TREEREG

Aerobotics is a company in the Precision Agriculture sector that provides agricultural trend analysis on tree orchards. They use multitemporal orthomosaic images of farmland to accomplish this. For successful trend analysis to occur, the input orthomosaics need to be correctly aligned (registered) automatically. This image registration process is currently performed manually with a human-in-the-loop approach. A feature extractor that detects individual tree canopies in the image has been developed as the input to a possible automatic registration method. The aim of this project is to use these feature detections to accurately and automatically register the orthomosaics. This project evaluated the applicability of three point cloud registration algorithms, Coherent Point Drift (CPD), Iterative Closest Point (ICP) and Robust Point Matching (RPM), as possible solutions to this problem. These methods were tested on input orchard datasets in which a wide range of rigid deformations have been applied. Their respective registration accuracies were recorded and compared to the current human-based registration method. Additionally, each method’s robustness to unforeseen input deformations (such as the addition of outliers and noise) were tested. The investigation showed that CPD and RPM were capable of automatically registering multitemporal orchard images using tree detections as inputs, despite any unforeseen deformations. With CPD producing accurate results in reasonable time. CPD and RPM consistently produced accuracy results comparable to the human-based method throughout the testing process. Thus, in their case, the aim of the project was met.