Maturity classification of sweet peppers using image datasets acquired in different times

This paper presents maturity classification algorithms developed for small datasets and methods to deal with the highly variable and continuously changing agricultural environment. The algorithms were applied to the maturity classification of red and yellow sweet peppers, with data acquired from two different datasets, including 296 images. The maturity classification achieved 98.2 % and 97.3 % accuracy for classifying into two classes, between mature and immature classes of red and yellow peppers, respectively, and 89.5 % and 97.3 % accuracy for classifying into four maturity classes. The random forest algorithm is very robust and incurs a low computational cost, and therefore is recommended for the highly variable agricultural domain. An improvement of 28.65 % in classification accuracy was achieved by applying the methods developed for adapting to new datasets.