机构地区:[1]华中农业大学工学院,湖北武汉430070 [2]农业农村部长江中下游农业装备重点实验室,湖北武汉430070 [3]农业农村部柑橘全程机械化科研基地,湖北武汉430070
出 处:《光谱学与光谱分析》2025年第2期492-500,共9页Spectroscopy and Spectral Analysis
基 金:国家自然科学基金青年科学基金项目(61205153)资助。
摘 要:针对沃柑品质快速检测需求,提出一种基于高光谱图像数据的沃柑可溶性固形物含量(SSC)检测方法,并分析随贮藏时间变化的SSC伪彩色分布图。分别获取307个整果样本和227个半果样本以及它们的SSC数据。比较标准正态变换(SNV)、多元散射校正(MSC)、Savitzky-Golay(SG)滤波、归一化(NM)、一阶导数(FD)、标准化(SD)和小波变换(WT)对偏最小二乘回归(PLSR)模型检测性能的影响以选择光谱预处理方法;比较PLSR、最小绝对值收缩与选择算子(LASSO)回归、支持向量机回归(SVR)、人工神经网络(ANN)、决策树(DT)、随机森林(RF)、轻量级梯度提升机(LightGBM)模型对独立验证集的检测能力以确定最佳模型建立方法,并利用遗传算法(GA)筛选特征波长以优化模型。结果表明:采用FD预处理结合LASSO回归算法所建模型对整果SSC检测效果最优,验证集决定系数(R_(p)^(2))和验证集均方根误差(RMSEP)分别为0.9257和0.9765;SD预处理结合RF模型对半果SSC检测效果最好,其R_(p)^(2)和RMSEP分别为0.8963和1.0630;GA能够滤除53.85%和50.58%的整果和半果波长变量数,基于选择变量的整果和半果最优建模算法分别为SVR和RF,其模型R_(p)^(2)、RMSEP分别为0.9189、1.0173和0.8953、1.0843。研究结果为沃柑SSC高通量无损检测提供了一种可行方案。To develop a rapid measurement method of SSC in Fortunella margarita,the detection models based on hyperspectral imaging data were established and optimized by employing various preprocess and regression algorithms,and the pseudo-color distribution of SSC with storage time was analyzed.The 307 whole citrus and 227 hemisected citrus samples were involved in hyperspectral data collection and the SSC values.The effects of preprocessing,including standard normal variate(SNV),multiplicative scatter correction(MSC),Savitzky-Golay(SG)filtering,normalization(NM),first derivative(FD),standardization(SD),and wavelet transformation(WT),on the performance of the partial least squares regression(PLSR)model were compared to select the appropriate preprocessing method.Then,the detection models were established by using PLSR,least absolute shrinkage and selection operator(LASSO)regression,support vector machine regression(SVR),artificial neural networks(ANN),decision trees(DT),random forest(RF)and light gradient boosting machine(Light GBM)algorithms.Furthermore,the models were optimized using genetic algorithms(GA)to select characteristic spectral wavelengths.The results indicated that for the whole citrus samples,the FD preprocessing could extract more features,and the LASSO regression model performed better than other models with 0.9257 and 0.9765 as the prediction determination coefficient(R_(p)^(2))and root mean square error of prediction(RMSEP),respectively.For the hemisected samples,the RF model based on the spectral after SD preprocessing had higher R_(p)^(2) at 0.8963 and lower RMSEP at 1.0630.The GA could remove 53.85% and 50.58% wavelength variables to reduce the computational complexity for the whole and hemisected sample spectral,of which the SVR model has R_(p)^(2) at 0.9189.RMSEP at 1.0173 RF model having R_(p)^(2) at 0.8953 and RMSEP at 1.0843 performed better than other models.The results provided a feasible solution for high-throughput,non-destructive detection of SSC of Fortunella margarita.
关 键 词:沃柑 高光谱 可溶性固形物含量 无损检测 集成学习
分 类 号:S24[农业科学—农业电气化与自动化]
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...