Vegetatıon Cover Extractıon And Monıtorıng Usıng Images Obtaıned From Aerıal And Satellıte Platforms

Abstract

Vegetation cover has an essential role in both urban and rural environments in several aspects, such as mitigating urban heat island effect, sustaining ecological balance, preserving biodiversity, improving the quality of life, etc. Vegetation change detection can be carried out by using image processing methods in order to distinguish differences. For this purpose, images taken at different moments need to be utilized. In urban areas, vegetation change can be a great indicator for the growth or decrease of cities green infrastructure. In contrast, the change in rural areas could mean the growth or decrease of agricultural fields. This thesis aimed to investigate the vegetation change detection methods by using multi-temporal and multi-platform data, i.e. obtained from aerial and satellite sensors at different seasons and years. Data from RGB cameras mounted on Unmanned Aerial Vehicles (UAV), and from multispectral Earth Observation (EO) satellite sensors (i.e. Gokturk-1 and Worldview-2) have been employed as data sources in this study. The study area was selected from the forest area in Akdeniz University Campus, Antalya, Turkey, due to data availability. A number of image pre-processing methods for radiometric and geometric improvements were the initial tasks prior to change detection. Different georeferencing methods were applied for accurate alignment of the images. Several vegetation indices such as normalized difference vegetation index (NDVI) and Green and Red ratio Vegetation Index (GRVI) were derived from the datasets to identify the vegetation better. These NDVI and GRVI images were employed in a supervised machine learning (ML) method, i.e. the random forest (RF), for vegetation classification. The validation results show that the RF method is a suitable method for vegetation mapping and decision level change detection for multi-temporal, multi-resolution and multi-platform datasets. In addition, for the purpose of unsupervised vegetation mapping, the Otsu thresholding method was found successful when applied to the GRVI data obtained from all sensors used here.