Study On Optimization Method Of Apple Quality Prediction Model In Different Growth Positions And Storage Period

Apples are the third most grown and consumed fruit in the world,after bananas and watermelons,and consumers have high demands on their quality,thus requiring research on their internal quality.According to statistics,the number of sunshine hours differs by up to 4hours/day for apples of different growing heights on the same apple tree,and the formation of fructose in apple tissues is closely related to the number of sunshine hours.Different growth heights of apples may lead to differences in rainfall,fertilizer application and fruit water uptake,which can seriously affect fruit quality and taste.It is especially important to find the optimal growing height of apples to select high-quality fruits from growing sources and to provide strong support for standardized planting and harvesting.Currently,the apple model built after collecting spectra is a mixture of apples of all growth heights,ignoring the differences between the spectra of apples of different growth heights.Therefore,the study of different growth heights is of guidance for fruit tree planting and model building,as well as important for studying the ripeness of fruits at different growth heights.Meanwhile,existing models are built by mixing fruits with different growth heights,and given the robustness and low accuracy of this model,building models with different growth heights is a key factor to improve model accuracy and robustness.And the quality of apples at the consumption end can be affected by different storage time and storage conditions.As a fruit with ripening period from July to early November,the quality of apples from May to June is especially important at the sales end of the year.Therefore,in this paper,we conducted a study on the internal quality of apples due to different growth heights and storage conditions,using the fruit quality dynamic online equipment and Fourier transform near-infrared spectrometer,with the following main findings.The main research contents and conclusions are as follows:（1）The differences in the internal quality of apples at different growth heights were investigated and the effects of apples at different growth heights on modeling were compared.Fruits at different growth heights are exposed to different levels of light and their ripeness varies.In this experiment,Fuji apples at different growth heights（fruit trees were divided into three heights with 300 samples at each height）were collected to measure soluble solids content（SSC）and hardness,and Fourier NIR spectra were collected in the range of4000-10000 cm^-1.Apples showed increased spectral intensity with slightly shifted peaks,increased SSC,and decreased hardness at higher heights,and decreased spectral intensity at lower concentrations and decreased spectral intensity and increased hardness at higher concentrations.In order to distinguish the spectral differences of apples at different growth heights,a discriminant model was established using partial least squares discriminant analysis.The model worked well with an accuracy higher than 90%.Partial least squares（PLS）was also used in combination with different preprocessing methods(multiple scattering correction（MSC）,standard normal variable transformation（SNV）,and different variable selection methods(uninformed variable elimination（UVE）and competitive adaptive reweighted sampling algorithm（CARS）to remove irrelevant information and noise,reduce the dimensionality,and optimize the prediction model.The results showed that the low-level SSC model worked best.the prediction correlation coefficient（Rp）of SSC was 0.941 and the root mean square error of prediction（RMSEP）was 0.660.the top-level hardness model had an Rp of 0.931 and an RMSEP of 0.344.the models built at different growth heights outperformed the hybrid model,and the top-level model was optimal.（2）Changes in the internal quality of four varieties of apples under different storage conditions were investigated.During the storage period,the internal respiration of apples changes,water loss,hardness decreases and taste deteriorates.This study aimed to investigate the changes of internal quality of different varieties of apples under different storage conditions during the spring-summer transition period in May and June of a year by near-infrared spectroscopy,and to establish near-infrared prediction models for brix and hardness,which are of guiding significance for selling apples in the market during the first two months of ripening.Firstly,the spectra of four apple varieties were collected under three storage conditions for 7 weeks using intelligent online fruit quality testing equipment to determine their sugar content（SSC）and hardness;the weekly changes of spectra,sugar content and hardness were analyzed,and it was found that the average spectra of different varieties of apples would usher in a peak at 3-5 weeks,and their sugar content and hardness would first increase and then decrease under refrigeration conditions,while the opposite was true at room temperature.Partial least squares discriminant analysis（PLS-DA）was used to develop a classification model for apple storage time with high accuracy.Partial least squares（PLS）was also used in combination with different preprocessing methods（normalization（Normalization）,multiple scattering correction（MSC）,standard normal variables transformation（SNV）,etc.）and different variable selection methods（uninformative variable elimination（UVE）and competitive adaptive reweighting algorithm（CARS））to remove irrelevant information and noise and reduce dimensionality to build SSC and hardness prediction models.The best results were found to be achieved using CARS after normalization of the models.the best combined results were obtained for the Normalization-CARS-PLS model for the sugar content of 1°C samples,which had the best prediction correlation coefficient（Rp）of 0.904 and its root mean square error of prediction（RMSEP）of 0.67,and the Normalization-CARS-PLS model for the hardness of samples at room temperature.The CARS-PLS model had the best combined effect with its best prediction correlation coefficient（Rp）of 0.823 and its root mean square error of prediction（RMSEP）of 0.809.