LULC Classification Performance of Supervised and Unsupervised Algorithms on UAV-Orthomosaics

Classification algorithms involve automatically categorising the response functions recorded in any image, i.e., the light reflected and recorded in the pixels by the sensors, as “virtual” representations of real-world objects in classes. These algorithms are finding increasing use in Remote Sensing (RS) applications and the scientific community is deeply engaged in investigating their performance. The substantial increase in spatial resolution that has been introduced through the onset of Unmanned Aerial Vehicle (UAV) platforms, even if equipped with low-cost sensors, has made classes recognition perform well for even the smallest scenario properties. Despite these advantages, it is emerged the need to address their limitations and to analyse their impacts in the data post-processing stages. It is necessary to validate the processing procedure so as to fully define their comparability and, in some cases, their interchangeability with other RS platforms. The aim of this research is to create a validated pixel-based classification procedure that will subsequently result in an automated method that is simple to apply, especially for end-users with limited knowledge of spatial data processing. Supervised and unsupervised approaches to testing performance were deemed more effective by adopting photogrammetric products based on UAV-acquisitions. This has mainly been accomplished by using statistical classifiers for Land Use/Land Cover (LULC) recognition based on several reflectance values across wavebands that compose an image and vegetation indexes in the visible bands. For these tests, the processing chains concerning classifications with supervised Random Forest (RF) and unsupervised K-Means Clustering algorithms were adopted.