Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Master → Belgique → Object based image analysis for Phoenix palm trees mapping : a regional case study

Katholieke Universiteit Leuven (2022)

Object based image analysis for Phoenix palm trees mapping : a regional case study

Zinbi, Sanaa

Titre : Object based image analysis for Phoenix palm trees mapping : a regional case study

Auteur : Zinbi, Sanaa

Université de soutenance : Katholieke Universiteit Leuven

Grade : Master of Bioscience Engineering. Agro- and Ecosystems Engineering 2022

Phoenix palm trees play a crucial role in arid regions as they are valued for ornamental purposes, besides bearing nutritious fruits. Unfortunately, these species of palms have been suffering worldwide, and more specifically in the Spanish province of Alicante, from the attack of an insect known as the Red Palm Weevil. To counter the advance of this pest, accurate information on the position of the palm trees is essential to monitor their health. Using images of the Earth’s surface and automatic detection techniques has the potential for rapid and accurate location of the palms, as opposed to resource-intensive field surveys. In this thesis, the feasibility of using an increasingly popular method for automatically detecting objects on the Earth’s surface is challenged to locate palm trees over the large area (5,816 km2) of the province of Alicante. To this end, multiple images taken by an airplane and hand-labeled palm trees were used to train a model to recognize Phoenix palms. Due to a lack of palm annotations over the whole province, the model was first trained on the images benefitting from a complete inventory. Subsequently, it was applied to the rest of the images that did not contain any palm labelling or had annotations limited to some environments. Next, the detection performance in different environments (an orchard, a nursery, a plantation, and a public park) containing palms was contrasted. Finally, the technique was compared in terms of efficiency in detecting palm trees to another method formerly designed to label palms in the same area. Palm trees could be detected with moderate accuracy over the image on which the model learned to detect them. For the images on which the model was applied and that contained labeled palms only in specific environments, the accuracy was poor, and the lowest accuracy was achieved in the plantation. The previously developed method also proved to be more efficient. The results of this research could be explained by several shortcomings linked to the methodology. Hence, recommendations are proposed to overcome them in future research.


Page publiée le 7 mars 2023