Hiperspektral Görüntü ile LiDAR Verisinin Koşullu Rastgele Alanlar Yöntemi ile Birleştirilmesi ve Sınıflandırılması

Abstract

With the advances in sensor technology and widespread usage of fast processors and large memories in computing in the recent years; research in computer image processing is in acceleration. Some of these researches are focused on semantic segmentation of hyperspectral images (HSI), which consist of images captured from hundreds of different frequencies. Thanks to another remote sensing technology called Light Detection and Ranging (LiDAR), the measurement of distances between objects can be remotely achieved. For example, producing three dimensional models of terrain or human-made objects on the surface of the earth is possible with the help of LiDAR.

The goal of this thesis is, through the fusion of data acquired from HSI and LiDAR images, to improve the overall accuracy in a semantic segmentation task. Fusion process is done with the help of years of research that are done in the topics of Markov Random Fields (MRF) and Conditional Random Fields (CRF). Based on the energy models proposed by these two approaches, unary and pairwise energy terms are obtained by means of several image processing methods. The unary term defines the spectral properties of an image. This information is extracted with the use of two different classifiers which are named Probabilistic Support Vector Machines (pSVM) and Sub-Space Multinomial Logistic Regression (MLRSub). The pairwise term on the other hand defines the spatial properties of the image pixels in hand. Novel energy functions for unary and pairwise terms are proposed in this work. These energy functions are minimized via Graph-Cuts and Mean Field Approximation methods separately. The proposed energy terms are also tested on real world datasets and compared against well-known methods in this area of expertise.