基于电学参数的苹果可溶性固形物含量预测

为实现苹果可溶性固形物含量的无损检测,该研究提出了一种长短期记忆编解码和多层感知机（LSTMED-MLP,long short-term memory encoder-decoder-multi-layer perceptron）融合的介电特征预测方法。在0.158～3 980 kHz频率范围内的9个频率点下,采用介电谱测量仪获取300个富士苹果的电学参数,其中每个频率点对应15项电学参数,即每个苹果对应135项电学特性参数,之后通过苹果基因组学理化分析方法,获取可溶性固形物含量;根据电学参数与可溶性固形物含量,构建苹果关键基因组学参数的回归预测模型。为简化模型输入,提取样本变量特征,使用主成分分析（principal component analysis,PCA）和LSTMED模型,提取每个样本的40项特征值,作为非线性回归模型多层感知机（MLP）和XGBoost的输入,建立可溶性固形物含量预测模型。试验结果表明,LSTMED具有更好的适用性,且LSTMED-MLP模型的预测效果最好,在校正集和预测集上,决定系数分别为0.95和0.90,均方根误差分别为0.77和0.84,且对不同种...

Prediction of the soluble solid contents for apple fruit using electrical parameters

Abstract: Soluble Solids Content (SSC) is one of the most important physiochemical parameters to evaluate the internal quality of apple fruits. However, the traditional SSC detection can cause some mechanical damages to the fruits under the static and dynamic forces. It is necessary to explore the non-destructive SSC prediction in modern agriculture. Fortunately, the dielectric spectrum detection has been widely used in the quality evaluation of fruits and vegetables in recent years, due to the simple procedure and measurement requirements for the samples. Nevertheless, the nondestructive quality detection can result in the high complexity of the prediction model, due mainly to the data redundancy in a large number of measured electrical parameters at different frequency points. In this study, the linear and nonlinear feature extraction of electrical parameters was proposed to simplify the input of regression model, in order to accurately and rapidly predict the SSC of apples. The electrical parameters of three hundred Fuji apples were measured at the nine frequencies in the range of 0.158-3980kHz by high-frequency LCR (Inductance (L), Capacitance (C), and Resistance (R)) meter. Among them, 15 electrical parameters were collected at each frequency, where each apple sample was contained 135 electrical parameters. All the electrical parameters of each apple were measured non-destructively. The SSC was also obtained by the special physical and chemical analysis machine after wholly destroying the samples. Finally, the regression prediction model was constructed using the key genomic parameters of apple. Two nonlinear regression models (namely Multi-layer Perceptron (MLP) and XGBoost model) were employed, according to the electrical characteristic parameters and SSC. Furthermore, 300 samples of apple were randomly divided into the calibration set and prediction set. The calibration set was used to construct the prediction model, whereas, the stability of the model was then tested by the prediction set. Firstly, the MLP and XGBoost were set as the input of nonlinear regression models. 40 dielectric characteristics were obtained by the linear feature extraction of Principal Component Analysis (PCA), with a cumulative contribution rate of 98.9%. More importantly, the evaluation index of PCA-XGBoost model was higher than that of PCA-MLP model in the calibration set and prediction set. The residual prediction deviation of PCA-MLP model was lower than PCA-XGBoost model. Furthermore, the nonlinear characteristics of the electrical parameters were ignored in the linear feature extraction, due to the nonlinear relationship between the electrical parameters of apple samples at different frequencies. Long Short-term Memory Encoder-Decoder model (LSTMED) was also utilized to extract the non-linear characteristics for the input of the nonlinear regression model of MLP and XGBoost. A comparison was made for the prediction accuracy of LSTMED-MLP and LSTMED-XGBoost. The experiments showed that the residual prediction deviations of LSTMED were 0.24 and 0.01 higher than those of PCA, respectively, indicating the better feasibility. The LSTMED-MLP model performed the best on the calibration sets and prediction sets, followed by the LSTMED-XGBoost model. The predicted correlation coefficients were 0.91 and 0.85, respectively. The predicted root mean square errors were 0.84 and 0.95, respectively. Therefore, the LSTMED demonstrated the effective performance of feature extraction and data reduction for the non-linear parameters. The wide prediction suitability of the improved model can be expected for the inner quality parameters of fruit and vegetables.