Hiperspektral ve Lidar Verilerinin Öznitelik Ve Karar Seviyelerinde Tümleştirilmesi Ve Derin Evrişimli Sinir Ağlarıyla Sınıflandırılması
Özet
With the evolving next generation remote sensing technology, hundreds of different
wavelength images can be captured in the electromagnetic spectrum. In addition to
this, light sensing technology can be used to determine distances between objects
and distant objects. This information obtained from two different sources constitutes
an input to the analysis of the semantic segmentation of a region.
In this thesis study, it is aimed to realize the semantic segmentation of two different
data sets belonging to the region with HSI (Hyperspectral Image) and LiDAR (Light
Detection and Ranging) data with high performance. In particular, classification with
deep convolutional neural networks (CNN) has been performed in recent years,
except for the classical methods used for classifying hyperspectral datasets, which
have had impressive results in the semantic segmentation of images.
In the study carried out within the scope of the thesis, the problem is addressed in
two steps. First, integration of the hyperspectral and LiDAR information was followediv
by finalization of the classification. In order to provide additional information on
spectral and height information, different structural elements and extended
morphological attribute profiles (EMAP) have been specially created for HSI and
LiDAR data. Within the scope of integrating the classifier decisions recommended
as the first integration method, CNN installations were made specifically for the
spectral and morphological profile maps and the results of the classifiers were
established. Using the knowledge that the test data provided by the classifier results
could be included possibility result in which class, the classifier results were
integrated and a general classifier was created. With this method, the best results
were obtained in classifying the Muufl dataset. In the dimensional integration, which
provides impressive results in the Houston dataset, which is the second method,
and in particular the competition dataset, the dimensional integration of different
height, morphological and spectral information is provided first. Subsequently,
attributes of this integrated data were extracted with filters in AlexNet's first
convolution layer. A special CNN setup was performed on this dataset, from which
the features were extracted, and the results of the classification were compared with
other studies. In the Houston dataset, the highest performance was 93.97% for the
cloud-based shadows. Among the studies that did not make any correction to the
cloud-based shadowing in the Houston dataset, the highest achievement value of
93.97% was obtained with this method.
As deep convolutional neural networks have become very popular in recent years,
the feasibility of these networks in many new areas is being questioned. It is one of
the important results of this thesis that it can give very effective results in the
classification of hyperspectral datasets. However, the need to ensure that formal
and spatial information to support spectral information is given as input to CNN is
one of the most important results that have been drawn from the study. In addition,
the use of filters in the first convolution layer of the AlexNet model classification with
CNN has been demonstrated in HSI’s.
