Hiperspektral Görüntülerde Eser Miktarda Kimyasal Madde Tespiti
Özet
Hyperspectral images, by their very nature, provide a great deal of information about the
contents of an image. In this way, they open the way to reveal the evidences that cannot be
distinguished or detected properly by the analysis of single or 3-band images. Shots made
with hyperspectral cameras, which provide an information vector consisting of hundreds
of bands for each pixel, provide spectral analysis even in cases where spatial analysis of
the objects in the image is not possible or can be made very limited. Spectral separation of
objects or backgrounds that can resemble each other in color or shape is made possible
with hyperspectral imaging technology.
There are many approaches that can be tried to analyze hyperspectral images, to classify
the objects in their content using these images, or to detect a searched object. These
approaches can be shaped based on different methods, depending on whether there is a
preliminary information about the searched object, whether there is a rich data set on the
analysis of the problem, or whether the data acquisition conditions change based on
different parameters settings. For example, anomaly detection-based methods are used for
the detection of objects that are not known exactly but separated from the background,
while spectral signature matching-oriented methods can be used in cases where the
preliminary information about the sought object is provided.
v
In this thesis, studies for the detection and identification of trace amounts of chemical
substances placed at levels measured in nanograms (500ng -10000ng) on different
backgrounds are described. Many different difficulties are faced such as target and
background diversity, placement of the very small amount of target, using raw data without
any radiometric correction, change of light and target position. The solutions developed to
avoid these difficulties and to produce a robust algorithm are presented step by step. First
of all, when the raw radiance data taken from the sensor is examined, it is observed that the
background and target separation cannot be made well in raw data. For this reason, a
number of data conversion algorithms are applied for data rectification with different
operations on the received radiance data. Thus, target and background separation is
performed more successfully. Afterwards, various signature-based target detection
algorithms have been tried and the methods that produce successful results for the problem
that is the subject of this thesis have been revealed.
Solving the problem of detection and identification of chemicals requires creating an
algorithm that takes images and other necessary information as input and gives target type
and location information as output. Hence, creating score maps is not enough. Steps that
produce the necessary outputs by using relevant score maps are also needed. For this
purpose, post-processing algorithms that lead to more robust results are also part of the
work done in this thesis. These steps eliminate false alarms as much as possible.
In summary, a hybrid method called "Multiple Target Detection" has been developed by
using different spectral signature matching-oriented methods in the literature to solve the
above-mentioned problem. First, the hyperspectral image taken in the ultraviolet and near
infrared spectrum and the preprocessing steps (whitening and Savitzky-Golay filter)
applied to the target reference signatures sought in the image are detailed. Afterwards, the
applied hybrid detection method is explained in detail. A two-stage score map generation
algorithm based on cross-correlation and spectral information divergence algorithms is
described. Then information is given about the final processing steps, which include
conditions such as thresholding according to the score and selecting final detected areas by
connective pixel analysis for potential target clusters.
In addition to the “Multiple Target Detection” method, a method named POSTNET, which
provides the final processing steps applied on the previously obtained score maps with deep
learning, has been developed and presented in detail. In addition, the great contribution of
vi
the pre-processing steps applied to the hyperspectral data and the post-processing steps
applied after creation of score maps are emphasized.
