The current study evaluates the potential of hyperspectral imaging combined with artificial neural networks to predict the moisture content in tomato fruit and to obtain spatial distribution maps. A total of 192 tomato samples, Solanum lycopersicum 'Pannovy', were harvested in the third and fourths stage of ripening according to the OECD color chart. Samples were immediately transmitted to the laboratory and stored at 15°C and 92% rH. Measurements were carried out after 1, 8, 15, 22, and 30 days of storage. After acquisition of hyperspectral images of 40 tomatoes on each date, moisture content of samples was calculated after oven drying (105°C). The electromagnetic spectrum between 400 to 1000 nm was recorded at 495 passbands by a hyperspectral imaging system. After applying preprocessing operations and removal of data in saturation, valid data sets were collected manually from the region of interest. Spectral dimensionality was reduced by selecting the wavelengths in which high correlation exists between the spectral data and fruit moisture content. A multilayer percepteron neural network has been trained with training dataset to predict the moisture content of samples. To prepare the spatial prediction maps, tomato samples were separated from the background. After that, 2D images were unfolded to vectors and the intensity values of each pixel in the selected wavelengths were fed into the fully trained neural network. The output matrix, containing the predicted values of moisture content in every pixel, was folded back to form a 2D matrix with the same spatial dimension of the hyperspectral images. Finally, the spatial distribution of moisture content was displayed as a color map, where colors represent different values of predicted attribute. Results proof the feasibility of the method for characterizing the spatial distribution of an attribute in horticultural produce.
