Vegetation indices and textures in object-based weed detection from UAV imagery

Abstract

The use of unmanned aerial vehicles (UAV) propelled the growing interest of small to medium farmers in precision agriculture applications. In this study, a methodology was developed to automatically map the land use in a vegetable farm with a very high resolution aerial image (5cm/pixel) taken at an altitude of 100 m. Using color index of vegetation extraction and Otsu's thresholding method, the soil was delineated from vegetation by object-based technique, and subsequently segmented the vegetation using multi-resolution algorithm. The Support Vector Machine classification on various vegetation indices produced agreeable results, where eggplant, corn, string beans and grass/weeds were classified. The classification was further improved by including Gray Level Co-Occurrence Measures or textural features, as indicated by the increase in overall accuracy and kappa index of agreement. The output map can serve as guide for farmers and other management agencies to know the correct farm interventions such as weed control. The output map can also be used in the periodic updating of the initial land cover map produced from high resolution LiDAR data.