Variety classification of date fruits in an orchard environment

Date variety is a key factor in determining the stage at which dates are harvested, unlike most fruits that only depend on the maturity stage of an individual fruit (i.e. ripe or not). For instance, some varieties of dates are harvested in the first maturity stage (Khalal) such as Barhi, whereas other date varieties are usually harvested later in the Rutab or Tamar stages such as Sullaj and Khalas. 

Classification of date fruits according to their varieties in an orchard environment is a challenging task due to the low inter-class variation between some date classes and the wide intra-class variation within one class, as shown in the following figure. Some varieties can easily be distinguished based on their visual appearance, whereas it is difficult, for even a specialist, to distinguish some varieties. In addition, there are large differences between date bunches of the same variety, as they vary in terms of maturity level, illumination, bagging state, scale, and angle.

Deep convolution neural networks CNNs are utilized with transfer learning and fine-tuning on pre-trained models , e.g. VGG-16, ResNet-50, ResNet-100, inception v3. To build a robust machine vision system, we used a rich image dataset of five date types for all maturity stages. The dataset was designed with a large degree of variation that represents the challenges in natural environments and date fruit orchards.