Title:
|
REMOTE SENSING CLASSIFICATION USING
MULTI-SENSOR SUPER-RESOLUTION ALGORITHM |
Author(s):
|
Alexander Belov and Anna Denisova |
ISBN:
|
978-989-8704-21-4 |
Editors:
|
Yingcai Xiao, Ajith P. Abraham and Jörg Roth |
Year:
|
2020 |
Edition:
|
Single |
Keywords:
|
Image Classification, Data Fusion, Super-Resolution, SVM, RF |
Type:
|
Full |
First Page:
|
131 |
Last Page:
|
140 |
Language:
|
English |
Cover:
|
|
Full Contents:
|
click to dowload
|
Paper Abstract:
|
Super-resolution image fusion aims to produce an image with finer spectral and spatial details than the input images.
However, the super-resolution fusion is mainly applied to enhance a visual representation of the images and its potential
benefits to the final thematic classification is an open question. In this paper, we present an experimental investigation of
the remote sensing image classification performance in the case of the multi-sensor super-resolution image fusion. The
research aims to compare classification performance obtained for the fused image and the low resolution input ones using
different standard-of-the-art classifiers and feature extraction methods. Input data are supposed to be multispectral data
obtained in visible and near infrared spectral ranges by the different remote sensing systems. To perform a multi-sensor
super-resolution image fusion, we used a gradient-descent optimization approach with a B-TV regularization successfully
adapted for remote sensing images with different spatial and spectral sampling characteristics by the authors of the paper.
As for features, we applied brightness in spectral channels, attribute profiles and local feature attribute profiles. The
classification was performed using support vector machines and random forest classifiers that have been proved to be
very effective for remote sensing data classification. The experimental research included the multi-sensor input data
simulation for four remote sensing systems, the super-resolution image fusion of all simulated images and the thematic
classification of the fused image and the images obtained as an average input for each of the simulated imaging systems.
The spatial resolution of the fused image was in 2, 3, 4 and 5 times better than the spatial resolution of the modeled input
images. The average bandwidth of the fused image was 29 nm whereas for the input low resolution images it was in the
range from 37 to 83 nm. Experimental results have shown that random forest classification is better to use with fusion,
whereas support vector machines demonstrated better results without fusion. The feature extraction test showed that
extended attribute profiles enhance the random forest classification accuracy of the fused image. Thus, the classification
results have shown that super-resolution image fusion leads to the classification accuracy increase in the case of random
forest classifier and there is no need to apply fusion in the case of support vector machines. |
|
|
|
|