Hyperspectral image spectra analysis

Một phần của tài liệu TSS water content and firmness in mango (Trang 80 - 84)

V. DETECTING MATURITY PARAMETERS OF MANGO USING

5.4.5. Hyperspectral image spectra analysis

Based on the feature wavelengths selected from the correlation and PLS models, spectral image were collected for the full area of the mango for all the samples. The spectral information from the whole area of the fruit for selected wave length bands were subjected to feed-forward neural network model for predicting the chemical attributes.

5.4.5.1 Total soluble solids content prediction

Table 5.7. shows the prediction of total soluble solids content by ANN model using the features of selected wavelengths from the simple correlation of original data, first derivative data and PLS regression models.

Table 5.7. Prediction of total soluble solids content

Wavelengths from SEC SEP r

Original data 0.98 0.64 0.69

First derivative data 0.86 0.41 0.78

PLS data 1.02 0.36 0.71

The calibration of the model using the samples resulted in a lowest standard error of calibration of 0.86 °brix between the feature values and the TSS of mangoes when the wavelengths from first derivative model were used. When the test samples were used for prediction, the results were r=0.78 with a standard error of prediction 0.41 °brix from the first derivative model. But the lowest SEP of 0.36 °brix was obtained from the ANN model using wavelengths from PLS regression with a correlation coefficient of 0.71.

Better results for correlation coefficient were obtained from first derivative model followed by PLS model and model used the wavelengths from the original data.

8 10 12 14 16 18 20

8 10 12 14 16 18 20

Measured TSS (%)

Predicted TSS (%)

8 10 12 14 16 18 20

8 10 12 14 16 18 20

Measured TSS (%)

Predicted TSS (%)

r=0.88 SEC=0.86

r=0.78 SEP=0.41

(a) Calibration results (b) Test results

Figure 5.7. Prediction results for total soluble solids content using feed-forward neural network with inputs of wavelengths from first derivative data and measured values

The correlation of calibration was high (r=0.88) with a standard error of calibration (SEC) of 0.86 for the first derivative model. The model predicted total soluble solids with a prediction correlation of 0.78 and SEP of 0.41 (Fig. 5.7).

5.4.5.2. Prediction of water content

Predicted results for water content by ANN using the features selected from the three models are presented in Table 5.8.

Table 5.8. Prediction of water content

Wavelengths from SEC SEP r

Original data 0.90 0.55 0.69

First derivative data 0.91 0.45 0.81

PLS data 0.76 0.33 0.79

Comparing the prediction results from the three models for water content, the results were similar to TSS predictions. The prediction results were better from the first

derivative model followed by PLS model and model using the original data. Figure 5.8.

shows the calibration and prediction results from the first derivative model for water content of the samples.

75 80 85 90

75 80 85 90

Measured water content (%)

Predicted water content (%)

75 80 85 90

75 80 85 90

Measured water content (%)

Predicted water content (%)

r=0.91

SEC=0.71 r=0.81

SEP=0.45

(a) Calibration results (b) Test results

Figure 5.8. Prediction results for water content using feed-forward neural network with inputs of wavelengths from first derivative data and measured values

High correlation coefficient of r=0.91 with a SEC of 0.71 was obtained while calibrating the model using the first derivative model. The model predicted the water content with a correlation coefficient of 0.81 and SEP of 0.45.

5.4.5.3 Firmness prediction

The calibration and prediction results from the feed-forward neural network model using the features from three models are presented in Table. 5.9.

Table 5.9. Prediction of firmness

Wavelengths from SEC SEP r

Original data 20.67 3.10 0.85

First derivative data 18.22 2.08 0.88

PLS data 24.22 0.83 0.79

Here also the first derivative model predictions were better than the other two models. The first derivative model predicted the firmness with a correlation coefficient of 0.88 and SEP of 2.08N. Next higher correlation coefficient was obtained from the model using the wavelengths from the original data with r=0.85 and SEP=3.1N. But the standard error of prediction was the lowest (SEP=0.83N) for PLS model. Figure 5.9.

shows the calibration and prediction results from the first derivative model for firmness.

0 50 100 150 200 250 300

0 50 100 150 200 250 300 Measured firmness (N)

Predicted firmness (N)

0 50 100 150 200 250 300

0 50 100 150 200 250 300 Measured firmness (N)

Predicted firmness (N)

r=0.99 SEC=18.22

r=0.88 SEP=2

(a) Calibration results (b) Test results

Figure 5.9. Prediction results for firmness using feed-forward neural network with inputs of wavelengths from first derivative data and measured values

The calibration of the model for firmness using the samples resulted in a correlation coefficient of 0.99 with a standard error of 18.22N. The test samples were predicted with an accuracy of r=0.88 and standard error of 2N.

Results from the three models showed higher correlation coefficient for predicting the firmness using features of selected wavelengths from the first derivative data and feed-forward neural network followed by water content and total soluble solids predictions. Considering only the NIR region of the spectra between 700-980 nm, the

wavelengths suggested for the prediction of firmness of mango are 728, 877, 890, 900 and 930 nm. Total soluble solids can be estimated at 715, 853, 916 and 950 nm wavelengths and the water content at 831, 923 and 950 nm.

Một phần của tài liệu TSS water content and firmness in mango (Trang 80 - 84)

Tải bản đầy đủ (PDF)

(93 trang)