11:30 AM - 12:30 PM
[5-1130-P-27] Classification for Fire Blight Disease Infection Area using Vegetation Index and Background Segmentation based on Multispectral Image
Keywords:Multispectral image, Point Cloud, Fire Blight, Vegetation index, Pear tree
Fire Blight (FB) is a bacterial virus called erwinia amylovora. The disease enters the flower or wounded area of the fruit tree, turning leaves and branches brown or black, and dies within one year. Leaves and branches dead by natural wind or pruning also fall into the orchard soil and become brown, similar to FB infection. In the aerial image for the FB discrimination of a wide orchard, there are naturally cut leaf and branches in addition to the desired FB area, which interferes with the FB discrimination.
In this study, we used the digital surface model (DSM) and vegetation index to remove unwanted areas and try to classify the FB infection area. The study area will be located on orchard A at Dokjeong-ri, Ipjang-myeon Cheonan-si, Chungcheongnam-do, Republic of Korea (36°92'42.0224"N, 127°22'70.6734"E) on June 7, 2018, and on June 20, it will be an orchard B at the National Institute of Horticultural & Herbal Science Pear Research Institute, Naju, Jeollanam-do, Republic of Korea (35°01'27.9912"N, 126°44'53.0412"E). Study equipment Unmanned aerial vehicles (UAVs) equipped with multispectral image sensors were used to acquire pear infection and non-infection multispectral images from two orchards. The acquired images were removed by using DSM generated by using the point cloud technique of Drone mapping software (Pix4D 4.3.31, Pix4D SA, Swiss) and GIS software (ArcGIS 10.5.1, Esri, USA), and the images were matched. The images were classified by FB area using vegetation index maps converted to spectral image software (ENVI 5.3, Exelis Visual Information, USA). Drone mapping software and GIS software were used to remove the background height of 100cm from the surface considering the FB area. As a result, an area of about 2,780 m² has been reduced to about 778 m². The area of the FB-infected area was estimated using the histogram and reflection values for the FB-infected and non-infected areas in the background-removed image. When histograms were used, the area of expected FB infection area was 142m² when Otsu's method was used at the NIR wavelength. When using the reflection values, a significant difference was found in the histograms of the red-red edge region and the red-NIR region, and only the overlapping regions were extracted by dividing the regions by Otsu's method. As a result, the estimated area of FB infection was reduced to 71m². As a result, removing the 100-cm-high background and then slinging certain areas of the reflection value could reduce the area of the FB-infected area the most.
In this study, we used the digital surface model (DSM) and vegetation index to remove unwanted areas and try to classify the FB infection area. The study area will be located on orchard A at Dokjeong-ri, Ipjang-myeon Cheonan-si, Chungcheongnam-do, Republic of Korea (36°92'42.0224"N, 127°22'70.6734"E) on June 7, 2018, and on June 20, it will be an orchard B at the National Institute of Horticultural & Herbal Science Pear Research Institute, Naju, Jeollanam-do, Republic of Korea (35°01'27.9912"N, 126°44'53.0412"E). Study equipment Unmanned aerial vehicles (UAVs) equipped with multispectral image sensors were used to acquire pear infection and non-infection multispectral images from two orchards. The acquired images were removed by using DSM generated by using the point cloud technique of Drone mapping software (Pix4D 4.3.31, Pix4D SA, Swiss) and GIS software (ArcGIS 10.5.1, Esri, USA), and the images were matched. The images were classified by FB area using vegetation index maps converted to spectral image software (ENVI 5.3, Exelis Visual Information, USA). Drone mapping software and GIS software were used to remove the background height of 100cm from the surface considering the FB area. As a result, an area of about 2,780 m² has been reduced to about 778 m². The area of the FB-infected area was estimated using the histogram and reflection values for the FB-infected and non-infected areas in the background-removed image. When histograms were used, the area of expected FB infection area was 142m² when Otsu's method was used at the NIR wavelength. When using the reflection values, a significant difference was found in the histograms of the red-red edge region and the red-NIR region, and only the overlapping regions were extracted by dividing the regions by Otsu's method. As a result, the estimated area of FB infection was reduced to 71m². As a result, removing the 100-cm-high background and then slinging certain areas of the reflection value could reduce the area of the FB-infected area the most.