عنوان مقاله [English]
Apples are one of those fruits that have a higher supply to demand ratio at fresh market in Iran. Therefore, for price adjustment in market, the availability of apple in all seasons and the balance between supply and demand, the product must be stored in the correct way to increase the shelf life of apples. The shelf life of apples in the cooling store depends on many factors. One of the most important factor is the time of harvest. In general, over- ripe fruit have shorter shelf life. Thus, the question that arises is how to determine the appropriate time to harvest the apple in orchard. Various methods including destructive and non-destructive have been used to determine the level of fruit maturity. Among them, near-infrared spectroscopy is one of the most widely used methods in the field of quality assessment of agricultural products.
Since determination of the most appropriate harvest time reduces waste and increases storage ability in apples, prediction of chemical properties e.g. starch, titratable acidity and total soluble solids provides very useful information to beneficiaries so that it can be properly managed in the storage and marketing of apples. Thus, present study attempt to non-destructive prediction of starch, TSS and TA using 1D convolutional neural network in regard to prediction of appropriate ripening time.
2. Material and Methods
2.1. Collect the samples to extract the spectral data
According to the purpose of this study, which is to predict the chemical properties of apples in different stages of ripening in Braeburn apples, the first step was determination of harvest time. Hereupon, local gardeners were polled. Then, a total of 187 apples were harvested at 4 stage including 20 days before maturity, 10 days before maturity, on the maturity time and 10 days after maturity.
2.2. Extraction of spectral properties of hyperspectral images
A hyperspectral camera with a spectral range of 400 to 1100 nm was used to capture images. The camera was placed inside a chamber and illuminated by two 10-watt tungsten halogen lamps (SLI-CAL (StellarNet, USA) to block out disturbing ambient light. In this research, a laptop with features (Intel Corei5, 2430M at 2.40GHz, 4GB of RAM, Windows 10) was used to process the data. The raw spectral data was corrected by multiplicative scatter correction (MSC) algorithm in ParLeS, and finally the smoothing operation was performed by wavelet filter algorithm. (Rossel, 2008).
2.3. Measurement of chemical properties by laboratory destructive method
The amount of starch was measured based on the method described by Martínez-Valdivieso et al. (2014).
2.3.2. Total soluble Solid (TSS)
One of the most common uses of Brix is to determine soluble solids (sugar levels) in fresh produce. Brix number is a measure of the mass ratio of sugar dissolved in a liquid. This amount can be found simply by reading a refractometer. For example, if the brix of a fruit is 32, which means that the fruit contains 32 gr of sugar (any type) in 100 gr of water.
2.3.3. Titratable Acidity (TA)
First, 10 ml of fruit juice is poured into beaker and diluted with 10 ml of distilled water. The beaker is placed on the shaker and under burette, while 4 to 5 drops of phenolphthalein reagent are added. The initial pH is measured before adding NaOH. Then the valve of burette is opened gradually to add the titrant into the juice. When more drops of NaOH fall into the juice, the color of the juice changes but is not stable. Whenever a pale pink color is created in the juice in a stable way, it means that we have reached the end point of the titration. In fact, when the pH is higher than 8.2, this indicates that we are close to the end of the titration. Eventually, the valve is closed and the amount of NaOH consumed is recorded and the following formula is used to calculate the acid in the juice.
TA= (VNaOH*180*100*NNaOH)/ (1000*W)
VNaOH*: Volume of NaOH consumed
NNaOH : Normality of NaOH
W: Weight of samples
2.4. Non-destructive estimation of chemical properties by CNN in apple
Deep learning is a category of machine learning and a set of algorithms that have a high ability to classified data due to their hierarchical structure. Deep structures can also provide a more comprehensive representation of functions than MLP structures. Deep learning based on artificial neural networks mimics the behavior of the brain when learning a set of samples. Deep learning algorithms have been further developed in the context of artificial neural networks. In common neural networks, the number of hidden layers is usually no more than two. In contrast, when the number of hidden layers increases, these networks are called deep networks. Convolutional neural network is known as one of the most popular types of deep neural networks (ConvNet). It does not require manual feature extraction, so automatic feature extraction makes deep learning models very popular for computer vision tasks such as object classification. CNNs learn to recognize different features of an image using tens or hundreds of hidden layers. Each hidden layer increases the complexity and features of the trained image.
The structure of the CNN used in this study is shown in Table 1. The input vector of the algorithm contains spectral data and the output is the actual amount of chemical properties of starch, soluble solids and titratable acidity. It should be noted that after extracting the spectral properties, 70% were randomly used as train data and 30% as test data. In other words, 187 apples were picked in 4 different stages of ripening. 30% of the data (57 apples) were randomly excluded for testing. Due to the fact that the remaining number (130 apples) is not enough to train the deep learning model and on the other hand, the cost of measurement the properties are very expensive, so the artificial data generation method was used and a total of 2500 train data were produced. Evaluation of the performance of model for estimation of chemical properties in apple
Some statistical parameters were used to evaluate the performance of the prediction model such as Max error (Douglas et al, 2010), Mean absolute error (Sabzi et al., 2020), coefficient of determination (Alibaba et al. 2020), score-variance (Kristof, 1969), Mean squared error (Sabzi et al, 2021), Median absolute error (Wang & Lu, 2018) and Mean squared logarithmic error (Jeong et al. 2020).
3. Results and Discussions
The results of performance evaluation of the prediction model for the chemical properties of starch can be seen in Table 2. Considering that the values of the values of variance score (Var) and coefficient of determination (R2) are close to one, it states that the proposed model for pre- Starch nose has been successful.
Table 2: Performance evaluation of model for prediction of starch
Var R2 MaxE MedE MSLE MSE MAE Min Max
0.954 0.944 16.58 5.02 0.12 37.86 4.80 2 89
3.2. Total Soluble Solid (TSS)
Table 3 shows the performance evaluation results of the prediction model for soluble solids. Considering that the values of Max error, Mean absolute error, Mean squared error, Median absolute error and Mean squared logarithmic error are close to zero and the values of variance score and coefficient of determination are close to one, it states that the proposed model for prediction of TSS has been successful.
Table 3: evaluation of Performance of model for predicting TSS in apple
Var R2 MaxE MedE MSLE MSE MAE Min Max
0.916 0.916 0.907 0.239 0.0008 0.122 0.284 9.1 13.8
3.3. Titratable Acidity (TA)
Table 4 shows the performance evaluation results of the predictive model for acidity. Considering that the values of Max error, Mean absolute error, Mean squared error, Median absolute error and Mean squared logarithmic error are close to zero and the values of variance score and coefficient of determination are close to one, it states that the proposed model for pre- Excess nitrogen nose has been 90% successful.
Table 6: Performance evaluation of model to predict TA
Var R2 MaxE MedAE MSLE MSE MAE Min Max
0.842 0.842 1.264 0.421 0.0074 0.280 0.424 3.35 8.33
Because CNN does not require manual extraction, it is one of the most popular types of neural networks. Automatic feature extraction makes work easy and timely especially in computer vision applications. Therefore, in this study, the chemical properties of starch, titratable acidity and total soluble solids as parameters involved in determining apple ripening, were predicted using the deep learning model that the results were as follows.
- The coefficient of determination and the mean squares error for the starch properties were 95.4% and 4.8.
- The coefficient of determination and the mean squares error for the properties of soluble solids was 91.6% and 0.284.
- The coefficient of determination and the mean squared error for the acidity characteristic were 84.2% and 0.424.