基于光声光谱和TCA迁移学习的稻种活力检测

种子活力是决定水稻产量的最重要因素之一，但目前水稻种子活力的近红外和高光谱等无损检测方法易受种子表皮颜色影响，且所建模型难以适应新品种。该研究提出基于光声光谱技术的稻种活力无损检测方法并结合迁移学习进行新品种稻种活力检测。首先，对Y两优、龙粳、南粳、宁粳、武运粳、新两优等具有区域代表性的典型6种水稻品种，进行高温高湿人工老化处理，得到0~7 d老化时间的水稻种子；再通过调制频率获得8种不同深度的光声光谱信息，用主成分分析、竞争性自适应重加权算法对光谱降维得到特征光谱后，对Y两优、龙粳、南粳、宁粳、武运粳分别建立偏最小二乘回归（Partial Least Squares Regression，PLSR）、反向传播神经网络（Back Propagation Neural Network，BP）、广义回归神经网络（Generalized Regression Neural Network，GRNN）、支持向量回归模型（Support Vector Regression，SVR）、深度卷积神经网络（Convolutional Netural Network，CNN）的稻种活力预测模型，并选择最优调制频率；最后，通过迁移学习将建立的模型迁移到新两优稻种进行活力预测。结果表明，光声光谱最佳扫描频率为300 Hz，CNN预测模型精度较高，相关系数和均方根误差分别优于0.990 9、低于0.967 5；且经过迁移学习，仅需通过对源域数据的训练，即可直接对新品种稻种的活力进行精确预测；通过TCA迁移学习后，新两优稻种活力预测的相关系数从0.718 5提高到0.990 3。研究表明，采用光声光谱深度扫描技术对不同种类稻种的活力进行高精度检测是可行的，且经过迁移学习，仅需80粒新品种稻种信息即可实现稻种活力的精确预测。

Detection of rice seeds vigor based on photoacoustic spectrum combined with TCA transfer learning

Abstract: Seed vigor is one of the most important factors in determining rice yield. The nondestructive test methods of rice seeds vigor, such as near infrared spectroscopy and hyperspectral spectroscopy, are easily affected by seed skin color, particularly where the current models are difficult to adapt to new varieties of rice. In this study, a non-destructive testing method was proposed to detect the seed vigor of new varieties of rice, using the photoacoustic spectroscopy technology, combined with the transfer learning. First, six typical regional representative rice varieties were selected in different latitudes in China, including Yliangyou, Longjing, Nanjing, Ningjing, Wuyunjing and Xinliangyou county. An aging box model RXZ-128A was used for the artificially ageing under high temperature and humidity. The temperature in the aging box was 45°C, and the relative humidity was maintained at 95%, to obtain rice seeds with the aging time of 0, 1, 2, 3, 4, 5, 6, and 7 days. A Nicolet Is50R infrared spectrometer (Thermal Fish, USA) was used in conjunction with the PA300 photoacoustic cell produced by MTEC Photoacoustics to establish a rice seed photoacoustic spectrum acquisition system, thereby to acquire 8 different depths of rice seed photoacoustic spectrum information. The germination test was conducted on rice seeds with different aging days, and the average germination rates of 0~7 days aging days were 95.34%, 91.56%, 89.56%, 87.71%, 84.35%, 78.22%, 72.21%, and 66.33% respectively. After pre-processing and ensemble empirical mode decomposition denoising, the principal component analysis and competitive adaptive reweighted sampling can be used to reduce the dimension of spectrum, and thereby obtained the characteristic spectrum. Then, the Partial Least Squares Regression(PLSR), Back Propagation Neural Network(BP), Generalized Regression Neural Network(GRNN), Support Vector Regression(SVR), Convolutional Netural Network(CNN) prediction models of rice seed vigor were established for Yliangyou, Longjing, Nanjing, Ningjing, and Wuyunjing county, and the optimal modulation frequency were selected. Finally, a new CNN prediction model was established for a new rice seed vigor using source domain data, concurrently, the photoacoustic spectroscopy target domain data of the Xinliangyou rice seed was input after transfer learning into the newly established CNN model for vigor prediction. The germination test showed that with the deepening aging of rice seeds, the vigor, germination rate, and germination potential of rice seeds gradually decreased, the plant height of seedlings decreased, the dry weight decreased, and the seedlings became thin and grow slowly. The modeling results showed that the best scanning frequency of photoacoustic spectrum was 300 Hz, competitive adaptive reweighted sampling had good spectral dimension reduction effect, where the prediction accuracy of CNN model was higher, the correlation coefficient and root mean square error were better than 0.990 9 and 0.507 7, respectively. After transfer learning, the vitality of new rice varieties can be directly and accurately predicted only by training the data of source domain. In TCA transfer learning, the correlation coefficient of prediction for the Xinliangyou rice seed vigor increased from 0.718 5 to 0.990 3. The usage of photoacoustic spectroscopy deep scanning technology can be proved to be feasible to detect the vigor of different types of rice with high precision. After transfer learning, only a small amount of information about new varieties of rice are required to be used to accurately predict rice seed vigor.