TY - JOUR
T1 - Application of maximum likelihood and spectral angle mapping classification techniques to evaluate forest fire severity from UAV multi-spectral images in South Korea
AU - Woo, Heesung
AU - Acuna, Mauricio
AU - Madurapperuma, Buddhika
AU - Jung, Geonhwi
AU - Woo, Choongshik
AU - Park, Joowon
N1 - Publisher Copyright:
© MYU K.K.
PY - 2021
Y1 - 2021
N2 - High-resolution unmanned aerial vehicle (UAV) multi-spectral sensor images can provide valuable information for mapping forest areas that have recently been burned. In this study, we investigate the use of multi-spectral images captured with a UAV to evaluate burn severity in areas affected by forest fires in Gumi-si, South Korea. Fire classification was performed using two supervised learning algorithms, maximum likelihood (ML) and spectral angle mapper (SAM). Three spectral indices, namely, normalized difference vegetation index (NDVI), RedEdge NDVI (RE-NDVI), and the visible-band difference vegetation index (VDVI), were used to create burn severity thresholds in ML and SAM classifiers. The classification results indicated that ML has higher overall accuracy (80–89%, Kappa coefficient = 0.8) than SAM (44–52%, Kappa coefficients ~0.27) in identifying fire severity classes. The ML classifier showed higher accuracy for both unburned and crown fire classes, whereas the SAM classifier exhibited moderate accuracy for all classes. Most of the misclassification was identified between the unburned area and the low heat-damaged area. This research revealed that distinguishing between the unburned area and low heat-damaged area is the most challenging task in fire severity classification. Also, further investigation is required to improve the accuracy of fire severity classification from multi-spectral images.
AB - High-resolution unmanned aerial vehicle (UAV) multi-spectral sensor images can provide valuable information for mapping forest areas that have recently been burned. In this study, we investigate the use of multi-spectral images captured with a UAV to evaluate burn severity in areas affected by forest fires in Gumi-si, South Korea. Fire classification was performed using two supervised learning algorithms, maximum likelihood (ML) and spectral angle mapper (SAM). Three spectral indices, namely, normalized difference vegetation index (NDVI), RedEdge NDVI (RE-NDVI), and the visible-band difference vegetation index (VDVI), were used to create burn severity thresholds in ML and SAM classifiers. The classification results indicated that ML has higher overall accuracy (80–89%, Kappa coefficient = 0.8) than SAM (44–52%, Kappa coefficients ~0.27) in identifying fire severity classes. The ML classifier showed higher accuracy for both unburned and crown fire classes, whereas the SAM classifier exhibited moderate accuracy for all classes. Most of the misclassification was identified between the unburned area and the low heat-damaged area. This research revealed that distinguishing between the unburned area and low heat-damaged area is the most challenging task in fire severity classification. Also, further investigation is required to improve the accuracy of fire severity classification from multi-spectral images.
KW - Accuracy assessment
KW - Forest fire
KW - Machine learning
KW - Remote sensing
KW - UAV
UR - http://www.scopus.com/inward/record.url?scp=85120611928&partnerID=8YFLogxK
U2 - 10.18494/SAM.2021.3365
DO - 10.18494/SAM.2021.3365
M3 - Article
AN - SCOPUS:85120611928
SN - 0914-4935
VL - 33
SP - 3745
EP - 3760
JO - Sensors and Materials
JF - Sensors and Materials
IS - 11
ER -