Through a project developed by the Andalusian Operative Group with the participation of Agrosap, Cooperativas Agroalimentarias and the University of Seville, a system is being developed to detect oranges with RGB images obtained by a flying UAV drone in combination with a machine learning model. This would be useful to estimate the production in citrus farms.
Given that citrus cultivation is currently one of the most relevant in the countries of the Mediterranean Arc, with 513,602 hectares of cultivated area, 57% of which (295,000 ha) are in Spain, and that the harvest is one of the costliest operations, due to it being done manually, choosing the right moment to carry out the harvest is extremely important.
Measuring the maturity, the fruit's color and the size are the determining factors, and by means of the so-called precision agriculture, they can be objectively estimated.
With this new project, the analysis of the RGB images obtained is combined with the use of machine learning (ML) techniques. These deep learning techniques allow the development of a model based on previously obtained data sets, making the automatic sorting or detection of elements in those images possible.
Therefore, unmanned aerial vehicles (UAV), capable of carrying very high resolution sensors and cameras provide an alternative that can be implemented quickly and cheaply. They also generate a large amount of data, usually in the form of images or video sequences, and allow flexible flight patterns adapted to the tasks demanded of them. The combination of the drones' technological potential and ML techniques provides unique perspectives and information that, otherwise, would be very expensive to obtain using traditional techniques.
The development of this new methodology to estimate the yield of orange trees with high precision, or at least with a lower margin of error than with visual methods, can contribute to improving profitability and reducing the logistics and operational costs.
Once the project was completed, it was found that the model's estimation of the yield was closer to the real yield than that determined by a professional technician using visual models, with an error of just 1.54%.