استخراج باغات از تصاوير ماهواره‌اي سنتينل-2 به‌كمك طبقه‌بندي در فضاهاي غيرRGB (مطالعۀ موردي شهرستان تيران و كرون اصفهان)

Extraction of Orchards from Sentinel 2 Satellite Images through Classification in Non RGB Spaces (Case Study: Tiran an‎d Karvan County, Isfahan Province)

مهندسي نقشه برداري

Vegetation covers play a determining role in maintaining an‎d sustaining the life of living organisms. Among these covers, orchards have a vital role in agricul-ture an‎d biodiversity conservation an‎d therefore require precise monitoring. To-day, remote sensing methods are superior in identifying an‎d extracting vegetation covers an‎d are preferred to ground mapping methods. Classification an‎d infor-mation extraction are usually performed in the RGB space; however, considering that in previous studies the use of non RGB spaces has resulted in increased ex-traction accuracy, an‎d given the challenges of orchard extraction from satellite im-ages—such as spectral similarity with other vegetation covers including range-lan‎ds, croplan‎ds, an‎d urban green areas; differences in orchard management; di-versity of plant species; an‎d dependence on seasonal characteristics—this study seeks to eva‎luate the potential of these spaces, including orthogonal transfor-mations (PCA, MNF, GRAM SCHMIDT, an‎d Per Class PCA based on training pixels of orchard an‎d non orchard classes in three ban‎d an‎d four ban‎d modes, an‎d the first three components out of the four components of the PCA, MNF, an‎d GRAM SCHMIDT transforms); non RGB color spaces (HSV, LAB, an‎d YCbCr in the RGB an‎d RNB ban‎d combinations); an‎d also the combination of these two (by applying PCA, MNF, an‎d GRAM SCHMIDT transforms to the output of each color space) in the field of orchard extraction in Tiran an‎d Karvan County, Isfahan Province. For this purpose, Sentinel 2 satellite images from the months of Mehr an‎d Aban 1398 (October an‎d November 2019), as well as ground truth data (cadaster maps) from the Agricultural Jihad Organization of Isfahan Province, were used. After performing the necessary preprocessing on the images, the intended spaces were applied to the October images in the RGB (red, green, an‎d blue), RNB (red, near infrared, an‎d blue), an‎d RGBN (red, green, blue, an‎d near infrared) ban‎d combinations. Subsequently, orchard extraction was carried out through a two class (orchard an‎d non orchard) classification in these spaces, an‎d the overall accuracy an‎d f rank coefficient of the orchard class were calculated for each space. In addition to the mentioned spaces, comparative methods (texture fea-tures, vegetation indices, decision level fusion methods, an‎d multitemporal classi-fication) were also employed so that their findings could be compared with the results obtained from the eva‎luated spaces. The results of this study indicate that in the three ban‎d mode, including RGB an‎d RNB, the combination produced from the Gram Schmidt transform an‎d the HSV color space achieved the highest overall accuracies (74.51% an‎d 78.63%) an‎d orchard class f rank coefficients (75.02% an‎d 79.55%). In the four ban‎d mode, the Per Class PCA transform based on training pixels of the non orchard class has the highest overall accuracy (80.56%) an‎d orchard class f rank coeffi-cient (81.08%). Among the comparative methods, multitemporal classification ap-plied to the eight ban‎d image resulting from independently applying PCA to the RGBN images of Mehr an‎d Aban achieved the highest overall accuracy (86.51%) an‎d orchard class f rank coefficient (86.82%) an‎d is introduced as the best space among all examined cases in this study.

5

158064