Area extraction of maize lodging based on remote sensing by small unmanned aerial vehicle

Abstract: The information of crop lodging, such as spatial distribution and area, is very critical for agricultural hazard assessment and agricultural insurance claims. It is hard work to measure the area of lodging in a ground survey. A survey method using remote sensing technology is fast and efficient, but it was limited by a lack of available satellite remote sensing data. In recent years, Unmanned Aerial Vehicle (UAV) has been rapidly developed in civil applications. A small UAV remote sensing system in which a UAV carries a digital camera is a portable, stable, and efficient tool for a crop field survey while there is no satellite remote sensing data, but only a few studies about a lodging survey using a UAV were published. There was no study of a survey of maize lodging using a RGB image. Therefore, the authors studied a survey method of maize lodging using some images derived from an UAV remote sensing experiment which was carried out in the Wan Zhuang agricultural high-tech industrial park of the Chinese Academy of Agricultural Sciences (Langfang City, Hebei Province of China) on Sept. 11th to 13th of 2012. In this experiment, some images of maize lodging were acquired after a lodging event on Sept. 12th of 2012. In this study, image features were calculated and summarized first. Three color features and 24 texture features were calculated by processing RGB images using HLS color transformation and co-occurrence texture filters. Mean, variance, coefficient of variation (CV), and relative difference (RD) of image features in normal and lodging maize were summarized. The optimum features for classification of normal and lodging maize were chosen from the 27 features by their coefficient of variation and relative difference. Finally, two methods of lodging area extraction, respectively based on RGB grey level and optimum features, were compared. The result of the image features summary showed that many features had a higher CV or lower RD compared to RGB grey levels, and were not suitable for classification of normal and lodging maize. According to CV and RD, three texture features, including the mean of red, the mean of green, and the mean of blue (RD:59.4%, 45.4%, 48.8%; CV of normal: 10.6%, 7.9%, 8.0%; CV of lodging: 7.5%, 5.6%, 7.2%), having a higher RD and a lower CV compared to a RGB grey level (RD:58.5%, 44.7%, 48.1%; CV of normal: 20.1%, 16.2%, 21.3%; CV of lodging: 14.1%, 12.1%, 16.2%), are optimum indicators for the classification. Compared with measurements of a lodging area, the method based on these optimum classification features (0.3%, 3.5%,6.9%) had lower errors than the method based on a RGB grey level (22.3%, 94.1%, 32.0%). The shadow of a high plant might influence the precision of the classification, but the error is negligible. According to the results of these studies, we may safely draw the conclusion that the method to extract lodging maize area using RGB images of UAV remote sensing based on optimum texture features is accurate.