Automatic localization of phoenix by satellite image analysis

Abstract EN

The Red palm weevil (RPW) Rhynchophorus ferrugineus is becoming one of the deadliest pests of the palms in the world.

In order to effectively implement a RPW control programme to achieve rapid regression of this pest, it is necessary to have GPS coordinates of each palm present on the control perimeter.

This location makes it possible to establish maps and databases which are essential for organizing, at the local and national level, the implementation and permanent monitoring of control measures.

It is difficult, time-consuming and expensive to locate palms by visually exploring the entire perimeter from the ground.

In the zone of regular plantations, this work can be processed but it becomes extremely heavy in the traditional oasis like in urban environment where the distribution of the palms is very irregular.

With advances in satellite imagery, it is possible to acquire high quality images at very short intervals of time from a standard format for a large part of the earth.

Combined with the progress of machine learning, particularly deep learning, this amount of data is able to feed a robust model.

It would allow to automate the detection of palms at large scale and monitor their evolution at very short intervals, which in the fight against RPW is valuable information.

This first work wants to test the interest in this solution.

We build and train a convolution neural network in order to find two species of palms Phoenix canariensis and Phoenix dactylifera (C&D) in a very heterogeneous area of 100 km².

Our model evaluation shows that 1/5 of the objects found are false positive and more than 2/3 of C&D are perfectly localized.

These first results could be improved greatly by implementing a more robust algorithm using more data and using larger colour spectrum (as near infra-red).

The question of the infested palms detection using satellite imagery and machine learning stays open.