Coffee Trees Segmentation in UAV-Acquired Images Using Deep Learning

Abstract

Following the line of the second Sustainable Development Goal, focused on exploring new agricultural technologies to strengthen food security, an algorithm is proposed for the segmentation of coffee trees that allows to study the trees at an individual level. This algorithm uses RGB images obtained through a UAV and relies on the Segmenting Anything Model (SAM) tool developed by META AI, designed for the application of deep learning models in image segmentation. Crucial feature extractions were performed from the segmentations, including information from various color spaces, textures such as Local Binary Pattern and Co-occurrence Matrix, as well as Hu invariant moments. Subsequently, supervised object classification was carried out to generate a binary dataset. This dataset was used to train various machine learning models, such as Random Forest, Support Vector Machine, and Decision Tree. The results highlighted the Random Forest model as the most effective, with a kappa value of 0.86 and an accuracy of 93%.