Improvement of land cover mapping using Sentinel 2 and Landsat 8 imageries via non-parametric classification
In Land Cover Mapping (LCM) studies, non-parametric classifiers are more accurate than parametric classifiers. However, its precision affected by numerous factors like data set type, spatial resolution, number of variables, etc. The overall objective of this study...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/85694/1/FK%202020%2084%20-%20ir.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In Land Cover Mapping (LCM) studies, non-parametric classifiers are more accurate than
parametric classifiers. However, its precision affected by numerous factors like data
set type, spatial resolution, number of variables, etc. The overall objective of this study
is to explore some factors affecting non- parametric LCM classification techniques in terms
of accuracy assessment and to compare their performance with a well-established
classification technique by implementing Landsat 8 and Sentinel satellites. The methodology
carried out in five main phases using Landsat 8 and Sentinel 2 imageries with different test areas
located in Kelantan state, Malaysia. In the first phase, the behavior of Spectral Angle Mapper
(SAM), Decision Tree classifiers (DTC), Support Vector Machines (SVM), and Maximum Likelihood
Classifier (MLC) and effect of some factors including training selection method, training sample
size, input variables and various user-defined parameters on LCM accuracy were
investigated using Landsat 8 data. The results indicated that good classification performance
depends on these factors. All algorithms showed more stability and accuracy when training
size applied is more than 6% by the Equal Sample Rate (ESR) method with six variables.
Furthermore, using SVM requires the user to set certain parameters values, including
kernel type and kernel parameters, since these parameters own a significant
effect on the effectiveness of SVM. The second stage involved assessing the spatial
resolution effect through utilizing Landsat 8 (30 m) and Sentinel (10 m) data on LCM accuracy using
SVM, K-Nearest Neighbor (K-NN), Random Forest (RF), and Neural Network (NN) algorithms. Based on
the concluding overall analysis, the classification accuracy derived from Sentinel 2 imagery
utilizing SVM and RF, Landsat 8 applying SVM donated higher than other methods
of classification. In addition, LCM derived from Sentinel 2 imagery is more accurate
almost about 12% than Landsat 8 by using the same classification technique. This implied
that when spatial resolution increase from 30m to 10m, LCM improved by 12%. The third phase studied
the reaction of the Radial Base Function-based SVM classifier to Modified Sun-Canopy-Sensor Correction (SCS+C) as a
Topographic Correction (TC) technique for Landsat 8 data. The result showed that the Modified
SCS+C method improved LCM's accuracies from 86.22% to 90.11% for the test area. Thus, TC
suggested being the main step in data pre-processing for mountainous terrain before the
RBF-based SVM classification process. At the fourth phase, a combination of Atmospheric Correction
(AC) and four fusion techniques; Brovey, Hue-Saturation-Value (HSV), Hue-Saturation-Value
(HSV), Principal Components (PC), and Gram– Schmidt (GS) spectral sharpening was evaluated.
Results revealed that the GS-sharpening method is more precise than other techniques
and fusion techniques play significant roles in LCM accuracy than AC. Nevertheless, AC is not
required for LCM if the original multi-spectral image is used. The last phase involves developing
a new fusion algorithm using SVM and Fuzzy K-Means Clustering (FKM) algorithms for Sentinel
2 data to enhance LCM accuracy. The highest LCM accuracy generated by a new approach with an
improvement of 8%, 14 %, and 22% compared to SVM, MLC, and S2+LS8-fused image-based
classifiers. |
|---|