Inspection of soluble solid content for tomatoes in different positions based on hyperspectral diffuse transmittance imaging

Abstract: Soluble solid content (SSC) is one of the most important indexes for quality evaluation of tomato products. Near infrared (NIR) spectroscopy and hyperspectral reflectance imaging have been widely used in quality evaluation of fruits and vegetables including tomatoes. But they have many disadvantages for inspection of SSC in tomato. For example, NIR spectroscopic assessments cannot get the spatial variability of sample materials. Although hyperspectral reflectance imaging can obtain both spatial and spectral information of tomatoes, it's almost impossible to avoid a serious influence of high specula patches on tomatoes. Diffuse transmittance is one kind of transmittance mode. Compared with transmittance, the influence of shape, size, and core of fruit can be reduced through adjusting the lighting angle in diffuse transmittance systems. So diffuse transmittance is more suitable to assess the components of fruits and vegetables. Hyperspectral imaging technique in a diffuse transmittance mode was used to measure the SSC of tomato. First, a hyperspectral imaging platform with diffuse transmittance illumination was set up, and then hyperspectral diffuse transmittance images of tomatoes were captured in different positions including BS, C1, C2, and C3. All images were resized to eliminate boundary noise. The position C1C2C3 was achieved through mosaicing images of position C1, C2, and C3. Then background segmentation on a single wavelength was operated on the images to extract regions of interest (ROIs). Afterwards, the mean diffuse transmittance spectra of tomatoes in each position were calculated and preprocessed using normalization, standard normal variate (SNV), and a quadratic linear removed baseline. Finally, partial least squares regression (PLSR) was used to establish predicting models among the SSC of tomatoes and mean diffuse transmittance spectra in different positions on three different wavebands (450~720 nm, 720~990 nm, and 450~990nm). The results indicated that the prediction precision of integrated position C1C2C3 was much better than that of the other positions on the above three wavebands. RMSEP of the C1C2C3 model on the three wavebands were 0.299%, 0.133% and 0.151%, and the correlation coefficients (rp) were 0.42, 0.89 and 0.90 respectively.