R_p, actually, it took about 5 min to get the accuracy more than 0.95 (R_p). The research results demonstrated the feasibility and high accuracy for predicting rice seed germination rate non-invasively by combining the fluorescent spectrometry and EEMD-DBN model, moreover, it adapts to the detection of rice seeds with different colors and contaminated levels.